Create OptionalSwitch node + EmptyInputSentinel to deal with inputs that should be seen as 'empty' when passed into nodes

* Fix bypass dtype/device moving

* Force offloading mode for training

* training context var

* offloading implementation in training node

* fix wrong input type

* Support bypass load lora model, correct adapter/offloading handling

.ci/update_windows/update.py

@@ -53,6 +53,16 @@ try:
     repo.stash(ident)
 except KeyError:
     print("nothing to stash")  # noqa: T201
+except:
+    print("Could not stash, cleaning index and trying again.")  # noqa: T201
+    repo.state_cleanup()
+    repo.index.read_tree(repo.head.peel().tree)
+    repo.index.write()
+    try:
+        repo.stash(ident)
+    except KeyError:
+        print("nothing to stash.")  # noqa: T201
+
 backup_branch_name = 'backup_branch_{}'.format(datetime.today().strftime('%Y-%m-%d_%H_%M_%S'))
 print("creating backup branch: {}".format(backup_branch_name))  # noqa: T201
 try:
@@ -66,8 +76,10 @@ if branch is None:
     try:
         ref = repo.lookup_reference('refs/remotes/origin/master')
     except:
-        print("pulling.")  # noqa: T201
-        pull(repo)
+        print("fetching.")  # noqa: T201
+        for remote in repo.remotes:
+            if remote.name == "origin":
+                remote.fetch()
         ref = repo.lookup_reference('refs/remotes/origin/master')
     repo.checkout(ref)
     branch = repo.lookup_branch('master')
@@ -149,3 +161,4 @@ try:
         shutil.copy(stable_update_script, stable_update_script_to)
 except:
     pass
+

.ci/windows_amd_base_files/README_VERY_IMPORTANT.txt

@@ -1,5 +1,5 @@
-As of the time of writing this you need this preview driver for best results:
-https://www.amd.com/en/resources/support-articles/release-notes/RN-AMDGPU-WINDOWS-PYTORCH-PREVIEW.html
+As of the time of writing this you need this driver for best results:
+https://www.amd.com/en/resources/support-articles/release-notes/RN-AMDGPU-WINDOWS-PYTORCH-7-1-1.html
 
 HOW TO RUN:
 
@@ -25,3 +25,4 @@ In the ComfyUI directory you will find a file: extra_model_paths.yaml.example
 Rename this file to: extra_model_paths.yaml and edit it with your favorite text editor.
 
 
+

.ci/windows_nvidia_base_files/advanced/run_nvidia_gpu_disable_api_nodes.bat

@@ -1,3 +1,3 @@
 ..\python_embeded\python.exe -s ..\ComfyUI\main.py --windows-standalone-build --disable-api-nodes
-echo If you see this and ComfyUI did not start try updating your Nvidia Drivers to the latest.
+echo If you see this and ComfyUI did not start try updating your Nvidia Drivers to the latest. If you get a c10.dll error you need to install vc redist that you can find: https://aka.ms/vc14/vc_redist.x64.exe
 pause

.ci/windows_nvidia_base_files/run_nvidia_gpu.bat

@@ -1,3 +1,3 @@
 .\python_embeded\python.exe -s ComfyUI\main.py --windows-standalone-build
-echo If you see this and ComfyUI did not start try updating your Nvidia Drivers to the latest.
+echo If you see this and ComfyUI did not start try updating your Nvidia Drivers to the latest. If you get a c10.dll error you need to install vc redist that you can find: https://aka.ms/vc14/vc_redist.x64.exe
 pause

.ci/windows_nvidia_base_files/run_nvidia_gpu_fast_fp16_accumulation.bat

@@ -1,3 +1,3 @@
 .\python_embeded\python.exe -s ComfyUI\main.py --windows-standalone-build --fast fp16_accumulation
-echo If you see this and ComfyUI did not start try updating your Nvidia Drivers to the latest.
+echo If you see this and ComfyUI did not start try updating your Nvidia Drivers to the latest. If you get a c10.dll error you need to install vc redist that you can find: https://aka.ms/vc14/vc_redist.x64.exe
 pause

.github/ISSUE_TEMPLATE/bug-report.yml

@@ -8,13 +8,15 @@ body:
         Before submitting a **Bug Report**, please ensure the following:
 
         - **1:** You are running the latest version of ComfyUI.
-        - **2:** You have looked at the existing bug reports and made sure this isn't already reported.
+        - **2:** You have your ComfyUI logs and relevant workflow on hand and will post them in this bug report.
         - **3:** You confirmed that the bug is not caused by a custom node. You can disable all custom nodes by passing
-        `--disable-all-custom-nodes` command line argument.
+        `--disable-all-custom-nodes` command line argument. If you have custom node try updating them to the latest version.
         - **4:** This is an actual bug in ComfyUI, not just a support question. A bug is when you can specify exact
         steps to replicate what went wrong and others will be able to repeat your steps and see the same issue happen.
 
-        If unsure, ask on the [ComfyUI Matrix Space](https://app.element.io/#/room/%23comfyui_space%3Amatrix.org) or the [Comfy Org Discord](https://discord.gg/comfyorg) first.
+        ## Very Important
+
+        Please make sure that you post ALL your ComfyUI logs in the bug report. A bug report without logs will likely be ignored.
   - type: checkboxes
     id: custom-nodes-test
     attributes:

.github/PULL_REQUEST_TEMPLATE/api-node.md

@@ -0,0 +1,21 @@
+<!-- API_NODE_PR_CHECKLIST: do not remove -->
+
+## API Node PR Checklist
+
+### Scope
+- [ ] **Is API Node Change**
+
+### Pricing & Billing
+- [ ] **Need pricing update**
+- [ ] **No pricing update**
+
+If **Need pricing update**:
+- [ ] Metronome rate cards updated
+- [ ] Auto‑billing tests updated and passing
+
+### QA
+- [ ] **QA done**
+- [ ] **QA not required**
+
+### Comms
+- [ ] Informed **Kosinkadink**

.github/workflows/api-node-template.yml

@@ -0,0 +1,58 @@
+name: Append API Node PR template
+
+on:
+  pull_request_target:
+    types: [opened, reopened, synchronize, ready_for_review]
+    paths:
+      - 'comfy_api_nodes/**'   # only run if these files changed
+
+permissions:
+  contents: read
+  pull-requests: write
+
+jobs:
+  inject:
+    runs-on: ubuntu-latest
+    steps:
+      - name: Ensure template exists and append to PR body
+        uses: actions/github-script@v7
+        with:
+          script: |
+            const { owner, repo } = context.repo;
+            const number = context.payload.pull_request.number;
+            const templatePath = '.github/PULL_REQUEST_TEMPLATE/api-node.md';
+            const marker = '<!-- API_NODE_PR_CHECKLIST: do not remove -->';
+
+            const { data: pr } = await github.rest.pulls.get({ owner, repo, pull_number: number });
+
+            let templateText;
+            try {
+              const res = await github.rest.repos.getContent({
+                owner,
+                repo,
+                path: templatePath,
+                ref: pr.base.ref
+              });
+              const buf = Buffer.from(res.data.content, res.data.encoding || 'base64');
+              templateText = buf.toString('utf8');
+            } catch (e) {
+              core.setFailed(`Required PR template not found at "${templatePath}" on ${pr.base.ref}. Please add it to the repo.`);
+              return;
+            }
+
+            // Enforce the presence of the marker inside the template (for idempotence)
+            if (!templateText.includes(marker)) {
+              core.setFailed(`Template at "${templatePath}" does not contain the required marker:\n${marker}\nAdd it so we can detect duplicates safely.`);
+              return;
+            }
+
+            // If the PR already contains the marker, do not append again.
+            const body = pr.body || '';
+            if (body.includes(marker)) {
+              core.info('Template already present in PR body; nothing to inject.');
+              return;
+            }
+
+            const newBody = (body ? body + '\n\n' : '') + templateText + '\n';
+            await github.rest.pulls.update({ owner, repo, pull_number: number, body: newBody });
+            core.notice('API Node template appended to PR description.');

.github/workflows/release-stable-all.yml

@@ -14,13 +14,13 @@ jobs:
       contents: "write"
       packages: "write"
       pull-requests: "read"
-    name: "Release NVIDIA Default (cu129)"
+    name: "Release NVIDIA Default (cu130)"
     uses: ./.github/workflows/stable-release.yml
     with:
       git_tag: ${{ inputs.git_tag }}
       cache_tag: "cu130"
       python_minor: "13"
-      python_patch: "9"
+      python_patch: "11"
       rel_name: "nvidia"
       rel_extra_name: ""
       test_release: true
@@ -43,16 +43,33 @@ jobs:
       test_release: true
     secrets: inherit
 
+  release_nvidia_cu126:
+    permissions:
+      contents: "write"
+      packages: "write"
+      pull-requests: "read"
+    name: "Release NVIDIA cu126"
+    uses: ./.github/workflows/stable-release.yml
+    with:
+      git_tag: ${{ inputs.git_tag }}
+      cache_tag: "cu126"
+      python_minor: "12"
+      python_patch: "10"
+      rel_name: "nvidia"
+      rel_extra_name: "_cu126"
+      test_release: true
+    secrets: inherit
+
   release_amd_rocm:
     permissions:
       contents: "write"
       packages: "write"
       pull-requests: "read"
-    name: "Release AMD ROCm 6.4.4"
+    name: "Release AMD ROCm 7.2"
     uses: ./.github/workflows/stable-release.yml
     with:
       git_tag: ${{ inputs.git_tag }}
-      cache_tag: "rocm644"
+      cache_tag: "rocm72"
       python_minor: "12"
       python_patch: "10"
       rel_name: "amd"

.github/workflows/stable-release.yml

@@ -117,7 +117,7 @@ jobs:
           ./python.exe get-pip.py
           ./python.exe -s -m pip install ../${{ inputs.cache_tag }}_python_deps/*
 
-          grep comfyui ../ComfyUI/requirements.txt > ./requirements_comfyui.txt
+          grep comfy ../ComfyUI/requirements.txt > ./requirements_comfyui.txt
           ./python.exe -s -m pip install -r requirements_comfyui.txt
           rm requirements_comfyui.txt
 

.github/workflows/test-build.yml

@@ -18,7 +18,7 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        python-version: ["3.9", "3.10", "3.11", "3.12", "3.13"]
+        python-version: ["3.10", "3.11", "3.12", "3.13", "3.14"]
     steps:
       - uses: actions/checkout@v4
       - name: Set up Python ${{ matrix.python-version }}

.github/workflows/test-ci.yml

@@ -5,6 +5,7 @@ on:
   push:
     branches:
       - master
+      - release/**
     paths-ignore:
       - 'app/**'
       - 'input/**'
@@ -21,14 +22,15 @@ jobs:
       fail-fast: false
       matrix:
         # os: [macos, linux, windows]
-        os: [macos, linux]
-        python_version: ["3.9", "3.10", "3.11", "3.12"]
+        # os: [macos, linux]
+        os: [linux]
+        python_version: ["3.10", "3.11", "3.12"]
         cuda_version: ["12.1"]
         torch_version: ["stable"]
         include:
-          - os: macos
-            runner_label: [self-hosted, macOS]
-            flags: "--use-pytorch-cross-attention"
+          # - os: macos
+          #   runner_label: [self-hosted, macOS]
+          #   flags: "--use-pytorch-cross-attention"
           - os: linux
             runner_label: [self-hosted, Linux]
             flags: ""
@@ -73,14 +75,15 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        os: [macos, linux]
+        # os: [macos, linux]
+        os: [linux]
         python_version: ["3.11"]
         cuda_version: ["12.1"]
         torch_version: ["nightly"]
         include:
-          - os: macos
-            runner_label: [self-hosted, macOS]
-            flags: "--use-pytorch-cross-attention"
+          # - os: macos
+          #   runner_label: [self-hosted, macOS]
+          #   flags: "--use-pytorch-cross-attention"
           - os: linux
             runner_label: [self-hosted, Linux]
             flags: ""

.github/workflows/test-execution.yml

@@ -2,9 +2,9 @@ name: Execution Tests
 
 on:
   push:
-    branches: [ main, master ]
+    branches: [ main, master, release/** ]
   pull_request:
-    branches: [ main, master ]
+    branches: [ main, master, release/** ]
 
 jobs:
   test:

.github/workflows/test-launch.yml

@@ -2,9 +2,9 @@ name: Test server launches without errors
 
 on:
   push:
-    branches: [ main, master ]
+    branches: [ main, master, release/** ]
   pull_request:
-    branches: [ main, master ]
+    branches: [ main, master, release/** ]
 
 jobs:
   test:
@@ -13,7 +13,7 @@ jobs:
     - name: Checkout ComfyUI
       uses: actions/checkout@v4
       with:
-        repository: "comfyanonymous/ComfyUI"
+        repository: "Comfy-Org/ComfyUI"
         path: "ComfyUI"
     - uses: actions/setup-python@v4
       with:
@@ -32,7 +32,9 @@ jobs:
       working-directory: ComfyUI
     - name: Check for unhandled exceptions in server log
       run: |
-        if grep -qE "Exception|Error" console_output.log; then
+        grep -v "Found comfy_kitchen backend triton: {'available': False, 'disabled': True, 'unavailable_reason': \"ImportError: No module named 'triton'\", 'capabilities': \[\]}" console_output.log | grep -v "Found comfy_kitchen backend triton: {'available': False, 'disabled': False, 'unavailable_reason': \"ImportError: No module named 'triton'\", 'capabilities': \[\]}" > console_output_filtered.log
+        cat console_output_filtered.log
+        if grep -qE "Exception|Error" console_output_filtered.log; then
           echo "Unhandled exception/error found in server log."
           exit 1
         fi

.github/workflows/test-unit.yml

@@ -2,9 +2,9 @@ name: Unit Tests
 
 on:
   push:
-    branches: [ main, master ]
+    branches: [ main, master, release/** ]
   pull_request:
-    branches: [ main, master ]
+    branches: [ main, master, release/** ]
 
 jobs:
   test:

.github/workflows/update-ci-container.yml

@@ -0,0 +1,59 @@
+name: "CI: Update CI Container"
+
+on:
+  release:
+    types: [published]
+  workflow_dispatch:
+    inputs:
+      version:
+        description: 'ComfyUI version (e.g., v0.7.0)'
+        required: true
+        type: string
+
+jobs:
+  update-ci-container:
+    runs-on: ubuntu-latest
+    # Skip pre-releases unless manually triggered
+    if: github.event_name == 'workflow_dispatch' || !github.event.release.prerelease
+    steps:
+      - name: Get version
+        id: version
+        run: |
+          if [ "${{ github.event_name }}" = "release" ]; then
+            VERSION="${{ github.event.release.tag_name }}"
+          else
+            VERSION="${{ inputs.version }}"
+          fi
+          echo "version=$VERSION" >> $GITHUB_OUTPUT
+
+      - name: Checkout comfyui-ci-container
+        uses: actions/checkout@v4
+        with:
+          repository: comfy-org/comfyui-ci-container
+          token: ${{ secrets.CI_CONTAINER_PAT }}
+
+      - name: Check current version
+        id: current
+        run: |
+          CURRENT=$(grep -oP 'ARG COMFYUI_VERSION=\K.*' Dockerfile || echo "unknown")
+          echo "current_version=$CURRENT" >> $GITHUB_OUTPUT
+
+      - name: Update Dockerfile
+        run: |
+          VERSION="${{ steps.version.outputs.version }}"
+          sed -i "s/^ARG COMFYUI_VERSION=.*/ARG COMFYUI_VERSION=${VERSION}/" Dockerfile
+
+      - name: Create Pull Request
+        id: create-pr
+        uses: peter-evans/create-pull-request@v7
+        with:
+          token: ${{ secrets.CI_CONTAINER_PAT }}
+          branch: automation/comfyui-${{ steps.version.outputs.version }}
+          title: "chore: bump ComfyUI to ${{ steps.version.outputs.version }}"
+          body: |
+            Updates ComfyUI version from `${{ steps.current.outputs.current_version }}` to `${{ steps.version.outputs.version }}`
+
+            **Triggered by:** ${{ github.event_name == 'release' && format('[Release {0}]({1})', github.event.release.tag_name, github.event.release.html_url) || 'Manual workflow dispatch' }}
+
+          labels: automation
+          commit-message: "chore: bump ComfyUI to ${{ steps.version.outputs.version }}"

.github/workflows/update-version.yml

@@ -6,6 +6,7 @@ on:
       - "pyproject.toml"
     branches:
       - master
+      - release/**
 
 jobs:
   update-version:

.github/workflows/windows_release_dependencies.yml

@@ -29,7 +29,7 @@ on:
         description: 'python patch version'
         required: true
         type: string
-        default: "9"
+        default: "11"
 #  push:
 #    branches:
 #      - master

CODEOWNERS

@@ -1,3 +1,2 @@
 # Admins
-* @comfyanonymous
-* @kosinkadink
+* @comfyanonymous @kosinkadink @guill

QUANTIZATION.md

@@ -0,0 +1,168 @@
+# The Comfy guide to Quantization
+
+
+## How does quantization work?
+
+Quantization aims to map a high-precision value x_f to a lower precision format with minimal loss in accuracy. These smaller formats then serve to reduce the models memory footprint and increase throughput by using specialized hardware.
+
+When simply converting a value from FP16 to FP8 using the round-nearest method we might hit two issues:
+- The dynamic range of FP16 (-65,504, 65,504) far exceeds FP8 formats like E4M3 (-448, 448) or E5M2 (-57,344, 57,344), potentially resulting in clipped values
+- The original values are concentrated in a small range (e.g. -1,1) leaving many FP8-bits "unused"
+
+By using a scaling factor, we aim to map these values into the quantized-dtype range, making use of the full spectrum. One of the easiest approaches, and common, is using per-tensor absolute-maximum scaling.
+
+```
+absmax = max(abs(tensor))
+scale = amax / max_dynamic_range_low_precision
+
+# Quantization
+tensor_q = (tensor / scale).to(low_precision_dtype)
+
+# De-Quantization
+tensor_dq = tensor_q.to(fp16) * scale
+
+tensor_dq ~ tensor
+```
+
+Given that additional information (scaling factor) is needed to "interpret" the quantized values, we describe those as derived datatypes.
+
+
+## Quantization in Comfy
+
+```
+QuantizedTensor (torch.Tensor subclass)
+  ↓ __torch_dispatch__
+Two-Level Registry (generic + layout handlers)
+  ↓
+MixedPrecisionOps + Metadata Detection
+```
+
+### Representation
+
+To represent these derived datatypes, ComfyUI uses a subclass of torch.Tensor to implements these using the `QuantizedTensor` class found in `comfy/quant_ops.py`
+
+A `Layout` class defines how a specific quantization format behaves:
+- Required parameters
+- Quantize method
+- De-Quantize method
+
+```python
+from comfy.quant_ops import QuantizedLayout
+
+class MyLayout(QuantizedLayout):
+    @classmethod
+    def quantize(cls, tensor, **kwargs):
+        # Convert to quantized format
+        qdata = ...
+        params = {'scale': ..., 'orig_dtype': tensor.dtype}
+        return qdata, params
+    
+    @staticmethod
+    def dequantize(qdata, scale, orig_dtype, **kwargs):
+        return qdata.to(orig_dtype) * scale
+```
+
+To then run operations using these QuantizedTensors we use two registry systems to define supported operations. 
+The first is a **generic registry** that handles operations common to all quantized formats (e.g., `.to()`, `.clone()`, `.reshape()`).
+
+The second registry is layout-specific and allows to implement fast-paths like nn.Linear.
+```python
+from comfy.quant_ops import register_layout_op
+
+@register_layout_op(torch.ops.aten.linear.default, MyLayout)
+def my_linear(func, args, kwargs):
+    # Extract tensors, call optimized kernel
+    ...
+```
+When `torch.nn.functional.linear()` is called with QuantizedTensor arguments, `__torch_dispatch__` automatically routes to the registered implementation.
+For any unsupported operation, QuantizedTensor will fallback to call `dequantize` and dispatch using the high-precision implementation.
+
+
+### Mixed Precision
+
+The `MixedPrecisionOps` class (lines 542-648 in `comfy/ops.py`) enables per-layer quantization decisions, allowing different layers in a model to use different precisions. This is activated when a model config contains a `layer_quant_config` dictionary that specifies which layers should be quantized and how.
+
+**Architecture:**
+
+```python
+class MixedPrecisionOps(disable_weight_init):
+    _layer_quant_config = {}  # Maps layer names to quantization configs
+    _compute_dtype = torch.bfloat16  # Default compute / dequantize precision
+```
+
+**Key mechanism:**
+
+The custom `Linear._load_from_state_dict()` method inspects each layer during model loading:
+- If the layer name is **not** in `_layer_quant_config`: load weight as regular tensor in `_compute_dtype`
+- If the layer name **is** in `_layer_quant_config`: 
+  - Load weight as `QuantizedTensor` with the specified layout (e.g., `TensorCoreFP8Layout`)
+  - Load associated quantization parameters (scales, block_size, etc.)
+
+**Why it's needed:**
+
+Not all layers tolerate quantization equally. Sensitive operations like final projections can be kept in higher precision, while compute-heavy matmuls are quantized. This provides most of the performance benefits while maintaining quality.
+
+The system is selected in `pick_operations()` when `model_config.layer_quant_config` is present, making it the highest-priority operation mode.
+
+
+## Checkpoint Format
+
+Quantized checkpoints are stored as standard safetensors files with quantized weight tensors and associated scaling parameters, plus a `_quantization_metadata` JSON entry describing the quantization scheme.
+
+The quantized checkpoint will contain the same layers as the original checkpoint but:
+- The weights are stored as quantized values, sometimes using a different storage datatype. E.g. uint8 container for fp8.
+- For each quantized weight a number of additional scaling parameters are stored alongside depending on the recipe.
+- We store a metadata.json in the metadata of the final safetensor containing the `_quantization_metadata` describing which layers are quantized and what layout has been used.
+
+### Scaling Parameters details
+We define 4 possible scaling parameters that should cover most recipes in the near-future:
+- **weight_scale**: quantization scalers for the weights
+- **weight_scale_2**: global scalers in the context of double scaling
+- **pre_quant_scale**: scalers used for smoothing salient weights
+- **input_scale**: quantization scalers for the activations
+
+| Format | Storage dtype | weight_scale | weight_scale_2 | pre_quant_scale | input_scale |
+|--------|---------------|--------------|----------------|-----------------|-------------|
+| float8_e4m3fn | float32 | float32 (scalar) | - | - | float32 (scalar) |
+
+You can find the defined formats in `comfy/quant_ops.py` (QUANT_ALGOS).
+
+### Quantization Metadata
+
+The metadata stored alongside the checkpoint contains:
+- **format_version**: String to define a version of the standard
+- **layers**: A dictionary mapping layer names to their quantization format. The format string maps to the definitions found in `QUANT_ALGOS`. 
+
+Example:
+```json
+{
+  "_quantization_metadata": {
+    "format_version": "1.0",
+    "layers": {
+      "model.layers.0.mlp.up_proj": "float8_e4m3fn",
+      "model.layers.0.mlp.down_proj": "float8_e4m3fn",
+      "model.layers.1.mlp.up_proj": "float8_e4m3fn"
+    }
+  }
+}
+```
+
+
+## Creating Quantized Checkpoints
+
+To create compatible checkpoints, use any quantization tool provided the output follows the checkpoint format described above and uses a layout defined in `QUANT_ALGOS`.
+
+### Weight Quantization
+
+Weight quantization is straightforward - compute the scaling factor directly from the weight tensor using the absolute maximum method described earlier. Each layer's weights are quantized independently and stored with their corresponding `weight_scale` parameter.
+
+### Calibration (for Activation Quantization)
+
+Activation quantization (e.g., for FP8 Tensor Core operations) requires `input_scale` parameters that cannot be determined from static weights alone. Since activation values depend on actual inputs, we use **post-training calibration (PTQ)**:
+
+1. **Collect statistics**: Run inference on N representative samples
+2. **Track activations**: Record the absolute maximum (`amax`) of inputs to each quantized layer
+3. **Compute scales**: Derive `input_scale` from collected statistics
+4. **Store in checkpoint**: Save `input_scale` parameters alongside weights
+
+The calibration dataset should be representative of your target use case. For diffusion models, this typically means a diverse set of prompts and generation parameters.

README.md

@@ -67,6 +67,8 @@ See what ComfyUI can do with the [example workflows](https://comfyanonymous.gith
    - [HiDream](https://comfyanonymous.github.io/ComfyUI_examples/hidream/)
    - [Qwen Image](https://comfyanonymous.github.io/ComfyUI_examples/qwen_image/)
    - [Hunyuan Image 2.1](https://comfyanonymous.github.io/ComfyUI_examples/hunyuan_image/)
+   - [Flux 2](https://comfyanonymous.github.io/ComfyUI_examples/flux2/)
+   - [Z Image](https://comfyanonymous.github.io/ComfyUI_examples/z_image/)
 - Image Editing Models
    - [Omnigen 2](https://comfyanonymous.github.io/ComfyUI_examples/omnigen/)
    - [Flux Kontext](https://comfyanonymous.github.io/ComfyUI_examples/flux/#flux-kontext-image-editing-model)
@@ -79,6 +81,7 @@ See what ComfyUI can do with the [example workflows](https://comfyanonymous.gith
    - [Hunyuan Video](https://comfyanonymous.github.io/ComfyUI_examples/hunyuan_video/)
    - [Wan 2.1](https://comfyanonymous.github.io/ComfyUI_examples/wan/)
    - [Wan 2.2](https://comfyanonymous.github.io/ComfyUI_examples/wan22/)
+   - [Hunyuan Video 1.5](https://docs.comfy.org/tutorials/video/hunyuan/hunyuan-video-1-5)
 - Audio Models
    - [Stable Audio](https://comfyanonymous.github.io/ComfyUI_examples/audio/)
    - [ACE Step](https://comfyanonymous.github.io/ComfyUI_examples/audio/)
@@ -105,17 +108,21 @@ See what ComfyUI can do with the [example workflows](https://comfyanonymous.gith
 - [LCM models and Loras](https://comfyanonymous.github.io/ComfyUI_examples/lcm/)
 - Latent previews with [TAESD](#how-to-show-high-quality-previews)
 - Works fully offline: core will never download anything unless you want to.
-- Optional API nodes to use paid models from external providers through the online [Comfy API](https://docs.comfy.org/tutorials/api-nodes/overview).
+- Optional API nodes to use paid models from external providers through the online [Comfy API](https://docs.comfy.org/tutorials/api-nodes/overview) disable with: `--disable-api-nodes`
 - [Config file](extra_model_paths.yaml.example) to set the search paths for models.
 
 Workflow examples can be found on the [Examples page](https://comfyanonymous.github.io/ComfyUI_examples/)
 
 ## Release Process
 
-ComfyUI follows a weekly release cycle targeting Friday but this regularly changes because of model releases or large changes to the codebase. There are three interconnected repositories:
+ComfyUI follows a weekly release cycle targeting Monday but this regularly changes because of model releases or large changes to the codebase. There are three interconnected repositories:
 
 1. **[ComfyUI Core](https://github.com/comfyanonymous/ComfyUI)**
-   - Releases a new stable version (e.g., v0.7.0)
+   - Releases a new stable version (e.g., v0.7.0) roughly every week.
+   - Starting from v0.4.0 patch versions will be used for fixes backported onto the current stable release.
+   - Minor versions will be used for releases off the master branch.
+   - Patch versions may still be used for releases on the master branch in cases where a backport would not make sense.
+   - Commits outside of the stable release tags may be very unstable and break many custom nodes.
    - Serves as the foundation for the desktop release
 
 2. **[ComfyUI Desktop](https://github.com/Comfy-Org/desktop)**
@@ -172,17 +179,19 @@ There is a portable standalone build for Windows that should work for running on
 
 ### [Direct link to download](https://github.com/comfyanonymous/ComfyUI/releases/latest/download/ComfyUI_windows_portable_nvidia.7z)
 
-Simply download, extract with [7-Zip](https://7-zip.org) and run. Make sure you put your Stable Diffusion checkpoints/models (the huge ckpt/safetensors files) in: ComfyUI\models\checkpoints
+Simply download, extract with [7-Zip](https://7-zip.org) or with the windows explorer on recent windows versions and run. For smaller models you normally only need to put the checkpoints (the huge ckpt/safetensors files) in: ComfyUI\models\checkpoints but many of the larger models have multiple files. Make sure to follow the instructions to know which subfolder to put them in ComfyUI\models\
 
 If you have trouble extracting it, right click the file -> properties -> unblock
 
-Update your Nvidia drivers if it doesn't start.
+The portable above currently comes with python 3.13 and pytorch cuda 13.0. Update your Nvidia drivers if it doesn't start.
 
 #### Alternative Downloads:
 
 [Experimental portable for AMD GPUs](https://github.com/comfyanonymous/ComfyUI/releases/latest/download/ComfyUI_windows_portable_amd.7z)
 
-[Portable with pytorch cuda 12.8 and python 3.12](https://github.com/comfyanonymous/ComfyUI/releases/latest/download/ComfyUI_windows_portable_nvidia_cu128.7z) (Supports Nvidia 10 series and older GPUs).
+[Portable with pytorch cuda 12.8 and python 3.12](https://github.com/comfyanonymous/ComfyUI/releases/latest/download/ComfyUI_windows_portable_nvidia_cu128.7z).
+
+[Portable with pytorch cuda 12.6 and python 3.12](https://github.com/comfyanonymous/ComfyUI/releases/latest/download/ComfyUI_windows_portable_nvidia_cu126.7z) (Supports Nvidia 10 series and older GPUs).
 
 #### How do I share models between another UI and ComfyUI?
 
@@ -199,10 +208,12 @@ comfy install
 
 ## Manual Install (Windows, Linux)
 
-Python 3.14 will work if you comment out the `kornia` dependency in the requirements.txt file (breaks the canny node) but it is not recommended.
+Python 3.14 works but some custom nodes may have issues. The free threaded variant works but some dependencies will enable the GIL so it's not fully supported.
 
 Python 3.13 is very well supported. If you have trouble with some custom node dependencies on 3.13 you can try 3.12
 
+torch 2.4 and above is supported but some features and optimizations might only work on newer versions. We generally recommend using the latest major version of pytorch with the latest cuda version unless it is less than 2 weeks old.
+
 ### Instructions:
 
 Git clone this repo.
@@ -218,9 +229,9 @@ AMD users can install rocm and pytorch with pip if you don't have it already ins
 
 ```pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/rocm6.4```
 
-This is the command to install the nightly with ROCm 7.0 which might have some performance improvements:
+This is the command to install the nightly with ROCm 7.1 which might have some performance improvements:
 
-```pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/rocm7.0```
+```pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/rocm7.1```
 
 
 ### AMD GPUs (Experimental: Windows and Linux), RDNA 3, 3.5 and 4 only.
@@ -229,7 +240,7 @@ These have less hardware support than the builds above but they work on windows.
 
 RDNA 3 (RX 7000 series):
 
-```pip install --pre torch torchvision torchaudio --index-url https://rocm.nightlies.amd.com/v2/gfx110X-dgpu/```
+```pip install --pre torch torchvision torchaudio --index-url https://rocm.nightlies.amd.com/v2/gfx110X-all/```
 
 RDNA 3.5 (Strix halo/Ryzen AI Max+ 365):
 
@@ -241,7 +252,7 @@ RDNA 4 (RX 9000 series):
 
 ### Intel GPUs (Windows and Linux)
 
-(Option 1) Intel Arc GPU users can install native PyTorch with torch.xpu support using pip. More information can be found [here](https://pytorch.org/docs/main/notes/get_start_xpu.html)
+Intel Arc GPU users can install native PyTorch with torch.xpu support using pip. More information can be found [here](https://pytorch.org/docs/main/notes/get_start_xpu.html)
 
 1. To install PyTorch xpu, use the following command:
 
@@ -251,10 +262,6 @@ This is the command to install the Pytorch xpu nightly which might have some per
 
 ```pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/xpu```
 
-(Option 2) Alternatively, Intel GPUs supported by Intel Extension for PyTorch (IPEX) can leverage IPEX for improved performance.
-
-1. visit [Installation](https://intel.github.io/intel-extension-for-pytorch/index.html#installation?platform=gpu) for more information.
-
 ### NVIDIA
 
 Nvidia users should install stable pytorch using this command:
@@ -318,6 +325,32 @@ For models compatible with Iluvatar Extension for PyTorch. Here's a step-by-step
 1. Install the Iluvatar Corex Toolkit by adhering to the platform-specific instructions on the [Installation](https://support.iluvatar.com/#/DocumentCentre?id=1&nameCenter=2&productId=520117912052801536)
 2. Launch ComfyUI by running `python main.py`
 
+
+## [ComfyUI-Manager](https://github.com/Comfy-Org/ComfyUI-Manager/tree/manager-v4)
+
+**ComfyUI-Manager** is an extension that allows you to easily install, update, and manage custom nodes for ComfyUI.
+
+### Setup
+
+1. Install the manager dependencies:
+   ```bash
+   pip install -r manager_requirements.txt
+   ```
+
+2. Enable the manager with the `--enable-manager` flag when running ComfyUI:
+   ```bash
+   python main.py --enable-manager
+   ```
+
+### Command Line Options
+
+| Flag | Description |
+|------|-------------|
+| `--enable-manager` | Enable ComfyUI-Manager |
+| `--enable-manager-legacy-ui` | Use the legacy manager UI instead of the new UI (requires `--enable-manager`) |
+| `--disable-manager-ui` | Disable the manager UI and endpoints while keeping background features like security checks and scheduled installation completion (requires `--enable-manager`) |
+
+
 # Running
 
 ```python main.py```

alembic_db/versions/0001_assets.py

@@ -0,0 +1,174 @@
+"""
+Initial assets schema
+Revision ID: 0001_assets
+Revises: None
+Create Date: 2025-12-10 00:00:00
+"""
+
+from alembic import op
+import sqlalchemy as sa
+
+revision = "0001_assets"
+down_revision = None
+branch_labels = None
+depends_on = None
+
+
+def upgrade() -> None:
+    # ASSETS: content identity
+    op.create_table(
+        "assets",
+        sa.Column("id", sa.String(length=36), primary_key=True),
+        sa.Column("hash", sa.String(length=256), nullable=True),
+        sa.Column("size_bytes", sa.BigInteger(), nullable=False, server_default="0"),
+        sa.Column("mime_type", sa.String(length=255), nullable=True),
+        sa.Column("created_at", sa.DateTime(timezone=False), nullable=False),
+        sa.CheckConstraint("size_bytes >= 0", name="ck_assets_size_nonneg"),
+    )
+    op.create_index("uq_assets_hash", "assets", ["hash"], unique=True)
+    op.create_index("ix_assets_mime_type", "assets", ["mime_type"])
+
+    # ASSETS_INFO: user-visible references
+    op.create_table(
+        "assets_info",
+        sa.Column("id", sa.String(length=36), primary_key=True),
+        sa.Column("owner_id", sa.String(length=128), nullable=False, server_default=""),
+        sa.Column("name", sa.String(length=512), nullable=False),
+        sa.Column("asset_id", sa.String(length=36), sa.ForeignKey("assets.id", ondelete="RESTRICT"), nullable=False),
+        sa.Column("preview_id", sa.String(length=36), sa.ForeignKey("assets.id", ondelete="SET NULL"), nullable=True),
+        sa.Column("user_metadata", sa.JSON(), nullable=True),
+        sa.Column("created_at", sa.DateTime(timezone=False), nullable=False),
+        sa.Column("updated_at", sa.DateTime(timezone=False), nullable=False),
+        sa.Column("last_access_time", sa.DateTime(timezone=False), nullable=False),
+        sa.UniqueConstraint("asset_id", "owner_id", "name", name="uq_assets_info_asset_owner_name"),
+    )
+    op.create_index("ix_assets_info_owner_id", "assets_info", ["owner_id"])
+    op.create_index("ix_assets_info_asset_id", "assets_info", ["asset_id"])
+    op.create_index("ix_assets_info_name", "assets_info", ["name"])
+    op.create_index("ix_assets_info_created_at", "assets_info", ["created_at"])
+    op.create_index("ix_assets_info_last_access_time", "assets_info", ["last_access_time"])
+    op.create_index("ix_assets_info_owner_name", "assets_info", ["owner_id", "name"])
+
+    # TAGS: normalized tag vocabulary
+    op.create_table(
+        "tags",
+        sa.Column("name", sa.String(length=512), primary_key=True),
+        sa.Column("tag_type", sa.String(length=32), nullable=False, server_default="user"),
+        sa.CheckConstraint("name = lower(name)", name="ck_tags_lowercase"),
+    )
+    op.create_index("ix_tags_tag_type", "tags", ["tag_type"])
+
+    # ASSET_INFO_TAGS: many-to-many for tags on AssetInfo
+    op.create_table(
+        "asset_info_tags",
+        sa.Column("asset_info_id", sa.String(length=36), sa.ForeignKey("assets_info.id", ondelete="CASCADE"), nullable=False),
+        sa.Column("tag_name", sa.String(length=512), sa.ForeignKey("tags.name", ondelete="RESTRICT"), nullable=False),
+        sa.Column("origin", sa.String(length=32), nullable=False, server_default="manual"),
+        sa.Column("added_at", sa.DateTime(timezone=False), nullable=False),
+        sa.PrimaryKeyConstraint("asset_info_id", "tag_name", name="pk_asset_info_tags"),
+    )
+    op.create_index("ix_asset_info_tags_tag_name", "asset_info_tags", ["tag_name"])
+    op.create_index("ix_asset_info_tags_asset_info_id", "asset_info_tags", ["asset_info_id"])
+
+    # ASSET_CACHE_STATE: N:1 local cache rows per Asset
+    op.create_table(
+        "asset_cache_state",
+        sa.Column("id", sa.Integer(), primary_key=True, autoincrement=True),
+        sa.Column("asset_id", sa.String(length=36), sa.ForeignKey("assets.id", ondelete="CASCADE"), nullable=False),
+        sa.Column("file_path", sa.Text(), nullable=False),  # absolute local path to cached file
+        sa.Column("mtime_ns", sa.BigInteger(), nullable=True),
+        sa.Column("needs_verify", sa.Boolean(), nullable=False, server_default=sa.text("false")),
+        sa.CheckConstraint("(mtime_ns IS NULL) OR (mtime_ns >= 0)", name="ck_acs_mtime_nonneg"),
+        sa.UniqueConstraint("file_path", name="uq_asset_cache_state_file_path"),
+    )
+    op.create_index("ix_asset_cache_state_file_path", "asset_cache_state", ["file_path"])
+    op.create_index("ix_asset_cache_state_asset_id", "asset_cache_state", ["asset_id"])
+
+    # ASSET_INFO_META: typed KV projection of user_metadata for filtering/sorting
+    op.create_table(
+        "asset_info_meta",
+        sa.Column("asset_info_id", sa.String(length=36), sa.ForeignKey("assets_info.id", ondelete="CASCADE"), nullable=False),
+        sa.Column("key", sa.String(length=256), nullable=False),
+        sa.Column("ordinal", sa.Integer(), nullable=False, server_default="0"),
+        sa.Column("val_str", sa.String(length=2048), nullable=True),
+        sa.Column("val_num", sa.Numeric(38, 10), nullable=True),
+        sa.Column("val_bool", sa.Boolean(), nullable=True),
+        sa.Column("val_json", sa.JSON(), nullable=True),
+        sa.PrimaryKeyConstraint("asset_info_id", "key", "ordinal", name="pk_asset_info_meta"),
+    )
+    op.create_index("ix_asset_info_meta_key", "asset_info_meta", ["key"])
+    op.create_index("ix_asset_info_meta_key_val_str", "asset_info_meta", ["key", "val_str"])
+    op.create_index("ix_asset_info_meta_key_val_num", "asset_info_meta", ["key", "val_num"])
+    op.create_index("ix_asset_info_meta_key_val_bool", "asset_info_meta", ["key", "val_bool"])
+
+    # Tags vocabulary
+    tags_table = sa.table(
+        "tags",
+        sa.column("name", sa.String(length=512)),
+        sa.column("tag_type", sa.String()),
+    )
+    op.bulk_insert(
+        tags_table,
+        [
+            {"name": "models", "tag_type": "system"},
+            {"name": "input", "tag_type": "system"},
+            {"name": "output", "tag_type": "system"},
+
+            {"name": "configs", "tag_type": "system"},
+            {"name": "checkpoints", "tag_type": "system"},
+            {"name": "loras", "tag_type": "system"},
+            {"name": "vae", "tag_type": "system"},
+            {"name": "text_encoders", "tag_type": "system"},
+            {"name": "diffusion_models", "tag_type": "system"},
+            {"name": "clip_vision", "tag_type": "system"},
+            {"name": "style_models", "tag_type": "system"},
+            {"name": "embeddings", "tag_type": "system"},
+            {"name": "diffusers", "tag_type": "system"},
+            {"name": "vae_approx", "tag_type": "system"},
+            {"name": "controlnet", "tag_type": "system"},
+            {"name": "gligen", "tag_type": "system"},
+            {"name": "upscale_models", "tag_type": "system"},
+            {"name": "hypernetworks", "tag_type": "system"},
+            {"name": "photomaker", "tag_type": "system"},
+            {"name": "classifiers", "tag_type": "system"},
+
+            {"name": "encoder", "tag_type": "system"},
+            {"name": "decoder", "tag_type": "system"},
+
+            {"name": "missing", "tag_type": "system"},
+            {"name": "rescan", "tag_type": "system"},
+        ],
+    )
+
+
+def downgrade() -> None:
+    op.drop_index("ix_asset_info_meta_key_val_bool", table_name="asset_info_meta")
+    op.drop_index("ix_asset_info_meta_key_val_num", table_name="asset_info_meta")
+    op.drop_index("ix_asset_info_meta_key_val_str", table_name="asset_info_meta")
+    op.drop_index("ix_asset_info_meta_key", table_name="asset_info_meta")
+    op.drop_table("asset_info_meta")
+
+    op.drop_index("ix_asset_cache_state_asset_id", table_name="asset_cache_state")
+    op.drop_index("ix_asset_cache_state_file_path", table_name="asset_cache_state")
+    op.drop_constraint("uq_asset_cache_state_file_path", table_name="asset_cache_state")
+    op.drop_table("asset_cache_state")
+
+    op.drop_index("ix_asset_info_tags_asset_info_id", table_name="asset_info_tags")
+    op.drop_index("ix_asset_info_tags_tag_name", table_name="asset_info_tags")
+    op.drop_table("asset_info_tags")
+
+    op.drop_index("ix_tags_tag_type", table_name="tags")
+    op.drop_table("tags")
+
+    op.drop_constraint("uq_assets_info_asset_owner_name", table_name="assets_info")
+    op.drop_index("ix_assets_info_owner_name", table_name="assets_info")
+    op.drop_index("ix_assets_info_last_access_time", table_name="assets_info")
+    op.drop_index("ix_assets_info_created_at", table_name="assets_info")
+    op.drop_index("ix_assets_info_name", table_name="assets_info")
+    op.drop_index("ix_assets_info_asset_id", table_name="assets_info")
+    op.drop_index("ix_assets_info_owner_id", table_name="assets_info")
+    op.drop_table("assets_info")
+
+    op.drop_index("uq_assets_hash", table_name="assets")
+    op.drop_index("ix_assets_mime_type", table_name="assets")
+    op.drop_table("assets")

api_server/routes/internal/internal_routes.py

@@ -58,8 +58,13 @@ class InternalRoutes:
                 return web.json_response({"error": "Invalid directory type"}, status=400)
 
             directory = get_directory_by_type(directory_type)
+
+            def is_visible_file(entry: os.DirEntry) -> bool:
+                """Filter out hidden files (e.g., .DS_Store on macOS)."""
+                return entry.is_file() and not entry.name.startswith('.')
+
             sorted_files = sorted(
-                (entry for entry in os.scandir(directory) if entry.is_file()),
+                (entry for entry in os.scandir(directory) if is_visible_file(entry)),
                 key=lambda entry: -entry.stat().st_mtime
             )
             return web.json_response([entry.name for entry in sorted_files], status=200)

app/assets/api/routes.py

@@ -0,0 +1,514 @@
+import logging
+import uuid
+import urllib.parse
+import os
+import contextlib
+from aiohttp import web
+
+from pydantic import ValidationError
+
+import app.assets.manager as manager
+from app import user_manager
+from app.assets.api import schemas_in
+from app.assets.helpers import get_query_dict
+from app.assets.scanner import seed_assets
+
+import folder_paths
+
+ROUTES = web.RouteTableDef()
+USER_MANAGER: user_manager.UserManager | None = None
+
+# UUID regex (canonical hyphenated form, case-insensitive)
+UUID_RE = r"[0-9a-fA-F]{8}-[0-9a-fA-F]{4}-[0-9a-fA-F]{4}-[0-9a-fA-F]{4}-[0-9a-fA-F]{12}"
+
+# Note to any custom node developers reading this code:
+# The assets system is not yet fully implemented, do not rely on the code in /app/assets remaining the same.
+
+def register_assets_system(app: web.Application, user_manager_instance: user_manager.UserManager) -> None:
+    global USER_MANAGER
+    USER_MANAGER = user_manager_instance
+    app.add_routes(ROUTES)
+
+def _error_response(status: int, code: str, message: str, details: dict | None = None) -> web.Response:
+    return web.json_response({"error": {"code": code, "message": message, "details": details or {}}}, status=status)
+
+
+def _validation_error_response(code: str, ve: ValidationError) -> web.Response:
+    return _error_response(400, code, "Validation failed.", {"errors": ve.json()})
+
+
+@ROUTES.head("/api/assets/hash/{hash}")
+async def head_asset_by_hash(request: web.Request) -> web.Response:
+    hash_str = request.match_info.get("hash", "").strip().lower()
+    if not hash_str or ":" not in hash_str:
+        return _error_response(400, "INVALID_HASH", "hash must be like 'blake3:<hex>'")
+    algo, digest = hash_str.split(":", 1)
+    if algo != "blake3" or not digest or any(c for c in digest if c not in "0123456789abcdef"):
+        return _error_response(400, "INVALID_HASH", "hash must be like 'blake3:<hex>'")
+    exists = manager.asset_exists(asset_hash=hash_str)
+    return web.Response(status=200 if exists else 404)
+
+
+@ROUTES.get("/api/assets")
+async def list_assets(request: web.Request) -> web.Response:
+    """
+    GET request to list assets.
+    """
+    query_dict = get_query_dict(request)
+    try:
+        q = schemas_in.ListAssetsQuery.model_validate(query_dict)
+    except ValidationError as ve:
+        return _validation_error_response("INVALID_QUERY", ve)
+
+    payload = manager.list_assets(
+        include_tags=q.include_tags,
+        exclude_tags=q.exclude_tags,
+        name_contains=q.name_contains,
+        metadata_filter=q.metadata_filter,
+        limit=q.limit,
+        offset=q.offset,
+        sort=q.sort,
+        order=q.order,
+        owner_id=USER_MANAGER.get_request_user_id(request),
+    )
+    return web.json_response(payload.model_dump(mode="json", exclude_none=True))
+
+
+@ROUTES.get(f"/api/assets/{{id:{UUID_RE}}}")
+async def get_asset(request: web.Request) -> web.Response:
+    """
+    GET request to get an asset's info as JSON.
+    """
+    asset_info_id = str(uuid.UUID(request.match_info["id"]))
+    try:
+        result = manager.get_asset(
+            asset_info_id=asset_info_id,
+            owner_id=USER_MANAGER.get_request_user_id(request),
+        )
+    except ValueError as e:
+        return _error_response(404, "ASSET_NOT_FOUND", str(e), {"id": asset_info_id})
+    except Exception:
+        logging.exception(
+            "get_asset failed for asset_info_id=%s, owner_id=%s",
+            asset_info_id,
+            USER_MANAGER.get_request_user_id(request),
+        )
+        return _error_response(500, "INTERNAL", "Unexpected server error.")
+    return web.json_response(result.model_dump(mode="json"), status=200)
+
+
+@ROUTES.get(f"/api/assets/{{id:{UUID_RE}}}/content")
+async def download_asset_content(request: web.Request) -> web.Response:
+    # question: do we need disposition? could we just stick with one of these?
+    disposition = request.query.get("disposition", "attachment").lower().strip()
+    if disposition not in {"inline", "attachment"}:
+        disposition = "attachment"
+
+    try:
+        abs_path, content_type, filename = manager.resolve_asset_content_for_download(
+            asset_info_id=str(uuid.UUID(request.match_info["id"])),
+            owner_id=USER_MANAGER.get_request_user_id(request),
+        )
+    except ValueError as ve:
+        return _error_response(404, "ASSET_NOT_FOUND", str(ve))
+    except NotImplementedError as nie:
+        return _error_response(501, "BACKEND_UNSUPPORTED", str(nie))
+    except FileNotFoundError:
+        return _error_response(404, "FILE_NOT_FOUND", "Underlying file not found on disk.")
+
+    quoted = (filename or "").replace("\r", "").replace("\n", "").replace('"', "'")
+    cd = f'{disposition}; filename="{quoted}"; filename*=UTF-8\'\'{urllib.parse.quote(filename)}'
+
+    file_size = os.path.getsize(abs_path)
+    logging.info(
+        "download_asset_content: path=%s, size=%d bytes (%.2f MB), content_type=%s, filename=%s",
+        abs_path,
+        file_size,
+        file_size / (1024 * 1024),
+        content_type,
+        filename,
+    )
+
+    async def file_sender():
+        chunk_size = 64 * 1024
+        with open(abs_path, "rb") as f:
+            while True:
+                chunk = f.read(chunk_size)
+                if not chunk:
+                    break
+                yield chunk
+
+    return web.Response(
+        body=file_sender(),
+        content_type=content_type,
+        headers={
+            "Content-Disposition": cd,
+            "Content-Length": str(file_size),
+        },
+    )
+
+
+@ROUTES.post("/api/assets/from-hash")
+async def create_asset_from_hash(request: web.Request) -> web.Response:
+    try:
+        payload = await request.json()
+        body = schemas_in.CreateFromHashBody.model_validate(payload)
+    except ValidationError as ve:
+        return _validation_error_response("INVALID_BODY", ve)
+    except Exception:
+        return _error_response(400, "INVALID_JSON", "Request body must be valid JSON.")
+
+    result = manager.create_asset_from_hash(
+        hash_str=body.hash,
+        name=body.name,
+        tags=body.tags,
+        user_metadata=body.user_metadata,
+        owner_id=USER_MANAGER.get_request_user_id(request),
+    )
+    if result is None:
+        return _error_response(404, "ASSET_NOT_FOUND", f"Asset content {body.hash} does not exist")
+    return web.json_response(result.model_dump(mode="json"), status=201)
+
+
+@ROUTES.post("/api/assets")
+async def upload_asset(request: web.Request) -> web.Response:
+    """Multipart/form-data endpoint for Asset uploads."""
+    if not (request.content_type or "").lower().startswith("multipart/"):
+        return _error_response(415, "UNSUPPORTED_MEDIA_TYPE", "Use multipart/form-data for uploads.")
+
+    reader = await request.multipart()
+
+    file_present = False
+    file_client_name: str | None = None
+    tags_raw: list[str] = []
+    provided_name: str | None = None
+    user_metadata_raw: str | None = None
+    provided_hash: str | None = None
+    provided_hash_exists: bool | None = None
+
+    file_written = 0
+    tmp_path: str | None = None
+    while True:
+        field = await reader.next()
+        if field is None:
+            break
+
+        fname = getattr(field, "name", "") or ""
+
+        if fname == "hash":
+            try:
+                s = ((await field.text()) or "").strip().lower()
+            except Exception:
+                return _error_response(400, "INVALID_HASH", "hash must be like 'blake3:<hex>'")
+
+            if s:
+                if ":" not in s:
+                    return _error_response(400, "INVALID_HASH", "hash must be like 'blake3:<hex>'")
+                algo, digest = s.split(":", 1)
+                if algo != "blake3" or not digest or any(c for c in digest if c not in "0123456789abcdef"):
+                    return _error_response(400, "INVALID_HASH", "hash must be like 'blake3:<hex>'")
+                provided_hash = f"{algo}:{digest}"
+                try:
+                    provided_hash_exists = manager.asset_exists(asset_hash=provided_hash)
+                except Exception:
+                    provided_hash_exists = None  # do not fail the whole request here
+
+        elif fname == "file":
+            file_present = True
+            file_client_name = (field.filename or "").strip()
+
+            if provided_hash and provided_hash_exists is True:
+                # If client supplied a hash that we know exists, drain but do not write to disk
+                try:
+                    while True:
+                        chunk = await field.read_chunk(8 * 1024 * 1024)
+                        if not chunk:
+                            break
+                        file_written += len(chunk)
+                except Exception:
+                    return _error_response(500, "UPLOAD_IO_ERROR", "Failed to receive uploaded file.")
+                continue  # Do not create temp file; we will create AssetInfo from the existing content
+
+            # Otherwise, store to temp for hashing/ingest
+            uploads_root = os.path.join(folder_paths.get_temp_directory(), "uploads")
+            unique_dir = os.path.join(uploads_root, uuid.uuid4().hex)
+            os.makedirs(unique_dir, exist_ok=True)
+            tmp_path = os.path.join(unique_dir, ".upload.part")
+
+            try:
+                with open(tmp_path, "wb") as f:
+                    while True:
+                        chunk = await field.read_chunk(8 * 1024 * 1024)
+                        if not chunk:
+                            break
+                        f.write(chunk)
+                        file_written += len(chunk)
+            except Exception:
+                try:
+                    if os.path.exists(tmp_path or ""):
+                        os.remove(tmp_path)
+                finally:
+                    return _error_response(500, "UPLOAD_IO_ERROR", "Failed to receive and store uploaded file.")
+        elif fname == "tags":
+            tags_raw.append((await field.text()) or "")
+        elif fname == "name":
+            provided_name = (await field.text()) or None
+        elif fname == "user_metadata":
+            user_metadata_raw = (await field.text()) or None
+
+    # If client did not send file, and we are not doing a from-hash fast path -> error
+    if not file_present and not (provided_hash and provided_hash_exists):
+        return _error_response(400, "MISSING_FILE", "Form must include a 'file' part or a known 'hash'.")
+
+    if file_present and file_written == 0 and not (provided_hash and provided_hash_exists):
+        # Empty upload is only acceptable if we are fast-pathing from existing hash
+        try:
+            if tmp_path and os.path.exists(tmp_path):
+                os.remove(tmp_path)
+        finally:
+            return _error_response(400, "EMPTY_UPLOAD", "Uploaded file is empty.")
+
+    try:
+        spec = schemas_in.UploadAssetSpec.model_validate({
+            "tags": tags_raw,
+            "name": provided_name,
+            "user_metadata": user_metadata_raw,
+            "hash": provided_hash,
+        })
+    except ValidationError as ve:
+        try:
+            if tmp_path and os.path.exists(tmp_path):
+                os.remove(tmp_path)
+        finally:
+            return _validation_error_response("INVALID_BODY", ve)
+
+    # Validate models category against configured folders (consistent with previous behavior)
+    if spec.tags and spec.tags[0] == "models":
+        if len(spec.tags) < 2 or spec.tags[1] not in folder_paths.folder_names_and_paths:
+            if tmp_path and os.path.exists(tmp_path):
+                os.remove(tmp_path)
+            return _error_response(
+                400, "INVALID_BODY", f"unknown models category '{spec.tags[1] if len(spec.tags) >= 2 else ''}'"
+            )
+
+    owner_id = USER_MANAGER.get_request_user_id(request)
+
+    # Fast path: if a valid provided hash exists, create AssetInfo without writing anything
+    if spec.hash and provided_hash_exists is True:
+        try:
+            result = manager.create_asset_from_hash(
+                hash_str=spec.hash,
+                name=spec.name or (spec.hash.split(":", 1)[1]),
+                tags=spec.tags,
+                user_metadata=spec.user_metadata or {},
+                owner_id=owner_id,
+            )
+        except Exception:
+            logging.exception("create_asset_from_hash failed for hash=%s, owner_id=%s", spec.hash, owner_id)
+            return _error_response(500, "INTERNAL", "Unexpected server error.")
+
+        if result is None:
+            return _error_response(404, "ASSET_NOT_FOUND", f"Asset content {spec.hash} does not exist")
+
+        # Drain temp if we accidentally saved (e.g., hash field came after file)
+        if tmp_path and os.path.exists(tmp_path):
+            with contextlib.suppress(Exception):
+                os.remove(tmp_path)
+
+        status = 200 if (not result.created_new) else 201
+        return web.json_response(result.model_dump(mode="json"), status=status)
+
+    # Otherwise, we must have a temp file path to ingest
+    if not tmp_path or not os.path.exists(tmp_path):
+        # The only case we reach here without a temp file is: client sent a hash that does not exist and no file
+        return _error_response(404, "ASSET_NOT_FOUND", "Provided hash not found and no file uploaded.")
+
+    try:
+        created = manager.upload_asset_from_temp_path(
+            spec,
+            temp_path=tmp_path,
+            client_filename=file_client_name,
+            owner_id=owner_id,
+            expected_asset_hash=spec.hash,
+        )
+        status = 201 if created.created_new else 200
+        return web.json_response(created.model_dump(mode="json"), status=status)
+    except ValueError as e:
+        if tmp_path and os.path.exists(tmp_path):
+            os.remove(tmp_path)
+        msg = str(e)
+        if "HASH_MISMATCH" in msg or msg.strip().upper() == "HASH_MISMATCH":
+            return _error_response(
+                400,
+                "HASH_MISMATCH",
+                "Uploaded file hash does not match provided hash.",
+            )
+        return _error_response(400, "BAD_REQUEST", "Invalid inputs.")
+    except Exception:
+        if tmp_path and os.path.exists(tmp_path):
+            os.remove(tmp_path)
+        logging.exception("upload_asset_from_temp_path failed for tmp_path=%s, owner_id=%s", tmp_path, owner_id)
+        return _error_response(500, "INTERNAL", "Unexpected server error.")
+
+
+@ROUTES.put(f"/api/assets/{{id:{UUID_RE}}}")
+async def update_asset(request: web.Request) -> web.Response:
+    asset_info_id = str(uuid.UUID(request.match_info["id"]))
+    try:
+        body = schemas_in.UpdateAssetBody.model_validate(await request.json())
+    except ValidationError as ve:
+        return _validation_error_response("INVALID_BODY", ve)
+    except Exception:
+        return _error_response(400, "INVALID_JSON", "Request body must be valid JSON.")
+
+    try:
+        result = manager.update_asset(
+            asset_info_id=asset_info_id,
+            name=body.name,
+            user_metadata=body.user_metadata,
+            owner_id=USER_MANAGER.get_request_user_id(request),
+        )
+    except (ValueError, PermissionError) as ve:
+        return _error_response(404, "ASSET_NOT_FOUND", str(ve), {"id": asset_info_id})
+    except Exception:
+        logging.exception(
+            "update_asset failed for asset_info_id=%s, owner_id=%s",
+            asset_info_id,
+            USER_MANAGER.get_request_user_id(request),
+        )
+        return _error_response(500, "INTERNAL", "Unexpected server error.")
+    return web.json_response(result.model_dump(mode="json"), status=200)
+
+
+@ROUTES.delete(f"/api/assets/{{id:{UUID_RE}}}")
+async def delete_asset(request: web.Request) -> web.Response:
+    asset_info_id = str(uuid.UUID(request.match_info["id"]))
+    delete_content = request.query.get("delete_content")
+    delete_content = True if delete_content is None else delete_content.lower() not in {"0", "false", "no"}
+
+    try:
+        deleted = manager.delete_asset_reference(
+            asset_info_id=asset_info_id,
+            owner_id=USER_MANAGER.get_request_user_id(request),
+            delete_content_if_orphan=delete_content,
+        )
+    except Exception:
+        logging.exception(
+            "delete_asset_reference failed for asset_info_id=%s, owner_id=%s",
+            asset_info_id,
+            USER_MANAGER.get_request_user_id(request),
+        )
+        return _error_response(500, "INTERNAL", "Unexpected server error.")
+
+    if not deleted:
+        return _error_response(404, "ASSET_NOT_FOUND", f"AssetInfo {asset_info_id} not found.")
+    return web.Response(status=204)
+
+
+@ROUTES.get("/api/tags")
+async def get_tags(request: web.Request) -> web.Response:
+    """
+    GET request to list all tags based on query parameters.
+    """
+    query_map = dict(request.rel_url.query)
+
+    try:
+        query = schemas_in.TagsListQuery.model_validate(query_map)
+    except ValidationError as e:
+        return web.json_response(
+            {"error": {"code": "INVALID_QUERY", "message": "Invalid query parameters", "details": e.errors()}},
+            status=400,
+        )
+
+    result = manager.list_tags(
+        prefix=query.prefix,
+        limit=query.limit,
+        offset=query.offset,
+        order=query.order,
+        include_zero=query.include_zero,
+        owner_id=USER_MANAGER.get_request_user_id(request),
+    )
+    return web.json_response(result.model_dump(mode="json"))
+
+
+@ROUTES.post(f"/api/assets/{{id:{UUID_RE}}}/tags")
+async def add_asset_tags(request: web.Request) -> web.Response:
+    asset_info_id = str(uuid.UUID(request.match_info["id"]))
+    try:
+        payload = await request.json()
+        data = schemas_in.TagsAdd.model_validate(payload)
+    except ValidationError as ve:
+        return _error_response(400, "INVALID_BODY", "Invalid JSON body for tags add.", {"errors": ve.errors()})
+    except Exception:
+        return _error_response(400, "INVALID_JSON", "Request body must be valid JSON.")
+
+    try:
+        result = manager.add_tags_to_asset(
+            asset_info_id=asset_info_id,
+            tags=data.tags,
+            origin="manual",
+            owner_id=USER_MANAGER.get_request_user_id(request),
+        )
+    except (ValueError, PermissionError) as ve:
+        return _error_response(404, "ASSET_NOT_FOUND", str(ve), {"id": asset_info_id})
+    except Exception:
+        logging.exception(
+            "add_tags_to_asset failed for asset_info_id=%s, owner_id=%s",
+            asset_info_id,
+            USER_MANAGER.get_request_user_id(request),
+        )
+        return _error_response(500, "INTERNAL", "Unexpected server error.")
+
+    return web.json_response(result.model_dump(mode="json"), status=200)
+
+
+@ROUTES.delete(f"/api/assets/{{id:{UUID_RE}}}/tags")
+async def delete_asset_tags(request: web.Request) -> web.Response:
+    asset_info_id = str(uuid.UUID(request.match_info["id"]))
+    try:
+        payload = await request.json()
+        data = schemas_in.TagsRemove.model_validate(payload)
+    except ValidationError as ve:
+        return _error_response(400, "INVALID_BODY", "Invalid JSON body for tags remove.", {"errors": ve.errors()})
+    except Exception:
+        return _error_response(400, "INVALID_JSON", "Request body must be valid JSON.")
+
+    try:
+        result = manager.remove_tags_from_asset(
+            asset_info_id=asset_info_id,
+            tags=data.tags,
+            owner_id=USER_MANAGER.get_request_user_id(request),
+        )
+    except ValueError as ve:
+        return _error_response(404, "ASSET_NOT_FOUND", str(ve), {"id": asset_info_id})
+    except Exception:
+        logging.exception(
+            "remove_tags_from_asset failed for asset_info_id=%s, owner_id=%s",
+            asset_info_id,
+            USER_MANAGER.get_request_user_id(request),
+        )
+        return _error_response(500, "INTERNAL", "Unexpected server error.")
+
+    return web.json_response(result.model_dump(mode="json"), status=200)
+
+
+@ROUTES.post("/api/assets/seed")
+async def seed_assets_endpoint(request: web.Request) -> web.Response:
+    """Trigger asset seeding for specified roots (models, input, output)."""
+    try:
+        payload = await request.json()
+        roots = payload.get("roots", ["models", "input", "output"])
+    except Exception:
+        roots = ["models", "input", "output"]
+
+    valid_roots = [r for r in roots if r in ("models", "input", "output")]
+    if not valid_roots:
+        return _error_response(400, "INVALID_BODY", "No valid roots specified")
+
+    try:
+        seed_assets(tuple(valid_roots))
+    except Exception:
+        logging.exception("seed_assets failed for roots=%s", valid_roots)
+        return _error_response(500, "INTERNAL", "Seed operation failed")
+
+    return web.json_response({"seeded": valid_roots}, status=200)

app/assets/api/schemas_in.py

@@ -0,0 +1,264 @@
+import json
+from typing import Any, Literal
+
+from pydantic import (
+    BaseModel,
+    ConfigDict,
+    Field,
+    conint,
+    field_validator,
+    model_validator,
+)
+
+class ListAssetsQuery(BaseModel):
+    include_tags: list[str] = Field(default_factory=list)
+    exclude_tags: list[str] = Field(default_factory=list)
+    name_contains: str | None = None
+
+    # Accept either a JSON string (query param) or a dict
+    metadata_filter: dict[str, Any] | None = None
+
+    limit: conint(ge=1, le=500) = 20
+    offset: conint(ge=0) = 0
+
+    sort: Literal["name", "created_at", "updated_at", "size", "last_access_time"] = "created_at"
+    order: Literal["asc", "desc"] = "desc"
+
+    @field_validator("include_tags", "exclude_tags", mode="before")
+    @classmethod
+    def _split_csv_tags(cls, v):
+        # Accept "a,b,c" or ["a","b"] (we are liberal in what we accept)
+        if v is None:
+            return []
+        if isinstance(v, str):
+            return [t.strip() for t in v.split(",") if t.strip()]
+        if isinstance(v, list):
+            out: list[str] = []
+            for item in v:
+                if isinstance(item, str):
+                    out.extend([t.strip() for t in item.split(",") if t.strip()])
+            return out
+        return v
+
+    @field_validator("metadata_filter", mode="before")
+    @classmethod
+    def _parse_metadata_json(cls, v):
+        if v is None or isinstance(v, dict):
+            return v
+        if isinstance(v, str) and v.strip():
+            try:
+                parsed = json.loads(v)
+            except Exception as e:
+                raise ValueError(f"metadata_filter must be JSON: {e}") from e
+            if not isinstance(parsed, dict):
+                raise ValueError("metadata_filter must be a JSON object")
+            return parsed
+        return None
+
+
+class UpdateAssetBody(BaseModel):
+    name: str | None = None
+    user_metadata: dict[str, Any] | None = None
+
+    @model_validator(mode="after")
+    def _at_least_one(self):
+        if self.name is None and self.user_metadata is None:
+            raise ValueError("Provide at least one of: name, user_metadata.")
+        return self
+
+
+class CreateFromHashBody(BaseModel):
+    model_config = ConfigDict(extra="ignore", str_strip_whitespace=True)
+
+    hash: str
+    name: str
+    tags: list[str] = Field(default_factory=list)
+    user_metadata: dict[str, Any] = Field(default_factory=dict)
+
+    @field_validator("hash")
+    @classmethod
+    def _require_blake3(cls, v):
+        s = (v or "").strip().lower()
+        if ":" not in s:
+            raise ValueError("hash must be 'blake3:<hex>'")
+        algo, digest = s.split(":", 1)
+        if algo != "blake3":
+            raise ValueError("only canonical 'blake3:<hex>' is accepted here")
+        if not digest or any(c for c in digest if c not in "0123456789abcdef"):
+            raise ValueError("hash digest must be lowercase hex")
+        return s
+
+    @field_validator("tags", mode="before")
+    @classmethod
+    def _tags_norm(cls, v):
+        if v is None:
+            return []
+        if isinstance(v, list):
+            out = [str(t).strip().lower() for t in v if str(t).strip()]
+            seen = set()
+            dedup = []
+            for t in out:
+                if t not in seen:
+                    seen.add(t)
+                    dedup.append(t)
+            return dedup
+        if isinstance(v, str):
+            return [t.strip().lower() for t in v.split(",") if t.strip()]
+        return []
+
+
+class TagsListQuery(BaseModel):
+    model_config = ConfigDict(extra="ignore", str_strip_whitespace=True)
+
+    prefix: str | None = Field(None, min_length=1, max_length=256)
+    limit: int = Field(100, ge=1, le=1000)
+    offset: int = Field(0, ge=0, le=10_000_000)
+    order: Literal["count_desc", "name_asc"] = "count_desc"
+    include_zero: bool = True
+
+    @field_validator("prefix")
+    @classmethod
+    def normalize_prefix(cls, v: str | None) -> str | None:
+        if v is None:
+            return v
+        v = v.strip()
+        return v.lower() or None
+
+
+class TagsAdd(BaseModel):
+    model_config = ConfigDict(extra="ignore")
+    tags: list[str] = Field(..., min_length=1)
+
+    @field_validator("tags")
+    @classmethod
+    def normalize_tags(cls, v: list[str]) -> list[str]:
+        out = []
+        for t in v:
+            if not isinstance(t, str):
+                raise TypeError("tags must be strings")
+            tnorm = t.strip().lower()
+            if tnorm:
+                out.append(tnorm)
+        seen = set()
+        deduplicated = []
+        for x in out:
+            if x not in seen:
+                seen.add(x)
+                deduplicated.append(x)
+        return deduplicated
+
+
+class TagsRemove(TagsAdd):
+    pass
+
+
+class UploadAssetSpec(BaseModel):
+    """Upload Asset operation.
+    - tags: ordered; first is root ('models'|'input'|'output');
+            if root == 'models', second must be a valid category from folder_paths.folder_names_and_paths
+    - name: display name
+    - user_metadata: arbitrary JSON object (optional)
+    - hash: optional canonical 'blake3:<hex>' provided by the client for validation / fast-path
+
+    Files created via this endpoint are stored on disk using the **content hash** as the filename stem
+    and the original extension is preserved when available.
+    """
+    model_config = ConfigDict(extra="ignore", str_strip_whitespace=True)
+
+    tags: list[str] = Field(..., min_length=1)
+    name: str | None = Field(default=None, max_length=512, description="Display Name")
+    user_metadata: dict[str, Any] = Field(default_factory=dict)
+    hash: str | None = Field(default=None)
+
+    @field_validator("hash", mode="before")
+    @classmethod
+    def _parse_hash(cls, v):
+        if v is None:
+            return None
+        s = str(v).strip().lower()
+        if not s:
+            return None
+        if ":" not in s:
+            raise ValueError("hash must be 'blake3:<hex>'")
+        algo, digest = s.split(":", 1)
+        if algo != "blake3":
+            raise ValueError("only canonical 'blake3:<hex>' is accepted here")
+        if not digest or any(c for c in digest if c not in "0123456789abcdef"):
+            raise ValueError("hash digest must be lowercase hex")
+        return f"{algo}:{digest}"
+
+    @field_validator("tags", mode="before")
+    @classmethod
+    def _parse_tags(cls, v):
+        """
+        Accepts a list of strings (possibly multiple form fields),
+        where each string can be:
+          - JSON array (e.g., '["models","loras","foo"]')
+          - comma-separated ('models, loras, foo')
+          - single token ('models')
+        Returns a normalized, deduplicated, ordered list.
+        """
+        items: list[str] = []
+        if v is None:
+            return []
+        if isinstance(v, str):
+            v = [v]
+
+        if isinstance(v, list):
+            for item in v:
+                if item is None:
+                    continue
+                s = str(item).strip()
+                if not s:
+                    continue
+                if s.startswith("["):
+                    try:
+                        arr = json.loads(s)
+                        if isinstance(arr, list):
+                            items.extend(str(x) for x in arr)
+                            continue
+                    except Exception:
+                        pass  # fallback to CSV parse below
+                items.extend([p for p in s.split(",") if p.strip()])
+        else:
+            return []
+
+        # normalize + dedupe
+        norm = []
+        seen = set()
+        for t in items:
+            tnorm = str(t).strip().lower()
+            if tnorm and tnorm not in seen:
+                seen.add(tnorm)
+                norm.append(tnorm)
+        return norm
+
+    @field_validator("user_metadata", mode="before")
+    @classmethod
+    def _parse_metadata_json(cls, v):
+        if v is None or isinstance(v, dict):
+            return v or {}
+        if isinstance(v, str):
+            s = v.strip()
+            if not s:
+                return {}
+            try:
+                parsed = json.loads(s)
+            except Exception as e:
+                raise ValueError(f"user_metadata must be JSON: {e}") from e
+            if not isinstance(parsed, dict):
+                raise ValueError("user_metadata must be a JSON object")
+            return parsed
+        return {}
+
+    @model_validator(mode="after")
+    def _validate_order(self):
+        if not self.tags:
+            raise ValueError("tags must be provided and non-empty")
+        root = self.tags[0]
+        if root not in {"models", "input", "output"}:
+            raise ValueError("first tag must be one of: models, input, output")
+        if root == "models":
+            if len(self.tags) < 2:
+                raise ValueError("models uploads require a category tag as the second tag")
+        return self

app/assets/api/schemas_out.py

@@ -0,0 +1,93 @@
+from datetime import datetime
+from typing import Any
+
+from pydantic import BaseModel, ConfigDict, Field, field_serializer
+
+
+class AssetSummary(BaseModel):
+    id: str
+    name: str
+    asset_hash: str | None = None
+    size: int | None = None
+    mime_type: str | None = None
+    tags: list[str] = Field(default_factory=list)
+    preview_url: str | None = None
+    created_at: datetime | None = None
+    updated_at: datetime | None = None
+    last_access_time: datetime | None = None
+
+    model_config = ConfigDict(from_attributes=True)
+
+    @field_serializer("created_at", "updated_at", "last_access_time")
+    def _ser_dt(self, v: datetime | None, _info):
+        return v.isoformat() if v else None
+
+
+class AssetsList(BaseModel):
+    assets: list[AssetSummary]
+    total: int
+    has_more: bool
+
+
+class AssetUpdated(BaseModel):
+    id: str
+    name: str
+    asset_hash: str | None = None
+    tags: list[str] = Field(default_factory=list)
+    user_metadata: dict[str, Any] = Field(default_factory=dict)
+    updated_at: datetime | None = None
+
+    model_config = ConfigDict(from_attributes=True)
+
+    @field_serializer("updated_at")
+    def _ser_updated(self, v: datetime | None, _info):
+        return v.isoformat() if v else None
+
+
+class AssetDetail(BaseModel):
+    id: str
+    name: str
+    asset_hash: str | None = None
+    size: int | None = None
+    mime_type: str | None = None
+    tags: list[str] = Field(default_factory=list)
+    user_metadata: dict[str, Any] = Field(default_factory=dict)
+    preview_id: str | None = None
+    created_at: datetime | None = None
+    last_access_time: datetime | None = None
+
+    model_config = ConfigDict(from_attributes=True)
+
+    @field_serializer("created_at", "last_access_time")
+    def _ser_dt(self, v: datetime | None, _info):
+        return v.isoformat() if v else None
+
+
+class AssetCreated(AssetDetail):
+    created_new: bool
+
+
+class TagUsage(BaseModel):
+    name: str
+    count: int
+    type: str
+
+
+class TagsList(BaseModel):
+    tags: list[TagUsage] = Field(default_factory=list)
+    total: int
+    has_more: bool
+
+
+class TagsAdd(BaseModel):
+    model_config = ConfigDict(str_strip_whitespace=True)
+    added: list[str] = Field(default_factory=list)
+    already_present: list[str] = Field(default_factory=list)
+    total_tags: list[str] = Field(default_factory=list)
+
+
+class TagsRemove(BaseModel):
+    model_config = ConfigDict(str_strip_whitespace=True)
+    removed: list[str] = Field(default_factory=list)
+    not_present: list[str] = Field(default_factory=list)
+    total_tags: list[str] = Field(default_factory=list)

app/assets/database/bulk_ops.py

@@ -0,0 +1,204 @@
+import os
+import uuid
+import sqlalchemy
+from typing import Iterable
+from sqlalchemy.orm import Session
+from sqlalchemy.dialects import sqlite
+
+from app.assets.helpers import utcnow
+from app.assets.database.models import Asset, AssetCacheState, AssetInfo, AssetInfoTag, AssetInfoMeta
+
+MAX_BIND_PARAMS = 800
+
+def _chunk_rows(rows: list[dict], cols_per_row: int, max_bind_params: int) -> Iterable[list[dict]]:
+    if not rows:
+        return []
+    rows_per_stmt = max(1, max_bind_params // max(1, cols_per_row))
+    for i in range(0, len(rows), rows_per_stmt):
+        yield rows[i:i + rows_per_stmt]
+
+def _iter_chunks(seq, n: int):
+    for i in range(0, len(seq), n):
+        yield seq[i:i + n]
+
+def _rows_per_stmt(cols: int) -> int:
+    return max(1, MAX_BIND_PARAMS // max(1, cols))
+
+
+def seed_from_paths_batch(
+    session: Session,
+    *,
+    specs: list[dict],
+    owner_id: str = "",
+) -> dict:
+    """Each spec is a dict with keys:
+      - abs_path: str
+      - size_bytes: int
+      - mtime_ns: int
+      - info_name: str
+      - tags: list[str]
+      - fname: Optional[str]
+    """
+    if not specs:
+        return {"inserted_infos": 0, "won_states": 0, "lost_states": 0}
+
+    now = utcnow()
+    asset_rows: list[dict] = []
+    state_rows: list[dict] = []
+    path_to_asset: dict[str, str] = {}
+    asset_to_info: dict[str, dict] = {}  # asset_id -> prepared info row
+    path_list: list[str] = []
+
+    for sp in specs:
+        ap = os.path.abspath(sp["abs_path"])
+        aid = str(uuid.uuid4())
+        iid = str(uuid.uuid4())
+        path_list.append(ap)
+        path_to_asset[ap] = aid
+
+        asset_rows.append(
+            {
+                "id": aid,
+                "hash": None,
+                "size_bytes": sp["size_bytes"],
+                "mime_type": None,
+                "created_at": now,
+            }
+        )
+        state_rows.append(
+            {
+                "asset_id": aid,
+                "file_path": ap,
+                "mtime_ns": sp["mtime_ns"],
+            }
+        )
+        asset_to_info[aid] = {
+            "id": iid,
+            "owner_id": owner_id,
+            "name": sp["info_name"],
+            "asset_id": aid,
+            "preview_id": None,
+            "user_metadata": {"filename": sp["fname"]} if sp["fname"] else None,
+            "created_at": now,
+            "updated_at": now,
+            "last_access_time": now,
+            "_tags": sp["tags"],
+            "_filename": sp["fname"],
+        }
+
+    # insert all seed Assets (hash=NULL)
+    ins_asset = sqlite.insert(Asset)
+    for chunk in _iter_chunks(asset_rows, _rows_per_stmt(5)):
+        session.execute(ins_asset, chunk)
+
+    # try to claim AssetCacheState (file_path)
+    # Insert with ON CONFLICT DO NOTHING, then query to find which paths were actually inserted
+    ins_state = (
+        sqlite.insert(AssetCacheState)
+        .on_conflict_do_nothing(index_elements=[AssetCacheState.file_path])
+    )
+    for chunk in _iter_chunks(state_rows, _rows_per_stmt(3)):
+        session.execute(ins_state, chunk)
+
+    # Query to find which of our paths won (were actually inserted)
+    winners_by_path: set[str] = set()
+    for chunk in _iter_chunks(path_list, MAX_BIND_PARAMS):
+        result = session.execute(
+            sqlalchemy.select(AssetCacheState.file_path)
+            .where(AssetCacheState.file_path.in_(chunk))
+            .where(AssetCacheState.asset_id.in_([path_to_asset[p] for p in chunk]))
+        )
+        winners_by_path.update(result.scalars().all())
+
+    all_paths_set = set(path_list)
+    losers_by_path = all_paths_set - winners_by_path
+    lost_assets = [path_to_asset[p] for p in losers_by_path]
+    if lost_assets:  # losers get their Asset removed
+        for id_chunk in _iter_chunks(lost_assets, MAX_BIND_PARAMS):
+            session.execute(sqlalchemy.delete(Asset).where(Asset.id.in_(id_chunk)))
+
+    if not winners_by_path:
+        return {"inserted_infos": 0, "won_states": 0, "lost_states": len(losers_by_path)}
+
+    # insert AssetInfo only for winners
+    # Insert with ON CONFLICT DO NOTHING, then query to find which were actually inserted
+    winner_info_rows = [asset_to_info[path_to_asset[p]] for p in winners_by_path]
+    ins_info = (
+        sqlite.insert(AssetInfo)
+        .on_conflict_do_nothing(index_elements=[AssetInfo.asset_id, AssetInfo.owner_id, AssetInfo.name])
+    )
+    for chunk in _iter_chunks(winner_info_rows, _rows_per_stmt(9)):
+        session.execute(ins_info, chunk)
+
+    # Query to find which info rows were actually inserted (by matching our generated IDs)
+    all_info_ids = [row["id"] for row in winner_info_rows]
+    inserted_info_ids: set[str] = set()
+    for chunk in _iter_chunks(all_info_ids, MAX_BIND_PARAMS):
+        result = session.execute(
+            sqlalchemy.select(AssetInfo.id).where(AssetInfo.id.in_(chunk))
+        )
+        inserted_info_ids.update(result.scalars().all())
+
+    # build and insert tag + meta rows for the AssetInfo
+    tag_rows: list[dict] = []
+    meta_rows: list[dict] = []
+    if inserted_info_ids:
+        for row in winner_info_rows:
+            iid = row["id"]
+            if iid not in inserted_info_ids:
+                continue
+            for t in row["_tags"]:
+                tag_rows.append({
+                    "asset_info_id": iid,
+                    "tag_name": t,
+                    "origin": "automatic",
+                    "added_at": now,
+                })
+            if row["_filename"]:
+                meta_rows.append(
+                    {
+                        "asset_info_id": iid,
+                        "key": "filename",
+                        "ordinal": 0,
+                        "val_str": row["_filename"],
+                        "val_num": None,
+                        "val_bool": None,
+                        "val_json": None,
+                    }
+                )
+
+    bulk_insert_tags_and_meta(session, tag_rows=tag_rows, meta_rows=meta_rows, max_bind_params=MAX_BIND_PARAMS)
+    return {
+        "inserted_infos": len(inserted_info_ids),
+        "won_states": len(winners_by_path),
+        "lost_states": len(losers_by_path),
+    }
+
+
+def bulk_insert_tags_and_meta(
+    session: Session,
+    *,
+    tag_rows: list[dict],
+    meta_rows: list[dict],
+    max_bind_params: int,
+) -> None:
+    """Batch insert into asset_info_tags and asset_info_meta with ON CONFLICT DO NOTHING.
+    - tag_rows keys: asset_info_id, tag_name, origin, added_at
+    - meta_rows keys: asset_info_id, key, ordinal, val_str, val_num, val_bool, val_json
+    """
+    if tag_rows:
+        ins_links = (
+            sqlite.insert(AssetInfoTag)
+            .on_conflict_do_nothing(index_elements=[AssetInfoTag.asset_info_id, AssetInfoTag.tag_name])
+        )
+        for chunk in _chunk_rows(tag_rows, cols_per_row=4, max_bind_params=max_bind_params):
+            session.execute(ins_links, chunk)
+    if meta_rows:
+        ins_meta = (
+            sqlite.insert(AssetInfoMeta)
+            .on_conflict_do_nothing(
+                index_elements=[AssetInfoMeta.asset_info_id, AssetInfoMeta.key, AssetInfoMeta.ordinal]
+            )
+        )
+        for chunk in _chunk_rows(meta_rows, cols_per_row=7, max_bind_params=max_bind_params):
+            session.execute(ins_meta, chunk)

app/assets/database/models.py

@@ -0,0 +1,233 @@
+from __future__ import annotations
+
+import uuid
+from datetime import datetime
+
+from typing import Any
+from sqlalchemy import (
+    JSON,
+    BigInteger,
+    Boolean,
+    CheckConstraint,
+    DateTime,
+    ForeignKey,
+    Index,
+    Integer,
+    Numeric,
+    String,
+    Text,
+    UniqueConstraint,
+)
+from sqlalchemy.orm import Mapped, foreign, mapped_column, relationship
+
+from app.assets.helpers import utcnow
+from app.database.models import to_dict, Base
+
+
+class Asset(Base):
+    __tablename__ = "assets"
+
+    id: Mapped[str] = mapped_column(String(36), primary_key=True, default=lambda: str(uuid.uuid4()))
+    hash: Mapped[str | None] = mapped_column(String(256), nullable=True)
+    size_bytes: Mapped[int] = mapped_column(BigInteger, nullable=False, default=0)
+    mime_type: Mapped[str | None] = mapped_column(String(255))
+    created_at: Mapped[datetime] = mapped_column(
+        DateTime(timezone=False), nullable=False, default=utcnow
+    )
+
+    infos: Mapped[list[AssetInfo]] = relationship(
+        "AssetInfo",
+        back_populates="asset",
+        primaryjoin=lambda: Asset.id == foreign(AssetInfo.asset_id),
+        foreign_keys=lambda: [AssetInfo.asset_id],
+        cascade="all,delete-orphan",
+        passive_deletes=True,
+    )
+
+    preview_of: Mapped[list[AssetInfo]] = relationship(
+        "AssetInfo",
+        back_populates="preview_asset",
+        primaryjoin=lambda: Asset.id == foreign(AssetInfo.preview_id),
+        foreign_keys=lambda: [AssetInfo.preview_id],
+        viewonly=True,
+    )
+
+    cache_states: Mapped[list[AssetCacheState]] = relationship(
+        back_populates="asset",
+        cascade="all, delete-orphan",
+        passive_deletes=True,
+    )
+
+    __table_args__ = (
+        Index("uq_assets_hash", "hash", unique=True),
+        Index("ix_assets_mime_type", "mime_type"),
+        CheckConstraint("size_bytes >= 0", name="ck_assets_size_nonneg"),
+    )
+
+    def to_dict(self, include_none: bool = False) -> dict[str, Any]:
+        return to_dict(self, include_none=include_none)
+
+    def __repr__(self) -> str:
+        return f"<Asset id={self.id} hash={(self.hash or '')[:12]}>"
+
+
+class AssetCacheState(Base):
+    __tablename__ = "asset_cache_state"
+
+    id: Mapped[int] = mapped_column(Integer, primary_key=True, autoincrement=True)
+    asset_id: Mapped[str] = mapped_column(String(36), ForeignKey("assets.id", ondelete="CASCADE"), nullable=False)
+    file_path: Mapped[str] = mapped_column(Text, nullable=False)
+    mtime_ns: Mapped[int | None] = mapped_column(BigInteger, nullable=True)
+    needs_verify: Mapped[bool] = mapped_column(Boolean, nullable=False, default=False)
+
+    asset: Mapped[Asset] = relationship(back_populates="cache_states")
+
+    __table_args__ = (
+        Index("ix_asset_cache_state_file_path", "file_path"),
+        Index("ix_asset_cache_state_asset_id", "asset_id"),
+        CheckConstraint("(mtime_ns IS NULL) OR (mtime_ns >= 0)", name="ck_acs_mtime_nonneg"),
+        UniqueConstraint("file_path", name="uq_asset_cache_state_file_path"),
+    )
+
+    def to_dict(self, include_none: bool = False) -> dict[str, Any]:
+        return to_dict(self, include_none=include_none)
+
+    def __repr__(self) -> str:
+        return f"<AssetCacheState id={self.id} asset_id={self.asset_id} path={self.file_path!r}>"
+
+
+class AssetInfo(Base):
+    __tablename__ = "assets_info"
+
+    id: Mapped[str] = mapped_column(String(36), primary_key=True, default=lambda: str(uuid.uuid4()))
+    owner_id: Mapped[str] = mapped_column(String(128), nullable=False, default="")
+    name: Mapped[str] = mapped_column(String(512), nullable=False)
+    asset_id: Mapped[str] = mapped_column(String(36), ForeignKey("assets.id", ondelete="RESTRICT"), nullable=False)
+    preview_id: Mapped[str | None] = mapped_column(String(36), ForeignKey("assets.id", ondelete="SET NULL"))
+    user_metadata: Mapped[dict[str, Any] | None] = mapped_column(JSON(none_as_null=True))
+    created_at: Mapped[datetime] = mapped_column(DateTime(timezone=False), nullable=False, default=utcnow)
+    updated_at: Mapped[datetime] = mapped_column(DateTime(timezone=False), nullable=False, default=utcnow)
+    last_access_time: Mapped[datetime] = mapped_column(DateTime(timezone=False), nullable=False, default=utcnow)
+
+    asset: Mapped[Asset] = relationship(
+        "Asset",
+        back_populates="infos",
+        foreign_keys=[asset_id],
+        lazy="selectin",
+    )
+    preview_asset: Mapped[Asset | None] = relationship(
+        "Asset",
+        back_populates="preview_of",
+        foreign_keys=[preview_id],
+    )
+
+    metadata_entries: Mapped[list[AssetInfoMeta]] = relationship(
+        back_populates="asset_info",
+        cascade="all,delete-orphan",
+        passive_deletes=True,
+    )
+
+    tag_links: Mapped[list[AssetInfoTag]] = relationship(
+        back_populates="asset_info",
+        cascade="all,delete-orphan",
+        passive_deletes=True,
+        overlaps="tags,asset_infos",
+    )
+
+    tags: Mapped[list[Tag]] = relationship(
+        secondary="asset_info_tags",
+        back_populates="asset_infos",
+        lazy="selectin",
+        viewonly=True,
+        overlaps="tag_links,asset_info_links,asset_infos,tag",
+    )
+
+    __table_args__ = (
+        UniqueConstraint("asset_id", "owner_id", "name", name="uq_assets_info_asset_owner_name"),
+        Index("ix_assets_info_owner_name", "owner_id", "name"),
+        Index("ix_assets_info_owner_id", "owner_id"),
+        Index("ix_assets_info_asset_id", "asset_id"),
+        Index("ix_assets_info_name", "name"),
+        Index("ix_assets_info_created_at", "created_at"),
+        Index("ix_assets_info_last_access_time", "last_access_time"),
+    )
+
+    def to_dict(self, include_none: bool = False) -> dict[str, Any]:
+        data = to_dict(self, include_none=include_none)
+        data["tags"] = [t.name for t in self.tags]
+        return data
+
+    def __repr__(self) -> str:
+        return f"<AssetInfo id={self.id} name={self.name!r} asset_id={self.asset_id}>"
+
+
+class AssetInfoMeta(Base):
+    __tablename__ = "asset_info_meta"
+
+    asset_info_id: Mapped[str] = mapped_column(
+        String(36), ForeignKey("assets_info.id", ondelete="CASCADE"), primary_key=True
+    )
+    key: Mapped[str] = mapped_column(String(256), primary_key=True)
+    ordinal: Mapped[int] = mapped_column(Integer, primary_key=True, default=0)
+
+    val_str: Mapped[str | None] = mapped_column(String(2048), nullable=True)
+    val_num: Mapped[float | None] = mapped_column(Numeric(38, 10), nullable=True)
+    val_bool: Mapped[bool | None] = mapped_column(Boolean, nullable=True)
+    val_json: Mapped[Any | None] = mapped_column(JSON(none_as_null=True), nullable=True)
+
+    asset_info: Mapped[AssetInfo] = relationship(back_populates="metadata_entries")
+
+    __table_args__ = (
+        Index("ix_asset_info_meta_key", "key"),
+        Index("ix_asset_info_meta_key_val_str", "key", "val_str"),
+        Index("ix_asset_info_meta_key_val_num", "key", "val_num"),
+        Index("ix_asset_info_meta_key_val_bool", "key", "val_bool"),
+    )
+
+
+class AssetInfoTag(Base):
+    __tablename__ = "asset_info_tags"
+
+    asset_info_id: Mapped[str] = mapped_column(
+        String(36), ForeignKey("assets_info.id", ondelete="CASCADE"), primary_key=True
+    )
+    tag_name: Mapped[str] = mapped_column(
+        String(512), ForeignKey("tags.name", ondelete="RESTRICT"), primary_key=True
+    )
+    origin: Mapped[str] = mapped_column(String(32), nullable=False, default="manual")
+    added_at: Mapped[datetime] = mapped_column(
+        DateTime(timezone=False), nullable=False, default=utcnow
+    )
+
+    asset_info: Mapped[AssetInfo] = relationship(back_populates="tag_links")
+    tag: Mapped[Tag] = relationship(back_populates="asset_info_links")
+
+    __table_args__ = (
+        Index("ix_asset_info_tags_tag_name", "tag_name"),
+        Index("ix_asset_info_tags_asset_info_id", "asset_info_id"),
+    )
+
+
+class Tag(Base):
+    __tablename__ = "tags"
+
+    name: Mapped[str] = mapped_column(String(512), primary_key=True)
+    tag_type: Mapped[str] = mapped_column(String(32), nullable=False, default="user")
+
+    asset_info_links: Mapped[list[AssetInfoTag]] = relationship(
+        back_populates="tag",
+        overlaps="asset_infos,tags",
+    )
+    asset_infos: Mapped[list[AssetInfo]] = relationship(
+        secondary="asset_info_tags",
+        back_populates="tags",
+        viewonly=True,
+        overlaps="asset_info_links,tag_links,tags,asset_info",
+    )
+
+    __table_args__ = (
+        Index("ix_tags_tag_type", "tag_type"),
+    )
+
+    def __repr__(self) -> str:
+        return f"<Tag {self.name}>"

app/assets/database/queries.py

@@ -0,0 +1,976 @@
+import os
+import logging
+import sqlalchemy as sa
+from collections import defaultdict
+from datetime import datetime
+from typing import Iterable, Any
+from sqlalchemy import select, delete, exists, func
+from sqlalchemy.dialects import sqlite
+from sqlalchemy.exc import IntegrityError
+from sqlalchemy.orm import Session, contains_eager, noload
+from app.assets.database.models import Asset, AssetInfo, AssetCacheState, AssetInfoMeta, AssetInfoTag, Tag
+from app.assets.helpers import (
+    compute_relative_filename, escape_like_prefix, normalize_tags, project_kv, utcnow
+)
+from typing import Sequence
+
+
+def visible_owner_clause(owner_id: str) -> sa.sql.ClauseElement:
+    """Build owner visibility predicate for reads. Owner-less rows are visible to everyone."""
+    owner_id = (owner_id or "").strip()
+    if owner_id == "":
+        return AssetInfo.owner_id == ""
+    return AssetInfo.owner_id.in_(["", owner_id])
+
+
+def pick_best_live_path(states: Sequence[AssetCacheState]) -> str:
+    """
+    Return the best on-disk path among cache states:
+      1) Prefer a path that exists with needs_verify == False (already verified).
+      2) Otherwise, pick the first path that exists.
+      3) Otherwise return empty string.
+    """
+    alive = [s for s in states if getattr(s, "file_path", None) and os.path.isfile(s.file_path)]
+    if not alive:
+        return ""
+    for s in alive:
+        if not getattr(s, "needs_verify", False):
+            return s.file_path
+    return alive[0].file_path
+
+
+def apply_tag_filters(
+    stmt: sa.sql.Select,
+    include_tags: Sequence[str] | None = None,
+    exclude_tags: Sequence[str] | None = None,
+) -> sa.sql.Select:
+    """include_tags: every tag must be present; exclude_tags: none may be present."""
+    include_tags = normalize_tags(include_tags)
+    exclude_tags = normalize_tags(exclude_tags)
+
+    if include_tags:
+        for tag_name in include_tags:
+            stmt = stmt.where(
+                exists().where(
+                    (AssetInfoTag.asset_info_id == AssetInfo.id)
+                    & (AssetInfoTag.tag_name == tag_name)
+                )
+            )
+
+    if exclude_tags:
+        stmt = stmt.where(
+            ~exists().where(
+                (AssetInfoTag.asset_info_id == AssetInfo.id)
+                & (AssetInfoTag.tag_name.in_(exclude_tags))
+            )
+        )
+    return stmt
+
+
+def apply_metadata_filter(
+    stmt: sa.sql.Select,
+    metadata_filter: dict | None = None,
+) -> sa.sql.Select:
+    """Apply filters using asset_info_meta projection table."""
+    if not metadata_filter:
+        return stmt
+
+    def _exists_for_pred(key: str, *preds) -> sa.sql.ClauseElement:
+        return sa.exists().where(
+            AssetInfoMeta.asset_info_id == AssetInfo.id,
+            AssetInfoMeta.key == key,
+            *preds,
+        )
+
+    def _exists_clause_for_value(key: str, value) -> sa.sql.ClauseElement:
+        if value is None:
+            no_row_for_key = sa.not_(
+                sa.exists().where(
+                    AssetInfoMeta.asset_info_id == AssetInfo.id,
+                    AssetInfoMeta.key == key,
+                )
+            )
+            null_row = _exists_for_pred(
+                key,
+                AssetInfoMeta.val_json.is_(None),
+                AssetInfoMeta.val_str.is_(None),
+                AssetInfoMeta.val_num.is_(None),
+                AssetInfoMeta.val_bool.is_(None),
+            )
+            return sa.or_(no_row_for_key, null_row)
+
+        if isinstance(value, bool):
+            return _exists_for_pred(key, AssetInfoMeta.val_bool == bool(value))
+        if isinstance(value, (int, float)):
+            from decimal import Decimal
+            num = value if isinstance(value, Decimal) else Decimal(str(value))
+            return _exists_for_pred(key, AssetInfoMeta.val_num == num)
+        if isinstance(value, str):
+            return _exists_for_pred(key, AssetInfoMeta.val_str == value)
+        return _exists_for_pred(key, AssetInfoMeta.val_json == value)
+
+    for k, v in metadata_filter.items():
+        if isinstance(v, list):
+            ors = [_exists_clause_for_value(k, elem) for elem in v]
+            if ors:
+                stmt = stmt.where(sa.or_(*ors))
+        else:
+            stmt = stmt.where(_exists_clause_for_value(k, v))
+    return stmt
+
+
+def asset_exists_by_hash(
+    session: Session,
+    *,
+    asset_hash: str,
+) -> bool:
+    """
+    Check if an asset with a given hash exists in database.
+    """
+    row = (
+        session.execute(
+            select(sa.literal(True)).select_from(Asset).where(Asset.hash == asset_hash).limit(1)
+        )
+    ).first()
+    return row is not None
+
+
+def asset_info_exists_for_asset_id(
+    session: Session,
+    *,
+    asset_id: str,
+) -> bool:
+    q = (
+        select(sa.literal(True))
+        .select_from(AssetInfo)
+        .where(AssetInfo.asset_id == asset_id)
+        .limit(1)
+    )
+    return (session.execute(q)).first() is not None
+
+
+def get_asset_by_hash(
+    session: Session,
+    *,
+    asset_hash: str,
+) -> Asset | None:
+    return (
+        session.execute(select(Asset).where(Asset.hash == asset_hash).limit(1))
+    ).scalars().first()
+
+
+def get_asset_info_by_id(
+    session: Session,
+    *,
+    asset_info_id: str,
+) -> AssetInfo | None:
+    return session.get(AssetInfo, asset_info_id)
+
+
+def list_asset_infos_page(
+    session: Session,
+    owner_id: str = "",
+    include_tags: Sequence[str] | None = None,
+    exclude_tags: Sequence[str] | None = None,
+    name_contains: str | None = None,
+    metadata_filter: dict | None = None,
+    limit: int = 20,
+    offset: int = 0,
+    sort: str = "created_at",
+    order: str = "desc",
+) -> tuple[list[AssetInfo], dict[str, list[str]], int]:
+    base = (
+        select(AssetInfo)
+        .join(Asset, Asset.id == AssetInfo.asset_id)
+        .options(contains_eager(AssetInfo.asset), noload(AssetInfo.tags))
+        .where(visible_owner_clause(owner_id))
+    )
+
+    if name_contains:
+        escaped, esc = escape_like_prefix(name_contains)
+        base = base.where(AssetInfo.name.ilike(f"%{escaped}%", escape=esc))
+
+    base = apply_tag_filters(base, include_tags, exclude_tags)
+    base = apply_metadata_filter(base, metadata_filter)
+
+    sort = (sort or "created_at").lower()
+    order = (order or "desc").lower()
+    sort_map = {
+        "name": AssetInfo.name,
+        "created_at": AssetInfo.created_at,
+        "updated_at": AssetInfo.updated_at,
+        "last_access_time": AssetInfo.last_access_time,
+        "size": Asset.size_bytes,
+    }
+    sort_col = sort_map.get(sort, AssetInfo.created_at)
+    sort_exp = sort_col.desc() if order == "desc" else sort_col.asc()
+
+    base = base.order_by(sort_exp).limit(limit).offset(offset)
+
+    count_stmt = (
+        select(sa.func.count())
+        .select_from(AssetInfo)
+        .join(Asset, Asset.id == AssetInfo.asset_id)
+        .where(visible_owner_clause(owner_id))
+    )
+    if name_contains:
+        escaped, esc = escape_like_prefix(name_contains)
+        count_stmt = count_stmt.where(AssetInfo.name.ilike(f"%{escaped}%", escape=esc))
+    count_stmt = apply_tag_filters(count_stmt, include_tags, exclude_tags)
+    count_stmt = apply_metadata_filter(count_stmt, metadata_filter)
+
+    total = int((session.execute(count_stmt)).scalar_one() or 0)
+
+    infos = (session.execute(base)).unique().scalars().all()
+
+    id_list: list[str] = [i.id for i in infos]
+    tag_map: dict[str, list[str]] = defaultdict(list)
+    if id_list:
+        rows = session.execute(
+            select(AssetInfoTag.asset_info_id, Tag.name)
+            .join(Tag, Tag.name == AssetInfoTag.tag_name)
+            .where(AssetInfoTag.asset_info_id.in_(id_list))
+            .order_by(AssetInfoTag.added_at)
+        )
+        for aid, tag_name in rows.all():
+            tag_map[aid].append(tag_name)
+
+    return infos, tag_map, total
+
+
+def fetch_asset_info_asset_and_tags(
+    session: Session,
+    asset_info_id: str,
+    owner_id: str = "",
+) -> tuple[AssetInfo, Asset, list[str]] | None:
+    stmt = (
+        select(AssetInfo, Asset, Tag.name)
+        .join(Asset, Asset.id == AssetInfo.asset_id)
+        .join(AssetInfoTag, AssetInfoTag.asset_info_id == AssetInfo.id, isouter=True)
+        .join(Tag, Tag.name == AssetInfoTag.tag_name, isouter=True)
+        .where(
+            AssetInfo.id == asset_info_id,
+            visible_owner_clause(owner_id),
+        )
+        .options(noload(AssetInfo.tags))
+        .order_by(Tag.name.asc())
+    )
+
+    rows = (session.execute(stmt)).all()
+    if not rows:
+        return None
+
+    first_info, first_asset, _ = rows[0]
+    tags: list[str] = []
+    seen: set[str] = set()
+    for _info, _asset, tag_name in rows:
+        if tag_name and tag_name not in seen:
+            seen.add(tag_name)
+            tags.append(tag_name)
+    return first_info, first_asset, tags
+
+
+def fetch_asset_info_and_asset(
+    session: Session,
+    *,
+    asset_info_id: str,
+    owner_id: str = "",
+) -> tuple[AssetInfo, Asset] | None:
+    stmt = (
+        select(AssetInfo, Asset)
+        .join(Asset, Asset.id == AssetInfo.asset_id)
+        .where(
+            AssetInfo.id == asset_info_id,
+            visible_owner_clause(owner_id),
+        )
+        .limit(1)
+        .options(noload(AssetInfo.tags))
+    )
+    row = session.execute(stmt)
+    pair = row.first()
+    if not pair:
+        return None
+    return pair[0], pair[1]
+
+def list_cache_states_by_asset_id(
+    session: Session, *, asset_id: str
+) -> Sequence[AssetCacheState]:
+    return (
+        session.execute(
+            select(AssetCacheState)
+            .where(AssetCacheState.asset_id == asset_id)
+            .order_by(AssetCacheState.id.asc())
+        )
+    ).scalars().all()
+
+
+def touch_asset_info_by_id(
+    session: Session,
+    *,
+    asset_info_id: str,
+    ts: datetime | None = None,
+    only_if_newer: bool = True,
+) -> None:
+    ts = ts or utcnow()
+    stmt = sa.update(AssetInfo).where(AssetInfo.id == asset_info_id)
+    if only_if_newer:
+        stmt = stmt.where(
+            sa.or_(AssetInfo.last_access_time.is_(None), AssetInfo.last_access_time < ts)
+        )
+    session.execute(stmt.values(last_access_time=ts))
+
+
+def create_asset_info_for_existing_asset(
+    session: Session,
+    *,
+    asset_hash: str,
+    name: str,
+    user_metadata: dict | None = None,
+    tags: Sequence[str] | None = None,
+    tag_origin: str = "manual",
+    owner_id: str = "",
+) -> AssetInfo:
+    """Create or return an existing AssetInfo for an Asset identified by asset_hash."""
+    now = utcnow()
+    asset = get_asset_by_hash(session, asset_hash=asset_hash)
+    if not asset:
+        raise ValueError(f"Unknown asset hash {asset_hash}")
+
+    info = AssetInfo(
+        owner_id=owner_id,
+        name=name,
+        asset_id=asset.id,
+        preview_id=None,
+        created_at=now,
+        updated_at=now,
+        last_access_time=now,
+    )
+    try:
+        with session.begin_nested():
+            session.add(info)
+            session.flush()
+    except IntegrityError:
+        existing = (
+            session.execute(
+                select(AssetInfo)
+                .options(noload(AssetInfo.tags))
+                .where(
+                    AssetInfo.asset_id == asset.id,
+                    AssetInfo.name == name,
+                    AssetInfo.owner_id == owner_id,
+                )
+                .limit(1)
+            )
+        ).unique().scalars().first()
+        if not existing:
+            raise RuntimeError("AssetInfo upsert failed to find existing row after conflict.")
+        return existing
+
+    # metadata["filename"] hack
+    new_meta = dict(user_metadata or {})
+    computed_filename = None
+    try:
+        p = pick_best_live_path(list_cache_states_by_asset_id(session, asset_id=asset.id))
+        if p:
+            computed_filename = compute_relative_filename(p)
+    except Exception:
+        computed_filename = None
+    if computed_filename:
+        new_meta["filename"] = computed_filename
+    if new_meta:
+        replace_asset_info_metadata_projection(
+            session,
+            asset_info_id=info.id,
+            user_metadata=new_meta,
+        )
+
+    if tags is not None:
+        set_asset_info_tags(
+            session,
+            asset_info_id=info.id,
+            tags=tags,
+            origin=tag_origin,
+        )
+    return info
+
+
+def set_asset_info_tags(
+    session: Session,
+    *,
+    asset_info_id: str,
+    tags: Sequence[str],
+    origin: str = "manual",
+) -> dict:
+    desired = normalize_tags(tags)
+
+    current = set(
+        tag_name for (tag_name,) in (
+            session.execute(select(AssetInfoTag.tag_name).where(AssetInfoTag.asset_info_id == asset_info_id))
+        ).all()
+    )
+
+    to_add = [t for t in desired if t not in current]
+    to_remove = [t for t in current if t not in desired]
+
+    if to_add:
+        ensure_tags_exist(session, to_add, tag_type="user")
+        session.add_all([
+            AssetInfoTag(asset_info_id=asset_info_id, tag_name=t, origin=origin, added_at=utcnow())
+            for t in to_add
+        ])
+        session.flush()
+
+    if to_remove:
+        session.execute(
+            delete(AssetInfoTag)
+            .where(AssetInfoTag.asset_info_id == asset_info_id, AssetInfoTag.tag_name.in_(to_remove))
+        )
+        session.flush()
+
+    return {"added": to_add, "removed": to_remove, "total": desired}
+
+
+def replace_asset_info_metadata_projection(
+    session: Session,
+    *,
+    asset_info_id: str,
+    user_metadata: dict | None = None,
+) -> None:
+    info = session.get(AssetInfo, asset_info_id)
+    if not info:
+        raise ValueError(f"AssetInfo {asset_info_id} not found")
+
+    info.user_metadata = user_metadata or {}
+    info.updated_at = utcnow()
+    session.flush()
+
+    session.execute(delete(AssetInfoMeta).where(AssetInfoMeta.asset_info_id == asset_info_id))
+    session.flush()
+
+    if not user_metadata:
+        return
+
+    rows: list[AssetInfoMeta] = []
+    for k, v in user_metadata.items():
+        for r in project_kv(k, v):
+            rows.append(
+                AssetInfoMeta(
+                    asset_info_id=asset_info_id,
+                    key=r["key"],
+                    ordinal=int(r["ordinal"]),
+                    val_str=r.get("val_str"),
+                    val_num=r.get("val_num"),
+                    val_bool=r.get("val_bool"),
+                    val_json=r.get("val_json"),
+                )
+            )
+    if rows:
+        session.add_all(rows)
+        session.flush()
+
+
+def ingest_fs_asset(
+    session: Session,
+    *,
+    asset_hash: str,
+    abs_path: str,
+    size_bytes: int,
+    mtime_ns: int,
+    mime_type: str | None = None,
+    info_name: str | None = None,
+    owner_id: str = "",
+    preview_id: str | None = None,
+    user_metadata: dict | None = None,
+    tags: Sequence[str] = (),
+    tag_origin: str = "manual",
+    require_existing_tags: bool = False,
+) -> dict:
+    """
+    Idempotently upsert:
+      - Asset by content hash (create if missing)
+      - AssetCacheState(file_path) pointing to asset_id
+      - Optionally AssetInfo + tag links and metadata projection
+    Returns flags and ids.
+    """
+    locator = os.path.abspath(abs_path)
+    now = utcnow()
+
+    if preview_id:
+        if not session.get(Asset, preview_id):
+            preview_id = None
+
+    out: dict[str, Any] = {
+        "asset_created": False,
+        "asset_updated": False,
+        "state_created": False,
+        "state_updated": False,
+        "asset_info_id": None,
+    }
+
+    # 1) Asset by hash
+    asset = (
+        session.execute(select(Asset).where(Asset.hash == asset_hash).limit(1))
+    ).scalars().first()
+    if not asset:
+        vals = {
+            "hash": asset_hash,
+            "size_bytes": int(size_bytes),
+            "mime_type": mime_type,
+            "created_at": now,
+        }
+        res = session.execute(
+            sqlite.insert(Asset)
+            .values(**vals)
+            .on_conflict_do_nothing(index_elements=[Asset.hash])
+        )
+        if int(res.rowcount or 0) > 0:
+            out["asset_created"] = True
+        asset = (
+            session.execute(
+                select(Asset).where(Asset.hash == asset_hash).limit(1)
+            )
+        ).scalars().first()
+        if not asset:
+            raise RuntimeError("Asset row not found after upsert.")
+    else:
+        changed = False
+        if asset.size_bytes != int(size_bytes) and int(size_bytes) > 0:
+            asset.size_bytes = int(size_bytes)
+            changed = True
+        if mime_type and asset.mime_type != mime_type:
+            asset.mime_type = mime_type
+            changed = True
+        if changed:
+            out["asset_updated"] = True
+
+    # 2) AssetCacheState upsert by file_path (unique)
+    vals = {
+        "asset_id": asset.id,
+        "file_path": locator,
+        "mtime_ns": int(mtime_ns),
+    }
+    ins = (
+        sqlite.insert(AssetCacheState)
+        .values(**vals)
+        .on_conflict_do_nothing(index_elements=[AssetCacheState.file_path])
+    )
+
+    res = session.execute(ins)
+    if int(res.rowcount or 0) > 0:
+        out["state_created"] = True
+    else:
+        upd = (
+            sa.update(AssetCacheState)
+            .where(AssetCacheState.file_path == locator)
+            .where(
+                sa.or_(
+                    AssetCacheState.asset_id != asset.id,
+                    AssetCacheState.mtime_ns.is_(None),
+                    AssetCacheState.mtime_ns != int(mtime_ns),
+                )
+            )
+            .values(asset_id=asset.id, mtime_ns=int(mtime_ns))
+        )
+        res2 = session.execute(upd)
+        if int(res2.rowcount or 0) > 0:
+            out["state_updated"] = True
+
+    # 3) Optional AssetInfo + tags + metadata
+    if info_name:
+        try:
+            with session.begin_nested():
+                info = AssetInfo(
+                    owner_id=owner_id,
+                    name=info_name,
+                    asset_id=asset.id,
+                    preview_id=preview_id,
+                    created_at=now,
+                    updated_at=now,
+                    last_access_time=now,
+                )
+                session.add(info)
+                session.flush()
+                out["asset_info_id"] = info.id
+        except IntegrityError:
+            pass
+
+        existing_info = (
+            session.execute(
+                select(AssetInfo)
+                .where(
+                    AssetInfo.asset_id == asset.id,
+                    AssetInfo.name == info_name,
+                    (AssetInfo.owner_id == owner_id),
+                )
+                .limit(1)
+            )
+        ).unique().scalar_one_or_none()
+        if not existing_info:
+            raise RuntimeError("Failed to update or insert AssetInfo.")
+
+        if preview_id and existing_info.preview_id != preview_id:
+            existing_info.preview_id = preview_id
+
+        existing_info.updated_at = now
+        if existing_info.last_access_time < now:
+            existing_info.last_access_time = now
+        session.flush()
+        out["asset_info_id"] = existing_info.id
+
+        norm = [t.strip().lower() for t in (tags or []) if (t or "").strip()]
+        if norm and out["asset_info_id"] is not None:
+            if not require_existing_tags:
+                ensure_tags_exist(session, norm, tag_type="user")
+
+            existing_tag_names = set(
+                name for (name,) in (session.execute(select(Tag.name).where(Tag.name.in_(norm)))).all()
+            )
+            missing = [t for t in norm if t not in existing_tag_names]
+            if missing and require_existing_tags:
+                raise ValueError(f"Unknown tags: {missing}")
+
+            existing_links = set(
+                tag_name
+                for (tag_name,) in (
+                    session.execute(
+                        select(AssetInfoTag.tag_name).where(AssetInfoTag.asset_info_id == out["asset_info_id"])
+                    )
+                ).all()
+            )
+            to_add = [t for t in norm if t in existing_tag_names and t not in existing_links]
+            if to_add:
+                session.add_all(
+                    [
+                        AssetInfoTag(
+                            asset_info_id=out["asset_info_id"],
+                            tag_name=t,
+                            origin=tag_origin,
+                            added_at=now,
+                        )
+                        for t in to_add
+                    ]
+                )
+                session.flush()
+
+        # metadata["filename"] hack
+        if out["asset_info_id"] is not None:
+            primary_path = pick_best_live_path(list_cache_states_by_asset_id(session, asset_id=asset.id))
+            computed_filename = compute_relative_filename(primary_path) if primary_path else None
+
+            current_meta = existing_info.user_metadata or {}
+            new_meta = dict(current_meta)
+            if user_metadata is not None:
+                for k, v in user_metadata.items():
+                    new_meta[k] = v
+            if computed_filename:
+                new_meta["filename"] = computed_filename
+
+            if new_meta != current_meta:
+                replace_asset_info_metadata_projection(
+                    session,
+                    asset_info_id=out["asset_info_id"],
+                    user_metadata=new_meta,
+                )
+
+    try:
+        remove_missing_tag_for_asset_id(session, asset_id=asset.id)
+    except Exception:
+        logging.exception("Failed to clear 'missing' tag for asset %s", asset.id)
+    return out
+
+
+def update_asset_info_full(
+    session: Session,
+    *,
+    asset_info_id: str,
+    name: str | None = None,
+    tags: Sequence[str] | None = None,
+    user_metadata: dict | None = None,
+    tag_origin: str = "manual",
+    asset_info_row: Any = None,
+) -> AssetInfo:
+    if not asset_info_row:
+        info = session.get(AssetInfo, asset_info_id)
+        if not info:
+            raise ValueError(f"AssetInfo {asset_info_id} not found")
+    else:
+        info = asset_info_row
+
+    touched = False
+    if name is not None and name != info.name:
+        info.name = name
+        touched = True
+
+    computed_filename = None
+    try:
+        p = pick_best_live_path(list_cache_states_by_asset_id(session, asset_id=info.asset_id))
+        if p:
+            computed_filename = compute_relative_filename(p)
+    except Exception:
+        computed_filename = None
+
+    if user_metadata is not None:
+        new_meta = dict(user_metadata)
+        if computed_filename:
+            new_meta["filename"] = computed_filename
+        replace_asset_info_metadata_projection(
+            session, asset_info_id=asset_info_id, user_metadata=new_meta
+        )
+        touched = True
+    else:
+        if computed_filename:
+            current_meta = info.user_metadata or {}
+            if current_meta.get("filename") != computed_filename:
+                new_meta = dict(current_meta)
+                new_meta["filename"] = computed_filename
+                replace_asset_info_metadata_projection(
+                    session, asset_info_id=asset_info_id, user_metadata=new_meta
+                )
+                touched = True
+
+    if tags is not None:
+        set_asset_info_tags(
+            session,
+            asset_info_id=asset_info_id,
+            tags=tags,
+            origin=tag_origin,
+        )
+        touched = True
+
+    if touched and user_metadata is None:
+        info.updated_at = utcnow()
+        session.flush()
+
+    return info
+
+
+def delete_asset_info_by_id(
+    session: Session,
+    *,
+    asset_info_id: str,
+    owner_id: str,
+) -> bool:
+    stmt = sa.delete(AssetInfo).where(
+        AssetInfo.id == asset_info_id,
+        visible_owner_clause(owner_id),
+    )
+    return int((session.execute(stmt)).rowcount or 0) > 0
+
+
+def list_tags_with_usage(
+    session: Session,
+    prefix: str | None = None,
+    limit: int = 100,
+    offset: int = 0,
+    include_zero: bool = True,
+    order: str = "count_desc",
+    owner_id: str = "",
+) -> tuple[list[tuple[str, str, int]], int]:
+    counts_sq = (
+        select(
+            AssetInfoTag.tag_name.label("tag_name"),
+            func.count(AssetInfoTag.asset_info_id).label("cnt"),
+        )
+        .select_from(AssetInfoTag)
+        .join(AssetInfo, AssetInfo.id == AssetInfoTag.asset_info_id)
+        .where(visible_owner_clause(owner_id))
+        .group_by(AssetInfoTag.tag_name)
+        .subquery()
+    )
+
+    q = (
+        select(
+            Tag.name,
+            Tag.tag_type,
+            func.coalesce(counts_sq.c.cnt, 0).label("count"),
+        )
+        .select_from(Tag)
+        .join(counts_sq, counts_sq.c.tag_name == Tag.name, isouter=True)
+    )
+
+    if prefix:
+        escaped, esc = escape_like_prefix(prefix.strip().lower())
+        q = q.where(Tag.name.like(escaped + "%", escape=esc))
+
+    if not include_zero:
+        q = q.where(func.coalesce(counts_sq.c.cnt, 0) > 0)
+
+    if order == "name_asc":
+        q = q.order_by(Tag.name.asc())
+    else:
+        q = q.order_by(func.coalesce(counts_sq.c.cnt, 0).desc(), Tag.name.asc())
+
+    total_q = select(func.count()).select_from(Tag)
+    if prefix:
+        escaped, esc = escape_like_prefix(prefix.strip().lower())
+        total_q = total_q.where(Tag.name.like(escaped + "%", escape=esc))
+    if not include_zero:
+        total_q = total_q.where(
+            Tag.name.in_(select(AssetInfoTag.tag_name).group_by(AssetInfoTag.tag_name))
+        )
+
+    rows = (session.execute(q.limit(limit).offset(offset))).all()
+    total = (session.execute(total_q)).scalar_one()
+
+    rows_norm = [(name, ttype, int(count or 0)) for (name, ttype, count) in rows]
+    return rows_norm, int(total or 0)
+
+
+def ensure_tags_exist(session: Session, names: Iterable[str], tag_type: str = "user") -> None:
+    wanted = normalize_tags(list(names))
+    if not wanted:
+        return
+    rows = [{"name": n, "tag_type": tag_type} for n in list(dict.fromkeys(wanted))]
+    ins = (
+        sqlite.insert(Tag)
+        .values(rows)
+        .on_conflict_do_nothing(index_elements=[Tag.name])
+    )
+    session.execute(ins)
+
+
+def get_asset_tags(session: Session, *, asset_info_id: str) -> list[str]:
+    return [
+        tag_name for (tag_name,) in (
+            session.execute(
+                select(AssetInfoTag.tag_name).where(AssetInfoTag.asset_info_id == asset_info_id)
+            )
+        ).all()
+    ]
+
+
+def add_tags_to_asset_info(
+    session: Session,
+    *,
+    asset_info_id: str,
+    tags: Sequence[str],
+    origin: str = "manual",
+    create_if_missing: bool = True,
+    asset_info_row: Any = None,
+) -> dict:
+    if not asset_info_row:
+        info = session.get(AssetInfo, asset_info_id)
+        if not info:
+            raise ValueError(f"AssetInfo {asset_info_id} not found")
+
+    norm = normalize_tags(tags)
+    if not norm:
+        total = get_asset_tags(session, asset_info_id=asset_info_id)
+        return {"added": [], "already_present": [], "total_tags": total}
+
+    if create_if_missing:
+        ensure_tags_exist(session, norm, tag_type="user")
+
+    current = {
+        tag_name
+        for (tag_name,) in (
+            session.execute(
+                sa.select(AssetInfoTag.tag_name).where(AssetInfoTag.asset_info_id == asset_info_id)
+            )
+        ).all()
+    }
+
+    want = set(norm)
+    to_add = sorted(want - current)
+
+    if to_add:
+        with session.begin_nested() as nested:
+            try:
+                session.add_all(
+                    [
+                        AssetInfoTag(
+                            asset_info_id=asset_info_id,
+                            tag_name=t,
+                            origin=origin,
+                            added_at=utcnow(),
+                        )
+                        for t in to_add
+                    ]
+                )
+                session.flush()
+            except IntegrityError:
+                nested.rollback()
+
+    after = set(get_asset_tags(session, asset_info_id=asset_info_id))
+    return {
+        "added": sorted(((after - current) & want)),
+        "already_present": sorted(want & current),
+        "total_tags": sorted(after),
+    }
+
+
+def remove_tags_from_asset_info(
+    session: Session,
+    *,
+    asset_info_id: str,
+    tags: Sequence[str],
+) -> dict:
+    info = session.get(AssetInfo, asset_info_id)
+    if not info:
+        raise ValueError(f"AssetInfo {asset_info_id} not found")
+
+    norm = normalize_tags(tags)
+    if not norm:
+        total = get_asset_tags(session, asset_info_id=asset_info_id)
+        return {"removed": [], "not_present": [], "total_tags": total}
+
+    existing = {
+        tag_name
+        for (tag_name,) in (
+            session.execute(
+                sa.select(AssetInfoTag.tag_name).where(AssetInfoTag.asset_info_id == asset_info_id)
+            )
+        ).all()
+    }
+
+    to_remove = sorted(set(t for t in norm if t in existing))
+    not_present = sorted(set(t for t in norm if t not in existing))
+
+    if to_remove:
+        session.execute(
+            delete(AssetInfoTag)
+            .where(
+                AssetInfoTag.asset_info_id == asset_info_id,
+                AssetInfoTag.tag_name.in_(to_remove),
+            )
+        )
+        session.flush()
+
+    total = get_asset_tags(session, asset_info_id=asset_info_id)
+    return {"removed": to_remove, "not_present": not_present, "total_tags": total}
+
+
+def remove_missing_tag_for_asset_id(
+    session: Session,
+    *,
+    asset_id: str,
+) -> None:
+    session.execute(
+        sa.delete(AssetInfoTag).where(
+            AssetInfoTag.asset_info_id.in_(sa.select(AssetInfo.id).where(AssetInfo.asset_id == asset_id)),
+            AssetInfoTag.tag_name == "missing",
+        )
+    )
+
+
+def set_asset_info_preview(
+    session: Session,
+    *,
+    asset_info_id: str,
+    preview_asset_id: str | None = None,
+) -> None:
+    """Set or clear preview_id and bump updated_at. Raises on unknown IDs."""
+    info = session.get(AssetInfo, asset_info_id)
+    if not info:
+        raise ValueError(f"AssetInfo {asset_info_id} not found")
+
+    if preview_asset_id is None:
+        info.preview_id = None
+    else:
+        # validate preview asset exists
+        if not session.get(Asset, preview_asset_id):
+            raise ValueError(f"Preview Asset {preview_asset_id} not found")
+        info.preview_id = preview_asset_id
+
+    info.updated_at = utcnow()
+    session.flush()

app/assets/database/tags.py

@@ -0,0 +1,62 @@
+from typing import Iterable
+
+import sqlalchemy
+from sqlalchemy.orm import Session
+from sqlalchemy.dialects import sqlite
+
+from app.assets.helpers import normalize_tags, utcnow
+from app.assets.database.models import Tag, AssetInfoTag, AssetInfo
+
+
+def ensure_tags_exist(session: Session, names: Iterable[str], tag_type: str = "user") -> None:
+    wanted = normalize_tags(list(names))
+    if not wanted:
+        return
+    rows = [{"name": n, "tag_type": tag_type} for n in list(dict.fromkeys(wanted))]
+    ins = (
+            sqlite.insert(Tag)
+            .values(rows)
+            .on_conflict_do_nothing(index_elements=[Tag.name])
+        )
+    return session.execute(ins)
+
+def add_missing_tag_for_asset_id(
+    session: Session,
+    *,
+    asset_id: str,
+    origin: str = "automatic",
+) -> None:
+    select_rows = (
+        sqlalchemy.select(
+            AssetInfo.id.label("asset_info_id"),
+            sqlalchemy.literal("missing").label("tag_name"),
+            sqlalchemy.literal(origin).label("origin"),
+            sqlalchemy.literal(utcnow()).label("added_at"),
+        )
+        .where(AssetInfo.asset_id == asset_id)
+        .where(
+            sqlalchemy.not_(
+                sqlalchemy.exists().where((AssetInfoTag.asset_info_id == AssetInfo.id) & (AssetInfoTag.tag_name == "missing"))
+            )
+        )
+    )
+    session.execute(
+        sqlite.insert(AssetInfoTag)
+        .from_select(
+            ["asset_info_id", "tag_name", "origin", "added_at"],
+            select_rows,
+        )
+        .on_conflict_do_nothing(index_elements=[AssetInfoTag.asset_info_id, AssetInfoTag.tag_name])
+    )
+
+def remove_missing_tag_for_asset_id(
+    session: Session,
+    *,
+    asset_id: str,
+) -> None:
+    session.execute(
+        sqlalchemy.delete(AssetInfoTag).where(
+            AssetInfoTag.asset_info_id.in_(sqlalchemy.select(AssetInfo.id).where(AssetInfo.asset_id == asset_id)),
+            AssetInfoTag.tag_name == "missing",
+        )
+    )

app/assets/hashing.py

@@ -0,0 +1,75 @@
+from blake3 import blake3
+from typing import IO
+import os
+import asyncio
+
+
+DEFAULT_CHUNK = 8 * 1024 *1024 # 8MB
+
+# NOTE: this allows hashing different representations of a file-like object
+def blake3_hash(
+    fp: str | IO[bytes],
+    chunk_size: int = DEFAULT_CHUNK,
+) -> str:
+    """
+    Returns a BLAKE3 hex digest for ``fp``, which may be:
+      - a filename (str/bytes) or PathLike
+      - an open binary file object
+    If ``fp`` is a file object, it must be opened in **binary** mode and support
+    ``read``, ``seek``, and ``tell``. The function will seek to the start before
+    reading and will attempt to restore the original position afterward.
+    """
+    # duck typing to check if input is a file-like object
+    if hasattr(fp, "read"):
+        return _hash_file_obj(fp, chunk_size)
+
+    with open(os.fspath(fp), "rb") as f:
+        return _hash_file_obj(f, chunk_size)
+
+
+async def blake3_hash_async(
+    fp: str | IO[bytes],
+    chunk_size: int = DEFAULT_CHUNK,
+) -> str:
+    """Async wrapper for ``blake3_hash_sync``.
+    Uses a worker thread so the event loop remains responsive.
+    """
+    # If it is a path, open inside the worker thread to keep I/O off the loop.
+    if hasattr(fp, "read"):
+        return await asyncio.to_thread(blake3_hash, fp, chunk_size)
+
+    def _worker() -> str:
+        with open(os.fspath(fp), "rb") as f:
+            return _hash_file_obj(f, chunk_size)
+
+    return await asyncio.to_thread(_worker)
+
+
+def _hash_file_obj(file_obj: IO, chunk_size: int = DEFAULT_CHUNK) -> str:
+    """
+    Hash an already-open binary file object by streaming in chunks.
+    - Seeks to the beginning before reading (if supported).
+    - Restores the original position afterward (if tell/seek are supported).
+    """
+    if chunk_size <= 0:
+        chunk_size = DEFAULT_CHUNK
+
+    # in case file object is already open and not at the beginning, track so can be restored after hashing
+    orig_pos = file_obj.tell()
+
+    try:
+        # seek to the beginning before reading
+        if orig_pos != 0:
+            file_obj.seek(0)
+
+        h = blake3()
+        while True:
+            chunk = file_obj.read(chunk_size)
+            if not chunk:
+                break
+            h.update(chunk)
+        return h.hexdigest()
+    finally:
+        # restore original position in file object, if needed
+        if orig_pos != 0:
+            file_obj.seek(orig_pos)

app/assets/helpers.py

@@ -0,0 +1,312 @@
+import contextlib
+import os
+from decimal import Decimal
+from aiohttp import web
+from datetime import datetime, timezone
+from pathlib import Path
+from typing import Literal, Any
+
+import folder_paths
+
+
+RootType = Literal["models", "input", "output"]
+ALLOWED_ROOTS: tuple[RootType, ...] = ("models", "input", "output")
+
+def get_query_dict(request: web.Request) -> dict[str, Any]:
+    """
+    Gets a dictionary of query parameters from the request.
+
+    'request.query' is a MultiMapping[str], needs to be converted to a dictionary to be validated by Pydantic.
+    """
+    query_dict = {
+        key: request.query.getall(key) if len(request.query.getall(key)) > 1 else request.query.get(key)
+        for key in request.query.keys()
+    }
+    return query_dict
+
+def list_tree(base_dir: str) -> list[str]:
+    out: list[str] = []
+    base_abs = os.path.abspath(base_dir)
+    if not os.path.isdir(base_abs):
+        return out
+    for dirpath, _subdirs, filenames in os.walk(base_abs, topdown=True, followlinks=False):
+        for name in filenames:
+            out.append(os.path.abspath(os.path.join(dirpath, name)))
+    return out
+
+def prefixes_for_root(root: RootType) -> list[str]:
+    if root == "models":
+        bases: list[str] = []
+        for _bucket, paths in get_comfy_models_folders():
+            bases.extend(paths)
+        return [os.path.abspath(p) for p in bases]
+    if root == "input":
+        return [os.path.abspath(folder_paths.get_input_directory())]
+    if root == "output":
+        return [os.path.abspath(folder_paths.get_output_directory())]
+    return []
+
+def escape_like_prefix(s: str, escape: str = "!") -> tuple[str, str]:
+    """Escapes %, _ and the escape char itself in a LIKE prefix.
+    Returns (escaped_prefix, escape_char). Caller should append '%' and pass escape=escape_char to .like().
+    """
+    s = s.replace(escape, escape + escape)  # escape the escape char first
+    s = s.replace("%", escape + "%").replace("_", escape + "_")  # escape LIKE wildcards
+    return s, escape
+
+def fast_asset_file_check(
+    *,
+    mtime_db: int | None,
+    size_db: int | None,
+    stat_result: os.stat_result,
+) -> bool:
+    if mtime_db is None:
+        return False
+    actual_mtime_ns = getattr(stat_result, "st_mtime_ns", int(stat_result.st_mtime * 1_000_000_000))
+    if int(mtime_db) != int(actual_mtime_ns):
+        return False
+    sz = int(size_db or 0)
+    if sz > 0:
+        return int(stat_result.st_size) == sz
+    return True
+
+def utcnow() -> datetime:
+    """Naive UTC timestamp (no tzinfo). We always treat DB datetimes as UTC."""
+    return datetime.now(timezone.utc).replace(tzinfo=None)
+
+def get_comfy_models_folders() -> list[tuple[str, list[str]]]:
+    """Build a list of (folder_name, base_paths[]) categories that are configured for model locations.
+
+    We trust `folder_paths.folder_names_and_paths` and include a category if
+    *any* of its base paths lies under the Comfy `models_dir`.
+    """
+    targets: list[tuple[str, list[str]]] = []
+    models_root = os.path.abspath(folder_paths.models_dir)
+    for name, values in folder_paths.folder_names_and_paths.items():
+        paths, _exts = values[0], values[1]  # NOTE: this prevents nodepacks that hackily edit folder_... from breaking ComfyUI
+        if any(os.path.abspath(p).startswith(models_root + os.sep) for p in paths):
+            targets.append((name, paths))
+    return targets
+
+def resolve_destination_from_tags(tags: list[str]) -> tuple[str, list[str]]:
+    """Validates and maps tags -> (base_dir, subdirs_for_fs)"""
+    root = tags[0]
+    if root == "models":
+        if len(tags) < 2:
+            raise ValueError("at least two tags required for model asset")
+        try:
+            bases = folder_paths.folder_names_and_paths[tags[1]][0]
+        except KeyError:
+            raise ValueError(f"unknown model category '{tags[1]}'")
+        if not bases:
+            raise ValueError(f"no base path configured for category '{tags[1]}'")
+        base_dir = os.path.abspath(bases[0])
+        raw_subdirs = tags[2:]
+    else:
+        base_dir = os.path.abspath(
+            folder_paths.get_input_directory() if root == "input" else folder_paths.get_output_directory()
+        )
+        raw_subdirs = tags[1:]
+    for i in raw_subdirs:
+        if i in (".", ".."):
+            raise ValueError("invalid path component in tags")
+
+    return base_dir, raw_subdirs if raw_subdirs else []
+
+def ensure_within_base(candidate: str, base: str) -> None:
+    cand_abs = os.path.abspath(candidate)
+    base_abs = os.path.abspath(base)
+    try:
+        if os.path.commonpath([cand_abs, base_abs]) != base_abs:
+            raise ValueError("destination escapes base directory")
+    except Exception:
+        raise ValueError("invalid destination path")
+
+def compute_relative_filename(file_path: str) -> str | None:
+    """
+    Return the model's path relative to the last well-known folder (the model category),
+    using forward slashes, eg:
+      /.../models/checkpoints/flux/123/flux.safetensors -> "flux/123/flux.safetensors"
+      /.../models/text_encoders/clip_g.safetensors -> "clip_g.safetensors"
+
+    For non-model paths, returns None.
+    NOTE: this is a temporary helper, used only for initializing metadata["filename"] field.
+    """
+    try:
+        root_category, rel_path = get_relative_to_root_category_path_of_asset(file_path)
+    except ValueError:
+        return None
+
+    p = Path(rel_path)
+    parts = [seg for seg in p.parts if seg not in (".", "..", p.anchor)]
+    if not parts:
+        return None
+
+    if root_category == "models":
+        # parts[0] is the category ("checkpoints", "vae", etc) – drop it
+        inside = parts[1:] if len(parts) > 1 else [parts[0]]
+        return "/".join(inside)
+    return "/".join(parts)  # input/output: keep all parts
+
+def get_relative_to_root_category_path_of_asset(file_path: str) -> tuple[Literal["input", "output", "models"], str]:
+    """Given an absolute or relative file path, determine which root category the path belongs to:
+      - 'input' if the file resides under `folder_paths.get_input_directory()`
+      - 'output' if the file resides under `folder_paths.get_output_directory()`
+      - 'models' if the file resides under any base path of categories returned by `get_comfy_models_folders()`
+
+    Returns:
+        (root_category, relative_path_inside_that_root)
+        For 'models', the relative path is prefixed with the category name:
+            e.g. ('models', 'vae/test/sub/ae.safetensors')
+
+    Raises:
+        ValueError: if the path does not belong to input, output, or configured model bases.
+    """
+    fp_abs = os.path.abspath(file_path)
+
+    def _is_within(child: str, parent: str) -> bool:
+        try:
+            return os.path.commonpath([child, parent]) == parent
+        except Exception:
+            return False
+
+    def _rel(child: str, parent: str) -> str:
+        return os.path.relpath(os.path.join(os.sep, os.path.relpath(child, parent)), os.sep)
+
+    # 1) input
+    input_base = os.path.abspath(folder_paths.get_input_directory())
+    if _is_within(fp_abs, input_base):
+        return "input", _rel(fp_abs, input_base)
+
+    # 2) output
+    output_base = os.path.abspath(folder_paths.get_output_directory())
+    if _is_within(fp_abs, output_base):
+        return "output", _rel(fp_abs, output_base)
+
+    # 3) models (check deepest matching base to avoid ambiguity)
+    best: tuple[int, str, str] | None = None  # (base_len, bucket, rel_inside_bucket)
+    for bucket, bases in get_comfy_models_folders():
+        for b in bases:
+            base_abs = os.path.abspath(b)
+            if not _is_within(fp_abs, base_abs):
+                continue
+            cand = (len(base_abs), bucket, _rel(fp_abs, base_abs))
+            if best is None or cand[0] > best[0]:
+                best = cand
+
+    if best is not None:
+        _, bucket, rel_inside = best
+        combined = os.path.join(bucket, rel_inside)
+        return "models", os.path.relpath(os.path.join(os.sep, combined), os.sep)
+
+    raise ValueError(f"Path is not within input, output, or configured model bases: {file_path}")
+
+def get_name_and_tags_from_asset_path(file_path: str) -> tuple[str, list[str]]:
+    """Return a tuple (name, tags) derived from a filesystem path.
+
+    Semantics:
+      - Root category is determined by `get_relative_to_root_category_path_of_asset`.
+      - The returned `name` is the base filename with extension from the relative path.
+      - The returned `tags` are:
+            [root_category] + parent folders of the relative path (in order)
+        For 'models', this means:
+            file '/.../ModelsDir/vae/test_tag/ae.safetensors'
+            -> root_category='models', some_path='vae/test_tag/ae.safetensors'
+            -> name='ae.safetensors', tags=['models', 'vae', 'test_tag']
+
+    Raises:
+        ValueError: if the path does not belong to input, output, or configured model bases.
+    """
+    root_category, some_path = get_relative_to_root_category_path_of_asset(file_path)
+    p = Path(some_path)
+    parent_parts = [part for part in p.parent.parts if part not in (".", "..", p.anchor)]
+    return p.name, list(dict.fromkeys(normalize_tags([root_category, *parent_parts])))
+
+def normalize_tags(tags: list[str] | None) -> list[str]:
+    """
+    Normalize a list of tags by:
+      - Stripping whitespace and converting to lowercase.
+      - Removing duplicates.
+    """
+    return [t.strip().lower() for t in (tags or []) if (t or "").strip()]
+
+def collect_models_files() -> list[str]:
+    out: list[str] = []
+    for folder_name, bases in get_comfy_models_folders():
+        rel_files = folder_paths.get_filename_list(folder_name) or []
+        for rel_path in rel_files:
+            abs_path = folder_paths.get_full_path(folder_name, rel_path)
+            if not abs_path:
+                continue
+            abs_path = os.path.abspath(abs_path)
+            allowed = False
+            for b in bases:
+                base_abs = os.path.abspath(b)
+                with contextlib.suppress(Exception):
+                    if os.path.commonpath([abs_path, base_abs]) == base_abs:
+                        allowed = True
+                        break
+            if allowed:
+                out.append(abs_path)
+    return out
+
+def is_scalar(v):
+    if v is None:
+        return True
+    if isinstance(v, bool):
+        return True
+    if isinstance(v, (int, float, Decimal, str)):
+        return True
+    return False
+
+def project_kv(key: str, value):
+    """
+    Turn a metadata key/value into typed projection rows.
+    Returns list[dict] with keys:
+      key, ordinal, and one of val_str / val_num / val_bool / val_json (others None)
+    """
+    rows: list[dict] = []
+
+    def _null_row(ordinal: int) -> dict:
+        return {
+            "key": key, "ordinal": ordinal,
+            "val_str": None, "val_num": None, "val_bool": None, "val_json": None
+        }
+
+    if value is None:
+        rows.append(_null_row(0))
+        return rows
+
+    if is_scalar(value):
+        if isinstance(value, bool):
+            rows.append({"key": key, "ordinal": 0, "val_bool": bool(value)})
+        elif isinstance(value, (int, float, Decimal)):
+            num = value if isinstance(value, Decimal) else Decimal(str(value))
+            rows.append({"key": key, "ordinal": 0, "val_num": num})
+        elif isinstance(value, str):
+            rows.append({"key": key, "ordinal": 0, "val_str": value})
+        else:
+            rows.append({"key": key, "ordinal": 0, "val_json": value})
+        return rows
+
+    if isinstance(value, list):
+        if all(is_scalar(x) for x in value):
+            for i, x in enumerate(value):
+                if x is None:
+                    rows.append(_null_row(i))
+                elif isinstance(x, bool):
+                    rows.append({"key": key, "ordinal": i, "val_bool": bool(x)})
+                elif isinstance(x, (int, float, Decimal)):
+                    num = x if isinstance(x, Decimal) else Decimal(str(x))
+                    rows.append({"key": key, "ordinal": i, "val_num": num})
+                elif isinstance(x, str):
+                    rows.append({"key": key, "ordinal": i, "val_str": x})
+                else:
+                    rows.append({"key": key, "ordinal": i, "val_json": x})
+            return rows
+        for i, x in enumerate(value):
+            rows.append({"key": key, "ordinal": i, "val_json": x})
+        return rows
+
+    rows.append({"key": key, "ordinal": 0, "val_json": value})
+    return rows

app/assets/manager.py

@@ -0,0 +1,516 @@
+import os
+import mimetypes
+import contextlib
+from typing import Sequence
+
+from app.database.db import create_session
+from app.assets.api import schemas_out, schemas_in
+from app.assets.database.queries import (
+    asset_exists_by_hash,
+    asset_info_exists_for_asset_id,
+    get_asset_by_hash,
+    get_asset_info_by_id,
+    fetch_asset_info_asset_and_tags,
+    fetch_asset_info_and_asset,
+    create_asset_info_for_existing_asset,
+    touch_asset_info_by_id,
+    update_asset_info_full,
+    delete_asset_info_by_id,
+    list_cache_states_by_asset_id,
+    list_asset_infos_page,
+    list_tags_with_usage,
+    get_asset_tags,
+    add_tags_to_asset_info,
+    remove_tags_from_asset_info,
+    pick_best_live_path,
+    ingest_fs_asset,
+    set_asset_info_preview,
+)
+from app.assets.helpers import resolve_destination_from_tags, ensure_within_base
+from app.assets.database.models import Asset
+
+
+def _safe_sort_field(requested: str | None) -> str:
+    if not requested:
+        return "created_at"
+    v = requested.lower()
+    if v in {"name", "created_at", "updated_at", "size", "last_access_time"}:
+        return v
+    return "created_at"
+
+
+def _get_size_mtime_ns(path: str) -> tuple[int, int]:
+    st = os.stat(path, follow_symlinks=True)
+    return st.st_size, getattr(st, "st_mtime_ns", int(st.st_mtime * 1_000_000_000))
+
+
+def _safe_filename(name: str | None, fallback: str) -> str:
+    n = os.path.basename((name or "").strip() or fallback)
+    if n:
+        return n
+    return fallback
+
+
+def asset_exists(*, asset_hash: str) -> bool:
+    """
+    Check if an asset with a given hash exists in database.
+    """
+    with create_session() as session:
+        return asset_exists_by_hash(session, asset_hash=asset_hash)
+
+
+def list_assets(
+    *,
+    include_tags: Sequence[str] | None = None,
+    exclude_tags: Sequence[str] | None = None,
+    name_contains: str | None = None,
+    metadata_filter: dict | None = None,
+    limit: int = 20,
+    offset: int = 0,
+    sort: str = "created_at",
+    order: str = "desc",
+    owner_id: str = "",
+) -> schemas_out.AssetsList:
+    sort = _safe_sort_field(sort)
+    order = "desc" if (order or "desc").lower() not in {"asc", "desc"} else order.lower()
+
+    with create_session() as session:
+        infos, tag_map, total = list_asset_infos_page(
+            session,
+            owner_id=owner_id,
+            include_tags=include_tags,
+            exclude_tags=exclude_tags,
+            name_contains=name_contains,
+            metadata_filter=metadata_filter,
+            limit=limit,
+            offset=offset,
+            sort=sort,
+            order=order,
+        )
+
+    summaries: list[schemas_out.AssetSummary] = []
+    for info in infos:
+        asset = info.asset
+        tags = tag_map.get(info.id, [])
+        summaries.append(
+            schemas_out.AssetSummary(
+                id=info.id,
+                name=info.name,
+                asset_hash=asset.hash if asset else None,
+                size=int(asset.size_bytes) if asset else None,
+                mime_type=asset.mime_type if asset else None,
+                tags=tags,
+                created_at=info.created_at,
+                updated_at=info.updated_at,
+                last_access_time=info.last_access_time,
+            )
+        )
+
+    return schemas_out.AssetsList(
+        assets=summaries,
+        total=total,
+        has_more=(offset + len(summaries)) < total,
+    )
+
+
+def get_asset(
+    *,
+    asset_info_id: str,
+    owner_id: str = "",
+) -> schemas_out.AssetDetail:
+    with create_session() as session:
+        res = fetch_asset_info_asset_and_tags(session, asset_info_id=asset_info_id, owner_id=owner_id)
+        if not res:
+            raise ValueError(f"AssetInfo {asset_info_id} not found")
+        info, asset, tag_names = res
+        preview_id = info.preview_id
+
+    return schemas_out.AssetDetail(
+        id=info.id,
+        name=info.name,
+        asset_hash=asset.hash if asset else None,
+        size=int(asset.size_bytes) if asset and asset.size_bytes is not None else None,
+        mime_type=asset.mime_type if asset else None,
+        tags=tag_names,
+        user_metadata=info.user_metadata or {},
+        preview_id=preview_id,
+        created_at=info.created_at,
+        last_access_time=info.last_access_time,
+    )
+
+
+def resolve_asset_content_for_download(
+    *,
+    asset_info_id: str,
+    owner_id: str = "",
+) -> tuple[str, str, str]:
+    with create_session() as session:
+        pair = fetch_asset_info_and_asset(session, asset_info_id=asset_info_id, owner_id=owner_id)
+        if not pair:
+            raise ValueError(f"AssetInfo {asset_info_id} not found")
+
+        info, asset = pair
+        states = list_cache_states_by_asset_id(session, asset_id=asset.id)
+        abs_path = pick_best_live_path(states)
+        if not abs_path:
+            raise FileNotFoundError
+
+        touch_asset_info_by_id(session, asset_info_id=asset_info_id)
+        session.commit()
+
+        ctype = asset.mime_type or mimetypes.guess_type(info.name or abs_path)[0] or "application/octet-stream"
+        download_name = info.name or os.path.basename(abs_path)
+        return abs_path, ctype, download_name
+
+
+def upload_asset_from_temp_path(
+    spec: schemas_in.UploadAssetSpec,
+    *,
+    temp_path: str,
+    client_filename: str | None = None,
+    owner_id: str = "",
+    expected_asset_hash: str | None = None,
+) -> schemas_out.AssetCreated:
+    """
+    Create new asset or update existing asset from a temporary file path.
+    """
+    try:
+        # NOTE: blake3 is not required right now, so this will fail if blake3 is not installed in local environment
+        import app.assets.hashing as hashing
+        digest = hashing.blake3_hash(temp_path)
+    except Exception as e:
+        raise RuntimeError(f"failed to hash uploaded file: {e}")
+    asset_hash = "blake3:" + digest
+
+    if expected_asset_hash and asset_hash != expected_asset_hash.strip().lower():
+        raise ValueError("HASH_MISMATCH")
+
+    with create_session() as session:
+        existing = get_asset_by_hash(session, asset_hash=asset_hash)
+        if existing is not None:
+            with contextlib.suppress(Exception):
+                if temp_path and os.path.exists(temp_path):
+                    os.remove(temp_path)
+
+            display_name = _safe_filename(spec.name or (client_filename or ""), fallback=digest)
+            info = create_asset_info_for_existing_asset(
+                session,
+                asset_hash=asset_hash,
+                name=display_name,
+                user_metadata=spec.user_metadata or {},
+                tags=spec.tags or [],
+                tag_origin="manual",
+                owner_id=owner_id,
+            )
+            tag_names = get_asset_tags(session, asset_info_id=info.id)
+            session.commit()
+
+            return schemas_out.AssetCreated(
+                id=info.id,
+                name=info.name,
+                asset_hash=existing.hash,
+                size=int(existing.size_bytes) if existing.size_bytes is not None else None,
+                mime_type=existing.mime_type,
+                tags=tag_names,
+                user_metadata=info.user_metadata or {},
+                preview_id=info.preview_id,
+                created_at=info.created_at,
+                last_access_time=info.last_access_time,
+                created_new=False,
+            )
+
+    base_dir, subdirs = resolve_destination_from_tags(spec.tags)
+    dest_dir = os.path.join(base_dir, *subdirs) if subdirs else base_dir
+    os.makedirs(dest_dir, exist_ok=True)
+
+    src_for_ext = (client_filename or spec.name or "").strip()
+    _ext = os.path.splitext(os.path.basename(src_for_ext))[1] if src_for_ext else ""
+    ext = _ext if 0 < len(_ext) <= 16 else ""
+    hashed_basename = f"{digest}{ext}"
+    dest_abs = os.path.abspath(os.path.join(dest_dir, hashed_basename))
+    ensure_within_base(dest_abs, base_dir)
+
+    content_type = (
+        mimetypes.guess_type(os.path.basename(src_for_ext), strict=False)[0]
+        or mimetypes.guess_type(hashed_basename, strict=False)[0]
+        or "application/octet-stream"
+    )
+
+    try:
+        os.replace(temp_path, dest_abs)
+    except Exception as e:
+        raise RuntimeError(f"failed to move uploaded file into place: {e}")
+
+    try:
+        size_bytes, mtime_ns = _get_size_mtime_ns(dest_abs)
+    except OSError as e:
+        raise RuntimeError(f"failed to stat destination file: {e}")
+
+    with create_session() as session:
+        result = ingest_fs_asset(
+            session,
+            asset_hash=asset_hash,
+            abs_path=dest_abs,
+            size_bytes=size_bytes,
+            mtime_ns=mtime_ns,
+            mime_type=content_type,
+            info_name=_safe_filename(spec.name or (client_filename or ""), fallback=digest),
+            owner_id=owner_id,
+            preview_id=None,
+            user_metadata=spec.user_metadata or {},
+            tags=spec.tags,
+            tag_origin="manual",
+            require_existing_tags=False,
+        )
+        info_id = result["asset_info_id"]
+        if not info_id:
+            raise RuntimeError("failed to create asset metadata")
+
+        pair = fetch_asset_info_and_asset(session, asset_info_id=info_id, owner_id=owner_id)
+        if not pair:
+            raise RuntimeError("inconsistent DB state after ingest")
+        info, asset = pair
+        tag_names = get_asset_tags(session, asset_info_id=info.id)
+        created_result = schemas_out.AssetCreated(
+            id=info.id,
+            name=info.name,
+            asset_hash=asset.hash,
+            size=int(asset.size_bytes),
+            mime_type=asset.mime_type,
+            tags=tag_names,
+            user_metadata=info.user_metadata or {},
+            preview_id=info.preview_id,
+            created_at=info.created_at,
+            last_access_time=info.last_access_time,
+            created_new=result["asset_created"],
+        )
+        session.commit()
+
+    return created_result
+
+
+def update_asset(
+    *,
+    asset_info_id: str,
+    name: str | None = None,
+    tags: list[str] | None = None,
+    user_metadata: dict | None = None,
+    owner_id: str = "",
+) -> schemas_out.AssetUpdated:
+    with create_session() as session:
+        info_row = get_asset_info_by_id(session, asset_info_id=asset_info_id)
+        if not info_row:
+            raise ValueError(f"AssetInfo {asset_info_id} not found")
+        if info_row.owner_id and info_row.owner_id != owner_id:
+            raise PermissionError("not owner")
+
+        info = update_asset_info_full(
+            session,
+            asset_info_id=asset_info_id,
+            name=name,
+            tags=tags,
+            user_metadata=user_metadata,
+            tag_origin="manual",
+            asset_info_row=info_row,
+        )
+
+        tag_names = get_asset_tags(session, asset_info_id=asset_info_id)
+        result = schemas_out.AssetUpdated(
+            id=info.id,
+            name=info.name,
+            asset_hash=info.asset.hash if info.asset else None,
+            tags=tag_names,
+            user_metadata=info.user_metadata or {},
+            updated_at=info.updated_at,
+        )
+        session.commit()
+
+    return result
+
+
+def set_asset_preview(
+    *,
+    asset_info_id: str,
+    preview_asset_id: str | None = None,
+    owner_id: str = "",
+) -> schemas_out.AssetDetail:
+    with create_session() as session:
+        info_row = get_asset_info_by_id(session, asset_info_id=asset_info_id)
+        if not info_row:
+            raise ValueError(f"AssetInfo {asset_info_id} not found")
+        if info_row.owner_id and info_row.owner_id != owner_id:
+            raise PermissionError("not owner")
+
+        set_asset_info_preview(
+            session,
+            asset_info_id=asset_info_id,
+            preview_asset_id=preview_asset_id,
+        )
+
+        res = fetch_asset_info_asset_and_tags(session, asset_info_id=asset_info_id, owner_id=owner_id)
+        if not res:
+            raise RuntimeError("State changed during preview update")
+        info, asset, tags = res
+        result = schemas_out.AssetDetail(
+            id=info.id,
+            name=info.name,
+            asset_hash=asset.hash if asset else None,
+            size=int(asset.size_bytes) if asset and asset.size_bytes is not None else None,
+            mime_type=asset.mime_type if asset else None,
+            tags=tags,
+            user_metadata=info.user_metadata or {},
+            preview_id=info.preview_id,
+            created_at=info.created_at,
+            last_access_time=info.last_access_time,
+        )
+        session.commit()
+
+    return result
+
+
+def delete_asset_reference(*, asset_info_id: str, owner_id: str, delete_content_if_orphan: bool = True) -> bool:
+    with create_session() as session:
+        info_row = get_asset_info_by_id(session, asset_info_id=asset_info_id)
+        asset_id = info_row.asset_id if info_row else None
+        deleted = delete_asset_info_by_id(session, asset_info_id=asset_info_id, owner_id=owner_id)
+        if not deleted:
+            session.commit()
+            return False
+
+        if not delete_content_if_orphan or not asset_id:
+            session.commit()
+            return True
+
+        still_exists = asset_info_exists_for_asset_id(session, asset_id=asset_id)
+        if still_exists:
+            session.commit()
+            return True
+
+        states = list_cache_states_by_asset_id(session, asset_id=asset_id)
+        file_paths = [s.file_path for s in (states or []) if getattr(s, "file_path", None)]
+
+        asset_row = session.get(Asset, asset_id)
+        if asset_row is not None:
+            session.delete(asset_row)
+
+        session.commit()
+        for p in file_paths:
+            with contextlib.suppress(Exception):
+                if p and os.path.isfile(p):
+                    os.remove(p)
+    return True
+
+
+def create_asset_from_hash(
+    *,
+    hash_str: str,
+    name: str,
+    tags: list[str] | None = None,
+    user_metadata: dict | None = None,
+    owner_id: str = "",
+) -> schemas_out.AssetCreated | None:
+    canonical = hash_str.strip().lower()
+    with create_session() as session:
+        asset = get_asset_by_hash(session, asset_hash=canonical)
+        if not asset:
+            return None
+
+        info = create_asset_info_for_existing_asset(
+            session,
+            asset_hash=canonical,
+            name=_safe_filename(name, fallback=canonical.split(":", 1)[1]),
+            user_metadata=user_metadata or {},
+            tags=tags or [],
+            tag_origin="manual",
+            owner_id=owner_id,
+        )
+        tag_names = get_asset_tags(session, asset_info_id=info.id)
+        result = schemas_out.AssetCreated(
+            id=info.id,
+            name=info.name,
+            asset_hash=asset.hash,
+            size=int(asset.size_bytes),
+            mime_type=asset.mime_type,
+            tags=tag_names,
+            user_metadata=info.user_metadata or {},
+            preview_id=info.preview_id,
+            created_at=info.created_at,
+            last_access_time=info.last_access_time,
+            created_new=False,
+        )
+        session.commit()
+
+    return result
+
+
+def add_tags_to_asset(
+    *,
+    asset_info_id: str,
+    tags: list[str],
+    origin: str = "manual",
+    owner_id: str = "",
+) -> schemas_out.TagsAdd:
+    with create_session() as session:
+        info_row = get_asset_info_by_id(session, asset_info_id=asset_info_id)
+        if not info_row:
+            raise ValueError(f"AssetInfo {asset_info_id} not found")
+        if info_row.owner_id and info_row.owner_id != owner_id:
+            raise PermissionError("not owner")
+        data = add_tags_to_asset_info(
+            session,
+            asset_info_id=asset_info_id,
+            tags=tags,
+            origin=origin,
+            create_if_missing=True,
+            asset_info_row=info_row,
+        )
+        session.commit()
+    return schemas_out.TagsAdd(**data)
+
+
+def remove_tags_from_asset(
+    *,
+    asset_info_id: str,
+    tags: list[str],
+    owner_id: str = "",
+) -> schemas_out.TagsRemove:
+    with create_session() as session:
+        info_row = get_asset_info_by_id(session, asset_info_id=asset_info_id)
+        if not info_row:
+            raise ValueError(f"AssetInfo {asset_info_id} not found")
+        if info_row.owner_id and info_row.owner_id != owner_id:
+            raise PermissionError("not owner")
+
+        data = remove_tags_from_asset_info(
+            session,
+            asset_info_id=asset_info_id,
+            tags=tags,
+        )
+        session.commit()
+    return schemas_out.TagsRemove(**data)
+
+
+def list_tags(
+    prefix: str | None = None,
+    limit: int = 100,
+    offset: int = 0,
+    order: str = "count_desc",
+    include_zero: bool = True,
+    owner_id: str = "",
+) -> schemas_out.TagsList:
+    limit = max(1, min(1000, limit))
+    offset = max(0, offset)
+
+    with create_session() as session:
+        rows, total = list_tags_with_usage(
+            session,
+            prefix=prefix,
+            limit=limit,
+            offset=offset,
+            include_zero=include_zero,
+            order=order,
+            owner_id=owner_id,
+        )
+
+    tags = [schemas_out.TagUsage(name=name, count=count, type=tag_type) for (name, tag_type, count) in rows]
+    return schemas_out.TagsList(tags=tags, total=total, has_more=(offset + len(tags)) < total)

app/assets/scanner.py

@@ -0,0 +1,263 @@
+import contextlib
+import time
+import logging
+import os
+import sqlalchemy
+
+import folder_paths
+from app.database.db import create_session, dependencies_available
+from app.assets.helpers import (
+    collect_models_files, compute_relative_filename, fast_asset_file_check, get_name_and_tags_from_asset_path,
+    list_tree,prefixes_for_root, escape_like_prefix,
+    RootType
+)
+from app.assets.database.tags import add_missing_tag_for_asset_id, ensure_tags_exist, remove_missing_tag_for_asset_id
+from app.assets.database.bulk_ops import seed_from_paths_batch
+from app.assets.database.models import Asset, AssetCacheState, AssetInfo
+
+
+def seed_assets(roots: tuple[RootType, ...], enable_logging: bool = False) -> None:
+    """
+    Scan the given roots and seed the assets into the database.
+    """
+    if not dependencies_available():
+        if enable_logging:
+            logging.warning("Database dependencies not available, skipping assets scan")
+        return
+    t_start = time.perf_counter()
+    created = 0
+    skipped_existing = 0
+    orphans_pruned = 0
+    paths: list[str] = []
+    try:
+        existing_paths: set[str] = set()
+        for r in roots:
+            try:
+                survivors: set[str] = _fast_db_consistency_pass(r, collect_existing_paths=True, update_missing_tags=True)
+                if survivors:
+                    existing_paths.update(survivors)
+            except Exception as e:
+                logging.exception("fast DB scan failed for %s: %s", r, e)
+
+        try:
+            orphans_pruned = _prune_orphaned_assets(roots)
+        except Exception as e:
+            logging.exception("orphan pruning failed: %s", e)
+
+        if "models" in roots:
+            paths.extend(collect_models_files())
+        if "input" in roots:
+            paths.extend(list_tree(folder_paths.get_input_directory()))
+        if "output" in roots:
+            paths.extend(list_tree(folder_paths.get_output_directory()))
+
+        specs: list[dict] = []
+        tag_pool: set[str] = set()
+        for p in paths:
+            abs_p = os.path.abspath(p)
+            if abs_p in existing_paths:
+                skipped_existing += 1
+                continue
+            try:
+                stat_p = os.stat(abs_p, follow_symlinks=False)
+            except OSError:
+                continue
+            # skip empty files
+            if not stat_p.st_size:
+                continue
+            name, tags = get_name_and_tags_from_asset_path(abs_p)
+            specs.append(
+                {
+                    "abs_path": abs_p,
+                    "size_bytes": stat_p.st_size,
+                    "mtime_ns": getattr(stat_p, "st_mtime_ns", int(stat_p.st_mtime * 1_000_000_000)),
+                    "info_name": name,
+                    "tags": tags,
+                    "fname": compute_relative_filename(abs_p),
+                }
+            )
+            for t in tags:
+                tag_pool.add(t)
+        # if no file specs, nothing to do
+        if not specs:
+            return
+        with create_session() as sess:
+            if tag_pool:
+                ensure_tags_exist(sess, tag_pool, tag_type="user")
+
+            result = seed_from_paths_batch(sess, specs=specs, owner_id="")
+            created += result["inserted_infos"]
+            sess.commit()
+    finally:
+        if enable_logging:
+            logging.info(
+                "Assets scan(roots=%s) completed in %.3fs (created=%d, skipped_existing=%d, orphans_pruned=%d, total_seen=%d)",
+                roots,
+                time.perf_counter() - t_start,
+                created,
+                skipped_existing,
+                orphans_pruned,
+                len(paths),
+            )
+
+
+def _prune_orphaned_assets(roots: tuple[RootType, ...]) -> int:
+    """Prune cache states outside configured prefixes, then delete orphaned seed assets."""
+    all_prefixes = [os.path.abspath(p) for r in roots for p in prefixes_for_root(r)]
+    if not all_prefixes:
+        return 0
+
+    def make_prefix_condition(prefix: str):
+        base = prefix if prefix.endswith(os.sep) else prefix + os.sep
+        escaped, esc = escape_like_prefix(base)
+        return AssetCacheState.file_path.like(escaped + "%", escape=esc)
+
+    matches_valid_prefix = sqlalchemy.or_(*[make_prefix_condition(p) for p in all_prefixes])
+
+    orphan_subq = (
+        sqlalchemy.select(Asset.id)
+        .outerjoin(AssetCacheState, AssetCacheState.asset_id == Asset.id)
+        .where(Asset.hash.is_(None), AssetCacheState.id.is_(None))
+    ).scalar_subquery()
+
+    with create_session() as sess:
+        sess.execute(sqlalchemy.delete(AssetCacheState).where(~matches_valid_prefix))
+        sess.execute(sqlalchemy.delete(AssetInfo).where(AssetInfo.asset_id.in_(orphan_subq)))
+        result = sess.execute(sqlalchemy.delete(Asset).where(Asset.id.in_(orphan_subq)))
+        sess.commit()
+        return result.rowcount
+
+
+def _fast_db_consistency_pass(
+    root: RootType,
+    *,
+    collect_existing_paths: bool = False,
+    update_missing_tags: bool = False,
+) -> set[str] | None:
+    """Fast DB+FS pass for a root:
+      - Toggle needs_verify per state using fast check
+      - For hashed assets with at least one fast-ok state in this root: delete stale missing states
+      - For seed assets with all states missing: delete Asset and its AssetInfos
+      - Optionally add/remove 'missing' tags based on fast-ok in this root
+      - Optionally return surviving absolute paths
+    """
+    prefixes = prefixes_for_root(root)
+    if not prefixes:
+        return set() if collect_existing_paths else None
+
+    conds = []
+    for p in prefixes:
+        base = os.path.abspath(p)
+        if not base.endswith(os.sep):
+            base += os.sep
+        escaped, esc = escape_like_prefix(base)
+        conds.append(AssetCacheState.file_path.like(escaped + "%", escape=esc))
+
+    with create_session() as sess:
+        rows = (
+            sess.execute(
+                sqlalchemy.select(
+                    AssetCacheState.id,
+                    AssetCacheState.file_path,
+                    AssetCacheState.mtime_ns,
+                    AssetCacheState.needs_verify,
+                    AssetCacheState.asset_id,
+                    Asset.hash,
+                    Asset.size_bytes,
+                )
+                .join(Asset, Asset.id == AssetCacheState.asset_id)
+                .where(sqlalchemy.or_(*conds))
+                .order_by(AssetCacheState.asset_id.asc(), AssetCacheState.id.asc())
+            )
+        ).all()
+
+        by_asset: dict[str, dict] = {}
+        for sid, fp, mtime_db, needs_verify, aid, a_hash, a_size in rows:
+            acc = by_asset.get(aid)
+            if acc is None:
+                acc = {"hash": a_hash, "size_db": int(a_size or 0), "states": []}
+                by_asset[aid] = acc
+
+            fast_ok = False
+            try:
+                exists = True
+                fast_ok = fast_asset_file_check(
+                    mtime_db=mtime_db,
+                    size_db=acc["size_db"],
+                    stat_result=os.stat(fp, follow_symlinks=True),
+                )
+            except FileNotFoundError:
+                exists = False
+            except OSError:
+                exists = False
+
+            acc["states"].append({
+                "sid": sid,
+                "fp": fp,
+                "exists": exists,
+                "fast_ok": fast_ok,
+                "needs_verify": bool(needs_verify),
+            })
+
+        to_set_verify: list[int] = []
+        to_clear_verify: list[int] = []
+        stale_state_ids: list[int] = []
+        survivors: set[str] = set()
+
+        for aid, acc in by_asset.items():
+            a_hash = acc["hash"]
+            states = acc["states"]
+            any_fast_ok = any(s["fast_ok"] for s in states)
+            all_missing = all(not s["exists"] for s in states)
+
+            for s in states:
+                if not s["exists"]:
+                    continue
+                if s["fast_ok"] and s["needs_verify"]:
+                    to_clear_verify.append(s["sid"])
+                if not s["fast_ok"] and not s["needs_verify"]:
+                    to_set_verify.append(s["sid"])
+
+            if a_hash is None:
+                if states and all_missing:  # remove seed Asset completely, if no valid AssetCache exists
+                    sess.execute(sqlalchemy.delete(AssetInfo).where(AssetInfo.asset_id == aid))
+                    asset = sess.get(Asset, aid)
+                    if asset:
+                        sess.delete(asset)
+                else:
+                    for s in states:
+                        if s["exists"]:
+                            survivors.add(os.path.abspath(s["fp"]))
+                continue
+
+            if any_fast_ok:  # if Asset has at least one valid AssetCache record, remove any invalid AssetCache records
+                for s in states:
+                    if not s["exists"]:
+                        stale_state_ids.append(s["sid"])
+                if update_missing_tags:
+                    with contextlib.suppress(Exception):
+                        remove_missing_tag_for_asset_id(sess, asset_id=aid)
+            elif update_missing_tags:
+                with contextlib.suppress(Exception):
+                    add_missing_tag_for_asset_id(sess, asset_id=aid, origin="automatic")
+
+            for s in states:
+                if s["exists"]:
+                    survivors.add(os.path.abspath(s["fp"]))
+
+        if stale_state_ids:
+            sess.execute(sqlalchemy.delete(AssetCacheState).where(AssetCacheState.id.in_(stale_state_ids)))
+        if to_set_verify:
+            sess.execute(
+                sqlalchemy.update(AssetCacheState)
+                .where(AssetCacheState.id.in_(to_set_verify))
+                .values(needs_verify=True)
+            )
+        if to_clear_verify:
+            sess.execute(
+                sqlalchemy.update(AssetCacheState)
+                .where(AssetCacheState.id.in_(to_clear_verify))
+                .values(needs_verify=False)
+            )
+        sess.commit()
+        return survivors if collect_existing_paths else None

app/database/models.py

@@ -1,14 +1,21 @@
-from sqlalchemy.orm import declarative_base
+from typing import Any
+from datetime import datetime
+from sqlalchemy.orm import DeclarativeBase
 
-Base = declarative_base()
+class Base(DeclarativeBase):
+    pass
 
-
-def to_dict(obj):
+def to_dict(obj: Any, include_none: bool = False) -> dict[str, Any]:
     fields = obj.__table__.columns.keys()
-    return {
-        field: (val.to_dict() if hasattr(val, "to_dict") else val)
-        for field in fields
-        if (val := getattr(obj, field))
-    }
+    out: dict[str, Any] = {}
+    for field in fields:
+        val = getattr(obj, field)
+        if val is None and not include_none:
+            continue
+        if isinstance(val, datetime):
+            out[field] = val.isoformat()
+        else:
+            out[field] = val
+    return out
 
 # TODO: Define models here

app/frontend_management.py

@@ -10,7 +10,8 @@ import importlib
 from dataclasses import dataclass
 from functools import cached_property
 from pathlib import Path
-from typing import TypedDict, Optional
+from typing import Dict, TypedDict, Optional
+from aiohttp import web
 from importlib.metadata import version
 
 import requests
@@ -257,7 +258,54 @@ comfyui-frontend-package is not installed.
             sys.exit(-1)
 
     @classmethod
-    def templates_path(cls) -> str:
+    def template_asset_map(cls) -> Optional[Dict[str, str]]:
+        """Return a mapping of template asset names to their absolute paths."""
+        try:
+            from comfyui_workflow_templates import (
+                get_asset_path,
+                iter_templates,
+            )
+        except ImportError:
+            logging.error(
+                f"""
+********** ERROR ***********
+
+comfyui-workflow-templates is not installed.
+
+{frontend_install_warning_message()}
+
+********** ERROR ***********
+""".strip()
+            )
+            return None
+
+        try:
+            template_entries = list(iter_templates())
+        except Exception as exc:
+            logging.error(f"Failed to enumerate workflow templates: {exc}")
+            return None
+
+        asset_map: Dict[str, str] = {}
+        try:
+            for entry in template_entries:
+                for asset in entry.assets:
+                    asset_map[asset.filename] = get_asset_path(
+                        entry.template_id, asset.filename
+                    )
+        except Exception as exc:
+            logging.error(f"Failed to resolve template asset paths: {exc}")
+            return None
+
+        if not asset_map:
+            logging.error("No workflow template assets found. Did the packages install correctly?")
+            return None
+
+        return asset_map
+
+
+    @classmethod
+    def legacy_templates_path(cls) -> Optional[str]:
+        """Return the legacy templates directory shipped inside the meta package."""
         try:
             import comfyui_workflow_templates
 
@@ -276,6 +324,7 @@ comfyui-workflow-templates is not installed.
 ********** ERROR ***********
 """.strip()
             )
+            return None
 
     @classmethod
     def embedded_docs_path(cls) -> str:
@@ -392,3 +441,17 @@ comfyui-workflow-templates is not installed.
             logging.info("Falling back to the default frontend.")
             check_frontend_version()
             return cls.default_frontend_path()
+    @classmethod
+    def template_asset_handler(cls):
+        assets = cls.template_asset_map()
+        if not assets:
+            return None
+
+        async def serve_template(request: web.Request) -> web.StreamResponse:
+            rel_path = request.match_info.get("path", "")
+            target = assets.get(rel_path)
+            if target is None:
+                raise web.HTTPNotFound()
+            return web.FileResponse(target)
+
+        return serve_template

app/model_manager.py

@@ -44,7 +44,7 @@ class ModelFileManager:
         @routes.get("/experiment/models/{folder}")
         async def get_all_models(request):
             folder = request.match_info.get("folder", None)
-            if not folder in folder_paths.folder_names_and_paths:
+            if folder not in folder_paths.folder_names_and_paths:
                 return web.Response(status=404)
             files = self.get_model_file_list(folder)
             return web.json_response(files)
@@ -55,7 +55,7 @@ class ModelFileManager:
             path_index = int(request.match_info.get("path_index", None))
             filename = request.match_info.get("filename", None)
 
-            if not folder_name in folder_paths.folder_names_and_paths:
+            if folder_name not in folder_paths.folder_names_and_paths:
                 return web.Response(status=404)
 
             folders = folder_paths.folder_names_and_paths[folder_name]

app/subgraph_manager.py

@@ -10,6 +10,7 @@ import hashlib
 
 class Source:
     custom_node = "custom_node"
+    templates = "templates"
 
 class SubgraphEntry(TypedDict):
     source: str
@@ -38,6 +39,18 @@ class CustomNodeSubgraphEntryInfo(TypedDict):
 class SubgraphManager:
     def __init__(self):
         self.cached_custom_node_subgraphs: dict[SubgraphEntry] | None = None
+        self.cached_blueprint_subgraphs: dict[SubgraphEntry] | None = None
+
+    def _create_entry(self, file: str, source: str, node_pack: str) -> tuple[str, SubgraphEntry]:
+        """Create a subgraph entry from a file path. Expects normalized path (forward slashes)."""
+        entry_id = hashlib.sha256(f"{source}{file}".encode()).hexdigest()
+        entry: SubgraphEntry = {
+            "source": source,
+            "name": os.path.splitext(os.path.basename(file))[0],
+            "path": file,
+            "info": {"node_pack": node_pack},
+        }
+        return entry_id, entry
 
     async def load_entry_data(self, entry: SubgraphEntry):
         with open(entry['path'], 'r') as f:
@@ -60,53 +73,60 @@ class SubgraphManager:
         return entries
 
     async def get_custom_node_subgraphs(self, loadedModules, force_reload=False):
-        # if not forced to reload and cached, return cache
+        """Load subgraphs from custom nodes."""
         if not force_reload and self.cached_custom_node_subgraphs is not None:
             return self.cached_custom_node_subgraphs
-        # Load subgraphs from custom nodes
-        subfolder = "subgraphs"
-        subgraphs_dict: dict[SubgraphEntry] = {}
 
+        subgraphs_dict: dict[SubgraphEntry] = {}
         for folder in folder_paths.get_folder_paths("custom_nodes"):
-            pattern = os.path.join(folder, f"*/{subfolder}/*.json")
-            matched_files = glob.glob(pattern)
-            for file in matched_files:
-                # replace backslashes with forward slashes
+            pattern = os.path.join(folder, "*/subgraphs/*.json")
+            for file in glob.glob(pattern):
                 file = file.replace('\\', '/')
-                info: CustomNodeSubgraphEntryInfo = {
-                    "node_pack": "custom_nodes." + file.split('/')[-3]
-                }
-                source = Source.custom_node
-                # hash source + path to make sure id will be as unique as possible, but
-                # reproducible across backend reloads
-                id = hashlib.sha256(f"{source}{file}".encode()).hexdigest()
-                entry: SubgraphEntry = {
-                    "source": Source.custom_node,
-                    "name": os.path.splitext(os.path.basename(file))[0],
-                    "path": file,
-                    "info": info,
-                }
-                subgraphs_dict[id] = entry
+                node_pack = "custom_nodes." + file.split('/')[-3]
+                entry_id, entry = self._create_entry(file, Source.custom_node, node_pack)
+                subgraphs_dict[entry_id] = entry
+
         self.cached_custom_node_subgraphs = subgraphs_dict
         return subgraphs_dict
 
-    async def get_custom_node_subgraph(self, id: str, loadedModules):
-        subgraphs = await self.get_custom_node_subgraphs(loadedModules)
-        entry: SubgraphEntry = subgraphs.get(id, None)
-        if entry is not None and entry.get('data', None) is None:
+    async def get_blueprint_subgraphs(self, force_reload=False):
+        """Load subgraphs from the blueprints directory."""
+        if not force_reload and self.cached_blueprint_subgraphs is not None:
+            return self.cached_blueprint_subgraphs
+
+        subgraphs_dict: dict[SubgraphEntry] = {}
+        blueprints_dir = os.path.join(os.path.dirname(os.path.dirname(__file__)), 'blueprints')
+
+        if os.path.exists(blueprints_dir):
+            for file in glob.glob(os.path.join(blueprints_dir, "*.json")):
+                file = file.replace('\\', '/')
+                entry_id, entry = self._create_entry(file, Source.templates, "comfyui")
+                subgraphs_dict[entry_id] = entry
+
+        self.cached_blueprint_subgraphs = subgraphs_dict
+        return subgraphs_dict
+
+    async def get_all_subgraphs(self, loadedModules, force_reload=False):
+        """Get all subgraphs from all sources (custom nodes and blueprints)."""
+        custom_node_subgraphs = await self.get_custom_node_subgraphs(loadedModules, force_reload)
+        blueprint_subgraphs = await self.get_blueprint_subgraphs(force_reload)
+        return {**custom_node_subgraphs, **blueprint_subgraphs}
+
+    async def get_subgraph(self, id: str, loadedModules):
+        """Get a specific subgraph by ID from any source."""
+        entry = (await self.get_all_subgraphs(loadedModules)).get(id)
+        if entry is not None and entry.get('data') is None:
             await self.load_entry_data(entry)
         return entry
 
     def add_routes(self, routes, loadedModules):
         @routes.get("/global_subgraphs")
         async def get_global_subgraphs(request):
-            subgraphs_dict = await self.get_custom_node_subgraphs(loadedModules)
-            # NOTE: we may want to include other sources of global subgraphs such as templates in the future;
-            # that's the reasoning for the current implementation
+            subgraphs_dict = await self.get_all_subgraphs(loadedModules)
             return web.json_response(await self.sanitize_entries(subgraphs_dict, remove_data=True))
 
         @routes.get("/global_subgraphs/{id}")
         async def get_global_subgraph(request):
             id = request.match_info.get("id", None)
-            subgraph = await self.get_custom_node_subgraph(id, loadedModules)
+            subgraph = await self.get_subgraph(id, loadedModules)
             return web.json_response(await self.sanitize_entry(subgraph))

app/user_manager.py

@@ -59,6 +59,9 @@ class UserManager():
         user = "default"
         if args.multi_user and "comfy-user" in request.headers:
             user = request.headers["comfy-user"]
+            # Block System Users (use same error message to prevent probing)
+            if user.startswith(folder_paths.SYSTEM_USER_PREFIX):
+                raise KeyError("Unknown user: " + user)
 
         if user not in self.users:
             raise KeyError("Unknown user: " + user)
@@ -66,15 +69,16 @@ class UserManager():
         return user
 
     def get_request_user_filepath(self, request, file, type="userdata", create_dir=True):
-        user_directory = folder_paths.get_user_directory()
-
         if type == "userdata":
-            root_dir = user_directory
+            root_dir = folder_paths.get_user_directory()
         else:
             raise KeyError("Unknown filepath type:" + type)
 
         user = self.get_request_user_id(request)
-        path = user_root = os.path.abspath(os.path.join(root_dir, user))
+        user_root = folder_paths.get_public_user_directory(user)
+        if user_root is None:
+            return None
+        path = user_root
 
         # prevent leaving /{type}
         if os.path.commonpath((root_dir, user_root)) != root_dir:
@@ -101,7 +105,11 @@ class UserManager():
         name = name.strip()
         if not name:
             raise ValueError("username not provided")
+        if name.startswith(folder_paths.SYSTEM_USER_PREFIX):
+            raise ValueError("System User prefix not allowed")
         user_id = re.sub("[^a-zA-Z0-9-_]+", '-', name)
+        if user_id.startswith(folder_paths.SYSTEM_USER_PREFIX):
+            raise ValueError("System User prefix not allowed")
         user_id = user_id + "_" + str(uuid.uuid4())
 
         self.users[user_id] = name
@@ -132,7 +140,10 @@ class UserManager():
             if username in self.users.values():
                 return web.json_response({"error": "Duplicate username."}, status=400)
 
-            user_id = self.add_user(username)
+            try:
+                user_id = self.add_user(username)
+            except ValueError as e:
+                return web.json_response({"error": str(e)}, status=400)
             return web.json_response(user_id)
 
         @routes.get("/userdata")
@@ -424,7 +435,7 @@ class UserManager():
                 return source
 
             dest = get_user_data_path(request, check_exists=False, param="dest")
-            if not isinstance(source, str):
+            if not isinstance(dest, str):
                 return dest
 
             overwrite = request.query.get("overwrite", 'true') != "false"

comfy/audio_encoders/audio_encoders.py

@@ -25,11 +25,11 @@ class AudioEncoderModel():
         elif model_type == "whisper3":
             self.model = WhisperLargeV3(**model_config)
         self.model.eval()
-        self.patcher = comfy.model_patcher.ModelPatcher(self.model, load_device=self.load_device, offload_device=offload_device)
+        self.patcher = comfy.model_patcher.CoreModelPatcher(self.model, load_device=self.load_device, offload_device=offload_device)
         self.model_sample_rate = 16000
 
     def load_sd(self, sd):
-        return self.model.load_state_dict(sd, strict=False)
+        return self.model.load_state_dict(sd, strict=False, assign=self.patcher.is_dynamic())
 
     def get_sd(self):
         return self.model.state_dict()

comfy/cldm/cldm.py

@@ -413,7 +413,8 @@ class ControlNet(nn.Module):
         out_middle = []
 
         if self.num_classes is not None:
-            assert y.shape[0] == x.shape[0]
+            if y is None:
+                raise ValueError("y is None, did you try using a controlnet for SDXL on SD1?")
             emb = emb + self.label_emb(y)
 
         h = x

comfy/cli_args.py

@@ -97,6 +97,13 @@ class LatentPreviewMethod(enum.Enum):
     Latent2RGB = "latent2rgb"
     TAESD = "taesd"
 
+    @classmethod
+    def from_string(cls, value: str):
+        for member in cls:
+            if member.value == value:
+                return member
+        return None
+
 parser.add_argument("--preview-method", type=LatentPreviewMethod, default=LatentPreviewMethod.NoPreviews, help="Default preview method for sampler nodes.", action=EnumAction)
 
 parser.add_argument("--preview-size", type=int, default=512, help="Sets the maximum preview size for sampler nodes.")
@@ -121,6 +128,12 @@ upcast.add_argument("--force-upcast-attention", action="store_true", help="Force
 upcast.add_argument("--dont-upcast-attention", action="store_true", help="Disable all upcasting of attention. Should be unnecessary except for debugging.")
 
 
+parser.add_argument("--enable-manager", action="store_true", help="Enable the ComfyUI-Manager feature.")
+manager_group = parser.add_mutually_exclusive_group()
+manager_group.add_argument("--disable-manager-ui", action="store_true", help="Disables only the ComfyUI-Manager UI and endpoints. Scheduled installations and similar background tasks will still operate.")
+manager_group.add_argument("--enable-manager-legacy-ui", action="store_true", help="Enables the legacy UI of ComfyUI-Manager")
+
+
 vram_group = parser.add_mutually_exclusive_group()
 vram_group.add_argument("--gpu-only", action="store_true", help="Store and run everything (text encoders/CLIP models, etc... on the GPU).")
 vram_group.add_argument("--highvram", action="store_true", help="By default models will be unloaded to CPU memory after being used. This option keeps them in GPU memory.")
@@ -131,7 +144,8 @@ vram_group.add_argument("--cpu", action="store_true", help="To use the CPU for e
 
 parser.add_argument("--reserve-vram", type=float, default=None, help="Set the amount of vram in GB you want to reserve for use by your OS/other software. By default some amount is reserved depending on your OS.")
 
-parser.add_argument("--async-offload", action="store_true", help="Use async weight offloading.")
+parser.add_argument("--async-offload", nargs='?', const=2, type=int, default=None, metavar="NUM_STREAMS", help="Use async weight offloading. An optional argument controls the amount of offload streams. Default is 2. Enabled by default on Nvidia.")
+parser.add_argument("--disable-async-offload", action="store_true", help="Disable async weight offloading.")
 
 parser.add_argument("--force-non-blocking", action="store_true", help="Force ComfyUI to use non-blocking operations for all applicable tensors. This may improve performance on some non-Nvidia systems but can cause issues with some workflows.")
 
@@ -145,10 +159,12 @@ class PerformanceFeature(enum.Enum):
     Fp8MatrixMultiplication = "fp8_matrix_mult"
     CublasOps = "cublas_ops"
     AutoTune = "autotune"
-    PinnedMem = "pinned_memory"
+    DynamicVRAM = "dynamic_vram"
 
 parser.add_argument("--fast", nargs="*", type=PerformanceFeature, help="Enable some untested and potentially quality deteriorating optimizations. This is used to test new features so using it might crash your comfyui. --fast with no arguments enables everything. You can pass a list specific optimizations if you only want to enable specific ones. Current valid optimizations: {}".format(" ".join(map(lambda c: c.value, PerformanceFeature))))
 
+parser.add_argument("--disable-pinned-memory", action="store_true", help="Disable pinned memory use.")
+
 parser.add_argument("--mmap-torch-files", action="store_true", help="Use mmap when loading ckpt/pt files.")
 parser.add_argument("--disable-mmap", action="store_true", help="Don't use mmap when loading safetensors.")
 
@@ -159,13 +175,14 @@ parser.add_argument("--windows-standalone-build", action="store_true", help="Win
 parser.add_argument("--disable-metadata", action="store_true", help="Disable saving prompt metadata in files.")
 parser.add_argument("--disable-all-custom-nodes", action="store_true", help="Disable loading all custom nodes.")
 parser.add_argument("--whitelist-custom-nodes", type=str, nargs='+', default=[], help="Specify custom node folders to load even when --disable-all-custom-nodes is enabled.")
-parser.add_argument("--disable-api-nodes", action="store_true", help="Disable loading all api nodes.")
+parser.add_argument("--disable-api-nodes", action="store_true", help="Disable loading all api nodes. Also prevents the frontend from communicating with the internet.")
 
 parser.add_argument("--multi-user", action="store_true", help="Enables per-user storage.")
 
 parser.add_argument("--verbose", default='INFO', const='DEBUG', nargs="?", choices=['DEBUG', 'INFO', 'WARNING', 'ERROR', 'CRITICAL'], help='Set the logging level')
 parser.add_argument("--log-stdout", action="store_true", help="Send normal process output to stdout instead of stderr (default).")
 
+
 # The default built-in provider hosted under web/
 DEFAULT_VERSION_STRING = "comfyanonymous/ComfyUI@latest"
 
@@ -215,6 +232,7 @@ database_default_path = os.path.abspath(
     os.path.join(os.path.dirname(__file__), "..", "user", "comfyui.db")
 )
 parser.add_argument("--database-url", type=str, default=f"sqlite:///{database_default_path}", help="Specify the database URL, e.g. for an in-memory database you can use 'sqlite:///:memory:'.")
+parser.add_argument("--disable-assets-autoscan", action="store_true", help="Disable asset scanning on startup for database synchronization.")
 
 if comfy.options.args_parsing:
     args = parser.parse_args()
@@ -240,3 +258,6 @@ elif args.fast == []:
 # '--fast' is provided with a list of performance features, use that list
 else:
     args.fast = set(args.fast)
+
+def enables_dynamic_vram():
+    return PerformanceFeature.DynamicVRAM in args.fast and not args.highvram and not args.gpu_only

comfy/clip_model.py

@@ -1,6 +1,59 @@
 import torch
 from comfy.ldm.modules.attention import optimized_attention_for_device
 import comfy.ops
+import math
+
+def clip_preprocess(image, size=224, mean=[0.48145466, 0.4578275, 0.40821073], std=[0.26862954, 0.26130258, 0.27577711], crop=True):
+    image = image[:, :, :, :3] if image.shape[3] > 3 else image
+    mean = torch.tensor(mean, device=image.device, dtype=image.dtype)
+    std = torch.tensor(std, device=image.device, dtype=image.dtype)
+    image = image.movedim(-1, 1)
+    if not (image.shape[2] == size and image.shape[3] == size):
+        if crop:
+            scale = (size / min(image.shape[2], image.shape[3]))
+            scale_size = (round(scale * image.shape[2]), round(scale * image.shape[3]))
+        else:
+            scale_size = (size, size)
+
+        image = torch.nn.functional.interpolate(image, size=scale_size, mode="bicubic", antialias=True)
+        h = (image.shape[2] - size)//2
+        w = (image.shape[3] - size)//2
+        image = image[:,:,h:h+size,w:w+size]
+    image = torch.clip((255. * image), 0, 255).round() / 255.0
+    return (image - mean.view([3,1,1])) / std.view([3,1,1])
+
+def siglip2_flex_calc_resolution(oh, ow, patch_size, max_num_patches, eps=1e-5):
+    def scale_dim(size, scale):
+        scaled = math.ceil(size * scale / patch_size) * patch_size
+        return max(patch_size, int(scaled))
+
+    # Binary search for optimal scale
+    lo, hi = eps / 10, 100.0
+    while hi - lo >= eps:
+        mid = (lo + hi) / 2
+        h, w = scale_dim(oh, mid), scale_dim(ow, mid)
+        if (h // patch_size) * (w // patch_size) <= max_num_patches:
+            lo = mid
+        else:
+            hi = mid
+
+    return scale_dim(oh, lo), scale_dim(ow, lo)
+
+def siglip2_preprocess(image, size, patch_size, num_patches, mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5], crop=True):
+    if size > 0:
+        return clip_preprocess(image, size=size, mean=mean, std=std, crop=crop)
+
+    image = image[:, :, :, :3] if image.shape[3] > 3 else image
+    mean = torch.tensor(mean, device=image.device, dtype=image.dtype)
+    std = torch.tensor(std, device=image.device, dtype=image.dtype)
+    image = image.movedim(-1, 1)
+
+    b, c, h, w = image.shape
+    h, w = siglip2_flex_calc_resolution(h, w, patch_size, num_patches)
+
+    image = torch.nn.functional.interpolate(image, size=(h, w), mode="bilinear", antialias=True)
+    image = torch.clip((255. * image), 0, 255).round() / 255.0
+    return (image - mean.view([3, 1, 1])) / std.view([3, 1, 1])
 
 class CLIPAttention(torch.nn.Module):
     def __init__(self, embed_dim, heads, dtype, device, operations):
@@ -156,6 +209,27 @@ class CLIPTextModel(torch.nn.Module):
         out = self.text_projection(x[2])
         return (x[0], x[1], out, x[2])
 
+def siglip2_pos_embed(embed_weight, embeds, orig_shape):
+    embed_weight_len = round(embed_weight.shape[0] ** 0.5)
+    embed_weight = comfy.ops.cast_to_input(embed_weight, embeds).movedim(1, 0).reshape(1, -1, embed_weight_len, embed_weight_len)
+    embed_weight = torch.nn.functional.interpolate(embed_weight, size=orig_shape, mode="bilinear", align_corners=False, antialias=True)
+    embed_weight = embed_weight.reshape(-1, embed_weight.shape[-2] * embed_weight.shape[-1]).movedim(0, 1)
+    return embeds + embed_weight
+
+class Siglip2Embeddings(torch.nn.Module):
+    def __init__(self, embed_dim, num_channels=3, patch_size=14, image_size=224, model_type="", num_patches=None, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.patch_embedding = operations.Linear(num_channels * patch_size * patch_size, embed_dim, dtype=dtype, device=device)
+        self.position_embedding = operations.Embedding(num_patches, embed_dim, dtype=dtype, device=device)
+        self.patch_size = patch_size
+
+    def forward(self, pixel_values):
+        b, c, h, w = pixel_values.shape
+        img = pixel_values.movedim(1, -1).reshape(b, h // self.patch_size, self.patch_size, w // self.patch_size, self.patch_size, c)
+        img = img.permute(0, 1, 3, 2, 4, 5)
+        img = img.reshape(b, img.shape[1] * img.shape[2], -1)
+        img = self.patch_embedding(img)
+        return siglip2_pos_embed(self.position_embedding.weight, img, (h // self.patch_size, w // self.patch_size))
 
 class CLIPVisionEmbeddings(torch.nn.Module):
     def __init__(self, embed_dim, num_channels=3, patch_size=14, image_size=224, model_type="", dtype=None, device=None, operations=None):
@@ -199,8 +273,11 @@ class CLIPVision(torch.nn.Module):
         intermediate_activation = config_dict["hidden_act"]
         model_type = config_dict["model_type"]
 
-        self.embeddings = CLIPVisionEmbeddings(embed_dim, config_dict["num_channels"], config_dict["patch_size"], config_dict["image_size"], model_type=model_type, dtype=dtype, device=device, operations=operations)
-        if model_type == "siglip_vision_model":
+        if model_type in ["siglip2_vision_model"]:
+            self.embeddings = Siglip2Embeddings(embed_dim, config_dict["num_channels"], config_dict["patch_size"], config_dict["image_size"], model_type=model_type, num_patches=config_dict.get("num_patches", None), dtype=dtype, device=device, operations=operations)
+        else:
+            self.embeddings = CLIPVisionEmbeddings(embed_dim, config_dict["num_channels"], config_dict["patch_size"], config_dict["image_size"], model_type=model_type, dtype=dtype, device=device, operations=operations)
+        if model_type in ["siglip_vision_model", "siglip2_vision_model"]:
             self.pre_layrnorm = lambda a: a
             self.output_layernorm = True
         else:

comfy/clip_vision.py

@@ -1,6 +1,5 @@
 from .utils import load_torch_file, transformers_convert, state_dict_prefix_replace
 import os
-import torch
 import json
 import logging
 
@@ -17,28 +16,12 @@ class Output:
     def __setitem__(self, key, item):
         setattr(self, key, item)
 
-def clip_preprocess(image, size=224, mean=[0.48145466, 0.4578275, 0.40821073], std=[0.26862954, 0.26130258, 0.27577711], crop=True):
-    image = image[:, :, :, :3] if image.shape[3] > 3 else image
-    mean = torch.tensor(mean, device=image.device, dtype=image.dtype)
-    std = torch.tensor(std, device=image.device, dtype=image.dtype)
-    image = image.movedim(-1, 1)
-    if not (image.shape[2] == size and image.shape[3] == size):
-        if crop:
-            scale = (size / min(image.shape[2], image.shape[3]))
-            scale_size = (round(scale * image.shape[2]), round(scale * image.shape[3]))
-        else:
-            scale_size = (size, size)
-
-        image = torch.nn.functional.interpolate(image, size=scale_size, mode="bicubic", antialias=True)
-        h = (image.shape[2] - size)//2
-        w = (image.shape[3] - size)//2
-        image = image[:,:,h:h+size,w:w+size]
-    image = torch.clip((255. * image), 0, 255).round() / 255.0
-    return (image - mean.view([3,1,1])) / std.view([3,1,1])
+clip_preprocess = comfy.clip_model.clip_preprocess  # Prevent some stuff from breaking, TODO: remove eventually
 
 IMAGE_ENCODERS = {
     "clip_vision_model": comfy.clip_model.CLIPVisionModelProjection,
     "siglip_vision_model": comfy.clip_model.CLIPVisionModelProjection,
+    "siglip2_vision_model": comfy.clip_model.CLIPVisionModelProjection,
     "dinov2": comfy.image_encoders.dino2.Dinov2Model,
 }
 
@@ -50,9 +33,10 @@ class ClipVisionModel():
         self.image_size = config.get("image_size", 224)
         self.image_mean = config.get("image_mean", [0.48145466, 0.4578275, 0.40821073])
         self.image_std = config.get("image_std", [0.26862954, 0.26130258, 0.27577711])
-        model_type = config.get("model_type", "clip_vision_model")
-        model_class = IMAGE_ENCODERS.get(model_type)
-        if model_type == "siglip_vision_model":
+        self.model_type = config.get("model_type", "clip_vision_model")
+        self.config = config.copy()
+        model_class = IMAGE_ENCODERS.get(self.model_type)
+        if self.model_type == "siglip_vision_model":
             self.return_all_hidden_states = True
         else:
             self.return_all_hidden_states = False
@@ -63,22 +47,26 @@ class ClipVisionModel():
         self.model = model_class(config, self.dtype, offload_device, comfy.ops.manual_cast)
         self.model.eval()
 
-        self.patcher = comfy.model_patcher.ModelPatcher(self.model, load_device=self.load_device, offload_device=offload_device)
+        self.patcher = comfy.model_patcher.CoreModelPatcher(self.model, load_device=self.load_device, offload_device=offload_device)
 
     def load_sd(self, sd):
-        return self.model.load_state_dict(sd, strict=False)
+        return self.model.load_state_dict(sd, strict=False, assign=self.patcher.is_dynamic())
 
     def get_sd(self):
         return self.model.state_dict()
 
     def encode_image(self, image, crop=True):
         comfy.model_management.load_model_gpu(self.patcher)
-        pixel_values = clip_preprocess(image.to(self.load_device), size=self.image_size, mean=self.image_mean, std=self.image_std, crop=crop).float()
+        if self.model_type == "siglip2_vision_model":
+            pixel_values = comfy.clip_model.siglip2_preprocess(image.to(self.load_device), size=self.image_size, patch_size=self.config.get("patch_size", 16), num_patches=self.config.get("num_patches", 256), mean=self.image_mean, std=self.image_std, crop=crop).float()
+        else:
+            pixel_values = comfy.clip_model.clip_preprocess(image.to(self.load_device), size=self.image_size, mean=self.image_mean, std=self.image_std, crop=crop).float()
         out = self.model(pixel_values=pixel_values, intermediate_output='all' if self.return_all_hidden_states else -2)
 
         outputs = Output()
         outputs["last_hidden_state"] = out[0].to(comfy.model_management.intermediate_device())
         outputs["image_embeds"] = out[2].to(comfy.model_management.intermediate_device())
+        outputs["image_sizes"] = [pixel_values.shape[1:]] * pixel_values.shape[0]
         if self.return_all_hidden_states:
             all_hs = out[1].to(comfy.model_management.intermediate_device())
             outputs["penultimate_hidden_states"] = all_hs[:, -2]
@@ -125,10 +113,14 @@ def load_clipvision_from_sd(sd, prefix="", convert_keys=False):
     elif "vision_model.encoder.layers.22.layer_norm1.weight" in sd:
         embed_shape = sd["vision_model.embeddings.position_embedding.weight"].shape[0]
         if sd["vision_model.encoder.layers.0.layer_norm1.weight"].shape[0] == 1152:
-            if embed_shape == 729:
-                json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "clip_vision_siglip_384.json")
-            elif embed_shape == 1024:
-                json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "clip_vision_siglip_512.json")
+            patch_embedding_shape = sd["vision_model.embeddings.patch_embedding.weight"].shape
+            if len(patch_embedding_shape) == 2:
+                json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "clip_vision_siglip2_base_naflex.json")
+            else:
+                if embed_shape == 729:
+                    json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "clip_vision_siglip_384.json")
+                elif embed_shape == 1024:
+                    json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "clip_vision_siglip_512.json")
         elif embed_shape == 577:
             if "multi_modal_projector.linear_1.bias" in sd:
                 json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "clip_vision_config_vitl_336_llava.json")

comfy/clip_vision_siglip2_base_naflex.json

@@ -0,0 +1,14 @@
+{
+  "num_channels": 3,
+  "hidden_act": "gelu_pytorch_tanh",
+  "hidden_size": 1152,
+  "image_size": -1,
+  "intermediate_size": 4304,
+  "model_type": "siglip2_vision_model",
+  "num_attention_heads": 16,
+  "num_hidden_layers": 27,
+  "patch_size": 16,
+  "num_patches": 256,
+  "image_mean": [0.5, 0.5, 0.5],
+  "image_std": [0.5, 0.5, 0.5]
+}

comfy/comfy_types/node_typing.py

@@ -236,6 +236,8 @@ class ComfyNodeABC(ABC):
     """Flags a node as experimental, informing users that it may change or not work as expected."""
     DEPRECATED: bool
     """Flags a node as deprecated, indicating to users that they should find alternatives to this node."""
+    DEV_ONLY: bool
+    """Flags a node as dev-only, hiding it from search/menus unless dev mode is enabled."""
     API_NODE: Optional[bool]
     """Flags a node as an API node. See: https://docs.comfy.org/tutorials/api-nodes/overview."""
 

comfy/context_windows.py

@@ -51,32 +51,43 @@ class ContextHandlerABC(ABC):
 
 
 class IndexListContextWindow(ContextWindowABC):
-    def __init__(self, index_list: list[int], dim: int=0):
+    def __init__(self, index_list: list[int], dim: int=0, total_frames: int=0):
         self.index_list = index_list
         self.context_length = len(index_list)
         self.dim = dim
+        self.total_frames = total_frames
+        self.center_ratio = (min(index_list) + max(index_list)) / (2 * total_frames)
 
-    def get_tensor(self, full: torch.Tensor, device=None, dim=None) -> torch.Tensor:
+    def get_tensor(self, full: torch.Tensor, device=None, dim=None, retain_index_list=[]) -> torch.Tensor:
         if dim is None:
             dim = self.dim
         if dim == 0 and full.shape[dim] == 1:
             return full
-        idx = [slice(None)] * dim + [self.index_list]
-        return full[idx].to(device)
+        idx = tuple([slice(None)] * dim + [self.index_list])
+        window = full[idx]
+        if retain_index_list:
+            idx = tuple([slice(None)] * dim + [retain_index_list])
+            window[idx] = full[idx]
+        return window.to(device)
 
     def add_window(self, full: torch.Tensor, to_add: torch.Tensor, dim=None) -> torch.Tensor:
         if dim is None:
             dim = self.dim
-        idx = [slice(None)] * dim + [self.index_list]
+        idx = tuple([slice(None)] * dim + [self.index_list])
         full[idx] += to_add
         return full
 
+    def get_region_index(self, num_regions: int) -> int:
+        region_idx = int(self.center_ratio * num_regions)
+        return min(max(region_idx, 0), num_regions - 1)
+
 
 class IndexListCallbacks:
     EVALUATE_CONTEXT_WINDOWS = "evaluate_context_windows"
     COMBINE_CONTEXT_WINDOW_RESULTS = "combine_context_window_results"
     EXECUTE_START = "execute_start"
     EXECUTE_CLEANUP = "execute_cleanup"
+    RESIZE_COND_ITEM = "resize_cond_item"
 
     def init_callbacks(self):
         return {}
@@ -94,7 +105,8 @@ class ContextFuseMethod:
 
 ContextResults = collections.namedtuple("ContextResults", ['window_idx', 'sub_conds_out', 'sub_conds', 'window'])
 class IndexListContextHandler(ContextHandlerABC):
-    def __init__(self, context_schedule: ContextSchedule, fuse_method: ContextFuseMethod, context_length: int=1, context_overlap: int=0, context_stride: int=1, closed_loop=False, dim=0):
+    def __init__(self, context_schedule: ContextSchedule, fuse_method: ContextFuseMethod, context_length: int=1, context_overlap: int=0, context_stride: int=1,
+                 closed_loop: bool=False, dim:int=0, freenoise: bool=False, cond_retain_index_list: list[int]=[], split_conds_to_windows: bool=False):
         self.context_schedule = context_schedule
         self.fuse_method = fuse_method
         self.context_length = context_length
@@ -103,13 +115,18 @@ class IndexListContextHandler(ContextHandlerABC):
         self.closed_loop = closed_loop
         self.dim = dim
         self._step = 0
+        self.freenoise = freenoise
+        self.cond_retain_index_list = [int(x.strip()) for x in cond_retain_index_list.split(",")] if cond_retain_index_list else []
+        self.split_conds_to_windows = split_conds_to_windows
 
         self.callbacks = {}
 
     def should_use_context(self, model: BaseModel, conds: list[list[dict]], x_in: torch.Tensor, timestep: torch.Tensor, model_options: dict[str]) -> bool:
         # for now, assume first dim is batch - should have stored on BaseModel in actual implementation
         if x_in.size(self.dim) > self.context_length:
-            logging.info(f"Using context windows {self.context_length} for {x_in.size(self.dim)} frames.")
+            logging.info(f"Using context windows {self.context_length} with overlap {self.context_overlap} for {x_in.size(self.dim)} frames.")
+            if self.cond_retain_index_list:
+                logging.info(f"Retaining original cond for indexes: {self.cond_retain_index_list}")
             return True
         return False
 
@@ -123,6 +140,11 @@ class IndexListContextHandler(ContextHandlerABC):
             return None
         # reuse or resize cond items to match context requirements
         resized_cond = []
+        # if multiple conds, split based on primary region
+        if self.split_conds_to_windows and len(cond_in) > 1:
+            region = window.get_region_index(len(cond_in))
+            logging.info(f"Splitting conds to windows; using region {region} for window {window.index_list[0]}-{window.index_list[-1]} with center ratio {window.center_ratio:.3f}")
+            cond_in = [cond_in[region]]
         # cond object is a list containing a dict - outer list is irrelevant, so just loop through it
         for actual_cond in cond_in:
             resized_actual_cond = actual_cond.copy()
@@ -145,13 +167,38 @@ class IndexListContextHandler(ContextHandlerABC):
                         new_cond_item = cond_item.copy()
                         # when in dictionary, look for tensors and CONDCrossAttn [comfy/conds.py] (has cond attr that is a tensor)
                         for cond_key, cond_value in new_cond_item.items():
+                            # Allow callbacks to handle custom conditioning items
+                            handled = False
+                            for callback in comfy.patcher_extension.get_all_callbacks(
+                                IndexListCallbacks.RESIZE_COND_ITEM, self.callbacks
+                            ):
+                                result = callback(cond_key, cond_value, window, x_in, device, new_cond_item)
+                                if result is not None:
+                                    new_cond_item[cond_key] = result
+                                    handled = True
+                                    break
+                            if handled:
+                                continue
                             if isinstance(cond_value, torch.Tensor):
-                                if cond_value.ndim < self.dim and cond_value.size(0) == x_in.size(self.dim):
+                                if (self.dim < cond_value.ndim and cond_value(self.dim) == x_in.size(self.dim)) or \
+                                   (cond_value.ndim < self.dim and cond_value.size(0) == x_in.size(self.dim)):
                                     new_cond_item[cond_key] = window.get_tensor(cond_value, device)
+                            # Handle audio_embed (temporal dim is 1)
+                            elif cond_key == "audio_embed" and hasattr(cond_value, "cond") and isinstance(cond_value.cond, torch.Tensor):
+                                audio_cond = cond_value.cond
+                                if audio_cond.ndim > 1 and audio_cond.size(1) == x_in.size(self.dim):
+                                    new_cond_item[cond_key] = cond_value._copy_with(window.get_tensor(audio_cond, device, dim=1))
+                            # Handle vace_context (temporal dim is 3)
+                            elif cond_key == "vace_context" and hasattr(cond_value, "cond") and isinstance(cond_value.cond, torch.Tensor):
+                                vace_cond = cond_value.cond
+                                if vace_cond.ndim >= 4 and vace_cond.size(3) == x_in.size(self.dim):
+                                    sliced_vace = window.get_tensor(vace_cond, device, dim=3, retain_index_list=self.cond_retain_index_list)
+                                    new_cond_item[cond_key] = cond_value._copy_with(sliced_vace)
                             # if has cond that is a Tensor, check if needs to be subset
                             elif hasattr(cond_value, "cond") and isinstance(cond_value.cond, torch.Tensor):
-                                if cond_value.cond.ndim < self.dim and cond_value.cond.size(0) == x_in.size(self.dim):
-                                    new_cond_item[cond_key] = cond_value._copy_with(window.get_tensor(cond_value.cond, device))
+                                if  (self.dim < cond_value.cond.ndim and cond_value.cond.size(self.dim) == x_in.size(self.dim)) or \
+                                    (cond_value.cond.ndim < self.dim and cond_value.cond.size(0) == x_in.size(self.dim)):
+                                    new_cond_item[cond_key] = cond_value._copy_with(window.get_tensor(cond_value.cond, device, retain_index_list=self.cond_retain_index_list))
                             elif cond_key == "num_video_frames": # for SVD
                                 new_cond_item[cond_key] = cond_value._copy_with(cond_value.cond)
                                 new_cond_item[cond_key].cond = window.context_length
@@ -164,7 +211,7 @@ class IndexListContextHandler(ContextHandlerABC):
         return resized_cond
 
     def set_step(self, timestep: torch.Tensor, model_options: dict[str]):
-        mask = torch.isclose(model_options["transformer_options"]["sample_sigmas"], timestep, rtol=0.0001)
+        mask = torch.isclose(model_options["transformer_options"]["sample_sigmas"], timestep[0], rtol=0.0001)
         matches = torch.nonzero(mask)
         if torch.numel(matches) == 0:
             raise Exception("No sample_sigmas matched current timestep; something went wrong.")
@@ -173,7 +220,7 @@ class IndexListContextHandler(ContextHandlerABC):
     def get_context_windows(self, model: BaseModel, x_in: torch.Tensor, model_options: dict[str]) -> list[IndexListContextWindow]:
         full_length = x_in.size(self.dim) # TODO: choose dim based on model
         context_windows = self.context_schedule.func(full_length, self, model_options)
-        context_windows = [IndexListContextWindow(window, dim=self.dim) for window in context_windows]
+        context_windows = [IndexListContextWindow(window, dim=self.dim, total_frames=full_length) for window in context_windows]
         return context_windows
 
     def execute(self, calc_cond_batch: Callable, model: BaseModel, conds: list[list[dict]], x_in: torch.Tensor, timestep: torch.Tensor, model_options: dict[str]):
@@ -250,8 +297,8 @@ class IndexListContextHandler(ContextHandlerABC):
                     prev_weight = (bias_total / (bias_total + bias))
                     new_weight = (bias / (bias_total + bias))
                     # account for dims of tensors
-                    idx_window = [slice(None)] * self.dim + [idx]
-                    pos_window = [slice(None)] * self.dim + [pos]
+                    idx_window = tuple([slice(None)] * self.dim + [idx])
+                    pos_window = tuple([slice(None)] * self.dim + [pos])
                     # apply new values
                     conds_final[i][idx_window] = conds_final[i][idx_window] * prev_weight + sub_conds_out[i][pos_window] * new_weight
                     biases_final[i][idx] = bias_total + bias
@@ -287,6 +334,28 @@ def create_prepare_sampling_wrapper(model: ModelPatcher):
     )
 
 
+def _sampler_sample_wrapper(executor, guider, sigmas, extra_args, callback, noise, *args, **kwargs):
+    model_options = extra_args.get("model_options", None)
+    if model_options is None:
+        raise Exception("model_options not found in sampler_sample_wrapper; this should never happen, something went wrong.")
+    handler: IndexListContextHandler = model_options.get("context_handler", None)
+    if handler is None:
+        raise Exception("context_handler not found in sampler_sample_wrapper; this should never happen, something went wrong.")
+    if not handler.freenoise:
+        return executor(guider, sigmas, extra_args, callback, noise, *args, **kwargs)
+    noise = apply_freenoise(noise, handler.dim, handler.context_length, handler.context_overlap, extra_args["seed"])
+
+    return executor(guider, sigmas, extra_args, callback, noise, *args, **kwargs)
+
+
+def create_sampler_sample_wrapper(model: ModelPatcher):
+    model.add_wrapper_with_key(
+        comfy.patcher_extension.WrappersMP.SAMPLER_SAMPLE,
+        "ContextWindows_sampler_sample",
+        _sampler_sample_wrapper
+    )
+
+
 def match_weights_to_dim(weights: list[float], x_in: torch.Tensor, dim: int, device=None) -> torch.Tensor:
     total_dims = len(x_in.shape)
     weights_tensor = torch.Tensor(weights).to(device=device)
@@ -538,3 +607,29 @@ def shift_window_to_end(window: list[int], num_frames: int):
     for i in range(len(window)):
         # 2) add end_delta to each val to slide windows to end
         window[i] = window[i] + end_delta
+
+
+# https://github.com/Kosinkadink/ComfyUI-AnimateDiff-Evolved/blob/90fb1331201a4b29488089e4fbffc0d82cc6d0a9/animatediff/sample_settings.py#L465
+def apply_freenoise(noise: torch.Tensor, dim: int, context_length: int, context_overlap: int, seed: int):
+    logging.info("Context windows: Applying FreeNoise")
+    generator = torch.Generator(device='cpu').manual_seed(seed)
+    latent_video_length = noise.shape[dim]
+    delta = context_length - context_overlap
+
+    for start_idx in range(0, latent_video_length - context_length, delta):
+        place_idx = start_idx + context_length
+
+        actual_delta = min(delta, latent_video_length - place_idx)
+        if actual_delta <= 0:
+            break
+
+        list_idx = torch.randperm(actual_delta, generator=generator, device='cpu') + start_idx
+
+        source_slice = [slice(None)] * noise.ndim
+        source_slice[dim] = list_idx
+        target_slice = [slice(None)] * noise.ndim
+        target_slice[dim] = slice(place_idx, place_idx + actual_delta)
+
+        noise[tuple(target_slice)] = noise[tuple(source_slice)]
+
+    return noise

comfy/controlnet.py

@@ -203,7 +203,7 @@ class ControlNet(ControlBase):
         self.control_model = control_model
         self.load_device = load_device
         if control_model is not None:
-            self.control_model_wrapped = comfy.model_patcher.ModelPatcher(self.control_model, load_device=load_device, offload_device=comfy.model_management.unet_offload_device())
+            self.control_model_wrapped = comfy.model_patcher.CoreModelPatcher(self.control_model, load_device=load_device, offload_device=comfy.model_management.unet_offload_device())
 
         self.compression_ratio = compression_ratio
         self.global_average_pooling = global_average_pooling

comfy/float.py

@@ -65,3 +65,147 @@ def stochastic_rounding(value, dtype, seed=0):
         return output
 
     return value.to(dtype=dtype)
+
+
+# TODO: improve this?
+def stochastic_float_to_fp4_e2m1(x, generator):
+    orig_shape = x.shape
+    sign = torch.signbit(x).to(torch.uint8)
+
+    exp = torch.floor(torch.log2(x.abs()) + 1.0).clamp(0, 3)
+    x += (torch.rand(x.size(), dtype=x.dtype, layout=x.layout, device=x.device, generator=generator) - 0.5) * (2 ** (exp - 2.0)) * 1.25
+
+    x = x.abs()
+    exp = torch.floor(torch.log2(x) + 1.1925).clamp(0, 3)
+
+    mantissa = torch.where(
+        exp > 0,
+        (x / (2.0 ** (exp - 1)) - 1.0) * 2.0,
+        (x * 2.0),
+        out=x
+    ).round().to(torch.uint8)
+    del x
+
+    exp = exp.to(torch.uint8)
+
+    fp4 = (sign << 3) | (exp << 1) | mantissa
+    del sign, exp, mantissa
+
+    fp4_flat = fp4.view(-1)
+    packed = (fp4_flat[0::2] << 4) | fp4_flat[1::2]
+    return packed.reshape(list(orig_shape)[:-1] + [-1])
+
+
+def to_blocked(input_matrix, flatten: bool = True) -> torch.Tensor:
+    """
+    Rearrange a large matrix by breaking it into blocks and applying the rearrangement pattern.
+    See:
+        https://docs.nvidia.com/cuda/cublas/index.html#d-block-scaling-factors-layout
+
+    Args:
+        input_matrix: Input tensor of shape (H, W)
+    Returns:
+        Rearranged tensor of shape (32*ceil_div(H,128), 16*ceil_div(W,4))
+    """
+
+    def ceil_div(a, b):
+        return (a + b - 1) // b
+
+    rows, cols = input_matrix.shape
+    n_row_blocks = ceil_div(rows, 128)
+    n_col_blocks = ceil_div(cols, 4)
+
+    # Calculate the padded shape
+    padded_rows = n_row_blocks * 128
+    padded_cols = n_col_blocks * 4
+
+    padded = input_matrix
+    if (rows, cols) != (padded_rows, padded_cols):
+        padded = torch.zeros(
+            (padded_rows, padded_cols),
+            device=input_matrix.device,
+            dtype=input_matrix.dtype,
+        )
+        padded[:rows, :cols] = input_matrix
+
+    # Rearrange the blocks
+    blocks = padded.view(n_row_blocks, 128, n_col_blocks, 4).permute(0, 2, 1, 3)
+    rearranged = blocks.reshape(-1, 4, 32, 4).transpose(1, 2).reshape(-1, 32, 16)
+    if flatten:
+        return rearranged.flatten()
+
+    return rearranged.reshape(padded_rows, padded_cols)
+
+
+def stochastic_round_quantize_nvfp4_block(x, per_tensor_scale, generator):
+    F4_E2M1_MAX = 6.0
+    F8_E4M3_MAX = 448.0
+
+    orig_shape = x.shape
+
+    block_size = 16
+
+    x = x.reshape(orig_shape[0], -1, block_size)
+    scaled_block_scales_fp8 = torch.clamp(((torch.amax(torch.abs(x), dim=-1)) / F4_E2M1_MAX) / per_tensor_scale.to(x.dtype), max=F8_E4M3_MAX).to(torch.float8_e4m3fn)
+    x = x / (per_tensor_scale.to(x.dtype) * scaled_block_scales_fp8.to(x.dtype)).unsqueeze(-1)
+
+    x = x.view(orig_shape).nan_to_num()
+    data_lp = stochastic_float_to_fp4_e2m1(x, generator=generator)
+    return data_lp, scaled_block_scales_fp8
+
+
+def stochastic_round_quantize_nvfp4(x, per_tensor_scale, pad_16x, seed=0):
+    def roundup(x: int, multiple: int) -> int:
+        """Round up x to the nearest multiple."""
+        return ((x + multiple - 1) // multiple) * multiple
+
+    generator = torch.Generator(device=x.device)
+    generator.manual_seed(seed)
+
+    # Handle padding
+    if pad_16x:
+        rows, cols = x.shape
+        padded_rows = roundup(rows, 16)
+        padded_cols = roundup(cols, 16)
+        if padded_rows != rows or padded_cols != cols:
+            x = torch.nn.functional.pad(x, (0, padded_cols - cols, 0, padded_rows - rows))
+
+    x, blocked_scaled = stochastic_round_quantize_nvfp4_block(x, per_tensor_scale, generator)
+    return x, to_blocked(blocked_scaled, flatten=False)
+
+
+def stochastic_round_quantize_nvfp4_by_block(x, per_tensor_scale, pad_16x, seed=0, block_size=4096 * 4096):
+    def roundup(x: int, multiple: int) -> int:
+        """Round up x to the nearest multiple."""
+        return ((x + multiple - 1) // multiple) * multiple
+
+    orig_shape = x.shape
+
+    # Handle padding
+    if pad_16x:
+        rows, cols = x.shape
+        padded_rows = roundup(rows, 16)
+        padded_cols = roundup(cols, 16)
+        if padded_rows != rows or padded_cols != cols:
+            x = torch.nn.functional.pad(x, (0, padded_cols - cols, 0, padded_rows - rows))
+            # Note: We update orig_shape because the output tensor logic below assumes x.shape matches
+            # what we want to produce. If we pad here, we want the padded output.
+            orig_shape = x.shape
+
+    orig_shape = list(orig_shape)
+
+    output_fp4 = torch.empty(orig_shape[:-1] + [orig_shape[-1] // 2], dtype=torch.uint8, device=x.device)
+    output_block = torch.empty(orig_shape[:-1] + [orig_shape[-1] // 16], dtype=torch.float8_e4m3fn, device=x.device)
+
+    generator = torch.Generator(device=x.device)
+    generator.manual_seed(seed)
+
+    num_slices = max(1, (x.numel() / block_size))
+    slice_size = max(1, (round(x.shape[0] / num_slices)))
+
+    for i in range(0, x.shape[0], slice_size):
+        fp4, block = stochastic_round_quantize_nvfp4_block(x[i: i + slice_size], per_tensor_scale, generator=generator)
+        output_fp4[i:i + slice_size].copy_(fp4)
+        output_block[i:i + slice_size].copy_(block)
+
+    return output_fp4, to_blocked(output_block, flatten=False)

comfy/hooks.py

@@ -527,7 +527,8 @@ class HookKeyframeGroup:
                         if self._current_keyframe.get_effective_guarantee_steps(max_sigma) > 0:
                             break
                     # if eval_c is outside the percent range, stop looking further
-                    else: break
+                    else:
+                        break
         # update steps current context is used
         self._current_used_steps += 1
         # update current timestep this was performed on

comfy/k_diffusion/sampling.py

@@ -1,11 +1,12 @@
 import math
+import time
 from functools import partial
 
 from scipy import integrate
 import torch
 from torch import nn
 import torchsde
-from tqdm.auto import trange, tqdm
+from tqdm.auto import trange as trange_, tqdm
 
 from . import utils
 from . import deis
@@ -13,6 +14,36 @@ from . import sa_solver
 import comfy.model_patcher
 import comfy.model_sampling
 
+import comfy.memory_management
+
+
+def trange(*args, **kwargs):
+    if comfy.memory_management.aimdo_allocator is None:
+        return trange_(*args, **kwargs)
+
+    pbar = trange_(*args, **kwargs, smoothing=1.0)
+    pbar._i = 0
+    pbar.set_postfix_str("  Model Initializing ...  ")
+
+    _update = pbar.update
+
+    def warmup_update(n=1):
+        pbar._i += 1
+        if pbar._i == 1:
+            pbar.i1_time = time.time()
+            pbar.set_postfix_str(" Model Initialization complete!  ")
+        elif pbar._i == 2:
+            #bring forward the effective start time based the the diff between first and second iteration
+            #to attempt to remove load overhead from the final step rate estimate.
+            pbar.start_t = pbar.i1_time - (time.time() - pbar.i1_time)
+            pbar.set_postfix_str("")
+
+        _update(n)
+
+    pbar.update = warmup_update
+    return pbar
+
+
 def append_zero(x):
     return torch.cat([x, x.new_zeros([1])])
 
@@ -74,6 +105,9 @@ def get_ancestral_step(sigma_from, sigma_to, eta=1.):
 
 def default_noise_sampler(x, seed=None):
     if seed is not None:
+        if x.device == torch.device("cpu"):
+            seed += 1
+
         generator = torch.Generator(device=x.device)
         generator.manual_seed(seed)
     else:
@@ -1557,10 +1591,13 @@ def sample_er_sde(model, x, sigmas, extra_args=None, callback=None, disable=None
 
 
 @torch.no_grad()
-def sample_seeds_2(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None, r=0.5):
+def sample_seeds_2(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None, r=0.5, solver_type="phi_1"):
     """SEEDS-2 - Stochastic Explicit Exponential Derivative-free Solvers (VP Data Prediction) stage 2.
     arXiv: https://arxiv.org/abs/2305.14267 (NeurIPS 2023)
     """
+    if solver_type not in {"phi_1", "phi_2"}:
+        raise ValueError("solver_type must be 'phi_1' or 'phi_2'")
+
     extra_args = {} if extra_args is None else extra_args
     seed = extra_args.get("seed", None)
     noise_sampler = default_noise_sampler(x, seed=seed) if noise_sampler is None else noise_sampler
@@ -1600,8 +1637,14 @@ def sample_seeds_2(model, x, sigmas, extra_args=None, callback=None, disable=Non
         denoised_2 = model(x_2, sigma_s_1 * s_in, **extra_args)
 
         # Step 2
-        denoised_d = torch.lerp(denoised, denoised_2, fac)
-        x = sigmas[i + 1] / sigmas[i] * (-h * eta).exp() * x - alpha_t * ei_h_phi_1(-h_eta) * denoised_d
+        if solver_type == "phi_1":
+            denoised_d = torch.lerp(denoised, denoised_2, fac)
+            x = sigmas[i + 1] / sigmas[i] * (-h * eta).exp() * x - alpha_t * ei_h_phi_1(-h_eta) * denoised_d
+        elif solver_type == "phi_2":
+            b2 = ei_h_phi_2(-h_eta) / r
+            b1 = ei_h_phi_1(-h_eta) - b2
+            x = sigmas[i + 1] / sigmas[i] * (-h * eta).exp() * x - alpha_t * (b1 * denoised + b2 * denoised_2)
+
         if inject_noise:
             segment_factor = (r - 1) * h * eta
             sde_noise = sde_noise * segment_factor.exp()
@@ -1609,6 +1652,17 @@ def sample_seeds_2(model, x, sigmas, extra_args=None, callback=None, disable=Non
             x = x + sde_noise * sigmas[i + 1] * s_noise
     return x
 
+@torch.no_grad()
+def sample_exp_heun_2_x0(model, x, sigmas, extra_args=None, callback=None, disable=None, solver_type="phi_2"):
+    """Deterministic exponential Heun second order method in data prediction (x0) and logSNR time."""
+    return sample_seeds_2(model, x, sigmas, extra_args=extra_args, callback=callback, disable=disable, eta=0.0, s_noise=0.0, noise_sampler=None, r=1.0, solver_type=solver_type)
+
+
+@torch.no_grad()
+def sample_exp_heun_2_x0_sde(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None, solver_type="phi_2"):
+    """Stochastic exponential Heun second order method in data prediction (x0) and logSNR time."""
+    return sample_seeds_2(model, x, sigmas, extra_args=extra_args, callback=callback, disable=disable, eta=eta, s_noise=s_noise, noise_sampler=noise_sampler, r=1.0, solver_type=solver_type)
+
 
 @torch.no_grad()
 def sample_seeds_3(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None, r_1=1./3, r_2=2./3):
@@ -1756,7 +1810,7 @@ def sample_sa_solver(model, x, sigmas, extra_args=None, callback=None, disable=F
         # Predictor
         if sigmas[i + 1] == 0:
             # Denoising step
-            x = denoised
+            x_pred = denoised
         else:
             tau_t = tau_func(sigmas[i + 1])
             curr_lambdas = lambdas[i - predictor_order_used + 1:i + 1]
@@ -1777,7 +1831,7 @@ def sample_sa_solver(model, x, sigmas, extra_args=None, callback=None, disable=F
             if tau_t > 0 and s_noise > 0:
                 noise = noise_sampler(sigmas[i], sigmas[i + 1]) * sigmas[i + 1] * (-2 * tau_t ** 2 * h).expm1().neg().sqrt() * s_noise
                 x_pred = x_pred + noise
-    return x
+    return x_pred
 
 
 @torch.no_grad()

comfy/latent_formats.py

@@ -6,7 +6,9 @@ class LatentFormat:
     latent_dimensions = 2
     latent_rgb_factors = None
     latent_rgb_factors_bias = None
+    latent_rgb_factors_reshape = None
     taesd_decoder_name = None
+    spacial_downscale_ratio = 8
 
     def process_in(self, latent):
         return latent * self.scale_factor
@@ -79,6 +81,7 @@ class SD_X4(LatentFormat):
 
 class SC_Prior(LatentFormat):
     latent_channels = 16
+    spacial_downscale_ratio = 42
     def __init__(self):
         self.scale_factor = 1.0
         self.latent_rgb_factors = [
@@ -101,6 +104,7 @@ class SC_Prior(LatentFormat):
         ]
 
 class SC_B(LatentFormat):
+    spacial_downscale_ratio = 4
     def __init__(self):
         self.scale_factor = 1.0 / 0.43
         self.latent_rgb_factors = [
@@ -178,6 +182,55 @@ class Flux(SD3):
     def process_out(self, latent):
         return (latent / self.scale_factor) + self.shift_factor
 
+class Flux2(LatentFormat):
+    latent_channels = 128
+    spacial_downscale_ratio = 16
+
+    def __init__(self):
+        self.latent_rgb_factors =[
+            [0.0058, 0.0113, 0.0073],
+            [0.0495, 0.0443, 0.0836],
+            [-0.0099, 0.0096, 0.0644],
+            [0.2144, 0.3009, 0.3652],
+            [0.0166, -0.0039, -0.0054],
+            [0.0157, 0.0103, -0.0160],
+            [-0.0398, 0.0902, -0.0235],
+            [-0.0052, 0.0095, 0.0109],
+            [-0.3527, -0.2712, -0.1666],
+            [-0.0301, -0.0356, -0.0180],
+            [-0.0107, 0.0078, 0.0013],
+            [0.0746, 0.0090, -0.0941],
+            [0.0156, 0.0169, 0.0070],
+            [-0.0034, -0.0040, -0.0114],
+            [0.0032, 0.0181, 0.0080],
+            [-0.0939, -0.0008, 0.0186],
+            [0.0018, 0.0043, 0.0104],
+            [0.0284, 0.0056, -0.0127],
+            [-0.0024, -0.0022, -0.0030],
+            [0.1207, -0.0026, 0.0065],
+            [0.0128, 0.0101, 0.0142],
+            [0.0137, -0.0072, -0.0007],
+            [0.0095, 0.0092, -0.0059],
+            [0.0000, -0.0077, -0.0049],
+            [-0.0465, -0.0204, -0.0312],
+            [0.0095, 0.0012, -0.0066],
+            [0.0290, -0.0034, 0.0025],
+            [0.0220, 0.0169, -0.0048],
+            [-0.0332, -0.0457, -0.0468],
+            [-0.0085, 0.0389, 0.0609],
+            [-0.0076, 0.0003, -0.0043],
+            [-0.0111, -0.0460, -0.0614],
+        ]
+
+        self.latent_rgb_factors_bias = [-0.0329, -0.0718, -0.0851]
+        self.latent_rgb_factors_reshape = lambda t: t.reshape(t.shape[0], 32, 2, 2, t.shape[-2], t.shape[-1]).permute(0, 1, 4, 2, 5, 3).reshape(t.shape[0], 32, t.shape[-2] * 2, t.shape[-1] * 2)
+
+    def process_in(self, latent):
+        return latent
+
+    def process_out(self, latent):
+        return latent
+
 class Mochi(LatentFormat):
     latent_channels = 12
     latent_dimensions = 3
@@ -223,6 +276,7 @@ class Mochi(LatentFormat):
 class LTXV(LatentFormat):
     latent_channels = 128
     latent_dimensions = 3
+    spacial_downscale_ratio = 32
 
     def __init__(self):
         self.latent_rgb_factors = [
@@ -358,6 +412,11 @@ class LTXV(LatentFormat):
 
         self.latent_rgb_factors_bias = [-0.0571, -0.1657, -0.2512]
 
+class LTXAV(LTXV):
+    def __init__(self):
+        self.latent_rgb_factors = None
+        self.latent_rgb_factors_bias = None
+
 class HunyuanVideo(LatentFormat):
     latent_channels = 16
     latent_dimensions = 3
@@ -382,6 +441,7 @@ class HunyuanVideo(LatentFormat):
     ]
 
     latent_rgb_factors_bias = [ 0.0259, -0.0192, -0.0761]
+    taesd_decoder_name = "taehv"
 
 class Cosmos1CV8x8x8(LatentFormat):
     latent_channels = 16
@@ -445,7 +505,7 @@ class Wan21(LatentFormat):
         ]).view(1, self.latent_channels, 1, 1, 1)
 
 
-        self.taesd_decoder_name = None #TODO
+        self.taesd_decoder_name = "lighttaew2_1"
 
     def process_in(self, latent):
         latents_mean = self.latents_mean.to(latent.device, latent.dtype)
@@ -460,6 +520,7 @@ class Wan21(LatentFormat):
 class Wan22(Wan21):
     latent_channels = 48
     latent_dimensions = 3
+    spacial_downscale_ratio = 16
 
     latent_rgb_factors = [
             [ 0.0119,  0.0103,  0.0046],
@@ -516,6 +577,7 @@ class Wan22(Wan21):
 
     def __init__(self):
         self.scale_factor = 1.0
+        self.taesd_decoder_name = "lighttaew2_2"
         self.latents_mean = torch.tensor([
                 -0.2289, -0.0052, -0.1323, -0.2339, -0.2799, 0.0174, 0.1838, 0.1557,
                 -0.1382, 0.0542, 0.2813, 0.0891, 0.1570, -0.0098, 0.0375, -0.1825,
@@ -536,6 +598,7 @@ class Wan22(Wan21):
 class HunyuanImage21(LatentFormat):
     latent_channels = 64
     latent_dimensions = 2
+    spacial_downscale_ratio = 32
     scale_factor = 0.75289
 
     latent_rgb_factors = [
@@ -611,6 +674,68 @@ class HunyuanImage21Refiner(LatentFormat):
     latent_dimensions = 3
     scale_factor = 1.03682
 
+    def process_in(self, latent):
+        out = latent * self.scale_factor
+        out = torch.cat((out[:, :, :1], out), dim=2)
+        out = out.permute(0, 2, 1, 3, 4)
+        b, f_times_2, c, h, w = out.shape
+        out = out.reshape(b, f_times_2 // 2, 2 * c, h, w)
+        out = out.permute(0, 2, 1, 3, 4).contiguous()
+        return out
+
+    def process_out(self, latent):
+        z = latent / self.scale_factor
+        z = z.permute(0, 2, 1, 3, 4)
+        b, f, c, h, w = z.shape
+        z = z.reshape(b, f, 2, c // 2, h, w)
+        z = z.permute(0, 1, 2, 3, 4, 5).reshape(b, f * 2, c // 2, h, w)
+        z = z.permute(0, 2, 1, 3, 4)
+        z = z[:, :, 1:]
+        return z
+
+class HunyuanVideo15(LatentFormat):
+    latent_rgb_factors = [
+        [ 0.0568, -0.0521, -0.0131],
+        [ 0.0014,  0.0735,  0.0326],
+        [ 0.0186,  0.0531, -0.0138],
+        [-0.0031,  0.0051,  0.0288],
+        [ 0.0110,  0.0556,  0.0432],
+        [-0.0041, -0.0023, -0.0485],
+        [ 0.0530,  0.0413,  0.0253],
+        [ 0.0283,  0.0251,  0.0339],
+        [ 0.0277, -0.0372, -0.0093],
+        [ 0.0393,  0.0944,  0.1131],
+        [ 0.0020,  0.0251,  0.0037],
+        [-0.0017,  0.0012,  0.0234],
+        [ 0.0468,  0.0436,  0.0203],
+        [ 0.0354,  0.0439, -0.0233],
+        [ 0.0090,  0.0123,  0.0346],
+        [ 0.0382,  0.0029,  0.0217],
+        [ 0.0261, -0.0300,  0.0030],
+        [-0.0088, -0.0220, -0.0283],
+        [-0.0272, -0.0121, -0.0363],
+        [-0.0664, -0.0622,  0.0144],
+        [ 0.0414,  0.0479,  0.0529],
+        [ 0.0355,  0.0612, -0.0247],
+        [ 0.0147,  0.0264,  0.0174],
+        [ 0.0438,  0.0038,  0.0542],
+        [ 0.0431, -0.0573, -0.0033],
+        [-0.0162, -0.0211, -0.0406],
+        [-0.0487, -0.0295, -0.0393],
+        [ 0.0005, -0.0109,  0.0253],
+        [ 0.0296,  0.0591,  0.0353],
+        [ 0.0119,  0.0181, -0.0306],
+        [-0.0085, -0.0362,  0.0229],
+        [ 0.0005, -0.0106,  0.0242]
+    ]
+
+    latent_rgb_factors_bias = [ 0.0456, -0.0202, -0.0644]
+    latent_channels = 32
+    latent_dimensions = 3
+    spacial_downscale_ratio = 16
+    scale_factor = 1.03682
+    taesd_decoder_name = "lighttaehy1_5"
+
 class Hunyuan3Dv2(LatentFormat):
     latent_channels = 64
     latent_dimensions = 1
@@ -630,8 +755,13 @@ class ACEAudio(LatentFormat):
     latent_channels = 8
     latent_dimensions = 2
 
+class ACEAudio15(LatentFormat):
+    latent_channels = 64
+    latent_dimensions = 1
+
 class ChromaRadiance(LatentFormat):
     latent_channels = 3
+    spacial_downscale_ratio = 1
 
     def __init__(self):
         self.latent_rgb_factors = [

comfy/ldm/anima/model.py

@@ -0,0 +1,214 @@
+from comfy.ldm.cosmos.predict2 import MiniTrainDIT
+import torch
+from torch import nn
+import torch.nn.functional as F
+
+
+def rotate_half(x):
+    x1 = x[..., : x.shape[-1] // 2]
+    x2 = x[..., x.shape[-1] // 2 :]
+    return torch.cat((-x2, x1), dim=-1)
+
+
+def apply_rotary_pos_emb(x, cos, sin, unsqueeze_dim=1):
+    cos = cos.unsqueeze(unsqueeze_dim)
+    sin = sin.unsqueeze(unsqueeze_dim)
+    x_embed = (x * cos) + (rotate_half(x) * sin)
+    return x_embed
+
+
+class RotaryEmbedding(nn.Module):
+    def __init__(self, head_dim):
+        super().__init__()
+        self.rope_theta = 10000
+        inv_freq = 1.0 / (self.rope_theta ** (torch.arange(0, head_dim, 2, dtype=torch.int64).to(dtype=torch.float) / head_dim))
+        self.register_buffer("inv_freq", inv_freq, persistent=False)
+
+    @torch.no_grad()
+    def forward(self, x, position_ids):
+        inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1).to(x.device)
+        position_ids_expanded = position_ids[:, None, :].float()
+
+        device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
+        with torch.autocast(device_type=device_type, enabled=False):  # Force float32
+            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
+            emb = torch.cat((freqs, freqs), dim=-1)
+            cos = emb.cos()
+            sin = emb.sin()
+
+        return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
+
+
+class Attention(nn.Module):
+    def __init__(self, query_dim, context_dim, n_heads, head_dim, device=None, dtype=None, operations=None):
+        super().__init__()
+
+        inner_dim = head_dim * n_heads
+        self.n_heads = n_heads
+        self.head_dim = head_dim
+        self.query_dim = query_dim
+        self.context_dim = context_dim
+
+        self.q_proj = operations.Linear(query_dim, inner_dim, bias=False, device=device, dtype=dtype)
+        self.q_norm = operations.RMSNorm(self.head_dim, eps=1e-6, device=device, dtype=dtype)
+
+        self.k_proj = operations.Linear(context_dim, inner_dim, bias=False, device=device, dtype=dtype)
+        self.k_norm = operations.RMSNorm(self.head_dim, eps=1e-6, device=device, dtype=dtype)
+
+        self.v_proj = operations.Linear(context_dim, inner_dim, bias=False, device=device, dtype=dtype)
+
+        self.o_proj = operations.Linear(inner_dim, query_dim, bias=False, device=device, dtype=dtype)
+
+    def forward(self, x, mask=None, context=None, position_embeddings=None, position_embeddings_context=None):
+        context = x if context is None else context
+        input_shape = x.shape[:-1]
+        q_shape = (*input_shape, self.n_heads, self.head_dim)
+        context_shape = context.shape[:-1]
+        kv_shape = (*context_shape, self.n_heads, self.head_dim)
+
+        query_states = self.q_norm(self.q_proj(x).view(q_shape)).transpose(1, 2)
+        key_states = self.k_norm(self.k_proj(context).view(kv_shape)).transpose(1, 2)
+        value_states = self.v_proj(context).view(kv_shape).transpose(1, 2)
+
+        if position_embeddings is not None:
+            assert position_embeddings_context is not None
+            cos, sin = position_embeddings
+            query_states = apply_rotary_pos_emb(query_states, cos, sin)
+            cos, sin = position_embeddings_context
+            key_states = apply_rotary_pos_emb(key_states, cos, sin)
+
+        attn_output = F.scaled_dot_product_attention(query_states, key_states, value_states, attn_mask=mask)
+
+        attn_output = attn_output.transpose(1, 2).reshape(*input_shape, -1).contiguous()
+        attn_output = self.o_proj(attn_output)
+        return attn_output
+
+    def init_weights(self):
+        torch.nn.init.zeros_(self.o_proj.weight)
+
+
+class TransformerBlock(nn.Module):
+    def __init__(self, source_dim, model_dim, num_heads=16, mlp_ratio=4.0, use_self_attn=False, layer_norm=False, device=None, dtype=None, operations=None):
+        super().__init__()
+        self.use_self_attn = use_self_attn
+
+        if self.use_self_attn:
+            self.norm_self_attn = operations.LayerNorm(model_dim, device=device, dtype=dtype) if layer_norm else operations.RMSNorm(model_dim, eps=1e-6, device=device, dtype=dtype)
+            self.self_attn = Attention(
+                query_dim=model_dim,
+                context_dim=model_dim,
+                n_heads=num_heads,
+                head_dim=model_dim//num_heads,
+                device=device,
+                dtype=dtype,
+                operations=operations,
+            )
+
+        self.norm_cross_attn = operations.LayerNorm(model_dim, device=device, dtype=dtype) if layer_norm else operations.RMSNorm(model_dim, eps=1e-6, device=device, dtype=dtype)
+        self.cross_attn = Attention(
+            query_dim=model_dim,
+            context_dim=source_dim,
+            n_heads=num_heads,
+            head_dim=model_dim//num_heads,
+            device=device,
+            dtype=dtype,
+            operations=operations,
+        )
+
+        self.norm_mlp = operations.LayerNorm(model_dim, device=device, dtype=dtype) if layer_norm else operations.RMSNorm(model_dim, eps=1e-6, device=device, dtype=dtype)
+        self.mlp = nn.Sequential(
+            operations.Linear(model_dim, int(model_dim * mlp_ratio), device=device, dtype=dtype),
+            nn.GELU(),
+            operations.Linear(int(model_dim * mlp_ratio), model_dim, device=device, dtype=dtype)
+        )
+
+    def forward(self, x, context, target_attention_mask=None, source_attention_mask=None, position_embeddings=None, position_embeddings_context=None):
+        if self.use_self_attn:
+            normed = self.norm_self_attn(x)
+            attn_out = self.self_attn(normed, mask=target_attention_mask, position_embeddings=position_embeddings, position_embeddings_context=position_embeddings)
+            x = x + attn_out
+
+        normed = self.norm_cross_attn(x)
+        attn_out = self.cross_attn(normed, mask=source_attention_mask, context=context, position_embeddings=position_embeddings, position_embeddings_context=position_embeddings_context)
+        x = x + attn_out
+
+        x = x + self.mlp(self.norm_mlp(x))
+        return x
+
+    def init_weights(self):
+        torch.nn.init.zeros_(self.mlp[2].weight)
+        self.cross_attn.init_weights()
+
+
+class LLMAdapter(nn.Module):
+    def __init__(
+            self,
+            source_dim=1024,
+            target_dim=1024,
+            model_dim=1024,
+            num_layers=6,
+            num_heads=16,
+            use_self_attn=True,
+            layer_norm=False,
+            device=None,
+            dtype=None,
+            operations=None,
+        ):
+        super().__init__()
+
+        self.embed = operations.Embedding(32128, target_dim, device=device, dtype=dtype)
+        if model_dim != target_dim:
+            self.in_proj = operations.Linear(target_dim, model_dim, device=device, dtype=dtype)
+        else:
+            self.in_proj = nn.Identity()
+        self.rotary_emb = RotaryEmbedding(model_dim//num_heads)
+        self.blocks = nn.ModuleList([
+            TransformerBlock(source_dim, model_dim, num_heads=num_heads, use_self_attn=use_self_attn, layer_norm=layer_norm, device=device, dtype=dtype, operations=operations) for _ in range(num_layers)
+        ])
+        self.out_proj = operations.Linear(model_dim, target_dim, device=device, dtype=dtype)
+        self.norm = operations.RMSNorm(target_dim, eps=1e-6, device=device, dtype=dtype)
+
+    def forward(self, source_hidden_states, target_input_ids, target_attention_mask=None, source_attention_mask=None):
+        if target_attention_mask is not None:
+            target_attention_mask = target_attention_mask.to(torch.bool)
+            if target_attention_mask.ndim == 2:
+                target_attention_mask = target_attention_mask.unsqueeze(1).unsqueeze(1)
+
+        if source_attention_mask is not None:
+            source_attention_mask = source_attention_mask.to(torch.bool)
+            if source_attention_mask.ndim == 2:
+                source_attention_mask = source_attention_mask.unsqueeze(1).unsqueeze(1)
+
+        x = self.in_proj(self.embed(target_input_ids))
+        context = source_hidden_states
+        position_ids = torch.arange(x.shape[1], device=x.device).unsqueeze(0)
+        position_ids_context = torch.arange(context.shape[1], device=x.device).unsqueeze(0)
+        position_embeddings = self.rotary_emb(x, position_ids)
+        position_embeddings_context = self.rotary_emb(x, position_ids_context)
+        for block in self.blocks:
+            x = block(x, context, target_attention_mask=target_attention_mask, source_attention_mask=source_attention_mask, position_embeddings=position_embeddings, position_embeddings_context=position_embeddings_context)
+        return self.norm(self.out_proj(x))
+
+
+class Anima(MiniTrainDIT):
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self.llm_adapter = LLMAdapter(device=kwargs.get("device"), dtype=kwargs.get("dtype"), operations=kwargs.get("operations"))
+
+    def preprocess_text_embeds(self, text_embeds, text_ids, t5xxl_weights=None):
+        if text_ids is not None:
+            out = self.llm_adapter(text_embeds, text_ids)
+            if t5xxl_weights is not None:
+                out = out * t5xxl_weights
+
+            if out.shape[1] < 512:
+                out = torch.nn.functional.pad(out, (0, 0, 0, 512 - out.shape[1]))
+            return out
+        else:
+            return text_embeds
+
+    def forward(self, x, timesteps, context, **kwargs):
+        t5xxl_ids = kwargs.pop("t5xxl_ids", None)
+        if t5xxl_ids is not None:
+            context = self.preprocess_text_embeds(context, t5xxl_ids, t5xxl_weights=kwargs.pop("t5xxl_weights", None))
+        return super().forward(x, timesteps, context, **kwargs)

comfy/ldm/chroma/layers.py

@@ -1,15 +1,15 @@
 import torch
 from torch import Tensor, nn
 
-from comfy.ldm.flux.math import attention
 from comfy.ldm.flux.layers import (
     MLPEmbedder,
     RMSNorm,
-    QKNorm,
-    SelfAttention,
     ModulationOut,
 )
 
+# TODO: remove this in a few months
+SingleStreamBlock = None
+DoubleStreamBlock = None
 
 
 class ChromaModulationOut(ModulationOut):
@@ -48,124 +48,6 @@ class Approximator(nn.Module):
         return x
 
 
-class DoubleStreamBlock(nn.Module):
-    def __init__(self, hidden_size: int, num_heads: int, mlp_ratio: float, qkv_bias: bool = False, flipped_img_txt=False, dtype=None, device=None, operations=None):
-        super().__init__()
-
-        mlp_hidden_dim = int(hidden_size * mlp_ratio)
-        self.num_heads = num_heads
-        self.hidden_size = hidden_size
-        self.img_norm1 = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
-        self.img_attn = SelfAttention(dim=hidden_size, num_heads=num_heads, qkv_bias=qkv_bias, dtype=dtype, device=device, operations=operations)
-
-        self.img_norm2 = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
-        self.img_mlp = nn.Sequential(
-            operations.Linear(hidden_size, mlp_hidden_dim, bias=True, dtype=dtype, device=device),
-            nn.GELU(approximate="tanh"),
-            operations.Linear(mlp_hidden_dim, hidden_size, bias=True, dtype=dtype, device=device),
-        )
-
-        self.txt_norm1 = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
-        self.txt_attn = SelfAttention(dim=hidden_size, num_heads=num_heads, qkv_bias=qkv_bias, dtype=dtype, device=device, operations=operations)
-
-        self.txt_norm2 = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
-        self.txt_mlp = nn.Sequential(
-            operations.Linear(hidden_size, mlp_hidden_dim, bias=True, dtype=dtype, device=device),
-            nn.GELU(approximate="tanh"),
-            operations.Linear(mlp_hidden_dim, hidden_size, bias=True, dtype=dtype, device=device),
-        )
-        self.flipped_img_txt = flipped_img_txt
-
-    def forward(self, img: Tensor, txt: Tensor, pe: Tensor, vec: Tensor, attn_mask=None, transformer_options={}):
-        (img_mod1, img_mod2), (txt_mod1, txt_mod2) = vec
-
-        # prepare image for attention
-        img_modulated = torch.addcmul(img_mod1.shift, 1 + img_mod1.scale, self.img_norm1(img))
-        img_qkv = self.img_attn.qkv(img_modulated)
-        img_q, img_k, img_v = img_qkv.view(img_qkv.shape[0], img_qkv.shape[1], 3, self.num_heads, -1).permute(2, 0, 3, 1, 4)
-        img_q, img_k = self.img_attn.norm(img_q, img_k, img_v)
-
-        # prepare txt for attention
-        txt_modulated = torch.addcmul(txt_mod1.shift, 1 + txt_mod1.scale, self.txt_norm1(txt))
-        txt_qkv = self.txt_attn.qkv(txt_modulated)
-        txt_q, txt_k, txt_v = txt_qkv.view(txt_qkv.shape[0], txt_qkv.shape[1], 3, self.num_heads, -1).permute(2, 0, 3, 1, 4)
-        txt_q, txt_k = self.txt_attn.norm(txt_q, txt_k, txt_v)
-
-        # run actual attention
-        attn = attention(torch.cat((txt_q, img_q), dim=2),
-                         torch.cat((txt_k, img_k), dim=2),
-                         torch.cat((txt_v, img_v), dim=2),
-                         pe=pe, mask=attn_mask, transformer_options=transformer_options)
-
-        txt_attn, img_attn = attn[:, : txt.shape[1]], attn[:, txt.shape[1] :]
-
-        # calculate the img bloks
-        img.addcmul_(img_mod1.gate, self.img_attn.proj(img_attn))
-        img.addcmul_(img_mod2.gate, self.img_mlp(torch.addcmul(img_mod2.shift, 1 + img_mod2.scale, self.img_norm2(img))))
-
-        # calculate the txt bloks
-        txt.addcmul_(txt_mod1.gate, self.txt_attn.proj(txt_attn))
-        txt.addcmul_(txt_mod2.gate, self.txt_mlp(torch.addcmul(txt_mod2.shift, 1 + txt_mod2.scale, self.txt_norm2(txt))))
-
-        if txt.dtype == torch.float16:
-            txt = torch.nan_to_num(txt, nan=0.0, posinf=65504, neginf=-65504)
-
-        return img, txt
-
-
-class SingleStreamBlock(nn.Module):
-    """
-    A DiT block with parallel linear layers as described in
-    https://arxiv.org/abs/2302.05442 and adapted modulation interface.
-    """
-
-    def __init__(
-        self,
-        hidden_size: int,
-        num_heads: int,
-        mlp_ratio: float = 4.0,
-        qk_scale: float = None,
-        dtype=None,
-        device=None,
-        operations=None
-    ):
-        super().__init__()
-        self.hidden_dim = hidden_size
-        self.num_heads = num_heads
-        head_dim = hidden_size // num_heads
-        self.scale = qk_scale or head_dim**-0.5
-
-        self.mlp_hidden_dim = int(hidden_size * mlp_ratio)
-        # qkv and mlp_in
-        self.linear1 = operations.Linear(hidden_size, hidden_size * 3 + self.mlp_hidden_dim, dtype=dtype, device=device)
-        # proj and mlp_out
-        self.linear2 = operations.Linear(hidden_size + self.mlp_hidden_dim, hidden_size, dtype=dtype, device=device)
-
-        self.norm = QKNorm(head_dim, dtype=dtype, device=device, operations=operations)
-
-        self.hidden_size = hidden_size
-        self.pre_norm = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
-
-        self.mlp_act = nn.GELU(approximate="tanh")
-
-    def forward(self, x: Tensor, pe: Tensor, vec: Tensor, attn_mask=None, transformer_options={}) -> Tensor:
-        mod = vec
-        x_mod = torch.addcmul(mod.shift, 1 + mod.scale, self.pre_norm(x))
-        qkv, mlp = torch.split(self.linear1(x_mod), [3 * self.hidden_size, self.mlp_hidden_dim], dim=-1)
-
-        q, k, v = qkv.view(qkv.shape[0], qkv.shape[1], 3, self.num_heads, -1).permute(2, 0, 3, 1, 4)
-        q, k = self.norm(q, k, v)
-
-        # compute attention
-        attn = attention(q, k, v, pe=pe, mask=attn_mask, transformer_options=transformer_options)
-        # compute activation in mlp stream, cat again and run second linear layer
-        output = self.linear2(torch.cat((attn, self.mlp_act(mlp)), 2))
-        x.addcmul_(mod.gate, output)
-        if x.dtype == torch.float16:
-            x = torch.nan_to_num(x, nan=0.0, posinf=65504, neginf=-65504)
-        return x
-
-
 class LastLayer(nn.Module):
     def __init__(self, hidden_size: int, patch_size: int, out_channels: int, dtype=None, device=None, operations=None):
         super().__init__()

comfy/ldm/chroma/model.py

@@ -11,12 +11,12 @@ import comfy.ldm.common_dit
 from comfy.ldm.flux.layers import (
     EmbedND,
     timestep_embedding,
+    DoubleStreamBlock,
+    SingleStreamBlock,
 )
 
 from .layers import (
-    DoubleStreamBlock,
     LastLayer,
-    SingleStreamBlock,
     Approximator,
     ChromaModulationOut,
 )
@@ -40,7 +40,8 @@ class ChromaParams:
     out_dim: int
     hidden_dim: int
     n_layers: int
-
+    txt_ids_dims: list
+    vec_in_dim: int
 
 
 
@@ -90,6 +91,7 @@ class Chroma(nn.Module):
                     self.num_heads,
                     mlp_ratio=params.mlp_ratio,
                     qkv_bias=params.qkv_bias,
+                    modulation=False,
                     dtype=dtype, device=device, operations=operations
                 )
                 for _ in range(params.depth)
@@ -98,7 +100,7 @@ class Chroma(nn.Module):
 
         self.single_blocks = nn.ModuleList(
             [
-                SingleStreamBlock(self.hidden_size, self.num_heads, mlp_ratio=params.mlp_ratio, dtype=dtype, device=device, operations=operations)
+                SingleStreamBlock(self.hidden_size, self.num_heads, mlp_ratio=params.mlp_ratio, modulation=False, dtype=dtype, device=device, operations=operations)
                 for _ in range(params.depth_single_blocks)
             ]
         )
@@ -178,7 +180,10 @@ class Chroma(nn.Module):
         pe = self.pe_embedder(ids)
 
         blocks_replace = patches_replace.get("dit", {})
+        transformer_options["total_blocks"] = len(self.double_blocks)
+        transformer_options["block_type"] = "double"
         for i, block in enumerate(self.double_blocks):
+            transformer_options["block_index"] = i
             if i not in self.skip_mmdit:
                 double_mod = (
                     self.get_modulations(mod_vectors, "double_img", idx=i),
@@ -221,7 +226,10 @@ class Chroma(nn.Module):
 
         img = torch.cat((txt, img), 1)
 
+        transformer_options["total_blocks"] = len(self.single_blocks)
+        transformer_options["block_type"] = "single"
         for i, block in enumerate(self.single_blocks):
+            transformer_options["block_index"] = i
             if i not in self.skip_dit:
                 single_mod = self.get_modulations(mod_vectors, "single", idx=i)
                 if ("single_block", i) in blocks_replace:

comfy/ldm/chroma_radiance/model.py

@@ -10,12 +10,10 @@ from torch import Tensor, nn
 from einops import repeat
 import comfy.ldm.common_dit
 
-from comfy.ldm.flux.layers import EmbedND
+from comfy.ldm.flux.layers import EmbedND, DoubleStreamBlock, SingleStreamBlock
 
 from comfy.ldm.chroma.model import Chroma, ChromaParams
 from comfy.ldm.chroma.layers import (
-    DoubleStreamBlock,
-    SingleStreamBlock,
     Approximator,
 )
 from .layers import (
@@ -39,7 +37,7 @@ class ChromaRadianceParams(ChromaParams):
     nerf_final_head_type: str
     # None means use the same dtype as the model.
     nerf_embedder_dtype: Optional[torch.dtype]
-
+    use_x0: bool
 
 class ChromaRadiance(Chroma):
     """
@@ -89,7 +87,6 @@ class ChromaRadiance(Chroma):
                     dtype=dtype, device=device, operations=operations
                 )
 
-
         self.double_blocks = nn.ModuleList(
             [
                 DoubleStreamBlock(
@@ -97,6 +94,7 @@ class ChromaRadiance(Chroma):
                     self.num_heads,
                     mlp_ratio=params.mlp_ratio,
                     qkv_bias=params.qkv_bias,
+                    modulation=False,
                     dtype=dtype, device=device, operations=operations
                 )
                 for _ in range(params.depth)
@@ -109,6 +107,7 @@ class ChromaRadiance(Chroma):
                     self.hidden_size,
                     self.num_heads,
                     mlp_ratio=params.mlp_ratio,
+                    modulation=False,
                     dtype=dtype, device=device, operations=operations,
                 )
                 for _ in range(params.depth_single_blocks)
@@ -160,6 +159,9 @@ class ChromaRadiance(Chroma):
         self.skip_dit = []
         self.lite = False
 
+        if params.use_x0:
+            self.register_buffer("__x0__", torch.tensor([]))
+
     @property
     def _nerf_final_layer(self) -> nn.Module:
         if self.params.nerf_final_head_type == "linear":
@@ -268,7 +270,7 @@ class ChromaRadiance(Chroma):
         bad_keys = tuple(
             k
             for k, v in overrides.items()
-            if type(v) != type(getattr(params, k)) and (v is not None or k not in nullable_keys)
+            if not isinstance(v, type(getattr(params, k))) and (v is not None or k not in nullable_keys)
         )
         if bad_keys:
             e = f"Invalid value(s) in transformer_options chroma_radiance_options: {', '.join(bad_keys)}"
@@ -277,6 +279,12 @@ class ChromaRadiance(Chroma):
         params_dict |= overrides
         return params.__class__(**params_dict)
 
+    def _apply_x0_residual(self, predicted, noisy, timesteps):
+
+        # non zero during training to prevent 0 div
+        eps = 0.0
+        return (noisy - predicted) / (timesteps.view(-1,1,1,1) + eps)
+
     def _forward(
         self,
         x: Tensor,
@@ -317,4 +325,11 @@ class ChromaRadiance(Chroma):
             transformer_options,
             attn_mask=kwargs.get("attention_mask", None),
         )
-        return self.forward_nerf(img, img_out, params)[:, :, :h, :w]
+
+        out = self.forward_nerf(img, img_out, params)[:, :, :h, :w]
+
+        # If x0 variant → v-pred, just return this instead
+        if hasattr(self, "__x0__"):
+            out = self._apply_x0_residual(out, img, timestep)
+        return out
+

comfy/ldm/cosmos/predict2.py

@@ -13,6 +13,7 @@ from torchvision import transforms
 
 import comfy.patcher_extension
 from comfy.ldm.modules.attention import optimized_attention
+import comfy.ldm.common_dit
 
 def apply_rotary_pos_emb(
     t: torch.Tensor,
@@ -334,7 +335,7 @@ class FinalLayer(nn.Module):
         device=None, dtype=None, operations=None
     ):
         super().__init__()
-        self.layer_norm = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
+        self.layer_norm = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
         self.linear = operations.Linear(
             hidden_size, spatial_patch_size * spatial_patch_size * temporal_patch_size * out_channels, bias=False, device=device, dtype=dtype
         )
@@ -462,6 +463,8 @@ class Block(nn.Module):
         extra_per_block_pos_emb: Optional[torch.Tensor] = None,
         transformer_options: Optional[dict] = {},
     ) -> torch.Tensor:
+        residual_dtype = x_B_T_H_W_D.dtype
+        compute_dtype = emb_B_T_D.dtype
         if extra_per_block_pos_emb is not None:
             x_B_T_H_W_D = x_B_T_H_W_D + extra_per_block_pos_emb
 
@@ -511,7 +514,7 @@ class Block(nn.Module):
         result_B_T_H_W_D = rearrange(
             self.self_attn(
                 # normalized_x_B_T_HW_D,
-                rearrange(normalized_x_B_T_H_W_D, "b t h w d -> b (t h w) d"),
+                rearrange(normalized_x_B_T_H_W_D.to(compute_dtype), "b t h w d -> b (t h w) d"),
                 None,
                 rope_emb=rope_emb_L_1_1_D,
                 transformer_options=transformer_options,
@@ -521,7 +524,7 @@ class Block(nn.Module):
             h=H,
             w=W,
         )
-        x_B_T_H_W_D = x_B_T_H_W_D + gate_self_attn_B_T_1_1_D * result_B_T_H_W_D
+        x_B_T_H_W_D = x_B_T_H_W_D + gate_self_attn_B_T_1_1_D.to(residual_dtype) * result_B_T_H_W_D.to(residual_dtype)
 
         def _x_fn(
             _x_B_T_H_W_D: torch.Tensor,
@@ -535,7 +538,7 @@ class Block(nn.Module):
             )
             _result_B_T_H_W_D = rearrange(
                 self.cross_attn(
-                    rearrange(_normalized_x_B_T_H_W_D, "b t h w d -> b (t h w) d"),
+                    rearrange(_normalized_x_B_T_H_W_D.to(compute_dtype), "b t h w d -> b (t h w) d"),
                     crossattn_emb,
                     rope_emb=rope_emb_L_1_1_D,
                     transformer_options=transformer_options,
@@ -554,7 +557,7 @@ class Block(nn.Module):
             shift_cross_attn_B_T_1_1_D,
             transformer_options=transformer_options,
         )
-        x_B_T_H_W_D = result_B_T_H_W_D * gate_cross_attn_B_T_1_1_D + x_B_T_H_W_D
+        x_B_T_H_W_D = result_B_T_H_W_D.to(residual_dtype) * gate_cross_attn_B_T_1_1_D.to(residual_dtype) + x_B_T_H_W_D
 
         normalized_x_B_T_H_W_D = _fn(
             x_B_T_H_W_D,
@@ -562,8 +565,8 @@ class Block(nn.Module):
             scale_mlp_B_T_1_1_D,
             shift_mlp_B_T_1_1_D,
         )
-        result_B_T_H_W_D = self.mlp(normalized_x_B_T_H_W_D)
-        x_B_T_H_W_D = x_B_T_H_W_D + gate_mlp_B_T_1_1_D * result_B_T_H_W_D
+        result_B_T_H_W_D = self.mlp(normalized_x_B_T_H_W_D.to(compute_dtype))
+        x_B_T_H_W_D = x_B_T_H_W_D + gate_mlp_B_T_1_1_D.to(residual_dtype) * result_B_T_H_W_D.to(residual_dtype)
         return x_B_T_H_W_D
 
 
@@ -835,6 +838,8 @@ class MiniTrainDIT(nn.Module):
         padding_mask: Optional[torch.Tensor] = None,
         **kwargs,
     ):
+        orig_shape = list(x.shape)
+        x = comfy.ldm.common_dit.pad_to_patch_size(x, (self.patch_temporal, self.patch_spatial, self.patch_spatial))
         x_B_C_T_H_W = x
         timesteps_B_T = timesteps
         crossattn_emb = context
@@ -873,6 +878,14 @@ class MiniTrainDIT(nn.Module):
             "extra_per_block_pos_emb": extra_pos_emb_B_T_H_W_D_or_T_H_W_B_D,
             "transformer_options": kwargs.get("transformer_options", {}),
         }
+
+        # The residual stream for this model has large values. To make fp16 compute_dtype work, we keep the residual stream
+        # in fp32, but run attention and MLP modules in fp16.
+        # An alternate method that clamps fp16 values "works" in the sense that it makes coherent images, but there is noticeable
+        # quality degradation and visual artifacts.
+        if x_B_T_H_W_D.dtype == torch.float16:
+            x_B_T_H_W_D = x_B_T_H_W_D.float()
+
         for block in self.blocks:
             x_B_T_H_W_D = block(
                 x_B_T_H_W_D,
@@ -881,6 +894,6 @@ class MiniTrainDIT(nn.Module):
                 **block_kwargs,
             )
 
-        x_B_T_H_W_O = self.final_layer(x_B_T_H_W_D, t_embedding_B_T_D, adaln_lora_B_T_3D=adaln_lora_B_T_3D)
-        x_B_C_Tt_Hp_Wp = self.unpatchify(x_B_T_H_W_O)
+        x_B_T_H_W_O = self.final_layer(x_B_T_H_W_D.to(crossattn_emb.dtype), t_embedding_B_T_D, adaln_lora_B_T_3D=adaln_lora_B_T_3D)
+        x_B_C_Tt_Hp_Wp = self.unpatchify(x_B_T_H_W_O)[:, :, :orig_shape[-3], :orig_shape[-2], :orig_shape[-1]]
         return x_B_C_Tt_Hp_Wp

comfy/ldm/flux/layers.py

@@ -48,15 +48,44 @@ def timestep_embedding(t: Tensor, dim, max_period=10000, time_factor: float = 10
     return embedding
 
 class MLPEmbedder(nn.Module):
-    def __init__(self, in_dim: int, hidden_dim: int, dtype=None, device=None, operations=None):
+    def __init__(self, in_dim: int, hidden_dim: int, bias=True, dtype=None, device=None, operations=None):
         super().__init__()
-        self.in_layer = operations.Linear(in_dim, hidden_dim, bias=True, dtype=dtype, device=device)
+        self.in_layer = operations.Linear(in_dim, hidden_dim, bias=bias, dtype=dtype, device=device)
         self.silu = nn.SiLU()
-        self.out_layer = operations.Linear(hidden_dim, hidden_dim, bias=True, dtype=dtype, device=device)
+        self.out_layer = operations.Linear(hidden_dim, hidden_dim, bias=bias, dtype=dtype, device=device)
 
     def forward(self, x: Tensor) -> Tensor:
         return self.out_layer(self.silu(self.in_layer(x)))
 
+class YakMLP(nn.Module):
+    def __init__(self, hidden_size: int, intermediate_size: int, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.gate_proj = operations.Linear(self.hidden_size, self.intermediate_size, bias=True, dtype=dtype, device=device)
+        self.up_proj = operations.Linear(self.hidden_size, self.intermediate_size, bias=True, dtype=dtype, device=device)
+        self.down_proj = operations.Linear(self.intermediate_size, self.hidden_size, bias=True, dtype=dtype, device=device)
+        self.act_fn = nn.SiLU()
+
+    def forward(self, x: Tensor) -> Tensor:
+        down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
+        return down_proj
+
+def build_mlp(hidden_size, mlp_hidden_dim, mlp_silu_act=False, yak_mlp=False, dtype=None, device=None, operations=None):
+    if yak_mlp:
+        return YakMLP(hidden_size, mlp_hidden_dim, dtype=dtype, device=device, operations=operations)
+    if mlp_silu_act:
+        return nn.Sequential(
+            operations.Linear(hidden_size, mlp_hidden_dim * 2, bias=False, dtype=dtype, device=device),
+            SiLUActivation(),
+            operations.Linear(mlp_hidden_dim, hidden_size, bias=False, dtype=dtype, device=device),
+        )
+    else:
+        return nn.Sequential(
+            operations.Linear(hidden_size, mlp_hidden_dim, bias=True, dtype=dtype, device=device),
+            nn.GELU(approximate="tanh"),
+            operations.Linear(mlp_hidden_dim, hidden_size, bias=True, dtype=dtype, device=device),
+        )
 
 class RMSNorm(torch.nn.Module):
     def __init__(self, dim: int, dtype=None, device=None, operations=None):
@@ -80,14 +109,14 @@ class QKNorm(torch.nn.Module):
 
 
 class SelfAttention(nn.Module):
-    def __init__(self, dim: int, num_heads: int = 8, qkv_bias: bool = False, dtype=None, device=None, operations=None):
+    def __init__(self, dim: int, num_heads: int = 8, qkv_bias: bool = False, proj_bias: bool = True, dtype=None, device=None, operations=None):
         super().__init__()
         self.num_heads = num_heads
         head_dim = dim // num_heads
 
         self.qkv = operations.Linear(dim, dim * 3, bias=qkv_bias, dtype=dtype, device=device)
         self.norm = QKNorm(head_dim, dtype=dtype, device=device, operations=operations)
-        self.proj = operations.Linear(dim, dim, dtype=dtype, device=device)
+        self.proj = operations.Linear(dim, dim, bias=proj_bias, dtype=dtype, device=device)
 
 
 @dataclass
@@ -98,11 +127,11 @@ class ModulationOut:
 
 
 class Modulation(nn.Module):
-    def __init__(self, dim: int, double: bool, dtype=None, device=None, operations=None):
+    def __init__(self, dim: int, double: bool, bias=True, dtype=None, device=None, operations=None):
         super().__init__()
         self.is_double = double
         self.multiplier = 6 if double else 3
-        self.lin = operations.Linear(dim, self.multiplier * dim, bias=True, dtype=dtype, device=device)
+        self.lin = operations.Linear(dim, self.multiplier * dim, bias=bias, dtype=dtype, device=device)
 
     def forward(self, vec: Tensor) -> tuple:
         if vec.ndim == 2:
@@ -129,77 +158,107 @@ def apply_mod(tensor, m_mult, m_add=None, modulation_dims=None):
         return tensor
 
 
+class SiLUActivation(nn.Module):
+    def __init__(self):
+        super().__init__()
+        self.gate_fn = nn.SiLU()
+
+    def forward(self, x: Tensor) -> Tensor:
+        x1, x2 = x.chunk(2, dim=-1)
+        return self.gate_fn(x1) * x2
+
+
 class DoubleStreamBlock(nn.Module):
-    def __init__(self, hidden_size: int, num_heads: int, mlp_ratio: float, qkv_bias: bool = False, flipped_img_txt=False, dtype=None, device=None, operations=None):
+    def __init__(self, hidden_size: int, num_heads: int, mlp_ratio: float, qkv_bias: bool = False, flipped_img_txt=False, modulation=True, mlp_silu_act=False, proj_bias=True, yak_mlp=False, dtype=None, device=None, operations=None):
         super().__init__()
 
         mlp_hidden_dim = int(hidden_size * mlp_ratio)
         self.num_heads = num_heads
         self.hidden_size = hidden_size
-        self.img_mod = Modulation(hidden_size, double=True, dtype=dtype, device=device, operations=operations)
+        self.modulation = modulation
+
+        if self.modulation:
+            self.img_mod = Modulation(hidden_size, double=True, dtype=dtype, device=device, operations=operations)
+
         self.img_norm1 = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
-        self.img_attn = SelfAttention(dim=hidden_size, num_heads=num_heads, qkv_bias=qkv_bias, dtype=dtype, device=device, operations=operations)
+        self.img_attn = SelfAttention(dim=hidden_size, num_heads=num_heads, qkv_bias=qkv_bias, proj_bias=proj_bias, dtype=dtype, device=device, operations=operations)
 
         self.img_norm2 = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
-        self.img_mlp = nn.Sequential(
-            operations.Linear(hidden_size, mlp_hidden_dim, bias=True, dtype=dtype, device=device),
-            nn.GELU(approximate="tanh"),
-            operations.Linear(mlp_hidden_dim, hidden_size, bias=True, dtype=dtype, device=device),
-        )
 
-        self.txt_mod = Modulation(hidden_size, double=True, dtype=dtype, device=device, operations=operations)
+        self.img_mlp = build_mlp(hidden_size, mlp_hidden_dim, mlp_silu_act=mlp_silu_act, yak_mlp=yak_mlp, dtype=dtype, device=device, operations=operations)
+
+        if self.modulation:
+            self.txt_mod = Modulation(hidden_size, double=True, dtype=dtype, device=device, operations=operations)
+
         self.txt_norm1 = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
-        self.txt_attn = SelfAttention(dim=hidden_size, num_heads=num_heads, qkv_bias=qkv_bias, dtype=dtype, device=device, operations=operations)
+        self.txt_attn = SelfAttention(dim=hidden_size, num_heads=num_heads, qkv_bias=qkv_bias, proj_bias=proj_bias, dtype=dtype, device=device, operations=operations)
 
         self.txt_norm2 = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
-        self.txt_mlp = nn.Sequential(
-            operations.Linear(hidden_size, mlp_hidden_dim, bias=True, dtype=dtype, device=device),
-            nn.GELU(approximate="tanh"),
-            operations.Linear(mlp_hidden_dim, hidden_size, bias=True, dtype=dtype, device=device),
-        )
+
+        self.txt_mlp = build_mlp(hidden_size, mlp_hidden_dim, mlp_silu_act=mlp_silu_act, yak_mlp=yak_mlp, dtype=dtype, device=device, operations=operations)
+
         self.flipped_img_txt = flipped_img_txt
 
     def forward(self, img: Tensor, txt: Tensor, vec: Tensor, pe: Tensor, attn_mask=None, modulation_dims_img=None, modulation_dims_txt=None, transformer_options={}):
-        img_mod1, img_mod2 = self.img_mod(vec)
-        txt_mod1, txt_mod2 = self.txt_mod(vec)
+        if self.modulation:
+            img_mod1, img_mod2 = self.img_mod(vec)
+            txt_mod1, txt_mod2 = self.txt_mod(vec)
+        else:
+            (img_mod1, img_mod2), (txt_mod1, txt_mod2) = vec
 
         # prepare image for attention
         img_modulated = self.img_norm1(img)
         img_modulated = apply_mod(img_modulated, (1 + img_mod1.scale), img_mod1.shift, modulation_dims_img)
         img_qkv = self.img_attn.qkv(img_modulated)
+        del img_modulated
         img_q, img_k, img_v = img_qkv.view(img_qkv.shape[0], img_qkv.shape[1], 3, self.num_heads, -1).permute(2, 0, 3, 1, 4)
+        del img_qkv
         img_q, img_k = self.img_attn.norm(img_q, img_k, img_v)
 
         # prepare txt for attention
         txt_modulated = self.txt_norm1(txt)
         txt_modulated = apply_mod(txt_modulated, (1 + txt_mod1.scale), txt_mod1.shift, modulation_dims_txt)
         txt_qkv = self.txt_attn.qkv(txt_modulated)
+        del txt_modulated
         txt_q, txt_k, txt_v = txt_qkv.view(txt_qkv.shape[0], txt_qkv.shape[1], 3, self.num_heads, -1).permute(2, 0, 3, 1, 4)
+        del txt_qkv
         txt_q, txt_k = self.txt_attn.norm(txt_q, txt_k, txt_v)
 
         if self.flipped_img_txt:
+            q = torch.cat((img_q, txt_q), dim=2)
+            del img_q, txt_q
+            k = torch.cat((img_k, txt_k), dim=2)
+            del img_k, txt_k
+            v = torch.cat((img_v, txt_v), dim=2)
+            del img_v, txt_v
             # run actual attention
-            attn = attention(torch.cat((img_q, txt_q), dim=2),
-                             torch.cat((img_k, txt_k), dim=2),
-                             torch.cat((img_v, txt_v), dim=2),
+            attn = attention(q, k, v,
                              pe=pe, mask=attn_mask, transformer_options=transformer_options)
+            del q, k, v
 
             img_attn, txt_attn = attn[:, : img.shape[1]], attn[:, img.shape[1]:]
         else:
+            q = torch.cat((txt_q, img_q), dim=2)
+            del txt_q, img_q
+            k = torch.cat((txt_k, img_k), dim=2)
+            del txt_k, img_k
+            v = torch.cat((txt_v, img_v), dim=2)
+            del txt_v, img_v
             # run actual attention
-            attn = attention(torch.cat((txt_q, img_q), dim=2),
-                             torch.cat((txt_k, img_k), dim=2),
-                             torch.cat((txt_v, img_v), dim=2),
+            attn = attention(q, k, v,
                              pe=pe, mask=attn_mask, transformer_options=transformer_options)
+            del q, k, v
 
             txt_attn, img_attn = attn[:, : txt.shape[1]], attn[:, txt.shape[1]:]
 
         # calculate the img bloks
-        img = img + apply_mod(self.img_attn.proj(img_attn), img_mod1.gate, None, modulation_dims_img)
-        img = img + apply_mod(self.img_mlp(apply_mod(self.img_norm2(img), (1 + img_mod2.scale), img_mod2.shift, modulation_dims_img)), img_mod2.gate, None, modulation_dims_img)
+        img += apply_mod(self.img_attn.proj(img_attn), img_mod1.gate, None, modulation_dims_img)
+        del img_attn
+        img += apply_mod(self.img_mlp(apply_mod(self.img_norm2(img), (1 + img_mod2.scale), img_mod2.shift, modulation_dims_img)), img_mod2.gate, None, modulation_dims_img)
 
         # calculate the txt bloks
         txt += apply_mod(self.txt_attn.proj(txt_attn), txt_mod1.gate, None, modulation_dims_txt)
+        del txt_attn
         txt += apply_mod(self.txt_mlp(apply_mod(self.txt_norm2(txt), (1 + txt_mod2.scale), txt_mod2.shift, modulation_dims_txt)), txt_mod2.gate, None, modulation_dims_txt)
 
         if txt.dtype == torch.float16:
@@ -220,6 +279,10 @@ class SingleStreamBlock(nn.Module):
         num_heads: int,
         mlp_ratio: float = 4.0,
         qk_scale: float = None,
+        modulation=True,
+        mlp_silu_act=False,
+        bias=True,
+        yak_mlp=False,
         dtype=None,
         device=None,
         operations=None
@@ -231,30 +294,55 @@ class SingleStreamBlock(nn.Module):
         self.scale = qk_scale or head_dim**-0.5
 
         self.mlp_hidden_dim = int(hidden_size * mlp_ratio)
+
+        self.mlp_hidden_dim_first = self.mlp_hidden_dim
+        self.yak_mlp = yak_mlp
+        if mlp_silu_act:
+            self.mlp_hidden_dim_first = int(hidden_size * mlp_ratio * 2)
+            self.mlp_act = SiLUActivation()
+        else:
+            self.mlp_act = nn.GELU(approximate="tanh")
+
+        if self.yak_mlp:
+            self.mlp_hidden_dim_first *= 2
+            self.mlp_act = nn.SiLU()
+
         # qkv and mlp_in
-        self.linear1 = operations.Linear(hidden_size, hidden_size * 3 + self.mlp_hidden_dim, dtype=dtype, device=device)
+        self.linear1 = operations.Linear(hidden_size, hidden_size * 3 + self.mlp_hidden_dim_first, bias=bias, dtype=dtype, device=device)
         # proj and mlp_out
-        self.linear2 = operations.Linear(hidden_size + self.mlp_hidden_dim, hidden_size, dtype=dtype, device=device)
+        self.linear2 = operations.Linear(hidden_size + self.mlp_hidden_dim, hidden_size, bias=bias, dtype=dtype, device=device)
 
         self.norm = QKNorm(head_dim, dtype=dtype, device=device, operations=operations)
 
         self.hidden_size = hidden_size
         self.pre_norm = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
 
-        self.mlp_act = nn.GELU(approximate="tanh")
-        self.modulation = Modulation(hidden_size, double=False, dtype=dtype, device=device, operations=operations)
+        if modulation:
+            self.modulation = Modulation(hidden_size, double=False, dtype=dtype, device=device, operations=operations)
+        else:
+            self.modulation = None
 
     def forward(self, x: Tensor, vec: Tensor, pe: Tensor, attn_mask=None, modulation_dims=None, transformer_options={}) -> Tensor:
-        mod, _ = self.modulation(vec)
-        qkv, mlp = torch.split(self.linear1(apply_mod(self.pre_norm(x), (1 + mod.scale), mod.shift, modulation_dims)), [3 * self.hidden_size, self.mlp_hidden_dim], dim=-1)
+        if self.modulation:
+            mod, _ = self.modulation(vec)
+        else:
+            mod = vec
+
+        qkv, mlp = torch.split(self.linear1(apply_mod(self.pre_norm(x), (1 + mod.scale), mod.shift, modulation_dims)), [3 * self.hidden_size, self.mlp_hidden_dim_first], dim=-1)
 
         q, k, v = qkv.view(qkv.shape[0], qkv.shape[1], 3, self.num_heads, -1).permute(2, 0, 3, 1, 4)
+        del qkv
         q, k = self.norm(q, k, v)
 
         # compute attention
         attn = attention(q, k, v, pe=pe, mask=attn_mask, transformer_options=transformer_options)
+        del q, k, v
         # compute activation in mlp stream, cat again and run second linear layer
-        output = self.linear2(torch.cat((attn, self.mlp_act(mlp)), 2))
+        if self.yak_mlp:
+            mlp = self.mlp_act(mlp[..., self.mlp_hidden_dim_first // 2:]) * mlp[..., :self.mlp_hidden_dim_first // 2]
+        else:
+            mlp = self.mlp_act(mlp)
+        output = self.linear2(torch.cat((attn, mlp), 2))
         x += apply_mod(output, mod.gate, None, modulation_dims)
         if x.dtype == torch.float16:
             x = torch.nan_to_num(x, nan=0.0, posinf=65504, neginf=-65504)
@@ -262,11 +350,11 @@ class SingleStreamBlock(nn.Module):
 
 
 class LastLayer(nn.Module):
-    def __init__(self, hidden_size: int, patch_size: int, out_channels: int, dtype=None, device=None, operations=None):
+    def __init__(self, hidden_size: int, patch_size: int, out_channels: int, bias=True, dtype=None, device=None, operations=None):
         super().__init__()
         self.norm_final = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
-        self.linear = operations.Linear(hidden_size, patch_size * patch_size * out_channels, bias=True, dtype=dtype, device=device)
-        self.adaLN_modulation = nn.Sequential(nn.SiLU(), operations.Linear(hidden_size, 2 * hidden_size, bias=True, dtype=dtype, device=device))
+        self.linear = operations.Linear(hidden_size, patch_size * patch_size * out_channels, bias=bias, dtype=dtype, device=device)
+        self.adaLN_modulation = nn.Sequential(nn.SiLU(), operations.Linear(hidden_size, 2 * hidden_size, bias=bias, dtype=dtype, device=device))
 
     def forward(self, x: Tensor, vec: Tensor, modulation_dims=None) -> Tensor:
         if vec.ndim == 2:

comfy/ldm/flux/math.py

@@ -4,23 +4,16 @@ from torch import Tensor
 
 from comfy.ldm.modules.attention import optimized_attention
 import comfy.model_management
+import logging
 
 
 def attention(q: Tensor, k: Tensor, v: Tensor, pe: Tensor, mask=None, transformer_options={}) -> Tensor:
-    q_shape = q.shape
-    k_shape = k.shape
-
     if pe is not None:
-        q = q.to(dtype=pe.dtype).reshape(*q.shape[:-1], -1, 1, 2)
-        k = k.to(dtype=pe.dtype).reshape(*k.shape[:-1], -1, 1, 2)
-        q = (pe[..., 0] * q[..., 0] + pe[..., 1] * q[..., 1]).reshape(*q_shape).type_as(v)
-        k = (pe[..., 0] * k[..., 0] + pe[..., 1] * k[..., 1]).reshape(*k_shape).type_as(v)
-
+        q, k = apply_rope(q, k, pe)
     heads = q.shape[1]
     x = optimized_attention(q, k, v, heads, skip_reshape=True, mask=mask, transformer_options=transformer_options)
     return x
 
-
 def rope(pos: Tensor, dim: int, theta: int) -> Tensor:
     assert dim % 2 == 0
     if comfy.model_management.is_device_mps(pos.device) or comfy.model_management.is_intel_xpu() or comfy.model_management.is_directml_enabled():
@@ -35,7 +28,8 @@ def rope(pos: Tensor, dim: int, theta: int) -> Tensor:
     out = rearrange(out, "b n d (i j) -> b n d i j", i=2, j=2)
     return out.to(dtype=torch.float32, device=pos.device)
 
-def apply_rope1(x: Tensor, freqs_cis: Tensor):
+
+def _apply_rope1(x: Tensor, freqs_cis: Tensor):
     x_ = x.to(dtype=freqs_cis.dtype).reshape(*x.shape[:-1], -1, 1, 2)
 
     x_out = freqs_cis[..., 0] * x_[..., 0]
@@ -43,5 +37,26 @@ def apply_rope1(x: Tensor, freqs_cis: Tensor):
 
     return x_out.reshape(*x.shape).type_as(x)
 
-def apply_rope(xq: Tensor, xk: Tensor, freqs_cis: Tensor):
+
+def _apply_rope(xq: Tensor, xk: Tensor, freqs_cis: Tensor):
     return apply_rope1(xq, freqs_cis), apply_rope1(xk, freqs_cis)
+
+
+try:
+    import comfy.quant_ops
+    q_apply_rope = comfy.quant_ops.ck.apply_rope
+    q_apply_rope1 = comfy.quant_ops.ck.apply_rope1
+    def apply_rope(xq, xk, freqs_cis):
+        if comfy.model_management.in_training:
+            return _apply_rope(xq, xk, freqs_cis)
+        else:
+            return apply_rope1(xq, freqs_cis), apply_rope1(xk, freqs_cis)
+    def apply_rope1(x, freqs_cis):
+        if comfy.model_management.in_training:
+            return _apply_rope1(x, freqs_cis)
+        else:
+            return q_apply_rope1(x, freqs_cis)
+except:
+    logging.warning("No comfy kitchen, using old apply_rope functions.")
+    apply_rope = _apply_rope
+    apply_rope1 = _apply_rope1

comfy/ldm/flux/model.py

@@ -15,6 +15,8 @@ from .layers import (
     MLPEmbedder,
     SingleStreamBlock,
     timestep_embedding,
+    Modulation,
+    RMSNorm
 )
 
 @dataclass
@@ -33,6 +35,14 @@ class FluxParams:
     patch_size: int
     qkv_bias: bool
     guidance_embed: bool
+    txt_ids_dims: list
+    global_modulation: bool = False
+    mlp_silu_act: bool = False
+    ops_bias: bool = True
+    default_ref_method: str = "offset"
+    ref_index_scale: float = 1.0
+    yak_mlp: bool = False
+    txt_norm: bool = False
 
 
 class Flux(nn.Module):
@@ -58,13 +68,22 @@ class Flux(nn.Module):
         self.hidden_size = params.hidden_size
         self.num_heads = params.num_heads
         self.pe_embedder = EmbedND(dim=pe_dim, theta=params.theta, axes_dim=params.axes_dim)
-        self.img_in = operations.Linear(self.in_channels, self.hidden_size, bias=True, dtype=dtype, device=device)
-        self.time_in = MLPEmbedder(in_dim=256, hidden_dim=self.hidden_size, dtype=dtype, device=device, operations=operations)
-        self.vector_in = MLPEmbedder(params.vec_in_dim, self.hidden_size, dtype=dtype, device=device, operations=operations)
+        self.img_in = operations.Linear(self.in_channels, self.hidden_size, bias=params.ops_bias, dtype=dtype, device=device)
+        self.time_in = MLPEmbedder(in_dim=256, hidden_dim=self.hidden_size, bias=params.ops_bias, dtype=dtype, device=device, operations=operations)
+        if params.vec_in_dim is not None:
+            self.vector_in = MLPEmbedder(params.vec_in_dim, self.hidden_size, dtype=dtype, device=device, operations=operations)
+        else:
+            self.vector_in = None
+
         self.guidance_in = (
-            MLPEmbedder(in_dim=256, hidden_dim=self.hidden_size, dtype=dtype, device=device, operations=operations) if params.guidance_embed else nn.Identity()
+            MLPEmbedder(in_dim=256, hidden_dim=self.hidden_size, bias=params.ops_bias, dtype=dtype, device=device, operations=operations) if params.guidance_embed else nn.Identity()
         )
-        self.txt_in = operations.Linear(params.context_in_dim, self.hidden_size, dtype=dtype, device=device)
+        self.txt_in = operations.Linear(params.context_in_dim, self.hidden_size, bias=params.ops_bias, dtype=dtype, device=device)
+
+        if params.txt_norm:
+            self.txt_norm = RMSNorm(params.context_in_dim, dtype=dtype, device=device, operations=operations)
+        else:
+            self.txt_norm = None
 
         self.double_blocks = nn.ModuleList(
             [
@@ -73,6 +92,10 @@ class Flux(nn.Module):
                     self.num_heads,
                     mlp_ratio=params.mlp_ratio,
                     qkv_bias=params.qkv_bias,
+                    modulation=params.global_modulation is False,
+                    mlp_silu_act=params.mlp_silu_act,
+                    proj_bias=params.ops_bias,
+                    yak_mlp=params.yak_mlp,
                     dtype=dtype, device=device, operations=operations
                 )
                 for _ in range(params.depth)
@@ -81,13 +104,30 @@ class Flux(nn.Module):
 
         self.single_blocks = nn.ModuleList(
             [
-                SingleStreamBlock(self.hidden_size, self.num_heads, mlp_ratio=params.mlp_ratio, dtype=dtype, device=device, operations=operations)
+                SingleStreamBlock(self.hidden_size, self.num_heads, mlp_ratio=params.mlp_ratio, modulation=params.global_modulation is False, mlp_silu_act=params.mlp_silu_act, bias=params.ops_bias, yak_mlp=params.yak_mlp, dtype=dtype, device=device, operations=operations)
                 for _ in range(params.depth_single_blocks)
             ]
         )
 
         if final_layer:
-            self.final_layer = LastLayer(self.hidden_size, 1, self.out_channels, dtype=dtype, device=device, operations=operations)
+            self.final_layer = LastLayer(self.hidden_size, 1, self.out_channels, bias=params.ops_bias, dtype=dtype, device=device, operations=operations)
+
+        if params.global_modulation:
+            self.double_stream_modulation_img = Modulation(
+                self.hidden_size,
+                double=True,
+                bias=False,
+                dtype=dtype, device=device, operations=operations
+            )
+            self.double_stream_modulation_txt = Modulation(
+                self.hidden_size,
+                double=True,
+                bias=False,
+                dtype=dtype, device=device, operations=operations
+            )
+            self.single_stream_modulation = Modulation(
+                self.hidden_size, double=False, bias=False, dtype=dtype, device=device, operations=operations
+            )
 
     def forward_orig(
         self,
@@ -103,9 +143,6 @@ class Flux(nn.Module):
         attn_mask: Tensor = None,
     ) -> Tensor:
 
-        if y is None:
-            y = torch.zeros((img.shape[0], self.params.vec_in_dim), device=img.device, dtype=img.dtype)
-
         patches = transformer_options.get("patches", {})
         patches_replace = transformer_options.get("patches_replace", {})
         if img.ndim != 3 or txt.ndim != 3:
@@ -118,9 +155,19 @@ class Flux(nn.Module):
             if guidance is not None:
                 vec = vec + self.guidance_in(timestep_embedding(guidance, 256).to(img.dtype))
 
-        vec = vec + self.vector_in(y[:, :self.params.vec_in_dim])
+        if self.vector_in is not None:
+            if y is None:
+                y = torch.zeros((img.shape[0], self.params.vec_in_dim), device=img.device, dtype=img.dtype)
+            vec = vec + self.vector_in(y[:, :self.params.vec_in_dim])
+
+        if self.txt_norm is not None:
+            txt = self.txt_norm(txt)
         txt = self.txt_in(txt)
 
+        vec_orig = vec
+        if self.params.global_modulation:
+            vec = (self.double_stream_modulation_img(vec_orig), self.double_stream_modulation_txt(vec_orig))
+
         if "post_input" in patches:
             for p in patches["post_input"]:
                 out = p({"img": img, "txt": txt, "img_ids": img_ids, "txt_ids": txt_ids})
@@ -136,7 +183,10 @@ class Flux(nn.Module):
             pe = None
 
         blocks_replace = patches_replace.get("dit", {})
+        transformer_options["total_blocks"] = len(self.double_blocks)
+        transformer_options["block_type"] = "double"
         for i, block in enumerate(self.double_blocks):
+            transformer_options["block_index"] = i
             if ("double_block", i) in blocks_replace:
                 def block_wrap(args):
                     out = {}
@@ -177,7 +227,13 @@ class Flux(nn.Module):
 
         img = torch.cat((txt, img), 1)
 
+        if self.params.global_modulation:
+            vec, _ = self.single_stream_modulation(vec_orig)
+
+        transformer_options["total_blocks"] = len(self.single_blocks)
+        transformer_options["block_type"] = "single"
         for i, block in enumerate(self.single_blocks):
+            transformer_options["block_index"] = i
             if ("single_block", i) in blocks_replace:
                 def block_wrap(args):
                     out = {}
@@ -207,10 +263,10 @@ class Flux(nn.Module):
 
         img = img[:, txt.shape[1] :, ...]
 
-        img = self.final_layer(img, vec)  # (N, T, patch_size ** 2 * out_channels)
+        img = self.final_layer(img, vec_orig)  # (N, T, patch_size ** 2 * out_channels)
         return img
 
-    def process_img(self, x, index=0, h_offset=0, w_offset=0):
+    def process_img(self, x, index=0, h_offset=0, w_offset=0, transformer_options={}):
         bs, c, h, w = x.shape
         patch_size = self.patch_size
         x = comfy.ldm.common_dit.pad_to_patch_size(x, (patch_size, patch_size))
@@ -222,10 +278,22 @@ class Flux(nn.Module):
         h_offset = ((h_offset + (patch_size // 2)) // patch_size)
         w_offset = ((w_offset + (patch_size // 2)) // patch_size)
 
-        img_ids = torch.zeros((h_len, w_len, 3), device=x.device, dtype=x.dtype)
+        steps_h = h_len
+        steps_w = w_len
+
+        rope_options = transformer_options.get("rope_options", None)
+        if rope_options is not None:
+            h_len = (h_len - 1.0) * rope_options.get("scale_y", 1.0) + 1.0
+            w_len = (w_len - 1.0) * rope_options.get("scale_x", 1.0) + 1.0
+
+            index += rope_options.get("shift_t", 0.0)
+            h_offset += rope_options.get("shift_y", 0.0)
+            w_offset += rope_options.get("shift_x", 0.0)
+
+        img_ids = torch.zeros((steps_h, steps_w, len(self.params.axes_dim)), device=x.device, dtype=torch.float32)
         img_ids[:, :, 0] = img_ids[:, :, 1] + index
-        img_ids[:, :, 1] = img_ids[:, :, 1] + torch.linspace(h_offset, h_len - 1 + h_offset, steps=h_len, device=x.device, dtype=x.dtype).unsqueeze(1)
-        img_ids[:, :, 2] = img_ids[:, :, 2] + torch.linspace(w_offset, w_len - 1 + w_offset, steps=w_len, device=x.device, dtype=x.dtype).unsqueeze(0)
+        img_ids[:, :, 1] = img_ids[:, :, 1] + torch.linspace(h_offset, h_len - 1 + h_offset, steps=steps_h, device=x.device, dtype=torch.float32).unsqueeze(1)
+        img_ids[:, :, 2] = img_ids[:, :, 2] + torch.linspace(w_offset, w_len - 1 + w_offset, steps=steps_w, device=x.device, dtype=torch.float32).unsqueeze(0)
         return img, repeat(img_ids, "h w c -> b (h w) c", b=bs)
 
     def forward(self, x, timestep, context, y=None, guidance=None, ref_latents=None, control=None, transformer_options={}, **kwargs):
@@ -241,16 +309,16 @@ class Flux(nn.Module):
 
         h_len = ((h_orig + (patch_size // 2)) // patch_size)
         w_len = ((w_orig + (patch_size // 2)) // patch_size)
-        img, img_ids = self.process_img(x)
+        img, img_ids = self.process_img(x, transformer_options=transformer_options)
         img_tokens = img.shape[1]
         if ref_latents is not None:
             h = 0
             w = 0
             index = 0
-            ref_latents_method = kwargs.get("ref_latents_method", "offset")
+            ref_latents_method = kwargs.get("ref_latents_method", self.params.default_ref_method)
             for ref in ref_latents:
                 if ref_latents_method == "index":
-                    index += 1
+                    index += self.params.ref_index_scale
                     h_offset = 0
                     w_offset = 0
                 elif ref_latents_method == "uxo":
@@ -274,7 +342,12 @@ class Flux(nn.Module):
                 img = torch.cat([img, kontext], dim=1)
                 img_ids = torch.cat([img_ids, kontext_ids], dim=1)
 
-        txt_ids = torch.zeros((bs, context.shape[1], 3), device=x.device, dtype=x.dtype)
+        txt_ids = torch.zeros((bs, context.shape[1], len(self.params.axes_dim)), device=x.device, dtype=torch.float32)
+
+        if len(self.params.txt_ids_dims) > 0:
+            for i in self.params.txt_ids_dims:
+                txt_ids[:, :, i] = torch.linspace(0, context.shape[1] - 1, steps=context.shape[1], device=x.device, dtype=torch.float32)
+
         out = self.forward_orig(img, img_ids, context, txt_ids, timestep, y, guidance, control, transformer_options, attn_mask=kwargs.get("attention_mask", None))
         out = out[:, :img_tokens]
-        return rearrange(out, "b (h w) (c ph pw) -> b c (h ph) (w pw)", h=h_len, w=w_len, ph=2, pw=2)[:,:,:h_orig,:w_orig]
+        return rearrange(out, "b (h w) (c ph pw) -> b c (h ph) (w pw)", h=h_len, w=w_len, ph=self.patch_size, pw=self.patch_size)[:,:,:h_orig,:w_orig]

comfy/ldm/hunyuan_video/model.py

@@ -6,7 +6,6 @@ import comfy.ldm.flux.layers
 import comfy.ldm.modules.diffusionmodules.mmdit
 from comfy.ldm.modules.attention import optimized_attention
 
-
 from dataclasses import dataclass
 from einops import repeat
 
@@ -42,6 +41,9 @@ class HunyuanVideoParams:
     guidance_embed: bool
     byt5: bool
     meanflow: bool
+    use_cond_type_embedding: bool
+    vision_in_dim: int
+    meanflow_sum: bool
 
 
 class SelfAttentionRef(nn.Module):
@@ -157,7 +159,10 @@ class TokenRefiner(nn.Module):
         t = self.t_embedder(timestep_embedding(timesteps, 256, time_factor=1.0).to(x.dtype))
         # m = mask.float().unsqueeze(-1)
         # c = (x.float() * m).sum(dim=1) / m.sum(dim=1) #TODO: the following works when the x.shape is the same length as the tokens but might break otherwise
-        c = x.sum(dim=1) / x.shape[1]
+        if x.dtype == torch.float16:
+            c = x.float().sum(dim=1) / x.shape[1]
+        else:
+            c = x.sum(dim=1) / x.shape[1]
 
         c = t + self.c_embedder(c.to(x.dtype))
         x = self.input_embedder(x)
@@ -196,11 +201,15 @@ class HunyuanVideo(nn.Module):
     def __init__(self, image_model=None, final_layer=True, dtype=None, device=None, operations=None, **kwargs):
         super().__init__()
         self.dtype = dtype
+        operation_settings = {"operations": operations, "device": device, "dtype": dtype}
+
         params = HunyuanVideoParams(**kwargs)
         self.params = params
         self.patch_size = params.patch_size
         self.in_channels = params.in_channels
         self.out_channels = params.out_channels
+        self.use_cond_type_embedding = params.use_cond_type_embedding
+        self.vision_in_dim = params.vision_in_dim
         if params.hidden_size % params.num_heads != 0:
             raise ValueError(
                 f"Hidden size {params.hidden_size} must be divisible by num_heads {params.num_heads}"
@@ -266,6 +275,18 @@ class HunyuanVideo(nn.Module):
         if final_layer:
             self.final_layer = LastLayer(self.hidden_size, self.patch_size[-1], self.out_channels, dtype=dtype, device=device, operations=operations)
 
+        # HunyuanVideo 1.5 specific modules
+        if self.vision_in_dim is not None:
+            from comfy.ldm.wan.model import MLPProj
+            self.vision_in = MLPProj(in_dim=self.vision_in_dim, out_dim=self.hidden_size, operation_settings=operation_settings)
+        else:
+            self.vision_in = None
+        if self.use_cond_type_embedding:
+            # 0: text_encoder feature 1: byt5 feature 2: vision_encoder feature
+            self.cond_type_embedding = nn.Embedding(3, self.hidden_size)
+        else:
+            self.cond_type_embedding = None
+
     def forward_orig(
         self,
         img: Tensor,
@@ -276,6 +297,7 @@ class HunyuanVideo(nn.Module):
         timesteps: Tensor,
         y: Tensor = None,
         txt_byt5=None,
+        clip_fea=None,
         guidance: Tensor = None,
         guiding_frame_index=None,
         ref_latent=None,
@@ -296,7 +318,7 @@ class HunyuanVideo(nn.Module):
                 timesteps_r = transformer_options['sample_sigmas'][w[0] + 1]
                 timesteps_r = timesteps_r.unsqueeze(0).to(device=timesteps.device, dtype=timesteps.dtype)
                 vec_r = self.time_r_in(timestep_embedding(timesteps_r, 256, time_factor=1000.0).to(img.dtype))
-                vec = (vec + vec_r) / 2
+                vec = (vec + vec_r) if self.params.meanflow_sum else (vec + vec_r) / 2
 
         if ref_latent is not None:
             ref_latent_ids = self.img_ids(ref_latent)
@@ -331,12 +353,31 @@ class HunyuanVideo(nn.Module):
 
         txt = self.txt_in(txt, timesteps, txt_mask, transformer_options=transformer_options)
 
+        if self.cond_type_embedding is not None:
+            self.cond_type_embedding.to(txt.device)
+            cond_emb = self.cond_type_embedding(torch.zeros_like(txt[:, :, 0], device=txt.device, dtype=torch.long))
+            txt = txt + cond_emb.to(txt.dtype)
+
         if self.byt5_in is not None and txt_byt5 is not None:
             txt_byt5 = self.byt5_in(txt_byt5)
+            if self.cond_type_embedding is not None:
+                cond_emb = self.cond_type_embedding(torch.ones_like(txt_byt5[:, :, 0], device=txt_byt5.device, dtype=torch.long))
+                txt_byt5 = txt_byt5 + cond_emb.to(txt_byt5.dtype)
+                txt = torch.cat((txt_byt5, txt), dim=1) # byt5 first for HunyuanVideo1.5
+            else:
+                txt = torch.cat((txt, txt_byt5), dim=1)
             txt_byt5_ids = torch.zeros((txt_ids.shape[0], txt_byt5.shape[1], txt_ids.shape[-1]), device=txt_ids.device, dtype=txt_ids.dtype)
-            txt = torch.cat((txt, txt_byt5), dim=1)
             txt_ids = torch.cat((txt_ids, txt_byt5_ids), dim=1)
 
+        if clip_fea is not None:
+            txt_vision_states = self.vision_in(clip_fea)
+            if self.cond_type_embedding is not None:
+                cond_emb = self.cond_type_embedding(2 * torch.ones_like(txt_vision_states[:, :, 0], dtype=torch.long, device=txt_vision_states.device))
+                txt_vision_states = txt_vision_states + cond_emb
+            txt = torch.cat((txt_vision_states.to(txt.dtype), txt), dim=1)
+            extra_txt_ids = torch.zeros((txt_ids.shape[0], txt_vision_states.shape[1], txt_ids.shape[-1]), device=txt_ids.device, dtype=txt_ids.dtype)
+            txt_ids = torch.cat((txt_ids, extra_txt_ids), dim=1)
+
         ids = torch.cat((img_ids, txt_ids), dim=1)
         pe = self.pe_embedder(ids)
 
@@ -349,7 +390,10 @@ class HunyuanVideo(nn.Module):
             attn_mask = None
 
         blocks_replace = patches_replace.get("dit", {})
+        transformer_options["total_blocks"] = len(self.double_blocks)
+        transformer_options["block_type"] = "double"
         for i, block in enumerate(self.double_blocks):
+            transformer_options["block_index"] = i
             if ("double_block", i) in blocks_replace:
                 def block_wrap(args):
                     out = {}
@@ -371,7 +415,10 @@ class HunyuanVideo(nn.Module):
 
         img = torch.cat((img, txt), 1)
 
+        transformer_options["total_blocks"] = len(self.single_blocks)
+        transformer_options["block_type"] = "single"
         for i, block in enumerate(self.single_blocks):
+            transformer_options["block_index"] = i
             if ("single_block", i) in blocks_replace:
                 def block_wrap(args):
                     out = {}
@@ -430,14 +477,14 @@ class HunyuanVideo(nn.Module):
         img_ids[:, :, 1] = img_ids[:, :, 1] + torch.linspace(0, w_len - 1, steps=w_len, device=x.device, dtype=x.dtype).unsqueeze(0)
         return repeat(img_ids, "h w c -> b (h w) c", b=bs)
 
-    def forward(self, x, timestep, context, y=None, txt_byt5=None, guidance=None, attention_mask=None, guiding_frame_index=None, ref_latent=None, disable_time_r=False, control=None, transformer_options={}, **kwargs):
+    def forward(self, x, timestep, context, y=None, txt_byt5=None, clip_fea=None, guidance=None, attention_mask=None, guiding_frame_index=None, ref_latent=None, disable_time_r=False, control=None, transformer_options={}, **kwargs):
         return comfy.patcher_extension.WrapperExecutor.new_class_executor(
             self._forward,
             self,
             comfy.patcher_extension.get_all_wrappers(comfy.patcher_extension.WrappersMP.DIFFUSION_MODEL, transformer_options)
-        ).execute(x, timestep, context, y, txt_byt5, guidance, attention_mask, guiding_frame_index, ref_latent, disable_time_r, control, transformer_options, **kwargs)
+        ).execute(x, timestep, context, y, txt_byt5, clip_fea, guidance, attention_mask, guiding_frame_index, ref_latent, disable_time_r, control, transformer_options, **kwargs)
 
-    def _forward(self, x, timestep, context, y=None, txt_byt5=None, guidance=None, attention_mask=None, guiding_frame_index=None, ref_latent=None, disable_time_r=False, control=None, transformer_options={}, **kwargs):
+    def _forward(self, x, timestep, context, y=None, txt_byt5=None, clip_fea=None, guidance=None, attention_mask=None, guiding_frame_index=None, ref_latent=None, disable_time_r=False, control=None, transformer_options={}, **kwargs):
         bs = x.shape[0]
         if len(self.patch_size) == 3:
             img_ids = self.img_ids(x)
@@ -445,5 +492,5 @@ class HunyuanVideo(nn.Module):
         else:
             img_ids = self.img_ids_2d(x)
             txt_ids = torch.zeros((bs, context.shape[1], 2), device=x.device, dtype=x.dtype)
-        out = self.forward_orig(x, img_ids, context, txt_ids, attention_mask, timestep, y, txt_byt5, guidance, guiding_frame_index, ref_latent, disable_time_r=disable_time_r, control=control, transformer_options=transformer_options)
+        out = self.forward_orig(x, img_ids, context, txt_ids, attention_mask, timestep, y, txt_byt5, clip_fea, guidance, guiding_frame_index, ref_latent, disable_time_r=disable_time_r, control=control, transformer_options=transformer_options)
         return out

comfy/ldm/hunyuan_video/upsampler.py

@@ -0,0 +1,122 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from comfy.ldm.modules.diffusionmodules.model import ResnetBlock, VideoConv3d
+from comfy.ldm.hunyuan_video.vae_refiner import RMS_norm
+import comfy.model_management
+import comfy.model_patcher
+
+class SRResidualCausalBlock3D(nn.Module):
+    def __init__(self, channels: int):
+        super().__init__()
+        self.block = nn.Sequential(
+            VideoConv3d(channels, channels, kernel_size=3),
+            nn.SiLU(inplace=True),
+            VideoConv3d(channels, channels, kernel_size=3),
+            nn.SiLU(inplace=True),
+            VideoConv3d(channels, channels, kernel_size=3),
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return x + self.block(x)
+
+class SRModel3DV2(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        hidden_channels: int = 64,
+        num_blocks: int = 6,
+        global_residual: bool = False,
+    ):
+        super().__init__()
+        self.in_conv = VideoConv3d(in_channels, hidden_channels, kernel_size=3)
+        self.blocks = nn.ModuleList([SRResidualCausalBlock3D(hidden_channels) for _ in range(num_blocks)])
+        self.out_conv = VideoConv3d(hidden_channels, out_channels, kernel_size=3)
+        self.global_residual = bool(global_residual)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        residual = x
+        y = self.in_conv(x)
+        for blk in self.blocks:
+            y = blk(y)
+        y = self.out_conv(y)
+        if self.global_residual and (y.shape == residual.shape):
+            y = y + residual
+        return y
+
+
+class Upsampler(nn.Module):
+    def __init__(
+        self,
+        z_channels: int,
+        out_channels: int,
+        block_out_channels: tuple[int, ...],
+        num_res_blocks: int = 2,
+    ):
+        super().__init__()
+        self.num_res_blocks = num_res_blocks
+        self.block_out_channels = block_out_channels
+        self.z_channels = z_channels
+
+        ch = block_out_channels[0]
+        self.conv_in = VideoConv3d(z_channels, ch, kernel_size=3)
+
+        self.up = nn.ModuleList()
+
+        for i, tgt in enumerate(block_out_channels):
+            stage = nn.Module()
+            stage.block = nn.ModuleList([ResnetBlock(in_channels=ch if j == 0 else tgt,
+                                                    out_channels=tgt,
+                                                    temb_channels=0,
+                                                    conv_shortcut=False,
+                                                    conv_op=VideoConv3d, norm_op=RMS_norm)
+                                        for j in range(num_res_blocks + 1)])
+            ch = tgt
+            self.up.append(stage)
+
+        self.norm_out = RMS_norm(ch)
+        self.conv_out = VideoConv3d(ch, out_channels, kernel_size=3)
+
+    def forward(self, z):
+        """
+        Args:
+            z: (B, C, T, H, W)
+            target_shape: (H, W)
+        """
+        # z to block_in
+        repeats = self.block_out_channels[0] // (self.z_channels)
+        x = self.conv_in(z) + z.repeat_interleave(repeats=repeats, dim=1)
+
+        # upsampling
+        for stage in self.up:
+            for blk in stage.block:
+                x = blk(x)
+
+        out = self.conv_out(F.silu(self.norm_out(x)))
+        return out
+
+UPSAMPLERS = {
+    "720p": SRModel3DV2,
+    "1080p": Upsampler,
+}
+
+class HunyuanVideo15SRModel():
+    def __init__(self, model_type, config):
+        self.load_device = comfy.model_management.vae_device()
+        offload_device = comfy.model_management.vae_offload_device()
+        self.dtype = comfy.model_management.vae_dtype(self.load_device)
+        self.model_class = UPSAMPLERS.get(model_type)
+        self.model = self.model_class(**config).eval()
+
+        self.patcher = comfy.model_patcher.CoreModelPatcher(self.model, load_device=self.load_device, offload_device=offload_device)
+
+    def load_sd(self, sd):
+        return self.model.load_state_dict(sd, strict=True, assign=self.patcher.is_dynamic())
+
+    def get_sd(self):
+        return self.model.state_dict()
+
+    def resample_latent(self, latent):
+        comfy.model_management.load_model_gpu(self.patcher)
+        return self.model(latent.to(self.load_device))

comfy/ldm/hunyuan_video/vae_refiner.py

@@ -1,11 +1,13 @@
 import torch
 import torch.nn as nn
 import torch.nn.functional as F
-from comfy.ldm.modules.diffusionmodules.model import ResnetBlock, AttnBlock, VideoConv3d, Normalize
+from comfy.ldm.modules.diffusionmodules.model import ResnetBlock, AttnBlock, CarriedConv3d, Normalize, conv_carry_causal_3d, torch_cat_if_needed
 import comfy.ops
 import comfy.ldm.models.autoencoder
+import comfy.model_management
 ops = comfy.ops.disable_weight_init
 
+
 class RMS_norm(nn.Module):
     def __init__(self, dim):
         super().__init__()
@@ -14,10 +16,10 @@ class RMS_norm(nn.Module):
         self.gamma = nn.Parameter(torch.empty(shape))
 
     def forward(self, x):
-        return F.normalize(x, dim=1) * self.scale * self.gamma
+        return F.normalize(x, dim=1) * self.scale * comfy.model_management.cast_to(self.gamma, dtype=x.dtype, device=x.device)
 
 class DnSmpl(nn.Module):
-    def __init__(self, ic, oc, tds=True, refiner_vae=True, op=VideoConv3d):
+    def __init__(self, ic, oc, tds, refiner_vae, op):
         super().__init__()
         fct = 2 * 2 * 2 if tds else 1 * 2 * 2
         assert oc % fct == 0
@@ -27,11 +29,12 @@ class DnSmpl(nn.Module):
         self.tds = tds
         self.gs = fct * ic // oc
 
-    def forward(self, x):
+    def forward(self, x, conv_carry_in=None, conv_carry_out=None):
         r1 = 2 if self.tds else 1
-        h = self.conv(x)
+        h = conv_carry_causal_3d([x], self.conv, conv_carry_in, conv_carry_out)
+
+        if self.tds and self.refiner_vae and conv_carry_in is None:
 
-        if self.tds and self.refiner_vae:
             hf = h[:, :, :1, :, :]
             b, c, f, ht, wd = hf.shape
             hf = hf.reshape(b, c, f, ht // 2, 2, wd // 2, 2)
@@ -39,14 +42,7 @@ class DnSmpl(nn.Module):
             hf = hf.reshape(b, 2 * 2 * c, f, ht // 2, wd // 2)
             hf = torch.cat([hf, hf], dim=1)
 
-            hn = h[:, :, 1:, :, :]
-            b, c, frms, ht, wd = hn.shape
-            nf = frms // r1
-            hn = hn.reshape(b, c, nf, r1, ht // 2, 2, wd // 2, 2)
-            hn = hn.permute(0, 3, 5, 7, 1, 2, 4, 6)
-            hn = hn.reshape(b, r1 * 2 * 2 * c, nf, ht // 2, wd // 2)
-
-            h = torch.cat([hf, hn], dim=2)
+            h = h[:, :, 1:, :, :]
 
             xf = x[:, :, :1, :, :]
             b, ci, f, ht, wd = xf.shape
@@ -54,38 +50,36 @@ class DnSmpl(nn.Module):
             xf = xf.permute(0, 4, 6, 1, 2, 3, 5)
             xf = xf.reshape(b, 2 * 2 * ci, f, ht // 2, wd // 2)
             B, C, T, H, W = xf.shape
-            xf = xf.view(B, h.shape[1], self.gs // 2, T, H, W).mean(dim=2)
+            xf = xf.view(B, hf.shape[1], self.gs // 2, T, H, W).mean(dim=2)
 
-            xn = x[:, :, 1:, :, :]
-            b, ci, frms, ht, wd = xn.shape
-            nf = frms // r1
-            xn = xn.reshape(b, ci, nf, r1, ht // 2, 2, wd // 2, 2)
-            xn = xn.permute(0, 3, 5, 7, 1, 2, 4, 6)
-            xn = xn.reshape(b, r1 * 2 * 2 * ci, nf, ht // 2, wd // 2)
-            B, C, T, H, W = xn.shape
-            xn = xn.view(B, h.shape[1], self.gs, T, H, W).mean(dim=2)
-            sc = torch.cat([xf, xn], dim=2)
-        else:
-            b, c, frms, ht, wd = h.shape
+            x = x[:, :, 1:, :, :]
 
-            nf = frms // r1
-            h = h.reshape(b, c, nf, r1, ht // 2, 2, wd // 2, 2)
-            h = h.permute(0, 3, 5, 7, 1, 2, 4, 6)
-            h = h.reshape(b, r1 * 2 * 2 * c, nf, ht // 2, wd // 2)
+        if h.shape[2] == 0:
+            return hf + xf
 
-            b, ci, frms, ht, wd = x.shape
-            nf = frms // r1
-            sc = x.reshape(b, ci, nf, r1, ht // 2, 2, wd // 2, 2)
-            sc = sc.permute(0, 3, 5, 7, 1, 2, 4, 6)
-            sc = sc.reshape(b, r1 * 2 * 2 * ci, nf, ht // 2, wd // 2)
-            B, C, T, H, W = sc.shape
-            sc = sc.view(B, h.shape[1], self.gs, T, H, W).mean(dim=2)
+        b, c, frms, ht, wd = h.shape
+        nf = frms // r1
+        h = h.reshape(b, c, nf, r1, ht // 2, 2, wd // 2, 2)
+        h = h.permute(0, 3, 5, 7, 1, 2, 4, 6)
+        h = h.reshape(b, r1 * 2 * 2 * c, nf, ht // 2, wd // 2)
 
-        return h + sc
+        b, ci, frms, ht, wd = x.shape
+        nf = frms // r1
+        x = x.reshape(b, ci, nf, r1, ht // 2, 2, wd // 2, 2)
+        x = x.permute(0, 3, 5, 7, 1, 2, 4, 6)
+        x = x.reshape(b, r1 * 2 * 2 * ci, nf, ht // 2, wd // 2)
+        B, C, T, H, W = x.shape
+        x = x.view(B, h.shape[1], self.gs, T, H, W).mean(dim=2)
+
+        if self.tds and self.refiner_vae and conv_carry_in is None:
+            h = torch.cat([hf, h], dim=2)
+            x = torch.cat([xf, x], dim=2)
+
+        return h + x
 
 
 class UpSmpl(nn.Module):
-    def __init__(self, ic, oc, tus=True, refiner_vae=True, op=VideoConv3d):
+    def __init__(self, ic, oc, tus, refiner_vae, op):
         super().__init__()
         fct = 2 * 2 * 2 if tus else 1 * 2 * 2
         self.conv = op(ic, oc * fct, kernel_size=3, stride=1, padding=1)
@@ -94,11 +88,11 @@ class UpSmpl(nn.Module):
         self.tus = tus
         self.rp = fct * oc // ic
 
-    def forward(self, x):
+    def forward(self, x, conv_carry_in=None, conv_carry_out=None):
         r1 = 2 if self.tus else 1
-        h = self.conv(x)
+        h = conv_carry_causal_3d([x], self.conv, conv_carry_in, conv_carry_out)
 
-        if self.tus and self.refiner_vae:
+        if self.tus and self.refiner_vae and conv_carry_in is None:
             hf = h[:, :, :1, :, :]
             b, c, f, ht, wd = hf.shape
             nc = c // (2 * 2)
@@ -107,14 +101,7 @@ class UpSmpl(nn.Module):
             hf = hf.reshape(b, nc, f, ht * 2, wd * 2)
             hf = hf[:, : hf.shape[1] // 2]
 
-            hn = h[:, :, 1:, :, :]
-            b, c, frms, ht, wd = hn.shape
-            nc = c // (r1 * 2 * 2)
-            hn = hn.reshape(b, r1, 2, 2, nc, frms, ht, wd)
-            hn = hn.permute(0, 4, 5, 1, 6, 2, 7, 3)
-            hn = hn.reshape(b, nc, frms * r1, ht * 2, wd * 2)
-
-            h = torch.cat([hf, hn], dim=2)
+            h = h[:, :, 1:, :, :]
 
             xf = x[:, :, :1, :, :]
             b, ci, f, ht, wd = xf.shape
@@ -125,29 +112,26 @@ class UpSmpl(nn.Module):
             xf = xf.permute(0, 3, 4, 5, 1, 6, 2)
             xf = xf.reshape(b, nc, f, ht * 2, wd * 2)
 
-            xn = x[:, :, 1:, :, :]
-            xn = xn.repeat_interleave(repeats=self.rp, dim=1)
-            b, c, frms, ht, wd = xn.shape
-            nc = c // (r1 * 2 * 2)
-            xn = xn.reshape(b, r1, 2, 2, nc, frms, ht, wd)
-            xn = xn.permute(0, 4, 5, 1, 6, 2, 7, 3)
-            xn = xn.reshape(b, nc, frms * r1, ht * 2, wd * 2)
-            sc = torch.cat([xf, xn], dim=2)
-        else:
-            b, c, frms, ht, wd = h.shape
-            nc = c // (r1 * 2 * 2)
-            h = h.reshape(b, r1, 2, 2, nc, frms, ht, wd)
-            h = h.permute(0, 4, 5, 1, 6, 2, 7, 3)
-            h = h.reshape(b, nc, frms * r1, ht * 2, wd * 2)
+            x = x[:, :, 1:, :, :]
 
-            sc = x.repeat_interleave(repeats=self.rp, dim=1)
-            b, c, frms, ht, wd = sc.shape
-            nc = c // (r1 * 2 * 2)
-            sc = sc.reshape(b, r1, 2, 2, nc, frms, ht, wd)
-            sc = sc.permute(0, 4, 5, 1, 6, 2, 7, 3)
-            sc = sc.reshape(b, nc, frms * r1, ht * 2, wd * 2)
+        b, c, frms, ht, wd = h.shape
+        nc = c // (r1 * 2 * 2)
+        h = h.reshape(b, r1, 2, 2, nc, frms, ht, wd)
+        h = h.permute(0, 4, 5, 1, 6, 2, 7, 3)
+        h = h.reshape(b, nc, frms * r1, ht * 2, wd * 2)
 
-        return h + sc
+        x = x.repeat_interleave(repeats=self.rp, dim=1)
+        b, c, frms, ht, wd = x.shape
+        nc = c // (r1 * 2 * 2)
+        x = x.reshape(b, r1, 2, 2, nc, frms, ht, wd)
+        x = x.permute(0, 4, 5, 1, 6, 2, 7, 3)
+        x = x.reshape(b, nc, frms * r1, ht * 2, wd * 2)
+
+        if self.tus and self.refiner_vae and conv_carry_in is None:
+            h = torch.cat([hf, h], dim=2)
+            x = torch.cat([xf, x], dim=2)
+
+        return h + x
 
 class Encoder(nn.Module):
     def __init__(self, in_channels, z_channels, block_out_channels, num_res_blocks,
@@ -160,7 +144,7 @@ class Encoder(nn.Module):
 
         self.refiner_vae = refiner_vae
         if self.refiner_vae:
-            conv_op = VideoConv3d
+            conv_op = CarriedConv3d
             norm_op = RMS_norm
         else:
             conv_op = ops.Conv3d
@@ -188,9 +172,9 @@ class Encoder(nn.Module):
             self.down.append(stage)
 
         self.mid = nn.Module()
-        self.mid.block_1 = ResnetBlock(in_channels=ch, out_channels=ch, temb_channels=0, conv_op=conv_op, norm_op=norm_op)
+        self.mid.block_1 = ResnetBlock(in_channels=ch, out_channels=ch, conv_op=conv_op, norm_op=norm_op)
         self.mid.attn_1 = AttnBlock(ch, conv_op=ops.Conv3d, norm_op=norm_op)
-        self.mid.block_2 = ResnetBlock(in_channels=ch, out_channels=ch, temb_channels=0, conv_op=conv_op, norm_op=norm_op)
+        self.mid.block_2 = ResnetBlock(in_channels=ch, out_channels=ch, conv_op=conv_op, norm_op=norm_op)
 
         self.norm_out = norm_op(ch)
         self.conv_out = conv_op(ch, z_channels << 1, 3, 1, 1)
@@ -201,31 +185,48 @@ class Encoder(nn.Module):
         if not self.refiner_vae and x.shape[2] == 1:
             x = x.expand(-1, -1, self.ffactor_temporal, -1, -1)
 
-        x = self.conv_in(x)
+        if self.refiner_vae:
+            xl = [x[:, :, :1, :, :]]
+            if x.shape[2] > self.ffactor_temporal:
+                xl += torch.split(x[:, :, 1: 1 + ((x.shape[2] - 1) // self.ffactor_temporal) * self.ffactor_temporal, :, :], self.ffactor_temporal * 2, dim=2)
+            x = xl
+        else:
+            x = [x]
+        out = []
 
-        for stage in self.down:
-            for blk in stage.block:
-                x = blk(x)
-            if hasattr(stage, 'downsample'):
-                x = stage.downsample(x)
+        conv_carry_in = None
 
-        x = self.mid.block_2(self.mid.attn_1(self.mid.block_1(x)))
+        for i, x1 in enumerate(x):
+            conv_carry_out = []
+            if i == len(x) - 1:
+                conv_carry_out = None
+
+            x1 = [ x1 ]
+            x1 = conv_carry_causal_3d(x1, self.conv_in, conv_carry_in, conv_carry_out)
+
+            for stage in self.down:
+                for blk in stage.block:
+                    x1 = blk(x1, None, conv_carry_in, conv_carry_out)
+                if hasattr(stage, 'downsample'):
+                    x1 = stage.downsample(x1, conv_carry_in, conv_carry_out)
+
+            out.append(x1)
+            conv_carry_in = conv_carry_out
+
+        out = torch_cat_if_needed(out, dim=2)
+
+        x = self.mid.block_2(self.mid.attn_1(self.mid.block_1(out)))
+        del out
 
         b, c, t, h, w = x.shape
         grp = c // (self.z_channels << 1)
         skip = x.view(b, c // grp, grp, t, h, w).mean(2)
 
-        out = self.conv_out(F.silu(self.norm_out(x))) + skip
+        out = conv_carry_causal_3d([F.silu(self.norm_out(x))], self.conv_out) + skip
 
         if self.refiner_vae:
             out = self.regul(out)[0]
 
-            out = torch.cat((out[:, :, :1], out), dim=2)
-            out = out.permute(0, 2, 1, 3, 4)
-            b, f_times_2, c, h, w = out.shape
-            out = out.reshape(b, f_times_2 // 2, 2 * c, h, w)
-            out = out.permute(0, 2, 1, 3, 4).contiguous()
-
         return out
 
 class Decoder(nn.Module):
@@ -239,7 +240,7 @@ class Decoder(nn.Module):
 
         self.refiner_vae = refiner_vae
         if self.refiner_vae:
-            conv_op = VideoConv3d
+            conv_op = CarriedConv3d
             norm_op = RMS_norm
         else:
             conv_op = ops.Conv3d
@@ -249,9 +250,9 @@ class Decoder(nn.Module):
         self.conv_in = conv_op(z_channels, ch, kernel_size=3, stride=1, padding=1)
 
         self.mid = nn.Module()
-        self.mid.block_1 = ResnetBlock(in_channels=ch, out_channels=ch, temb_channels=0, conv_op=conv_op, norm_op=norm_op)
+        self.mid.block_1 = ResnetBlock(in_channels=ch, out_channels=ch, conv_op=conv_op, norm_op=norm_op)
         self.mid.attn_1 = AttnBlock(ch, conv_op=ops.Conv3d, norm_op=norm_op)
-        self.mid.block_2 = ResnetBlock(in_channels=ch, out_channels=ch, temb_channels=0, conv_op=conv_op, norm_op=norm_op)
+        self.mid.block_2 = ResnetBlock(in_channels=ch, out_channels=ch,  conv_op=conv_op, norm_op=norm_op)
 
         self.up = nn.ModuleList()
         depth = (ffactor_spatial >> 1).bit_length()
@@ -275,27 +276,38 @@ class Decoder(nn.Module):
         self.conv_out = conv_op(ch, out_channels, 3, stride=1, padding=1)
 
     def forward(self, z):
-        if self.refiner_vae:
-            z = z.permute(0, 2, 1, 3, 4)
-            b, f, c, h, w = z.shape
-            z = z.reshape(b, f, 2, c // 2, h, w)
-            z = z.permute(0, 1, 2, 3, 4, 5).reshape(b, f * 2, c // 2, h, w)
-            z = z.permute(0, 2, 1, 3, 4)
-            z = z[:, :, 1:]
-
-        x = self.conv_in(z) + z.repeat_interleave(self.block_out_channels[0] // self.z_channels, 1)
+        x = conv_carry_causal_3d([z], self.conv_in) + z.repeat_interleave(self.block_out_channels[0] // self.z_channels, 1)
         x = self.mid.block_2(self.mid.attn_1(self.mid.block_1(x)))
 
-        for stage in self.up:
-            for blk in stage.block:
-                x = blk(x)
-            if hasattr(stage, 'upsample'):
-                x = stage.upsample(x)
+        if self.refiner_vae:
+            x = torch.split(x, 2, dim=2)
+        else:
+            x = [ x ]
+        out = []
 
-        out = self.conv_out(F.silu(self.norm_out(x)))
+        conv_carry_in = None
+
+        for i, x1 in enumerate(x):
+            conv_carry_out = []
+            if i == len(x) - 1:
+                conv_carry_out = None
+            for stage in self.up:
+                for blk in stage.block:
+                    x1 = blk(x1, None, conv_carry_in, conv_carry_out)
+                if hasattr(stage, 'upsample'):
+                    x1 = stage.upsample(x1, conv_carry_in, conv_carry_out)
+
+            x1 = [ F.silu(self.norm_out(x1)) ]
+            x1 = conv_carry_causal_3d(x1, self.conv_out, conv_carry_in, conv_carry_out)
+            out.append(x1)
+            conv_carry_in = conv_carry_out
+        del x
+
+        out = torch_cat_if_needed(out, dim=2)
 
         if not self.refiner_vae:
             if z.shape[-3] == 1:
                 out = out[:, :, -1:]
 
         return out
+

comfy/ldm/kandinsky5/model.py

@@ -0,0 +1,413 @@
+import torch
+from torch import nn
+import math
+
+import comfy.ldm.common_dit
+from comfy.ldm.modules.attention import optimized_attention
+from comfy.ldm.flux.math import apply_rope1
+from comfy.ldm.flux.layers import EmbedND
+
+def attention(q, k, v, heads, transformer_options={}):
+    return optimized_attention(
+        q.transpose(1, 2),
+        k.transpose(1, 2),
+        v.transpose(1, 2),
+        heads=heads,
+        skip_reshape=True,
+        transformer_options=transformer_options
+    )
+
+def apply_scale_shift_norm(norm, x, scale, shift):
+    return torch.addcmul(shift, norm(x), scale + 1.0)
+
+def apply_gate_sum(x, out, gate):
+    return torch.addcmul(x, gate, out)
+
+def get_shift_scale_gate(params):
+    shift, scale, gate = torch.chunk(params, 3, dim=-1)
+    return tuple(x.unsqueeze(1) for x in (shift, scale, gate))
+
+def get_freqs(dim, max_period=10000.0):
+    return torch.exp(-math.log(max_period) * torch.arange(start=0, end=dim, dtype=torch.float32) / dim)
+
+
+class TimeEmbeddings(nn.Module):
+    def __init__(self, model_dim, time_dim, max_period=10000.0, operation_settings=None):
+        super().__init__()
+        assert model_dim % 2 == 0
+        self.model_dim = model_dim
+        self.max_period = max_period
+        self.register_buffer("freqs", get_freqs(model_dim // 2, max_period), persistent=False)
+        operations = operation_settings.get("operations")
+        self.in_layer = operations.Linear(model_dim, time_dim, bias=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.activation = nn.SiLU()
+        self.out_layer = operations.Linear(time_dim, time_dim, bias=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+
+    def forward(self, timestep, dtype):
+        args = torch.outer(timestep, self.freqs.to(device=timestep.device))
+        time_embed = torch.cat([torch.cos(args), torch.sin(args)], dim=-1).to(dtype)
+        time_embed = self.out_layer(self.activation(self.in_layer(time_embed)))
+        return time_embed
+
+
+class TextEmbeddings(nn.Module):
+    def __init__(self, text_dim, model_dim, operation_settings=None):
+        super().__init__()
+        operations = operation_settings.get("operations")
+        self.in_layer = operations.Linear(text_dim, model_dim, bias=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.norm = operations.LayerNorm(model_dim, elementwise_affine=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+
+    def forward(self, text_embed):
+        text_embed = self.in_layer(text_embed)
+        return self.norm(text_embed).type_as(text_embed)
+
+
+class VisualEmbeddings(nn.Module):
+    def __init__(self, visual_dim, model_dim, patch_size, operation_settings=None):
+        super().__init__()
+        self.patch_size = patch_size
+        operations = operation_settings.get("operations")
+        self.in_layer = operations.Linear(visual_dim, model_dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+
+    def forward(self, x):
+        x = x.movedim(1, -1)  # B C T H W -> B T H W C
+        B, T, H, W, dim = x.shape
+        pt, ph, pw = self.patch_size
+
+        x = x.view(
+            B,
+            T // pt, pt,
+            H // ph, ph,
+            W // pw, pw,
+            dim,
+        ).permute(0, 1, 3, 5, 2, 4, 6, 7).flatten(4, 7)
+
+        return self.in_layer(x)
+
+
+class Modulation(nn.Module):
+    def __init__(self, time_dim, model_dim, num_params, operation_settings=None):
+        super().__init__()
+        self.activation = nn.SiLU()
+        self.out_layer = operation_settings.get("operations").Linear(time_dim, num_params * model_dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+
+    def forward(self, x):
+        return self.out_layer(self.activation(x))
+
+
+class SelfAttention(nn.Module):
+    def __init__(self, num_channels, head_dim, operation_settings=None):
+        super().__init__()
+        assert num_channels % head_dim == 0
+        self.num_heads = num_channels // head_dim
+        self.head_dim = head_dim
+
+        operations = operation_settings.get("operations")
+        self.to_query = operations.Linear(num_channels, num_channels, bias=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.to_key = operations.Linear(num_channels, num_channels, bias=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.to_value = operations.Linear(num_channels, num_channels, bias=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.query_norm = operations.RMSNorm(head_dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.key_norm = operations.RMSNorm(head_dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+
+        self.out_layer = operations.Linear(num_channels, num_channels, bias=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.num_chunks = 2
+
+    def _compute_qk(self, x, freqs, proj_fn, norm_fn):
+        result = proj_fn(x).view(*x.shape[:-1], self.num_heads, -1)
+        return apply_rope1(norm_fn(result), freqs)
+
+    def _forward(self, x, freqs, transformer_options={}):
+        q = self._compute_qk(x, freqs, self.to_query, self.query_norm)
+        k = self._compute_qk(x, freqs, self.to_key, self.key_norm)
+        v = self.to_value(x).view(*x.shape[:-1], self.num_heads, -1)
+        out = attention(q, k, v, self.num_heads, transformer_options=transformer_options)
+        return self.out_layer(out)
+
+    def _forward_chunked(self, x, freqs, transformer_options={}):
+        def process_chunks(proj_fn, norm_fn):
+            x_chunks = torch.chunk(x, self.num_chunks, dim=1)
+            freqs_chunks = torch.chunk(freqs, self.num_chunks, dim=1)
+            chunks = []
+            for x_chunk, freqs_chunk in zip(x_chunks, freqs_chunks):
+                chunks.append(self._compute_qk(x_chunk, freqs_chunk, proj_fn, norm_fn))
+            return torch.cat(chunks, dim=1)
+
+        q = process_chunks(self.to_query, self.query_norm)
+        k = process_chunks(self.to_key, self.key_norm)
+        v = self.to_value(x).view(*x.shape[:-1], self.num_heads, -1)
+        out = attention(q, k, v, self.num_heads, transformer_options=transformer_options)
+        return self.out_layer(out)
+
+    def forward(self, x, freqs, transformer_options={}):
+        if x.shape[1] > 8192:
+            return self._forward_chunked(x, freqs, transformer_options=transformer_options)
+        else:
+            return self._forward(x, freqs, transformer_options=transformer_options)
+
+
+class CrossAttention(SelfAttention):
+    def get_qkv(self, x, context):
+        q = self.to_query(x).view(*x.shape[:-1], self.num_heads, -1)
+        k = self.to_key(context).view(*context.shape[:-1], self.num_heads, -1)
+        v = self.to_value(context).view(*context.shape[:-1], self.num_heads, -1)
+        return q, k, v
+
+    def forward(self, x, context, transformer_options={}):
+        q, k, v = self.get_qkv(x, context)
+        out = attention(self.query_norm(q), self.key_norm(k), v, self.num_heads, transformer_options=transformer_options)
+        return self.out_layer(out)
+
+
+class FeedForward(nn.Module):
+    def __init__(self, dim, ff_dim, operation_settings=None):
+        super().__init__()
+        operations = operation_settings.get("operations")
+        self.in_layer = operations.Linear(dim, ff_dim, bias=False, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.activation = nn.GELU()
+        self.out_layer = operations.Linear(ff_dim, dim, bias=False, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.num_chunks = 4
+
+    def _forward(self, x):
+        return self.out_layer(self.activation(self.in_layer(x)))
+
+    def _forward_chunked(self, x):
+        chunks = torch.chunk(x, self.num_chunks, dim=1)
+        output_chunks = []
+        for chunk in chunks:
+            output_chunks.append(self._forward(chunk))
+        return torch.cat(output_chunks, dim=1)
+
+    def forward(self, x):
+        if x.shape[1] > 8192:
+            return self._forward_chunked(x)
+        else:
+            return self._forward(x)
+
+
+class OutLayer(nn.Module):
+    def __init__(self, model_dim, time_dim, visual_dim, patch_size, operation_settings=None):
+        super().__init__()
+        self.patch_size = patch_size
+        self.modulation = Modulation(time_dim, model_dim, 2, operation_settings=operation_settings)
+        operations = operation_settings.get("operations")
+        self.norm = operations.LayerNorm(model_dim, elementwise_affine=False, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.out_layer = operations.Linear(model_dim, math.prod(patch_size) * visual_dim, bias=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+
+    def forward(self, visual_embed, time_embed):
+        B, T, H, W, _ = visual_embed.shape
+        shift, scale = torch.chunk(self.modulation(time_embed), 2, dim=-1)
+        scale = scale[:, None, None, None, :]
+        shift = shift[:, None, None, None, :]
+        visual_embed = apply_scale_shift_norm(self.norm, visual_embed, scale, shift)
+        x = self.out_layer(visual_embed)
+
+        out_dim = x.shape[-1] // (self.patch_size[0] * self.patch_size[1] * self.patch_size[2])
+        x = x.view(
+            B, T, H, W,
+            out_dim,
+            self.patch_size[0], self.patch_size[1], self.patch_size[2]
+        )
+        return x.permute(0, 4, 1, 5, 2, 6, 3, 7).flatten(2, 3).flatten(3, 4).flatten(4, 5)
+
+
+class TransformerEncoderBlock(nn.Module):
+    def __init__(self, model_dim, time_dim, ff_dim, head_dim, operation_settings=None):
+        super().__init__()
+        self.text_modulation = Modulation(time_dim, model_dim, 6, operation_settings=operation_settings)
+        operations = operation_settings.get("operations")
+
+        self.self_attention_norm = operations.LayerNorm(model_dim, elementwise_affine=False, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.self_attention = SelfAttention(model_dim, head_dim, operation_settings=operation_settings)
+
+        self.feed_forward_norm = operations.LayerNorm(model_dim, elementwise_affine=False, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.feed_forward = FeedForward(model_dim, ff_dim, operation_settings=operation_settings)
+
+    def forward(self, x, time_embed, freqs, transformer_options={}):
+        self_attn_params, ff_params = torch.chunk(self.text_modulation(time_embed), 2, dim=-1)
+        shift, scale, gate = get_shift_scale_gate(self_attn_params)
+        out = apply_scale_shift_norm(self.self_attention_norm, x, scale, shift)
+        out = self.self_attention(out, freqs, transformer_options=transformer_options)
+        x = apply_gate_sum(x, out, gate)
+
+        shift, scale, gate = get_shift_scale_gate(ff_params)
+        out = apply_scale_shift_norm(self.feed_forward_norm, x, scale, shift)
+        out = self.feed_forward(out)
+        x = apply_gate_sum(x, out, gate)
+        return x
+
+
+class TransformerDecoderBlock(nn.Module):
+    def __init__(self, model_dim, time_dim, ff_dim, head_dim, operation_settings=None):
+        super().__init__()
+        self.visual_modulation = Modulation(time_dim, model_dim, 9, operation_settings=operation_settings)
+
+        operations = operation_settings.get("operations")
+        self.self_attention_norm = operations.LayerNorm(model_dim, elementwise_affine=False, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.self_attention = SelfAttention(model_dim, head_dim, operation_settings=operation_settings)
+
+        self.cross_attention_norm = operations.LayerNorm(model_dim, elementwise_affine=False, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.cross_attention = CrossAttention(model_dim, head_dim, operation_settings=operation_settings)
+
+        self.feed_forward_norm = operations.LayerNorm(model_dim, elementwise_affine=False, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.feed_forward = FeedForward(model_dim, ff_dim, operation_settings=operation_settings)
+
+    def forward(self, visual_embed, text_embed, time_embed, freqs, transformer_options={}):
+        self_attn_params, cross_attn_params, ff_params = torch.chunk(self.visual_modulation(time_embed), 3, dim=-1)
+        # self attention
+        shift, scale, gate = get_shift_scale_gate(self_attn_params)
+        visual_out = apply_scale_shift_norm(self.self_attention_norm, visual_embed, scale, shift)
+        visual_out = self.self_attention(visual_out, freqs, transformer_options=transformer_options)
+        visual_embed = apply_gate_sum(visual_embed, visual_out, gate)
+        # cross attention
+        shift, scale, gate = get_shift_scale_gate(cross_attn_params)
+        visual_out = apply_scale_shift_norm(self.cross_attention_norm, visual_embed, scale, shift)
+        visual_out = self.cross_attention(visual_out, text_embed, transformer_options=transformer_options)
+        visual_embed = apply_gate_sum(visual_embed, visual_out, gate)
+        # feed forward
+        shift, scale, gate = get_shift_scale_gate(ff_params)
+        visual_out = apply_scale_shift_norm(self.feed_forward_norm, visual_embed, scale, shift)
+        visual_out = self.feed_forward(visual_out)
+        visual_embed = apply_gate_sum(visual_embed, visual_out, gate)
+        return visual_embed
+
+
+class Kandinsky5(nn.Module):
+    def __init__(
+        self,
+        in_visual_dim=16, out_visual_dim=16, in_text_dim=3584, in_text_dim2=768, time_dim=512,
+        model_dim=1792, ff_dim=7168, visual_embed_dim=132, patch_size=(1, 2, 2), num_text_blocks=2, num_visual_blocks=32,
+        axes_dims=(16, 24, 24), rope_scale_factor=(1.0, 2.0, 2.0),
+        dtype=None, device=None, operations=None, **kwargs
+    ):
+        super().__init__()
+        head_dim = sum(axes_dims)
+        self.rope_scale_factor = rope_scale_factor
+        self.in_visual_dim = in_visual_dim
+        self.model_dim = model_dim
+        self.patch_size = patch_size
+        self.visual_embed_dim = visual_embed_dim
+        self.dtype = dtype
+        self.device = device
+        operation_settings = {"operations": operations, "device": device, "dtype": dtype}
+
+        self.time_embeddings = TimeEmbeddings(model_dim, time_dim, operation_settings=operation_settings)
+        self.text_embeddings = TextEmbeddings(in_text_dim, model_dim, operation_settings=operation_settings)
+        self.pooled_text_embeddings = TextEmbeddings(in_text_dim2, time_dim, operation_settings=operation_settings)
+        self.visual_embeddings = VisualEmbeddings(visual_embed_dim, model_dim, patch_size, operation_settings=operation_settings)
+
+        self.text_transformer_blocks = nn.ModuleList(
+            [TransformerEncoderBlock(model_dim, time_dim, ff_dim, head_dim, operation_settings=operation_settings) for _ in range(num_text_blocks)]
+        )
+
+        self.visual_transformer_blocks = nn.ModuleList(
+            [TransformerDecoderBlock(model_dim, time_dim, ff_dim, head_dim, operation_settings=operation_settings) for _ in range(num_visual_blocks)]
+        )
+
+        self.out_layer = OutLayer(model_dim, time_dim, out_visual_dim, patch_size, operation_settings=operation_settings)
+
+        self.rope_embedder_3d = EmbedND(dim=head_dim, theta=10000.0, axes_dim=axes_dims)
+        self.rope_embedder_1d = EmbedND(dim=head_dim, theta=10000.0, axes_dim=[head_dim])
+
+    def rope_encode_1d(self, seq_len, seq_start=0, steps=None, device=None, dtype=None, transformer_options={}):
+        steps = seq_len if steps is None else steps
+        seq_ids = torch.linspace(seq_start, seq_start + (seq_len - 1), steps=steps, device=device, dtype=dtype)
+        seq_ids = seq_ids.reshape(-1, 1).unsqueeze(0)  # Shape: (1, steps, 1)
+        freqs = self.rope_embedder_1d(seq_ids).movedim(1, 2)
+        return freqs
+
+    def rope_encode_3d(self, t, h, w, t_start=0, steps_t=None, steps_h=None, steps_w=None, device=None, dtype=None, transformer_options={}):
+
+        patch_size = self.patch_size
+        t_len = ((t + (patch_size[0] // 2)) // patch_size[0])
+        h_len = ((h + (patch_size[1] // 2)) // patch_size[1])
+        w_len = ((w + (patch_size[2] // 2)) // patch_size[2])
+
+        if steps_t is None:
+            steps_t = t_len
+        if steps_h is None:
+            steps_h = h_len
+        if steps_w is None:
+            steps_w = w_len
+
+        h_start = 0
+        w_start = 0
+        rope_options = transformer_options.get("rope_options", None)
+        if rope_options is not None:
+            t_len = (t_len - 1.0) * rope_options.get("scale_t", 1.0) + 1.0
+            h_len = (h_len - 1.0) * rope_options.get("scale_y", 1.0) + 1.0
+            w_len = (w_len - 1.0) * rope_options.get("scale_x", 1.0) + 1.0
+
+            t_start += rope_options.get("shift_t", 0.0)
+            h_start += rope_options.get("shift_y", 0.0)
+            w_start += rope_options.get("shift_x", 0.0)
+        else:
+            rope_scale_factor = self.rope_scale_factor
+            if self.model_dim == 4096: # pro video model uses different rope scaling at higher resolutions
+                if h * w >= 14080:
+                    rope_scale_factor = (1.0, 3.16, 3.16)
+
+            t_len = (t_len - 1.0) / rope_scale_factor[0] + 1.0
+            h_len = (h_len - 1.0) / rope_scale_factor[1] + 1.0
+            w_len = (w_len - 1.0) / rope_scale_factor[2] + 1.0
+
+        img_ids = torch.zeros((steps_t, steps_h, steps_w, 3), device=device, dtype=dtype)
+        img_ids[:, :, :, 0] = img_ids[:, :, :, 0] + torch.linspace(t_start, t_start + (t_len - 1), steps=steps_t, device=device, dtype=dtype).reshape(-1, 1, 1)
+        img_ids[:, :, :, 1] = img_ids[:, :, :, 1] + torch.linspace(h_start, h_start + (h_len - 1), steps=steps_h, device=device, dtype=dtype).reshape(1, -1, 1)
+        img_ids[:, :, :, 2] = img_ids[:, :, :, 2] + torch.linspace(w_start, w_start + (w_len - 1), steps=steps_w, device=device, dtype=dtype).reshape(1, 1, -1)
+        img_ids = img_ids.reshape(1, -1, img_ids.shape[-1])
+
+        freqs = self.rope_embedder_3d(img_ids).movedim(1, 2)
+        return freqs
+
+    def forward_orig(self, x, timestep, context, y, freqs, freqs_text, transformer_options={}, **kwargs):
+        patches_replace = transformer_options.get("patches_replace", {})
+        context = self.text_embeddings(context)
+        time_embed = self.time_embeddings(timestep, x.dtype) + self.pooled_text_embeddings(y)
+
+        for block in self.text_transformer_blocks:
+            context = block(context, time_embed, freqs_text, transformer_options=transformer_options)
+
+        visual_embed = self.visual_embeddings(x)
+        visual_shape = visual_embed.shape[:-1]
+        visual_embed = visual_embed.flatten(1, -2)
+
+        blocks_replace = patches_replace.get("dit", {})
+        transformer_options["total_blocks"] = len(self.visual_transformer_blocks)
+        transformer_options["block_type"] = "double"
+        for i, block in enumerate(self.visual_transformer_blocks):
+            transformer_options["block_index"] = i
+            if ("double_block", i) in blocks_replace:
+                def block_wrap(args):
+                    return block(x=args["x"], context=args["context"], time_embed=args["time_embed"], freqs=args["freqs"], transformer_options=args.get("transformer_options"))
+                visual_embed = blocks_replace[("double_block", i)]({"x": visual_embed, "context": context, "time_embed": time_embed, "freqs": freqs, "transformer_options": transformer_options}, {"original_block": block_wrap})["x"]
+            else:
+                visual_embed = block(visual_embed, context, time_embed, freqs=freqs, transformer_options=transformer_options)
+
+        visual_embed = visual_embed.reshape(*visual_shape, -1)
+        return self.out_layer(visual_embed, time_embed)
+
+    def _forward(self, x, timestep, context, y, time_dim_replace=None, transformer_options={}, **kwargs):
+        original_dims = x.ndim
+        if original_dims == 4:
+            x = x.unsqueeze(2)
+        bs, c, t_len, h, w = x.shape
+        x = comfy.ldm.common_dit.pad_to_patch_size(x, self.patch_size)
+
+        if time_dim_replace is not None:
+            time_dim_replace = comfy.ldm.common_dit.pad_to_patch_size(time_dim_replace, self.patch_size)
+            x[:, :time_dim_replace.shape[1], :time_dim_replace.shape[2]] = time_dim_replace
+
+        freqs = self.rope_encode_3d(t_len, h, w, device=x.device, dtype=x.dtype, transformer_options=transformer_options)
+        freqs_text = self.rope_encode_1d(context.shape[1], device=x.device, dtype=x.dtype, transformer_options=transformer_options)
+
+        out = self.forward_orig(x, timestep, context, y, freqs, freqs_text, transformer_options=transformer_options, **kwargs)
+        if original_dims == 4:
+            out = out.squeeze(2)
+        return out
+
+    def forward(self, x, timestep, context, y, time_dim_replace=None, transformer_options={}, **kwargs):
+        return comfy.patcher_extension.WrapperExecutor.new_class_executor(
+            self._forward,
+            self,
+            comfy.patcher_extension.get_all_wrappers(comfy.patcher_extension.WrappersMP.DIFFUSION_MODEL, transformer_options)
+        ).execute(x, timestep, context, y, time_dim_replace=time_dim_replace, transformer_options=transformer_options, **kwargs)

comfy/ldm/lightricks/av_model.py

@@ -0,0 +1,871 @@
+from typing import Tuple
+import torch
+import torch.nn as nn
+from comfy.ldm.lightricks.model import (
+    CrossAttention,
+    FeedForward,
+    AdaLayerNormSingle,
+    PixArtAlphaTextProjection,
+    LTXVModel,
+)
+from comfy.ldm.lightricks.symmetric_patchifier import AudioPatchifier
+import comfy.ldm.common_dit
+
+class CompressedTimestep:
+    """Store video timestep embeddings in compressed form using per-frame indexing."""
+    __slots__ = ('data', 'batch_size', 'num_frames', 'patches_per_frame', 'feature_dim')
+
+    def __init__(self, tensor: torch.Tensor, patches_per_frame: int):
+        """
+        tensor: [batch_size, num_tokens, feature_dim] tensor where num_tokens = num_frames * patches_per_frame
+        patches_per_frame: Number of spatial patches per frame (height * width in latent space), or None to disable compression
+        """
+        self.batch_size, num_tokens, self.feature_dim = tensor.shape
+
+        # Check if compression is valid (num_tokens must be divisible by patches_per_frame)
+        if patches_per_frame is not None and num_tokens % patches_per_frame == 0 and num_tokens >= patches_per_frame:
+            self.patches_per_frame = patches_per_frame
+            self.num_frames = num_tokens // patches_per_frame
+
+            # Reshape to [batch, frames, patches_per_frame, feature_dim] and store one value per frame
+            # All patches in a frame are identical, so we only keep the first one
+            reshaped = tensor.view(self.batch_size, self.num_frames, patches_per_frame, self.feature_dim)
+            self.data = reshaped[:, :, 0, :].contiguous()  # [batch, frames, feature_dim]
+        else:
+            # Not divisible or too small - store directly without compression
+            self.patches_per_frame = 1
+            self.num_frames = num_tokens
+            self.data = tensor
+
+    def expand(self):
+        """Expand back to original tensor."""
+        if self.patches_per_frame == 1:
+            return self.data
+
+        # [batch, frames, feature_dim] -> [batch, frames, patches_per_frame, feature_dim] -> [batch, tokens, feature_dim]
+        expanded = self.data.unsqueeze(2).expand(self.batch_size, self.num_frames, self.patches_per_frame, self.feature_dim)
+        return expanded.reshape(self.batch_size, -1, self.feature_dim)
+
+    def expand_for_computation(self, scale_shift_table: torch.Tensor, batch_size: int, indices: slice = slice(None, None)):
+        """Compute ada values on compressed per-frame data, then expand spatially."""
+        num_ada_params = scale_shift_table.shape[0]
+
+        # No compression - compute directly
+        if self.patches_per_frame == 1:
+            num_tokens = self.data.shape[1]
+            dim_per_param = self.feature_dim // num_ada_params
+            reshaped = self.data.reshape(batch_size, num_tokens, num_ada_params, dim_per_param)[:, :, indices, :]
+            table_values = scale_shift_table[indices].unsqueeze(0).unsqueeze(0).to(device=self.data.device, dtype=self.data.dtype)
+            ada_values = (table_values + reshaped).unbind(dim=2)
+            return ada_values
+
+        # Compressed: compute on per-frame data then expand spatially
+        # Reshape: [batch, frames, feature_dim] -> [batch, frames, num_ada_params, dim_per_param]
+        frame_reshaped = self.data.reshape(batch_size, self.num_frames, num_ada_params, -1)[:, :, indices, :]
+        table_values = scale_shift_table[indices].unsqueeze(0).unsqueeze(0).to(
+            device=self.data.device, dtype=self.data.dtype
+        )
+        frame_ada = (table_values + frame_reshaped).unbind(dim=2)
+
+        # Expand each ada parameter spatially: [batch, frames, dim] -> [batch, frames, patches, dim] -> [batch, tokens, dim]
+        return tuple(
+            frame_val.unsqueeze(2).expand(batch_size, self.num_frames, self.patches_per_frame, -1)
+            .reshape(batch_size, -1, frame_val.shape[-1])
+            for frame_val in frame_ada
+        )
+
+class BasicAVTransformerBlock(nn.Module):
+    def __init__(
+        self,
+        v_dim,
+        a_dim,
+        v_heads,
+        a_heads,
+        vd_head,
+        ad_head,
+        v_context_dim=None,
+        a_context_dim=None,
+        attn_precision=None,
+        dtype=None,
+        device=None,
+        operations=None,
+    ):
+        super().__init__()
+
+        self.attn_precision = attn_precision
+
+        self.attn1 = CrossAttention(
+            query_dim=v_dim,
+            heads=v_heads,
+            dim_head=vd_head,
+            context_dim=None,
+            attn_precision=self.attn_precision,
+            dtype=dtype,
+            device=device,
+            operations=operations,
+        )
+        self.audio_attn1 = CrossAttention(
+            query_dim=a_dim,
+            heads=a_heads,
+            dim_head=ad_head,
+            context_dim=None,
+            attn_precision=self.attn_precision,
+            dtype=dtype,
+            device=device,
+            operations=operations,
+        )
+
+        self.attn2 = CrossAttention(
+            query_dim=v_dim,
+            context_dim=v_context_dim,
+            heads=v_heads,
+            dim_head=vd_head,
+            attn_precision=self.attn_precision,
+            dtype=dtype,
+            device=device,
+            operations=operations,
+        )
+        self.audio_attn2 = CrossAttention(
+            query_dim=a_dim,
+            context_dim=a_context_dim,
+            heads=a_heads,
+            dim_head=ad_head,
+            attn_precision=self.attn_precision,
+            dtype=dtype,
+            device=device,
+            operations=operations,
+        )
+
+        # Q: Video, K,V: Audio
+        self.audio_to_video_attn = CrossAttention(
+            query_dim=v_dim,
+            context_dim=a_dim,
+            heads=a_heads,
+            dim_head=ad_head,
+            attn_precision=self.attn_precision,
+            dtype=dtype,
+            device=device,
+            operations=operations,
+        )
+
+        # Q: Audio, K,V: Video
+        self.video_to_audio_attn = CrossAttention(
+            query_dim=a_dim,
+            context_dim=v_dim,
+            heads=a_heads,
+            dim_head=ad_head,
+            attn_precision=self.attn_precision,
+            dtype=dtype,
+            device=device,
+            operations=operations,
+        )
+
+        self.ff = FeedForward(
+            v_dim, dim_out=v_dim, glu=True, dtype=dtype, device=device, operations=operations
+        )
+        self.audio_ff = FeedForward(
+            a_dim, dim_out=a_dim, glu=True, dtype=dtype, device=device, operations=operations
+        )
+
+        self.scale_shift_table = nn.Parameter(torch.empty(6, v_dim, device=device, dtype=dtype))
+        self.audio_scale_shift_table = nn.Parameter(
+            torch.empty(6, a_dim, device=device, dtype=dtype)
+        )
+
+        self.scale_shift_table_a2v_ca_audio = nn.Parameter(
+            torch.empty(5, a_dim, device=device, dtype=dtype)
+        )
+        self.scale_shift_table_a2v_ca_video = nn.Parameter(
+            torch.empty(5, v_dim, device=device, dtype=dtype)
+        )
+
+    def get_ada_values(
+        self, scale_shift_table: torch.Tensor, batch_size: int, timestep: torch.Tensor, indices: slice = slice(None, None)
+    ):
+        if isinstance(timestep, CompressedTimestep):
+            return timestep.expand_for_computation(scale_shift_table, batch_size, indices)
+
+        num_ada_params = scale_shift_table.shape[0]
+
+        ada_values = (
+            scale_shift_table[indices].unsqueeze(0).unsqueeze(0).to(device=timestep.device, dtype=timestep.dtype)
+            + timestep.reshape(batch_size, timestep.shape[1], num_ada_params, -1)[:, :, indices, :]
+        ).unbind(dim=2)
+        return ada_values
+
+    def get_av_ca_ada_values(
+        self,
+        scale_shift_table: torch.Tensor,
+        batch_size: int,
+        scale_shift_timestep: torch.Tensor,
+        gate_timestep: torch.Tensor,
+        num_scale_shift_values: int = 4,
+    ):
+        scale_shift_ada_values = self.get_ada_values(
+            scale_shift_table[:num_scale_shift_values, :],
+            batch_size,
+            scale_shift_timestep,
+        )
+        gate_ada_values = self.get_ada_values(
+            scale_shift_table[num_scale_shift_values:, :],
+            batch_size,
+            gate_timestep,
+        )
+
+        return (*scale_shift_ada_values, *gate_ada_values)
+
+    def forward(
+        self, x: Tuple[torch.Tensor, torch.Tensor], v_context=None, a_context=None, attention_mask=None, v_timestep=None, a_timestep=None,
+        v_pe=None, a_pe=None, v_cross_pe=None, a_cross_pe=None, v_cross_scale_shift_timestep=None, a_cross_scale_shift_timestep=None,
+        v_cross_gate_timestep=None, a_cross_gate_timestep=None, transformer_options=None,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        run_vx = transformer_options.get("run_vx", True)
+        run_ax = transformer_options.get("run_ax", True)
+
+        vx, ax = x
+        run_ax = run_ax and ax.numel() > 0
+        run_a2v = run_vx and transformer_options.get("a2v_cross_attn", True) and ax.numel() > 0
+        run_v2a = run_ax and transformer_options.get("v2a_cross_attn", True)
+
+        # video
+        if run_vx:
+            # video self-attention
+            vshift_msa, vscale_msa = (self.get_ada_values(self.scale_shift_table, vx.shape[0], v_timestep, slice(0, 2)))
+            norm_vx = comfy.ldm.common_dit.rms_norm(vx) * (1 + vscale_msa) + vshift_msa
+            del vshift_msa, vscale_msa
+            attn1_out = self.attn1(norm_vx, pe=v_pe, transformer_options=transformer_options)
+            del norm_vx
+            # video cross-attention
+            vgate_msa = self.get_ada_values(self.scale_shift_table, vx.shape[0], v_timestep, slice(2, 3))[0]
+            vx.addcmul_(attn1_out, vgate_msa)
+            del vgate_msa, attn1_out
+            vx.add_(self.attn2(comfy.ldm.common_dit.rms_norm(vx), context=v_context, mask=attention_mask, transformer_options=transformer_options))
+
+        # audio
+        if run_ax:
+            # audio self-attention
+            ashift_msa, ascale_msa = (self.get_ada_values(self.audio_scale_shift_table, ax.shape[0], a_timestep, slice(0, 2)))
+            norm_ax = comfy.ldm.common_dit.rms_norm(ax) * (1 + ascale_msa) + ashift_msa
+            del ashift_msa, ascale_msa
+            attn1_out = self.audio_attn1(norm_ax, pe=a_pe, transformer_options=transformer_options)
+            del norm_ax
+            # audio cross-attention
+            agate_msa = self.get_ada_values(self.audio_scale_shift_table, ax.shape[0], a_timestep, slice(2, 3))[0]
+            ax.addcmul_(attn1_out, agate_msa)
+            del agate_msa, attn1_out
+            ax.add_(self.audio_attn2(comfy.ldm.common_dit.rms_norm(ax), context=a_context, mask=attention_mask, transformer_options=transformer_options))
+
+        # video - audio cross attention.
+        if run_a2v or run_v2a:
+            vx_norm3 = comfy.ldm.common_dit.rms_norm(vx)
+            ax_norm3 = comfy.ldm.common_dit.rms_norm(ax)
+
+            # audio to video cross attention
+            if run_a2v:
+                scale_ca_audio_hidden_states_a2v, shift_ca_audio_hidden_states_a2v = self.get_ada_values(
+                    self.scale_shift_table_a2v_ca_audio[:4, :], ax.shape[0], a_cross_scale_shift_timestep)[:2]
+                scale_ca_video_hidden_states_a2v_v, shift_ca_video_hidden_states_a2v_v = self.get_ada_values(
+                    self.scale_shift_table_a2v_ca_video[:4, :], vx.shape[0], v_cross_scale_shift_timestep)[:2]
+
+                vx_scaled = vx_norm3 * (1 + scale_ca_video_hidden_states_a2v_v) + shift_ca_video_hidden_states_a2v_v
+                ax_scaled = ax_norm3 * (1 + scale_ca_audio_hidden_states_a2v) + shift_ca_audio_hidden_states_a2v
+                del scale_ca_video_hidden_states_a2v_v, shift_ca_video_hidden_states_a2v_v, scale_ca_audio_hidden_states_a2v, shift_ca_audio_hidden_states_a2v
+
+                a2v_out = self.audio_to_video_attn(vx_scaled, context=ax_scaled, pe=v_cross_pe, k_pe=a_cross_pe, transformer_options=transformer_options)
+                del vx_scaled, ax_scaled
+
+                gate_out_a2v = self.get_ada_values(self.scale_shift_table_a2v_ca_video[4:, :], vx.shape[0], v_cross_gate_timestep)[0]
+                vx.addcmul_(a2v_out, gate_out_a2v)
+                del gate_out_a2v, a2v_out
+
+            # video to audio cross attention
+            if run_v2a:
+                scale_ca_audio_hidden_states_v2a, shift_ca_audio_hidden_states_v2a = self.get_ada_values(
+                    self.scale_shift_table_a2v_ca_audio[:4, :], ax.shape[0], a_cross_scale_shift_timestep)[2:4]
+                scale_ca_video_hidden_states_v2a, shift_ca_video_hidden_states_v2a = self.get_ada_values(
+                    self.scale_shift_table_a2v_ca_video[:4, :], vx.shape[0], v_cross_scale_shift_timestep)[2:4]
+
+                ax_scaled = ax_norm3 * (1 + scale_ca_audio_hidden_states_v2a) + shift_ca_audio_hidden_states_v2a
+                vx_scaled = vx_norm3 * (1 + scale_ca_video_hidden_states_v2a) + shift_ca_video_hidden_states_v2a
+                del scale_ca_video_hidden_states_v2a, shift_ca_video_hidden_states_v2a, scale_ca_audio_hidden_states_v2a, shift_ca_audio_hidden_states_v2a
+
+                v2a_out = self.video_to_audio_attn(ax_scaled, context=vx_scaled, pe=a_cross_pe, k_pe=v_cross_pe, transformer_options=transformer_options)
+                del ax_scaled, vx_scaled
+
+                gate_out_v2a = self.get_ada_values(self.scale_shift_table_a2v_ca_audio[4:, :], ax.shape[0], a_cross_gate_timestep)[0]
+                ax.addcmul_(v2a_out, gate_out_v2a)
+                del gate_out_v2a, v2a_out
+
+            del vx_norm3, ax_norm3
+
+        # video feedforward
+        if run_vx:
+            vshift_mlp, vscale_mlp = self.get_ada_values(self.scale_shift_table, vx.shape[0], v_timestep, slice(3, 5))
+            vx_scaled = comfy.ldm.common_dit.rms_norm(vx) * (1 + vscale_mlp) + vshift_mlp
+            del vshift_mlp, vscale_mlp
+
+            ff_out = self.ff(vx_scaled)
+            del vx_scaled
+
+            vgate_mlp = self.get_ada_values(self.scale_shift_table, vx.shape[0], v_timestep, slice(5, 6))[0]
+            vx.addcmul_(ff_out, vgate_mlp)
+            del vgate_mlp, ff_out
+
+        # audio feedforward
+        if run_ax:
+            ashift_mlp, ascale_mlp = self.get_ada_values(self.audio_scale_shift_table, ax.shape[0], a_timestep, slice(3, 5))
+            ax_scaled = comfy.ldm.common_dit.rms_norm(ax) * (1 + ascale_mlp) + ashift_mlp
+            del ashift_mlp, ascale_mlp
+
+            ff_out = self.audio_ff(ax_scaled)
+            del ax_scaled
+
+            agate_mlp = self.get_ada_values(self.audio_scale_shift_table, ax.shape[0], a_timestep, slice(5, 6))[0]
+            ax.addcmul_(ff_out, agate_mlp)
+            del agate_mlp, ff_out
+
+        return vx, ax
+
+
+class LTXAVModel(LTXVModel):
+    """LTXAV model for audio-video generation."""
+
+    def __init__(
+        self,
+        in_channels=128,
+        audio_in_channels=128,
+        cross_attention_dim=4096,
+        audio_cross_attention_dim=2048,
+        attention_head_dim=128,
+        audio_attention_head_dim=64,
+        num_attention_heads=32,
+        audio_num_attention_heads=32,
+        caption_channels=3840,
+        num_layers=48,
+        positional_embedding_theta=10000.0,
+        positional_embedding_max_pos=[20, 2048, 2048],
+        audio_positional_embedding_max_pos=[20],
+        causal_temporal_positioning=False,
+        vae_scale_factors=(8, 32, 32),
+        use_middle_indices_grid=False,
+        timestep_scale_multiplier=1000.0,
+        av_ca_timestep_scale_multiplier=1.0,
+        dtype=None,
+        device=None,
+        operations=None,
+        **kwargs,
+    ):
+        # Store audio-specific parameters
+        self.audio_in_channels = audio_in_channels
+        self.audio_cross_attention_dim = audio_cross_attention_dim
+        self.audio_attention_head_dim = audio_attention_head_dim
+        self.audio_num_attention_heads = audio_num_attention_heads
+        self.audio_positional_embedding_max_pos = audio_positional_embedding_max_pos
+
+        # Calculate audio dimensions
+        self.audio_inner_dim = audio_num_attention_heads * audio_attention_head_dim
+        self.audio_out_channels = audio_in_channels
+
+        # Audio-specific constants
+        self.num_audio_channels = 8
+        self.audio_frequency_bins = 16
+
+        self.av_ca_timestep_scale_multiplier = av_ca_timestep_scale_multiplier
+
+        super().__init__(
+            in_channels=in_channels,
+            cross_attention_dim=cross_attention_dim,
+            attention_head_dim=attention_head_dim,
+            num_attention_heads=num_attention_heads,
+            caption_channels=caption_channels,
+            num_layers=num_layers,
+            positional_embedding_theta=positional_embedding_theta,
+            positional_embedding_max_pos=positional_embedding_max_pos,
+            causal_temporal_positioning=causal_temporal_positioning,
+            vae_scale_factors=vae_scale_factors,
+            use_middle_indices_grid=use_middle_indices_grid,
+            timestep_scale_multiplier=timestep_scale_multiplier,
+            dtype=dtype,
+            device=device,
+            operations=operations,
+            **kwargs,
+        )
+
+    def _init_model_components(self, device, dtype, **kwargs):
+        """Initialize LTXAV-specific components."""
+        # Audio-specific projections
+        self.audio_patchify_proj = self.operations.Linear(
+            self.audio_in_channels, self.audio_inner_dim, bias=True, dtype=dtype, device=device
+        )
+
+        # Audio-specific AdaLN
+        self.audio_adaln_single = AdaLayerNormSingle(
+            self.audio_inner_dim,
+            use_additional_conditions=False,
+            dtype=dtype,
+            device=device,
+            operations=self.operations,
+        )
+
+        num_scale_shift_values = 4
+        self.av_ca_video_scale_shift_adaln_single = AdaLayerNormSingle(
+            self.inner_dim,
+            use_additional_conditions=False,
+            embedding_coefficient=num_scale_shift_values,
+            dtype=dtype,
+            device=device,
+            operations=self.operations,
+        )
+        self.av_ca_a2v_gate_adaln_single = AdaLayerNormSingle(
+            self.inner_dim,
+            use_additional_conditions=False,
+            embedding_coefficient=1,
+            dtype=dtype,
+            device=device,
+            operations=self.operations,
+        )
+        self.av_ca_audio_scale_shift_adaln_single = AdaLayerNormSingle(
+            self.audio_inner_dim,
+            use_additional_conditions=False,
+            embedding_coefficient=num_scale_shift_values,
+            dtype=dtype,
+            device=device,
+            operations=self.operations,
+        )
+        self.av_ca_v2a_gate_adaln_single = AdaLayerNormSingle(
+            self.audio_inner_dim,
+            use_additional_conditions=False,
+            embedding_coefficient=1,
+            dtype=dtype,
+            device=device,
+            operations=self.operations,
+        )
+
+        # Audio caption projection
+        self.audio_caption_projection = PixArtAlphaTextProjection(
+            in_features=self.caption_channels,
+            hidden_size=self.audio_inner_dim,
+            dtype=dtype,
+            device=device,
+            operations=self.operations,
+        )
+
+    def _init_transformer_blocks(self, device, dtype, **kwargs):
+        """Initialize transformer blocks for LTXAV."""
+        self.transformer_blocks = nn.ModuleList(
+            [
+                BasicAVTransformerBlock(
+                    v_dim=self.inner_dim,
+                    a_dim=self.audio_inner_dim,
+                    v_heads=self.num_attention_heads,
+                    a_heads=self.audio_num_attention_heads,
+                    vd_head=self.attention_head_dim,
+                    ad_head=self.audio_attention_head_dim,
+                    v_context_dim=self.cross_attention_dim,
+                    a_context_dim=self.audio_cross_attention_dim,
+                    dtype=dtype,
+                    device=device,
+                    operations=self.operations,
+                )
+                for _ in range(self.num_layers)
+            ]
+        )
+
+    def _init_output_components(self, device, dtype):
+        """Initialize output components for LTXAV."""
+        # Video output components
+        super()._init_output_components(device, dtype)
+        # Audio output components
+        self.audio_scale_shift_table = nn.Parameter(
+            torch.empty(2, self.audio_inner_dim, dtype=dtype, device=device)
+        )
+        self.audio_norm_out = self.operations.LayerNorm(
+            self.audio_inner_dim, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device
+        )
+        self.audio_proj_out = self.operations.Linear(
+            self.audio_inner_dim, self.audio_out_channels, dtype=dtype, device=device
+        )
+        self.a_patchifier = AudioPatchifier(1, start_end=True)
+
+    def separate_audio_and_video_latents(self, x, audio_length):
+        """Separate audio and video latents from combined input."""
+        # vx = x[:, : self.in_channels]
+        # ax = x[:, self.in_channels :]
+        #
+        # ax = ax.reshape(ax.shape[0], -1)
+        # ax = ax[:, : audio_length * self.num_audio_channels * self.audio_frequency_bins]
+        #
+        # ax = ax.reshape(
+        #     ax.shape[0], self.num_audio_channels, audio_length, self.audio_frequency_bins
+        # )
+
+        vx = x[0]
+        ax = x[1] if len(x) > 1 else torch.zeros(
+            (vx.shape[0], self.num_audio_channels, 0, self.audio_frequency_bins),
+            device=vx.device, dtype=vx.dtype
+        )
+        return vx, ax
+
+    def recombine_audio_and_video_latents(self, vx, ax, target_shape=None):
+        if ax.numel() == 0:
+            return vx
+        else:
+            return [vx, ax]
+        """Recombine audio and video latents for output."""
+        # if ax.device != vx.device or ax.dtype != vx.dtype:
+        #     logging.warning("Audio and video latents are on different devices or dtypes.")
+        #     ax = ax.to(device=vx.device, dtype=vx.dtype)
+        #     logging.warning(f"Audio audio latent moved to device: {ax.device}, dtype: {ax.dtype}")
+        #
+        # ax = ax.reshape(ax.shape[0], -1)
+        # # pad to f x h x w of the video latents
+        # divisor = vx.shape[-1] * vx.shape[-2] * vx.shape[-3]
+        # if target_shape is None:
+        #     repetitions = math.ceil(ax.shape[-1] / divisor)
+        # else:
+        #     repetitions = target_shape[1] - vx.shape[1]
+        # padded_len = repetitions * divisor
+        # ax = F.pad(ax, (0, padded_len - ax.shape[-1]))
+        # ax = ax.reshape(ax.shape[0], -1, vx.shape[-3], vx.shape[-2], vx.shape[-1])
+        # return torch.cat([vx, ax], dim=1)
+
+    def _process_input(self, x, keyframe_idxs, denoise_mask, **kwargs):
+        """Process input for LTXAV - separate audio and video, then patchify."""
+        audio_length = kwargs.get("audio_length", 0)
+        # Separate audio and video latents
+        vx, ax = self.separate_audio_and_video_latents(x, audio_length)
+
+        has_spatial_mask = False
+        if denoise_mask is not None:
+            # check if any frame has spatial variation (inpainting)
+            for frame_idx in range(denoise_mask.shape[2]):
+                frame_mask = denoise_mask[0, 0, frame_idx]
+                if frame_mask.numel() > 0 and frame_mask.min() != frame_mask.max():
+                    has_spatial_mask = True
+                    break
+
+        [vx, v_pixel_coords, additional_args] = super()._process_input(
+            vx, keyframe_idxs, denoise_mask, **kwargs
+        )
+        additional_args["has_spatial_mask"] = has_spatial_mask
+
+        ax, a_latent_coords = self.a_patchifier.patchify(ax)
+        ax = self.audio_patchify_proj(ax)
+
+        # additional_args.update({"av_orig_shape": list(x.shape)})
+        return [vx, ax], [v_pixel_coords, a_latent_coords], additional_args
+
+    def _prepare_timestep(self, timestep, batch_size, hidden_dtype, **kwargs):
+        """Prepare timestep embeddings."""
+        # TODO: some code reuse is needed here.
+        grid_mask = kwargs.get("grid_mask", None)
+        if grid_mask is not None:
+            timestep = timestep[:, grid_mask]
+
+        timestep_scaled = timestep * self.timestep_scale_multiplier
+
+        v_timestep, v_embedded_timestep = self.adaln_single(
+            timestep_scaled.flatten(),
+            {"resolution": None, "aspect_ratio": None},
+            batch_size=batch_size,
+            hidden_dtype=hidden_dtype,
+        )
+
+        # Calculate patches_per_frame from orig_shape: [batch, channels, frames, height, width]
+        # Video tokens are arranged as (frames * height * width), so patches_per_frame = height * width
+        orig_shape = kwargs.get("orig_shape")
+        has_spatial_mask = kwargs.get("has_spatial_mask", None)
+        v_patches_per_frame = None
+        if not has_spatial_mask and orig_shape is not None and len(orig_shape) == 5:
+            # orig_shape[3] = height, orig_shape[4] = width (in latent space)
+            v_patches_per_frame = orig_shape[3] * orig_shape[4]
+
+        # Reshape to [batch_size, num_tokens, dim] and compress for storage
+        v_timestep = CompressedTimestep(v_timestep.view(batch_size, -1, v_timestep.shape[-1]), v_patches_per_frame)
+        v_embedded_timestep = CompressedTimestep(v_embedded_timestep.view(batch_size, -1, v_embedded_timestep.shape[-1]), v_patches_per_frame)
+
+        # Prepare audio timestep
+        a_timestep = kwargs.get("a_timestep")
+        if a_timestep is not None:
+            a_timestep_scaled = a_timestep * self.timestep_scale_multiplier
+            a_timestep_flat = a_timestep_scaled.flatten()
+            timestep_flat = timestep_scaled.flatten()
+            av_ca_factor = self.av_ca_timestep_scale_multiplier / self.timestep_scale_multiplier
+
+            # Cross-attention timesteps - compress these too
+            av_ca_audio_scale_shift_timestep, _ = self.av_ca_audio_scale_shift_adaln_single(
+                a_timestep_flat,
+                {"resolution": None, "aspect_ratio": None},
+                batch_size=batch_size,
+                hidden_dtype=hidden_dtype,
+            )
+            av_ca_video_scale_shift_timestep, _ = self.av_ca_video_scale_shift_adaln_single(
+                timestep_flat,
+                {"resolution": None, "aspect_ratio": None},
+                batch_size=batch_size,
+                hidden_dtype=hidden_dtype,
+            )
+            av_ca_a2v_gate_noise_timestep, _ = self.av_ca_a2v_gate_adaln_single(
+                timestep_flat * av_ca_factor,
+                {"resolution": None, "aspect_ratio": None},
+                batch_size=batch_size,
+                hidden_dtype=hidden_dtype,
+            )
+            av_ca_v2a_gate_noise_timestep, _ = self.av_ca_v2a_gate_adaln_single(
+                a_timestep_flat * av_ca_factor,
+                {"resolution": None, "aspect_ratio": None},
+                batch_size=batch_size,
+                hidden_dtype=hidden_dtype,
+            )
+
+            # Compress cross-attention timesteps (only video side, audio is too small to benefit)
+            # v_patches_per_frame is None for spatial masks, set for temporal masks or no mask
+            cross_av_timestep_ss = [
+                av_ca_audio_scale_shift_timestep.view(batch_size, -1, av_ca_audio_scale_shift_timestep.shape[-1]),
+                CompressedTimestep(av_ca_video_scale_shift_timestep.view(batch_size, -1, av_ca_video_scale_shift_timestep.shape[-1]), v_patches_per_frame),  # video - compressed if possible
+                CompressedTimestep(av_ca_a2v_gate_noise_timestep.view(batch_size, -1, av_ca_a2v_gate_noise_timestep.shape[-1]), v_patches_per_frame),  # video - compressed if possible
+                av_ca_v2a_gate_noise_timestep.view(batch_size, -1, av_ca_v2a_gate_noise_timestep.shape[-1]),
+            ]
+
+            a_timestep, a_embedded_timestep = self.audio_adaln_single(
+                a_timestep_flat,
+                {"resolution": None, "aspect_ratio": None},
+                batch_size=batch_size,
+                hidden_dtype=hidden_dtype,
+            )
+            # Audio timesteps
+            a_timestep = a_timestep.view(batch_size, -1, a_timestep.shape[-1])
+            a_embedded_timestep = a_embedded_timestep.view(batch_size, -1, a_embedded_timestep.shape[-1])
+        else:
+            a_timestep = timestep_scaled
+            a_embedded_timestep = kwargs.get("embedded_timestep")
+            cross_av_timestep_ss = []
+
+        return [v_timestep, a_timestep, cross_av_timestep_ss], [
+            v_embedded_timestep,
+            a_embedded_timestep,
+        ]
+
+    def _prepare_context(self, context, batch_size, x, attention_mask=None):
+        vx = x[0]
+        ax = x[1]
+        v_context, a_context = torch.split(
+            context, int(context.shape[-1] / 2), len(context.shape) - 1
+        )
+
+        v_context, attention_mask = super()._prepare_context(
+            v_context, batch_size, vx, attention_mask
+        )
+        if self.audio_caption_projection is not None:
+            a_context = self.audio_caption_projection(a_context)
+            a_context = a_context.view(batch_size, -1, ax.shape[-1])
+
+        return [v_context, a_context], attention_mask
+
+    def _prepare_positional_embeddings(self, pixel_coords, frame_rate, x_dtype):
+        v_pixel_coords = pixel_coords[0]
+        v_pe = super()._prepare_positional_embeddings(v_pixel_coords, frame_rate, x_dtype)
+
+        a_latent_coords = pixel_coords[1]
+        a_pe = self._precompute_freqs_cis(
+            a_latent_coords,
+            dim=self.audio_inner_dim,
+            out_dtype=x_dtype,
+            max_pos=self.audio_positional_embedding_max_pos,
+            use_middle_indices_grid=self.use_middle_indices_grid,
+            num_attention_heads=self.audio_num_attention_heads,
+        )
+
+        # calculate positional embeddings for the middle of the token duration, to use in av cross attention layers.
+        max_pos = max(
+            self.positional_embedding_max_pos[0], self.audio_positional_embedding_max_pos[0]
+        )
+        v_pixel_coords = v_pixel_coords.to(torch.float32)
+        v_pixel_coords[:, 0] = v_pixel_coords[:, 0] * (1.0 / frame_rate)
+        av_cross_video_freq_cis = self._precompute_freqs_cis(
+            v_pixel_coords[:, 0:1, :],
+            dim=self.audio_cross_attention_dim,
+            out_dtype=x_dtype,
+            max_pos=[max_pos],
+            use_middle_indices_grid=True,
+            num_attention_heads=self.audio_num_attention_heads,
+        )
+        av_cross_audio_freq_cis = self._precompute_freqs_cis(
+            a_latent_coords[:, 0:1, :],
+            dim=self.audio_cross_attention_dim,
+            out_dtype=x_dtype,
+            max_pos=[max_pos],
+            use_middle_indices_grid=True,
+            num_attention_heads=self.audio_num_attention_heads,
+        )
+
+        return [(v_pe, av_cross_video_freq_cis), (a_pe, av_cross_audio_freq_cis)]
+
+    def _process_transformer_blocks(
+        self, x, context, attention_mask, timestep, pe, transformer_options={}, **kwargs
+    ):
+        vx = x[0]
+        ax = x[1]
+        v_context = context[0]
+        a_context = context[1]
+        v_timestep = timestep[0]
+        a_timestep = timestep[1]
+        v_pe, av_cross_video_freq_cis = pe[0]
+        a_pe, av_cross_audio_freq_cis = pe[1]
+
+        (
+            av_ca_audio_scale_shift_timestep,
+            av_ca_video_scale_shift_timestep,
+            av_ca_a2v_gate_noise_timestep,
+            av_ca_v2a_gate_noise_timestep,
+        ) = timestep[2]
+
+        """Process transformer blocks for LTXAV."""
+        patches_replace = transformer_options.get("patches_replace", {})
+        blocks_replace = patches_replace.get("dit", {})
+
+        # Process transformer blocks
+        for i, block in enumerate(self.transformer_blocks):
+            if ("double_block", i) in blocks_replace:
+
+                def block_wrap(args):
+                    out = {}
+                    out["img"] = block(
+                        args["img"],
+                        v_context=args["v_context"],
+                        a_context=args["a_context"],
+                        attention_mask=args["attention_mask"],
+                        v_timestep=args["v_timestep"],
+                        a_timestep=args["a_timestep"],
+                        v_pe=args["v_pe"],
+                        a_pe=args["a_pe"],
+                        v_cross_pe=args["v_cross_pe"],
+                        a_cross_pe=args["a_cross_pe"],
+                        v_cross_scale_shift_timestep=args["v_cross_scale_shift_timestep"],
+                        a_cross_scale_shift_timestep=args["a_cross_scale_shift_timestep"],
+                        v_cross_gate_timestep=args["v_cross_gate_timestep"],
+                        a_cross_gate_timestep=args["a_cross_gate_timestep"],
+                        transformer_options=args["transformer_options"],
+                    )
+                    return out
+
+                out = blocks_replace[("double_block", i)](
+                    {
+                        "img": (vx, ax),
+                        "v_context": v_context,
+                        "a_context": a_context,
+                        "attention_mask": attention_mask,
+                        "v_timestep": v_timestep,
+                        "a_timestep": a_timestep,
+                        "v_pe": v_pe,
+                        "a_pe": a_pe,
+                        "v_cross_pe": av_cross_video_freq_cis,
+                        "a_cross_pe": av_cross_audio_freq_cis,
+                        "v_cross_scale_shift_timestep": av_ca_video_scale_shift_timestep,
+                        "a_cross_scale_shift_timestep": av_ca_audio_scale_shift_timestep,
+                        "v_cross_gate_timestep": av_ca_a2v_gate_noise_timestep,
+                        "a_cross_gate_timestep": av_ca_v2a_gate_noise_timestep,
+                        "transformer_options": transformer_options,
+                    },
+                    {"original_block": block_wrap},
+                )
+                vx, ax = out["img"]
+            else:
+                vx, ax = block(
+                    (vx, ax),
+                    v_context=v_context,
+                    a_context=a_context,
+                    attention_mask=attention_mask,
+                    v_timestep=v_timestep,
+                    a_timestep=a_timestep,
+                    v_pe=v_pe,
+                    a_pe=a_pe,
+                    v_cross_pe=av_cross_video_freq_cis,
+                    a_cross_pe=av_cross_audio_freq_cis,
+                    v_cross_scale_shift_timestep=av_ca_video_scale_shift_timestep,
+                    a_cross_scale_shift_timestep=av_ca_audio_scale_shift_timestep,
+                    v_cross_gate_timestep=av_ca_a2v_gate_noise_timestep,
+                    a_cross_gate_timestep=av_ca_v2a_gate_noise_timestep,
+                    transformer_options=transformer_options,
+                )
+
+        return [vx, ax]
+
+    def _process_output(self, x, embedded_timestep, keyframe_idxs, **kwargs):
+        vx = x[0]
+        ax = x[1]
+        v_embedded_timestep = embedded_timestep[0]
+        a_embedded_timestep = embedded_timestep[1]
+
+        # Expand compressed video timestep if needed
+        if isinstance(v_embedded_timestep, CompressedTimestep):
+            v_embedded_timestep = v_embedded_timestep.expand()
+
+        vx = super()._process_output(vx, v_embedded_timestep, keyframe_idxs, **kwargs)
+
+        # Process audio output
+        a_scale_shift_values = (
+            self.audio_scale_shift_table[None, None].to(device=a_embedded_timestep.device, dtype=a_embedded_timestep.dtype)
+            + a_embedded_timestep[:, :, None]
+        )
+        a_shift, a_scale = a_scale_shift_values[:, :, 0], a_scale_shift_values[:, :, 1]
+
+        ax = self.audio_norm_out(ax)
+        ax = ax * (1 + a_scale) + a_shift
+        ax = self.audio_proj_out(ax)
+
+        # Unpatchify audio
+        ax = self.a_patchifier.unpatchify(
+            ax, channels=self.num_audio_channels, freq=self.audio_frequency_bins
+        )
+
+        # Recombine audio and video
+        original_shape = kwargs.get("av_orig_shape")
+        return self.recombine_audio_and_video_latents(vx, ax, original_shape)
+
+    def forward(
+        self,
+        x,
+        timestep,
+        context,
+        attention_mask=None,
+        frame_rate=25,
+        transformer_options={},
+        keyframe_idxs=None,
+        **kwargs,
+    ):
+        """
+        Forward pass for LTXAV model.
+
+        Args:
+            x: Combined audio-video input tensor
+            timestep: Tuple of (video_timestep, audio_timestep) or single timestep
+            context: Context tensor (e.g., text embeddings)
+            attention_mask: Attention mask tensor
+            frame_rate: Frame rate for temporal processing
+            transformer_options: Additional options for transformer blocks
+            keyframe_idxs: Keyframe indices for temporal processing
+            **kwargs: Additional keyword arguments including audio_length
+
+        Returns:
+            Combined audio-video output tensor
+        """
+        # Handle timestep format
+        if isinstance(timestep, (tuple, list)) and len(timestep) == 2:
+            v_timestep, a_timestep = timestep
+            kwargs["a_timestep"] = a_timestep
+            timestep = v_timestep
+        else:
+            kwargs["a_timestep"] = timestep
+
+        # Call parent forward method
+        return super().forward(
+            x,
+            timestep,
+            context,
+            attention_mask,
+            frame_rate,
+            transformer_options,
+            keyframe_idxs,
+            **kwargs,
+        )

comfy/ldm/lightricks/embeddings_connector.py

@@ -0,0 +1,305 @@
+import math
+from typing import Optional
+
+import comfy.ldm.common_dit
+import torch
+from comfy.ldm.lightricks.model import (
+    CrossAttention,
+    FeedForward,
+    generate_freq_grid_np,
+    interleaved_freqs_cis,
+    split_freqs_cis,
+)
+from torch import nn
+
+
+class BasicTransformerBlock1D(nn.Module):
+    r"""
+    A basic Transformer block.
+
+    Parameters:
+
+        dim (`int`): The number of channels in the input and output.
+        num_attention_heads (`int`): The number of heads to use for multi-head attention.
+        attention_head_dim (`int`): The number of channels in each head.
+        dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
+        activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward.
+        attention_bias (:
+            obj: `bool`, *optional*, defaults to `False`): Configure if the attentions should contain a bias parameter.
+        upcast_attention (`bool`, *optional*):
+            Whether to upcast the attention computation to float32. This is useful for mixed precision training.
+        norm_elementwise_affine (`bool`, *optional*, defaults to `True`):
+            Whether to use learnable elementwise affine parameters for normalization.
+        standardization_norm (`str`, *optional*, defaults to `"layer_norm"`): The type of pre-normalization to use. Can be `"layer_norm"` or `"rms_norm"`.
+        norm_eps (`float`, *optional*, defaults to 1e-5): Epsilon value for normalization layers.
+        qk_norm (`str`, *optional*, defaults to None):
+            Set to 'layer_norm' or `rms_norm` to perform query and key normalization.
+        final_dropout (`bool` *optional*, defaults to False):
+            Whether to apply a final dropout after the last feed-forward layer.
+        ff_inner_dim (`int`, *optional*): Dimension of the inner feed-forward layer. If not provided, defaults to `dim * 4`.
+        ff_bias (`bool`, *optional*, defaults to `True`): Whether to use bias in the feed-forward layer.
+        attention_out_bias (`bool`, *optional*, defaults to `True`): Whether to use bias in the attention output layer.
+        use_rope (`bool`, *optional*, defaults to `False`): Whether to use Rotary Position Embeddings (RoPE).
+        ffn_dim_mult (`int`, *optional*, defaults to 4): Multiplier for the inner dimension of the feed-forward layer.
+    """
+
+    def __init__(
+        self,
+        dim,
+        n_heads,
+        d_head,
+        context_dim=None,
+        attn_precision=None,
+        dtype=None,
+        device=None,
+        operations=None,
+    ):
+        super().__init__()
+
+        # Define 3 blocks. Each block has its own normalization layer.
+        # 1. Self-Attn
+        self.attn1 = CrossAttention(
+            query_dim=dim,
+            heads=n_heads,
+            dim_head=d_head,
+            context_dim=None,
+            dtype=dtype,
+            device=device,
+            operations=operations,
+        )
+
+        # 3. Feed-forward
+        self.ff = FeedForward(
+            dim,
+            dim_out=dim,
+            glu=True,
+            dtype=dtype,
+            device=device,
+            operations=operations,
+        )
+
+    def forward(self, hidden_states, attention_mask=None, pe=None) -> torch.FloatTensor:
+
+        # Notice that normalization is always applied before the real computation in the following blocks.
+
+        # 1. Normalization Before Self-Attention
+        norm_hidden_states = comfy.ldm.common_dit.rms_norm(hidden_states)
+
+        norm_hidden_states = norm_hidden_states.squeeze(1)
+
+        # 2. Self-Attention
+        attn_output = self.attn1(norm_hidden_states, mask=attention_mask, pe=pe)
+
+        hidden_states = attn_output + hidden_states
+        if hidden_states.ndim == 4:
+            hidden_states = hidden_states.squeeze(1)
+
+        # 3. Normalization before Feed-Forward
+        norm_hidden_states = comfy.ldm.common_dit.rms_norm(hidden_states)
+
+        # 4. Feed-forward
+        ff_output = self.ff(norm_hidden_states)
+
+        hidden_states = ff_output + hidden_states
+        if hidden_states.ndim == 4:
+            hidden_states = hidden_states.squeeze(1)
+
+        return hidden_states
+
+
+class Embeddings1DConnector(nn.Module):
+    _supports_gradient_checkpointing = True
+
+    def __init__(
+        self,
+        in_channels=128,
+        cross_attention_dim=2048,
+        attention_head_dim=128,
+        num_attention_heads=30,
+        num_layers=2,
+        positional_embedding_theta=10000.0,
+        positional_embedding_max_pos=[4096],
+        causal_temporal_positioning=False,
+        num_learnable_registers: Optional[int] = 128,
+        dtype=None,
+        device=None,
+        operations=None,
+        split_rope=False,
+        double_precision_rope=False,
+        **kwargs,
+    ):
+        super().__init__()
+        self.dtype = dtype
+        self.out_channels = in_channels
+        self.num_attention_heads = num_attention_heads
+        self.inner_dim = num_attention_heads * attention_head_dim
+        self.causal_temporal_positioning = causal_temporal_positioning
+        self.positional_embedding_theta = positional_embedding_theta
+        self.positional_embedding_max_pos = positional_embedding_max_pos
+        self.split_rope = split_rope
+        self.double_precision_rope = double_precision_rope
+        self.transformer_1d_blocks = nn.ModuleList(
+            [
+                BasicTransformerBlock1D(
+                    self.inner_dim,
+                    num_attention_heads,
+                    attention_head_dim,
+                    context_dim=cross_attention_dim,
+                    dtype=dtype,
+                    device=device,
+                    operations=operations,
+                )
+                for _ in range(num_layers)
+            ]
+        )
+
+        inner_dim = num_attention_heads * attention_head_dim
+        self.num_learnable_registers = num_learnable_registers
+        if self.num_learnable_registers:
+            self.learnable_registers = nn.Parameter(
+                torch.rand(
+                    self.num_learnable_registers, inner_dim, dtype=dtype, device=device
+                )
+                * 2.0
+                - 1.0
+            )
+
+    def get_fractional_positions(self, indices_grid):
+        fractional_positions = torch.stack(
+            [
+                indices_grid[:, i] / self.positional_embedding_max_pos[i]
+                for i in range(1)
+            ],
+            dim=-1,
+        )
+        return fractional_positions
+
+    def precompute_freqs(self, indices_grid, spacing):
+        source_dtype = indices_grid.dtype
+        dtype = (
+            torch.float32
+            if source_dtype in (torch.bfloat16, torch.float16)
+            else source_dtype
+        )
+
+        fractional_positions = self.get_fractional_positions(indices_grid)
+        indices = (
+            generate_freq_grid_np(
+                self.positional_embedding_theta,
+                indices_grid.shape[1],
+                self.inner_dim,
+            )
+            if self.double_precision_rope
+            else self.generate_freq_grid(spacing, dtype, fractional_positions.device)
+        ).to(device=fractional_positions.device)
+
+        if spacing == "exp_2":
+            freqs = (
+                (indices * fractional_positions.unsqueeze(-1))
+                .transpose(-1, -2)
+                .flatten(2)
+            )
+        else:
+            freqs = (
+                (indices * (fractional_positions.unsqueeze(-1) * 2 - 1))
+                .transpose(-1, -2)
+                .flatten(2)
+            )
+        return freqs
+
+    def generate_freq_grid(self, spacing, dtype, device):
+        dim = self.inner_dim
+        theta = self.positional_embedding_theta
+        n_pos_dims = 1
+        n_elem = 2 * n_pos_dims  # 2 for cos and sin e.g. x 3 = 6
+        start = 1
+        end = theta
+
+        if spacing == "exp":
+            indices = theta ** (torch.arange(0, dim, n_elem, device="cpu", dtype=torch.float32) / (dim - n_elem))
+            indices = indices.to(dtype=dtype, device=device)
+        elif spacing == "exp_2":
+            indices = 1.0 / theta ** (torch.arange(0, dim, n_elem, device=device) / dim)
+            indices = indices.to(dtype=dtype)
+        elif spacing == "linear":
+            indices = torch.linspace(
+                start, end, dim // n_elem, device=device, dtype=dtype
+            )
+        elif spacing == "sqrt":
+            indices = torch.linspace(
+                start**2, end**2, dim // n_elem, device=device, dtype=dtype
+            ).sqrt()
+
+        indices = indices * math.pi / 2
+
+        return indices
+
+    def precompute_freqs_cis(self, indices_grid, spacing="exp"):
+        dim = self.inner_dim
+        n_elem = 2  # 2 because of cos and sin
+        freqs = self.precompute_freqs(indices_grid, spacing)
+        if self.split_rope:
+            expected_freqs = dim // 2
+            current_freqs = freqs.shape[-1]
+            pad_size = expected_freqs - current_freqs
+            cos_freq, sin_freq = split_freqs_cis(
+                freqs, pad_size, self.num_attention_heads
+            )
+        else:
+            cos_freq, sin_freq = interleaved_freqs_cis(freqs, dim % n_elem)
+        return cos_freq.to(self.dtype), sin_freq.to(self.dtype), self.split_rope
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+    ):
+        """
+        The [`Transformer2DModel`] forward method.
+
+        Args:
+            hidden_states (`torch.LongTensor` of shape `(batch size, num latent pixels)` if discrete, `torch.FloatTensor` of shape `(batch size, channel, height, width)` if continuous):
+                Input `hidden_states`.
+            indices_grid (`torch.LongTensor` of shape `(batch size, 3, num latent pixels)`):
+            attention_mask ( `torch.Tensor`, *optional*):
+                An attention mask of shape `(batch, key_tokens)` is applied to `encoder_hidden_states`. If `1` the mask
+                is kept, otherwise if `0` it is discarded. Mask will be converted into a bias, which adds large
+                negative values to the attention scores corresponding to "discard" tokens.
+        Returns:
+            If `return_dict` is True, an [`~models.transformer_2d.Transformer2DModelOutput`] is returned, otherwise a
+            `tuple` where the first element is the sample tensor.
+        """
+        # 1. Input
+
+        if self.num_learnable_registers:
+            num_registers_duplications = math.ceil(
+                max(1024, hidden_states.shape[1]) / self.num_learnable_registers
+            )
+            learnable_registers = torch.tile(
+                self.learnable_registers.to(hidden_states), (num_registers_duplications, 1)
+            )
+
+            hidden_states = torch.cat((hidden_states, learnable_registers[hidden_states.shape[1]:].unsqueeze(0).repeat(hidden_states.shape[0], 1, 1)), dim=1)
+
+            if attention_mask is not None:
+                attention_mask = torch.zeros([1, 1, 1, hidden_states.shape[1]], dtype=attention_mask.dtype, device=attention_mask.device)
+
+        indices_grid = torch.arange(
+            hidden_states.shape[1], dtype=torch.float32, device=hidden_states.device
+        )
+        indices_grid = indices_grid[None, None, :]
+        freqs_cis = self.precompute_freqs_cis(indices_grid)
+
+        # 2. Blocks
+        for block_idx, block in enumerate(self.transformer_1d_blocks):
+            hidden_states = block(
+                hidden_states, attention_mask=attention_mask, pe=freqs_cis
+            )
+
+        # 3. Output
+        # if self.output_scale is not None:
+        #     hidden_states = hidden_states / self.output_scale
+
+        hidden_states = comfy.ldm.common_dit.rms_norm(hidden_states)
+
+        return hidden_states, attention_mask

comfy/ldm/lightricks/latent_upsampler.py

@@ -0,0 +1,292 @@
+from typing import Optional, Tuple
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from einops import rearrange
+
+
+def _rational_for_scale(scale: float) -> Tuple[int, int]:
+    mapping = {0.75: (3, 4), 1.5: (3, 2), 2.0: (2, 1), 4.0: (4, 1)}
+    if float(scale) not in mapping:
+        raise ValueError(
+            f"Unsupported spatial_scale {scale}. Choose from {list(mapping.keys())}"
+        )
+    return mapping[float(scale)]
+
+
+class PixelShuffleND(nn.Module):
+    def __init__(self, dims, upscale_factors=(2, 2, 2)):
+        super().__init__()
+        assert dims in [1, 2, 3], "dims must be 1, 2, or 3"
+        self.dims = dims
+        self.upscale_factors = upscale_factors
+
+    def forward(self, x):
+        if self.dims == 3:
+            return rearrange(
+                x,
+                "b (c p1 p2 p3) d h w -> b c (d p1) (h p2) (w p3)",
+                p1=self.upscale_factors[0],
+                p2=self.upscale_factors[1],
+                p3=self.upscale_factors[2],
+            )
+        elif self.dims == 2:
+            return rearrange(
+                x,
+                "b (c p1 p2) h w -> b c (h p1) (w p2)",
+                p1=self.upscale_factors[0],
+                p2=self.upscale_factors[1],
+            )
+        elif self.dims == 1:
+            return rearrange(
+                x,
+                "b (c p1) f h w -> b c (f p1) h w",
+                p1=self.upscale_factors[0],
+            )
+
+
+class BlurDownsample(nn.Module):
+    """
+    Anti-aliased spatial downsampling by integer stride using a fixed separable binomial kernel.
+    Applies only on H,W. Works for dims=2 or dims=3 (per-frame).
+    """
+
+    def __init__(self, dims: int, stride: int):
+        super().__init__()
+        assert dims in (2, 3)
+        assert stride >= 1 and isinstance(stride, int)
+        self.dims = dims
+        self.stride = stride
+
+        # 5x5 separable binomial kernel [1,4,6,4,1] (outer product), normalized
+        k = torch.tensor([1.0, 4.0, 6.0, 4.0, 1.0])
+        k2d = k[:, None] @ k[None, :]
+        k2d = (k2d / k2d.sum()).float()  # shape (5,5)
+        self.register_buffer("kernel", k2d[None, None, :, :])  # (1,1,5,5)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        if self.stride == 1:
+            return x
+
+        def _apply_2d(x2d: torch.Tensor) -> torch.Tensor:
+            # x2d: (B, C, H, W)
+            B, C, H, W = x2d.shape
+            weight = self.kernel.expand(C, 1, 5, 5)  # depthwise
+            x2d = F.conv2d(
+                x2d, weight=weight, bias=None, stride=self.stride, padding=2, groups=C
+            )
+            return x2d
+
+        if self.dims == 2:
+            return _apply_2d(x)
+        else:
+            # dims == 3: apply per-frame on H,W
+            b, c, f, h, w = x.shape
+            x = rearrange(x, "b c f h w -> (b f) c h w")
+            x = _apply_2d(x)
+            h2, w2 = x.shape[-2:]
+            x = rearrange(x, "(b f) c h w -> b c f h w", b=b, f=f, h=h2, w=w2)
+            return x
+
+
+class SpatialRationalResampler(nn.Module):
+    """
+    Fully-learned rational spatial scaling: up by 'num' via PixelShuffle, then anti-aliased
+    downsample by 'den' using fixed blur + stride. Operates on H,W only.
+
+    For dims==3, work per-frame for spatial scaling (temporal axis untouched).
+    """
+
+    def __init__(self, mid_channels: int, scale: float):
+        super().__init__()
+        self.scale = float(scale)
+        self.num, self.den = _rational_for_scale(self.scale)
+        self.conv = nn.Conv2d(
+            mid_channels, (self.num**2) * mid_channels, kernel_size=3, padding=1
+        )
+        self.pixel_shuffle = PixelShuffleND(2, upscale_factors=(self.num, self.num))
+        self.blur_down = BlurDownsample(dims=2, stride=self.den)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        b, c, f, h, w = x.shape
+        x = rearrange(x, "b c f h w -> (b f) c h w")
+        x = self.conv(x)
+        x = self.pixel_shuffle(x)
+        x = self.blur_down(x)
+        x = rearrange(x, "(b f) c h w -> b c f h w", b=b, f=f)
+        return x
+
+
+class ResBlock(nn.Module):
+    def __init__(
+        self, channels: int, mid_channels: Optional[int] = None, dims: int = 3
+    ):
+        super().__init__()
+        if mid_channels is None:
+            mid_channels = channels
+
+        Conv = nn.Conv2d if dims == 2 else nn.Conv3d
+
+        self.conv1 = Conv(channels, mid_channels, kernel_size=3, padding=1)
+        self.norm1 = nn.GroupNorm(32, mid_channels)
+        self.conv2 = Conv(mid_channels, channels, kernel_size=3, padding=1)
+        self.norm2 = nn.GroupNorm(32, channels)
+        self.activation = nn.SiLU()
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        residual = x
+        x = self.conv1(x)
+        x = self.norm1(x)
+        x = self.activation(x)
+        x = self.conv2(x)
+        x = self.norm2(x)
+        x = self.activation(x + residual)
+        return x
+
+
+class LatentUpsampler(nn.Module):
+    """
+    Model to spatially upsample VAE latents.
+
+    Args:
+        in_channels (`int`): Number of channels in the input latent
+        mid_channels (`int`): Number of channels in the middle layers
+        num_blocks_per_stage (`int`): Number of ResBlocks to use in each stage (pre/post upsampling)
+        dims (`int`): Number of dimensions for convolutions (2 or 3)
+        spatial_upsample (`bool`): Whether to spatially upsample the latent
+        temporal_upsample (`bool`): Whether to temporally upsample the latent
+    """
+
+    def __init__(
+        self,
+        in_channels: int = 128,
+        mid_channels: int = 512,
+        num_blocks_per_stage: int = 4,
+        dims: int = 3,
+        spatial_upsample: bool = True,
+        temporal_upsample: bool = False,
+        spatial_scale: float = 2.0,
+        rational_resampler: bool = False,
+    ):
+        super().__init__()
+
+        self.in_channels = in_channels
+        self.mid_channels = mid_channels
+        self.num_blocks_per_stage = num_blocks_per_stage
+        self.dims = dims
+        self.spatial_upsample = spatial_upsample
+        self.temporal_upsample = temporal_upsample
+        self.spatial_scale = float(spatial_scale)
+        self.rational_resampler = rational_resampler
+
+        Conv = nn.Conv2d if dims == 2 else nn.Conv3d
+
+        self.initial_conv = Conv(in_channels, mid_channels, kernel_size=3, padding=1)
+        self.initial_norm = nn.GroupNorm(32, mid_channels)
+        self.initial_activation = nn.SiLU()
+
+        self.res_blocks = nn.ModuleList(
+            [ResBlock(mid_channels, dims=dims) for _ in range(num_blocks_per_stage)]
+        )
+
+        if spatial_upsample and temporal_upsample:
+            self.upsampler = nn.Sequential(
+                nn.Conv3d(mid_channels, 8 * mid_channels, kernel_size=3, padding=1),
+                PixelShuffleND(3),
+            )
+        elif spatial_upsample:
+            if rational_resampler:
+                self.upsampler = SpatialRationalResampler(
+                    mid_channels=mid_channels, scale=self.spatial_scale
+                )
+            else:
+                self.upsampler = nn.Sequential(
+                    nn.Conv2d(mid_channels, 4 * mid_channels, kernel_size=3, padding=1),
+                    PixelShuffleND(2),
+                )
+        elif temporal_upsample:
+            self.upsampler = nn.Sequential(
+                nn.Conv3d(mid_channels, 2 * mid_channels, kernel_size=3, padding=1),
+                PixelShuffleND(1),
+            )
+        else:
+            raise ValueError(
+                "Either spatial_upsample or temporal_upsample must be True"
+            )
+
+        self.post_upsample_res_blocks = nn.ModuleList(
+            [ResBlock(mid_channels, dims=dims) for _ in range(num_blocks_per_stage)]
+        )
+
+        self.final_conv = Conv(mid_channels, in_channels, kernel_size=3, padding=1)
+
+    def forward(self, latent: torch.Tensor) -> torch.Tensor:
+        b, c, f, h, w = latent.shape
+
+        if self.dims == 2:
+            x = rearrange(latent, "b c f h w -> (b f) c h w")
+            x = self.initial_conv(x)
+            x = self.initial_norm(x)
+            x = self.initial_activation(x)
+
+            for block in self.res_blocks:
+                x = block(x)
+
+            x = self.upsampler(x)
+
+            for block in self.post_upsample_res_blocks:
+                x = block(x)
+
+            x = self.final_conv(x)
+            x = rearrange(x, "(b f) c h w -> b c f h w", b=b, f=f)
+        else:
+            x = self.initial_conv(latent)
+            x = self.initial_norm(x)
+            x = self.initial_activation(x)
+
+            for block in self.res_blocks:
+                x = block(x)
+
+            if self.temporal_upsample:
+                x = self.upsampler(x)
+                x = x[:, :, 1:, :, :]
+            else:
+                if isinstance(self.upsampler, SpatialRationalResampler):
+                    x = self.upsampler(x)
+                else:
+                    x = rearrange(x, "b c f h w -> (b f) c h w")
+                    x = self.upsampler(x)
+                    x = rearrange(x, "(b f) c h w -> b c f h w", b=b, f=f)
+
+            for block in self.post_upsample_res_blocks:
+                x = block(x)
+
+            x = self.final_conv(x)
+
+        return x
+
+    @classmethod
+    def from_config(cls, config):
+        return cls(
+            in_channels=config.get("in_channels", 4),
+            mid_channels=config.get("mid_channels", 128),
+            num_blocks_per_stage=config.get("num_blocks_per_stage", 4),
+            dims=config.get("dims", 2),
+            spatial_upsample=config.get("spatial_upsample", True),
+            temporal_upsample=config.get("temporal_upsample", False),
+            spatial_scale=config.get("spatial_scale", 2.0),
+            rational_resampler=config.get("rational_resampler", False),
+        )
+
+    def config(self):
+        return {
+            "_class_name": "LatentUpsampler",
+            "in_channels": self.in_channels,
+            "mid_channels": self.mid_channels,
+            "num_blocks_per_stage": self.num_blocks_per_stage,
+            "dims": self.dims,
+            "spatial_upsample": self.spatial_upsample,
+            "temporal_upsample": self.temporal_upsample,
+            "spatial_scale": self.spatial_scale,
+            "rational_resampler": self.rational_resampler,
+        }

comfy/ldm/lightricks/model.py

@@ -1,14 +1,47 @@
+from abc import ABC, abstractmethod
+from enum import Enum
+import functools
+import math
+from typing import Dict, Optional, Tuple
+
+from einops import rearrange
+import numpy as np
 import torch
 from torch import nn
 import comfy.patcher_extension
 import comfy.ldm.modules.attention
 import comfy.ldm.common_dit
-from einops import rearrange
-import math
-from typing import Dict, Optional, Tuple
 
 from .symmetric_patchifier import SymmetricPatchifier, latent_to_pixel_coords
 
+def _log_base(x, base):
+    return np.log(x) / np.log(base)
+
+class LTXRopeType(str, Enum):
+    INTERLEAVED = "interleaved"
+    SPLIT = "split"
+
+    KEY = "rope_type"
+
+    @classmethod
+    def from_dict(cls, kwargs, default=None):
+        if default is None:
+            default = cls.INTERLEAVED
+        return cls(kwargs.get(cls.KEY, default))
+
+
+class LTXFrequenciesPrecision(str, Enum):
+    FLOAT32 = "float32"
+    FLOAT64 = "float64"
+
+    KEY = "frequencies_precision"
+
+    @classmethod
+    def from_dict(cls, kwargs, default=None):
+        if default is None:
+            default = cls.FLOAT32
+        return cls(kwargs.get(cls.KEY, default))
+
 
 def get_timestep_embedding(
     timesteps: torch.Tensor,
@@ -40,9 +73,7 @@ def get_timestep_embedding(
     assert len(timesteps.shape) == 1, "Timesteps should be a 1d-array"
 
     half_dim = embedding_dim // 2
-    exponent = -math.log(max_period) * torch.arange(
-        start=0, end=half_dim, dtype=torch.float32, device=timesteps.device
-    )
+    exponent = -math.log(max_period) * torch.arange(start=0, end=half_dim, dtype=torch.float32, device=timesteps.device)
     exponent = exponent / (half_dim - downscale_freq_shift)
 
     emb = torch.exp(exponent)
@@ -74,7 +105,9 @@ class TimestepEmbedding(nn.Module):
         post_act_fn: Optional[str] = None,
         cond_proj_dim=None,
         sample_proj_bias=True,
-        dtype=None, device=None, operations=None,
+        dtype=None,
+        device=None,
+        operations=None,
     ):
         super().__init__()
 
@@ -91,7 +124,9 @@ class TimestepEmbedding(nn.Module):
             time_embed_dim_out = out_dim
         else:
             time_embed_dim_out = time_embed_dim
-        self.linear_2 = operations.Linear(time_embed_dim, time_embed_dim_out, sample_proj_bias, dtype=dtype, device=device)
+        self.linear_2 = operations.Linear(
+            time_embed_dim, time_embed_dim_out, sample_proj_bias, dtype=dtype, device=device
+        )
 
         if post_act_fn is None:
             self.post_act = None
@@ -140,12 +175,22 @@ class PixArtAlphaCombinedTimestepSizeEmbeddings(nn.Module):
     https://github.com/PixArt-alpha/PixArt-alpha/blob/0f55e922376d8b797edd44d25d0e7464b260dcab/diffusion/model/nets/PixArtMS.py#L164C9-L168C29
     """
 
-    def __init__(self, embedding_dim, size_emb_dim, use_additional_conditions: bool = False, dtype=None, device=None, operations=None):
+    def __init__(
+        self,
+        embedding_dim,
+        size_emb_dim,
+        use_additional_conditions: bool = False,
+        dtype=None,
+        device=None,
+        operations=None,
+    ):
         super().__init__()
 
         self.outdim = size_emb_dim
         self.time_proj = Timesteps(num_channels=256, flip_sin_to_cos=True, downscale_freq_shift=0)
-        self.timestep_embedder = TimestepEmbedding(in_channels=256, time_embed_dim=embedding_dim, dtype=dtype, device=device, operations=operations)
+        self.timestep_embedder = TimestepEmbedding(
+            in_channels=256, time_embed_dim=embedding_dim, dtype=dtype, device=device, operations=operations
+        )
 
     def forward(self, timestep, resolution, aspect_ratio, batch_size, hidden_dtype):
         timesteps_proj = self.time_proj(timestep)
@@ -164,15 +209,22 @@ class AdaLayerNormSingle(nn.Module):
         use_additional_conditions (`bool`): To use additional conditions for normalization or not.
     """
 
-    def __init__(self, embedding_dim: int, use_additional_conditions: bool = False, dtype=None, device=None, operations=None):
+    def __init__(
+        self, embedding_dim: int, embedding_coefficient: int = 6, use_additional_conditions: bool = False, dtype=None, device=None, operations=None
+    ):
         super().__init__()
 
         self.emb = PixArtAlphaCombinedTimestepSizeEmbeddings(
-            embedding_dim, size_emb_dim=embedding_dim // 3, use_additional_conditions=use_additional_conditions, dtype=dtype, device=device, operations=operations
+            embedding_dim,
+            size_emb_dim=embedding_dim // 3,
+            use_additional_conditions=use_additional_conditions,
+            dtype=dtype,
+            device=device,
+            operations=operations,
         )
 
         self.silu = nn.SiLU()
-        self.linear = operations.Linear(embedding_dim, 6 * embedding_dim, bias=True, dtype=dtype, device=device)
+        self.linear = operations.Linear(embedding_dim, embedding_coefficient * embedding_dim, bias=True, dtype=dtype, device=device)
 
     def forward(
         self,
@@ -186,6 +238,7 @@ class AdaLayerNormSingle(nn.Module):
         embedded_timestep = self.emb(timestep, **added_cond_kwargs, batch_size=batch_size, hidden_dtype=hidden_dtype)
         return self.linear(self.silu(embedded_timestep)), embedded_timestep
 
+
 class PixArtAlphaTextProjection(nn.Module):
     """
     Projects caption embeddings. Also handles dropout for classifier-free guidance.
@@ -193,18 +246,24 @@ class PixArtAlphaTextProjection(nn.Module):
     Adapted from https://github.com/PixArt-alpha/PixArt-alpha/blob/master/diffusion/model/nets/PixArt_blocks.py
     """
 
-    def __init__(self, in_features, hidden_size, out_features=None, act_fn="gelu_tanh", dtype=None, device=None, operations=None):
+    def __init__(
+        self, in_features, hidden_size, out_features=None, act_fn="gelu_tanh", dtype=None, device=None, operations=None
+    ):
         super().__init__()
         if out_features is None:
             out_features = hidden_size
-        self.linear_1 = operations.Linear(in_features=in_features, out_features=hidden_size, bias=True, dtype=dtype, device=device)
+        self.linear_1 = operations.Linear(
+            in_features=in_features, out_features=hidden_size, bias=True, dtype=dtype, device=device
+        )
         if act_fn == "gelu_tanh":
             self.act_1 = nn.GELU(approximate="tanh")
         elif act_fn == "silu":
             self.act_1 = nn.SiLU()
         else:
             raise ValueError(f"Unknown activation function: {act_fn}")
-        self.linear_2 = operations.Linear(in_features=hidden_size, out_features=out_features, bias=True, dtype=dtype, device=device)
+        self.linear_2 = operations.Linear(
+            in_features=hidden_size, out_features=out_features, bias=True, dtype=dtype, device=device
+        )
 
     def forward(self, caption):
         hidden_states = self.linear_1(caption)
@@ -223,25 +282,28 @@ class GELU_approx(nn.Module):
 
 
 class FeedForward(nn.Module):
-    def __init__(self, dim, dim_out, mult=4, glu=False, dropout=0., dtype=None, device=None, operations=None):
+    def __init__(self, dim, dim_out, mult=4, glu=False, dropout=0.0, dtype=None, device=None, operations=None):
         super().__init__()
         inner_dim = int(dim * mult)
         project_in = GELU_approx(dim, inner_dim, dtype=dtype, device=device, operations=operations)
 
         self.net = nn.Sequential(
-            project_in,
-            nn.Dropout(dropout),
-            operations.Linear(inner_dim, dim_out, dtype=dtype, device=device)
+            project_in, nn.Dropout(dropout), operations.Linear(inner_dim, dim_out, dtype=dtype, device=device)
         )
 
     def forward(self, x):
         return self.net(x)
 
+def apply_rotary_emb(input_tensor, freqs_cis):
+    cos_freqs, sin_freqs = freqs_cis[0], freqs_cis[1]
+    split_pe = freqs_cis[2] if len(freqs_cis) > 2 else False
+    return (
+        apply_split_rotary_emb(input_tensor, cos_freqs, sin_freqs)
+        if split_pe else
+        apply_interleaved_rotary_emb(input_tensor, cos_freqs, sin_freqs)
+    )
 
-def apply_rotary_emb(input_tensor, freqs_cis): #TODO: remove duplicate funcs and pick the best/fastest one
-    cos_freqs = freqs_cis[0]
-    sin_freqs = freqs_cis[1]
-
+def apply_interleaved_rotary_emb(input_tensor, cos_freqs, sin_freqs):  # TODO: remove duplicate funcs and pick the best/fastest one
     t_dup = rearrange(input_tensor, "... (d r) -> ... d r", r=2)
     t1, t2 = t_dup.unbind(dim=-1)
     t_dup = torch.stack((-t2, t1), dim=-1)
@@ -251,9 +313,37 @@ def apply_rotary_emb(input_tensor, freqs_cis): #TODO: remove duplicate funcs and
 
     return out
 
+def apply_split_rotary_emb(input_tensor, cos, sin):
+    needs_reshape = False
+    if input_tensor.ndim != 4 and cos.ndim == 4:
+        B, H, T, _ = cos.shape
+        input_tensor = input_tensor.reshape(B, T, H, -1).swapaxes(1, 2)
+        needs_reshape = True
+    split_input = rearrange(input_tensor, "... (d r) -> ... d r", d=2)
+    first_half_input = split_input[..., :1, :]
+    second_half_input = split_input[..., 1:, :]
+    output = split_input * cos.unsqueeze(-2)
+    first_half_output = output[..., :1, :]
+    second_half_output = output[..., 1:, :]
+    first_half_output.addcmul_(-sin.unsqueeze(-2), second_half_input)
+    second_half_output.addcmul_(sin.unsqueeze(-2), first_half_input)
+    output = rearrange(output, "... d r -> ... (d r)")
+    return output.swapaxes(1, 2).reshape(B, T, -1) if needs_reshape else output
+
 
 class CrossAttention(nn.Module):
-    def __init__(self, query_dim, context_dim=None, heads=8, dim_head=64, dropout=0., attn_precision=None, dtype=None, device=None, operations=None):
+    def __init__(
+        self,
+        query_dim,
+        context_dim=None,
+        heads=8,
+        dim_head=64,
+        dropout=0.0,
+        attn_precision=None,
+        dtype=None,
+        device=None,
+        operations=None,
+    ):
         super().__init__()
         inner_dim = dim_head * heads
         context_dim = query_dim if context_dim is None else context_dim
@@ -269,9 +359,11 @@ class CrossAttention(nn.Module):
         self.to_k = operations.Linear(context_dim, inner_dim, bias=True, dtype=dtype, device=device)
         self.to_v = operations.Linear(context_dim, inner_dim, bias=True, dtype=dtype, device=device)
 
-        self.to_out = nn.Sequential(operations.Linear(inner_dim, query_dim, dtype=dtype, device=device), nn.Dropout(dropout))
+        self.to_out = nn.Sequential(
+            operations.Linear(inner_dim, query_dim, dtype=dtype, device=device), nn.Dropout(dropout)
+        )
 
-    def forward(self, x, context=None, mask=None, pe=None, transformer_options={}):
+    def forward(self, x, context=None, mask=None, pe=None, k_pe=None, transformer_options={}):
         q = self.to_q(x)
         context = x if context is None else context
         k = self.to_k(context)
@@ -282,7 +374,7 @@ class CrossAttention(nn.Module):
 
         if pe is not None:
             q = apply_rotary_emb(q, pe)
-            k = apply_rotary_emb(k, pe)
+            k = apply_rotary_emb(k, pe if k_pe is None else k_pe)
 
         if mask is None:
             out = comfy.ldm.modules.attention.optimized_attention(q, k, v, self.heads, attn_precision=self.attn_precision, transformer_options=transformer_options)
@@ -292,146 +384,495 @@ class CrossAttention(nn.Module):
 
 
 class BasicTransformerBlock(nn.Module):
-    def __init__(self, dim, n_heads, d_head, context_dim=None, attn_precision=None, dtype=None, device=None, operations=None):
+    def __init__(
+        self, dim, n_heads, d_head, context_dim=None, attn_precision=None, dtype=None, device=None, operations=None
+    ):
         super().__init__()
 
         self.attn_precision = attn_precision
-        self.attn1 = CrossAttention(query_dim=dim, heads=n_heads, dim_head=d_head, context_dim=None, attn_precision=self.attn_precision, dtype=dtype, device=device, operations=operations)
+        self.attn1 = CrossAttention(
+            query_dim=dim,
+            heads=n_heads,
+            dim_head=d_head,
+            context_dim=None,
+            attn_precision=self.attn_precision,
+            dtype=dtype,
+            device=device,
+            operations=operations,
+        )
         self.ff = FeedForward(dim, dim_out=dim, glu=True, dtype=dtype, device=device, operations=operations)
 
-        self.attn2 = CrossAttention(query_dim=dim, context_dim=context_dim, heads=n_heads, dim_head=d_head, attn_precision=self.attn_precision, dtype=dtype, device=device, operations=operations)
+        self.attn2 = CrossAttention(
+            query_dim=dim,
+            context_dim=context_dim,
+            heads=n_heads,
+            dim_head=d_head,
+            attn_precision=self.attn_precision,
+            dtype=dtype,
+            device=device,
+            operations=operations,
+        )
 
         self.scale_shift_table = nn.Parameter(torch.empty(6, dim, device=device, dtype=dtype))
 
     def forward(self, x, context=None, attention_mask=None, timestep=None, pe=None, transformer_options={}):
         shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = (self.scale_shift_table[None, None].to(device=x.device, dtype=x.dtype) + timestep.reshape(x.shape[0], timestep.shape[1], self.scale_shift_table.shape[0], -1)).unbind(dim=2)
 
-        x += self.attn1(comfy.ldm.common_dit.rms_norm(x) * (1 + scale_msa) + shift_msa, pe=pe, transformer_options=transformer_options) * gate_msa
+        attn1_input = comfy.ldm.common_dit.rms_norm(x)
+        attn1_input = torch.addcmul(attn1_input, attn1_input, scale_msa).add_(shift_msa)
+        attn1_input = self.attn1(attn1_input, pe=pe, transformer_options=transformer_options)
+        x.addcmul_(attn1_input, gate_msa)
+        del attn1_input
 
         x += self.attn2(x, context=context, mask=attention_mask, transformer_options=transformer_options)
 
-        y = comfy.ldm.common_dit.rms_norm(x) * (1 + scale_mlp) + shift_mlp
-        x += self.ff(y) * gate_mlp
+        y = comfy.ldm.common_dit.rms_norm(x)
+        y = torch.addcmul(y, y, scale_mlp).add_(shift_mlp)
+        x.addcmul_(self.ff(y), gate_mlp)
 
         return x
 
 def get_fractional_positions(indices_grid, max_pos):
+    n_pos_dims = indices_grid.shape[1]
+    assert n_pos_dims == len(max_pos), f'Number of position dimensions ({n_pos_dims}) must match max_pos length ({len(max_pos)})'
     fractional_positions = torch.stack(
-        [
-            indices_grid[:, i] / max_pos[i]
-            for i in range(3)
-        ],
-        dim=-1,
+        [indices_grid[:, i] / max_pos[i] for i in range(n_pos_dims)],
+        axis=-1,
     )
     return fractional_positions
 
 
-def precompute_freqs_cis(indices_grid, dim, out_dtype, theta=10000.0, max_pos=[20, 2048, 2048]):
-    dtype = torch.float32 #self.dtype
-
-    fractional_positions = get_fractional_positions(indices_grid, max_pos)
-
+@functools.lru_cache(maxsize=5)
+def generate_freq_grid_np(positional_embedding_theta, positional_embedding_max_pos_count, inner_dim, _ = None):
+    theta = positional_embedding_theta
     start = 1
     end = theta
-    device = fractional_positions.device
+
+    n_elem = 2 * positional_embedding_max_pos_count
+    pow_indices = np.power(
+        theta,
+        np.linspace(
+            _log_base(start, theta),
+            _log_base(end, theta),
+            inner_dim // n_elem,
+            dtype=np.float64,
+        ),
+    )
+    return torch.tensor(pow_indices * math.pi / 2, dtype=torch.float32)
+
+def generate_freq_grid_pytorch(positional_embedding_theta, positional_embedding_max_pos_count, inner_dim, device):
+    theta = positional_embedding_theta
+    start = 1
+    end = theta
+    n_elem = 2 * positional_embedding_max_pos_count
 
     indices = theta ** (
         torch.linspace(
             math.log(start, theta),
             math.log(end, theta),
-            dim // 6,
+            inner_dim // n_elem,
             device=device,
-            dtype=dtype,
+            dtype=torch.float32,
         )
     )
-    indices = indices.to(dtype=dtype)
+    indices = indices.to(dtype=torch.float32)
 
     indices = indices * math.pi / 2
 
+    return indices
+
+def generate_freqs(indices, indices_grid, max_pos, use_middle_indices_grid):
+    if use_middle_indices_grid:
+        assert(len(indices_grid.shape) == 4 and indices_grid.shape[-1] ==2)
+        indices_grid_start, indices_grid_end = indices_grid[..., 0], indices_grid[..., 1]
+        indices_grid = (indices_grid_start + indices_grid_end) / 2.0
+    elif len(indices_grid.shape) == 4:
+        indices_grid = indices_grid[..., 0]
+
+    # Get fractional positions and compute frequency indices
+    fractional_positions = get_fractional_positions(indices_grid, max_pos)
+    indices = indices.to(device=fractional_positions.device)
+
     freqs = (
         (indices * (fractional_positions.unsqueeze(-1) * 2 - 1))
         .transpose(-1, -2)
         .flatten(2)
     )
+    return freqs
 
+def interleaved_freqs_cis(freqs, pad_size):
     cos_freq = freqs.cos().repeat_interleave(2, dim=-1)
     sin_freq = freqs.sin().repeat_interleave(2, dim=-1)
-    if dim % 6 != 0:
-        cos_padding = torch.ones_like(cos_freq[:, :, : dim % 6])
-        sin_padding = torch.zeros_like(cos_freq[:, :, : dim % 6])
+    if pad_size != 0:
+        cos_padding = torch.ones_like(cos_freq[:, :, : pad_size])
+        sin_padding = torch.zeros_like(cos_freq[:, :, : pad_size])
         cos_freq = torch.cat([cos_padding, cos_freq], dim=-1)
         sin_freq = torch.cat([sin_padding, sin_freq], dim=-1)
-    return cos_freq.to(out_dtype), sin_freq.to(out_dtype)
+    return cos_freq, sin_freq
 
+def split_freqs_cis(freqs, pad_size, num_attention_heads):
+    cos_freq = freqs.cos()
+    sin_freq = freqs.sin()
 
-class LTXVModel(torch.nn.Module):
-    def __init__(self,
-                 in_channels=128,
-                 cross_attention_dim=2048,
-                 attention_head_dim=64,
-                 num_attention_heads=32,
+    if pad_size != 0:
+        cos_padding = torch.ones_like(cos_freq[:, :, :pad_size])
+        sin_padding = torch.zeros_like(sin_freq[:, :, :pad_size])
 
-                 caption_channels=4096,
-                 num_layers=28,
+        cos_freq = torch.concatenate([cos_padding, cos_freq], axis=-1)
+        sin_freq = torch.concatenate([sin_padding, sin_freq], axis=-1)
 
+    # Reshape freqs to be compatible with multi-head attention
+    B , T, half_HD = cos_freq.shape
 
-                 positional_embedding_theta=10000.0,
-                 positional_embedding_max_pos=[20, 2048, 2048],
-                 causal_temporal_positioning=False,
-                 vae_scale_factors=(8, 32, 32),
-                 dtype=None, device=None, operations=None, **kwargs):
+    cos_freq = cos_freq.reshape(B, T, num_attention_heads, half_HD // num_attention_heads)
+    sin_freq = sin_freq.reshape(B, T, num_attention_heads, half_HD // num_attention_heads)
+
+    cos_freq = torch.swapaxes(cos_freq, 1, 2)  # (B,H,T,D//2)
+    sin_freq = torch.swapaxes(sin_freq, 1, 2)  # (B,H,T,D//2)
+    return cos_freq, sin_freq
+
+class LTXBaseModel(torch.nn.Module, ABC):
+    """
+    Abstract base class for LTX models (Lightricks Transformer models).
+
+    This class defines the common interface and shared functionality for all LTX models,
+    including LTXV (video) and LTXAV (audio-video) variants.
+    """
+
+    def __init__(
+        self,
+        in_channels: int,
+        cross_attention_dim: int,
+        attention_head_dim: int,
+        num_attention_heads: int,
+        caption_channels: int,
+        num_layers: int,
+        positional_embedding_theta: float = 10000.0,
+        positional_embedding_max_pos: list = [20, 2048, 2048],
+        causal_temporal_positioning: bool = False,
+        vae_scale_factors: tuple = (8, 32, 32),
+        use_middle_indices_grid=False,
+        timestep_scale_multiplier = 1000.0,
+        dtype=None,
+        device=None,
+        operations=None,
+        **kwargs,
+    ):
         super().__init__()
         self.generator = None
         self.vae_scale_factors = vae_scale_factors
+        self.use_middle_indices_grid = use_middle_indices_grid
         self.dtype = dtype
-        self.out_channels = in_channels
-        self.inner_dim = num_attention_heads * attention_head_dim
+        self.in_channels = in_channels
+        self.cross_attention_dim = cross_attention_dim
+        self.attention_head_dim = attention_head_dim
+        self.num_attention_heads = num_attention_heads
+        self.caption_channels = caption_channels
+        self.num_layers = num_layers
+        self.positional_embedding_theta = positional_embedding_theta
+        self.positional_embedding_max_pos = positional_embedding_max_pos
+        self.split_positional_embedding = LTXRopeType.from_dict(kwargs)
+        self.freq_grid_generator = (
+            generate_freq_grid_np if LTXFrequenciesPrecision.from_dict(kwargs) == LTXFrequenciesPrecision.FLOAT64
+            else generate_freq_grid_pytorch
+        )
         self.causal_temporal_positioning = causal_temporal_positioning
+        self.operations = operations
+        self.timestep_scale_multiplier = timestep_scale_multiplier
 
-        self.patchify_proj = operations.Linear(in_channels, self.inner_dim, bias=True, dtype=dtype, device=device)
+        # Common dimensions
+        self.inner_dim = num_attention_heads * attention_head_dim
+        self.out_channels = in_channels
+
+        # Initialize common components
+        self._init_common_components(device, dtype)
+
+        # Initialize model-specific components
+        self._init_model_components(device, dtype, **kwargs)
+
+        # Initialize transformer blocks
+        self._init_transformer_blocks(device, dtype, **kwargs)
+
+        # Initialize output components
+        self._init_output_components(device, dtype)
+
+    def _init_common_components(self, device, dtype):
+        """Initialize components common to all LTX models
+        - patchify_proj: Linear projection for patchifying input
+        - adaln_single: AdaLN layer for timestep embedding
+        - caption_projection: Linear projection for caption embedding
+        """
+        self.patchify_proj = self.operations.Linear(
+            self.in_channels, self.inner_dim, bias=True, dtype=dtype, device=device
+        )
 
         self.adaln_single = AdaLayerNormSingle(
-            self.inner_dim, use_additional_conditions=False, dtype=dtype, device=device, operations=operations
+            self.inner_dim, use_additional_conditions=False, dtype=dtype, device=device, operations=self.operations
         )
 
-        # self.adaln_single.linear = operations.Linear(self.inner_dim, 4 * self.inner_dim, bias=True, dtype=dtype, device=device)
-
         self.caption_projection = PixArtAlphaTextProjection(
-            in_features=caption_channels, hidden_size=self.inner_dim, dtype=dtype, device=device, operations=operations
+            in_features=self.caption_channels,
+            hidden_size=self.inner_dim,
+            dtype=dtype,
+            device=device,
+            operations=self.operations,
         )
 
+    @abstractmethod
+    def _init_model_components(self, device, dtype, **kwargs):
+        """Initialize model-specific components. Must be implemented by subclasses."""
+        pass
+
+    @abstractmethod
+    def _init_transformer_blocks(self, device, dtype, **kwargs):
+        """Initialize transformer blocks. Must be implemented by subclasses."""
+        pass
+
+    @abstractmethod
+    def _init_output_components(self, device, dtype):
+        """Initialize output components. Must be implemented by subclasses."""
+        pass
+
+    @abstractmethod
+    def _process_input(self, x, keyframe_idxs, denoise_mask, **kwargs):
+        """Process input data. Must be implemented by subclasses."""
+        pass
+
+    @abstractmethod
+    def _process_transformer_blocks(self, x, context, attention_mask, timestep, pe, **kwargs):
+        """Process transformer blocks. Must be implemented by subclasses."""
+        pass
+
+    @abstractmethod
+    def _process_output(self, x, embedded_timestep, keyframe_idxs, **kwargs):
+        """Process output data. Must be implemented by subclasses."""
+        pass
+
+    def _prepare_timestep(self, timestep, batch_size, hidden_dtype, **kwargs):
+        """Prepare timestep embeddings."""
+        grid_mask = kwargs.get("grid_mask", None)
+        if grid_mask is not None:
+            timestep = timestep[:, grid_mask]
+
+        timestep = timestep * self.timestep_scale_multiplier
+        timestep, embedded_timestep = self.adaln_single(
+            timestep.flatten(),
+            {"resolution": None, "aspect_ratio": None},
+            batch_size=batch_size,
+            hidden_dtype=hidden_dtype,
+        )
+
+        # Second dimension is 1 or number of tokens (if timestep_per_token)
+        timestep = timestep.view(batch_size, -1, timestep.shape[-1])
+        embedded_timestep = embedded_timestep.view(batch_size, -1, embedded_timestep.shape[-1])
+
+        return timestep, embedded_timestep
+
+    def _prepare_context(self, context, batch_size, x, attention_mask=None):
+        """Prepare context for transformer blocks."""
+        if self.caption_projection is not None:
+            context = self.caption_projection(context)
+            context = context.view(batch_size, -1, x.shape[-1])
+
+        return context, attention_mask
+
+    def _precompute_freqs_cis(
+        self,
+        indices_grid,
+        dim,
+        out_dtype,
+        theta=10000.0,
+        max_pos=[20, 2048, 2048],
+        use_middle_indices_grid=False,
+        num_attention_heads=32,
+    ):
+        split_mode = self.split_positional_embedding == LTXRopeType.SPLIT
+        indices = self.freq_grid_generator(theta, indices_grid.shape[1], dim, indices_grid.device)
+        freqs = generate_freqs(indices, indices_grid, max_pos, use_middle_indices_grid)
+
+        if split_mode:
+            expected_freqs = dim // 2
+            current_freqs = freqs.shape[-1]
+            pad_size = expected_freqs - current_freqs
+            cos_freq, sin_freq = split_freqs_cis(freqs, pad_size, num_attention_heads)
+        else:
+            # 2 because of cos and sin by 3 for (t, x, y), 1 for temporal only
+            n_elem = 2 * indices_grid.shape[1]
+            cos_freq, sin_freq = interleaved_freqs_cis(freqs, dim % n_elem)
+        return cos_freq.to(out_dtype), sin_freq.to(out_dtype), split_mode
+
+    def _prepare_positional_embeddings(self, pixel_coords, frame_rate, x_dtype):
+        """Prepare positional embeddings."""
+        fractional_coords = pixel_coords.to(torch.float32)
+        fractional_coords[:, 0] = fractional_coords[:, 0] * (1.0 / frame_rate)
+        pe = self._precompute_freqs_cis(
+            fractional_coords,
+            dim=self.inner_dim,
+            out_dtype=x_dtype,
+            max_pos=self.positional_embedding_max_pos,
+            use_middle_indices_grid=self.use_middle_indices_grid,
+            num_attention_heads=self.num_attention_heads,
+        )
+        return pe
+
+    def _prepare_attention_mask(self, attention_mask, x_dtype):
+        """Prepare attention mask."""
+        if attention_mask is not None and not torch.is_floating_point(attention_mask):
+            attention_mask = (attention_mask - 1).to(x_dtype).reshape(
+                (attention_mask.shape[0], 1, -1, attention_mask.shape[-1])
+            ) * torch.finfo(x_dtype).max
+        return attention_mask
+
+    def forward(
+        self, x, timestep, context, attention_mask, frame_rate=25, transformer_options={}, keyframe_idxs=None, denoise_mask=None, **kwargs
+    ):
+        """
+        Forward pass for LTX models.
+
+        Args:
+            x: Input tensor
+            timestep: Timestep tensor
+            context: Context tensor (e.g., text embeddings)
+            attention_mask: Attention mask tensor
+            frame_rate: Frame rate for temporal processing
+            transformer_options: Additional options for transformer blocks
+            keyframe_idxs: Keyframe indices for temporal processing
+            **kwargs: Additional keyword arguments
+
+        Returns:
+            Processed output tensor
+        """
+        return comfy.patcher_extension.WrapperExecutor.new_class_executor(
+            self._forward,
+            self,
+            comfy.patcher_extension.get_all_wrappers(
+                comfy.patcher_extension.WrappersMP.DIFFUSION_MODEL, transformer_options
+            ),
+        ).execute(x, timestep, context, attention_mask, frame_rate, transformer_options, keyframe_idxs, denoise_mask=denoise_mask, **kwargs)
+
+    def _forward(
+        self, x, timestep, context, attention_mask, frame_rate=25, transformer_options={}, keyframe_idxs=None, denoise_mask=None, **kwargs
+    ):
+        """
+        Internal forward pass for LTX models.
+
+        Args:
+            x: Input tensor
+            timestep: Timestep tensor
+            context: Context tensor (e.g., text embeddings)
+            attention_mask: Attention mask tensor
+            frame_rate: Frame rate for temporal processing
+            transformer_options: Additional options for transformer blocks
+            keyframe_idxs: Keyframe indices for temporal processing
+            **kwargs: Additional keyword arguments
+
+        Returns:
+            Processed output tensor
+        """
+        if isinstance(x, list):
+            input_dtype = x[0].dtype
+            batch_size = x[0].shape[0]
+        else:
+            input_dtype = x.dtype
+            batch_size = x.shape[0]
+        # Process input
+        merged_args = {**transformer_options, **kwargs}
+        x, pixel_coords, additional_args = self._process_input(x, keyframe_idxs, denoise_mask, **merged_args)
+        merged_args.update(additional_args)
+
+        # Prepare timestep and context
+        timestep, embedded_timestep = self._prepare_timestep(timestep, batch_size, input_dtype, **merged_args)
+        context, attention_mask = self._prepare_context(context, batch_size, x, attention_mask)
+
+        # Prepare attention mask and positional embeddings
+        attention_mask = self._prepare_attention_mask(attention_mask, input_dtype)
+        pe = self._prepare_positional_embeddings(pixel_coords, frame_rate, input_dtype)
+
+        # Process transformer blocks
+        x = self._process_transformer_blocks(
+            x, context, attention_mask, timestep, pe, transformer_options=transformer_options, **merged_args
+        )
+
+        # Process output
+        x = self._process_output(x, embedded_timestep, keyframe_idxs, **merged_args)
+        return x
+
+
+class LTXVModel(LTXBaseModel):
+    """LTXV model for video generation."""
+
+    def __init__(
+        self,
+        in_channels=128,
+        cross_attention_dim=2048,
+        attention_head_dim=64,
+        num_attention_heads=32,
+        caption_channels=4096,
+        num_layers=28,
+        positional_embedding_theta=10000.0,
+        positional_embedding_max_pos=[20, 2048, 2048],
+        causal_temporal_positioning=False,
+        vae_scale_factors=(8, 32, 32),
+        use_middle_indices_grid=False,
+        timestep_scale_multiplier = 1000.0,
+        dtype=None,
+        device=None,
+        operations=None,
+        **kwargs,
+    ):
+        super().__init__(
+            in_channels=in_channels,
+            cross_attention_dim=cross_attention_dim,
+            attention_head_dim=attention_head_dim,
+            num_attention_heads=num_attention_heads,
+            caption_channels=caption_channels,
+            num_layers=num_layers,
+            positional_embedding_theta=positional_embedding_theta,
+            positional_embedding_max_pos=positional_embedding_max_pos,
+            causal_temporal_positioning=causal_temporal_positioning,
+            vae_scale_factors=vae_scale_factors,
+            use_middle_indices_grid=use_middle_indices_grid,
+            timestep_scale_multiplier=timestep_scale_multiplier,
+            dtype=dtype,
+            device=device,
+            operations=operations,
+            **kwargs,
+        )
+
+    def _init_model_components(self, device, dtype, **kwargs):
+        """Initialize LTXV-specific components."""
+        # No additional components needed for LTXV beyond base class
+        pass
+
+    def _init_transformer_blocks(self, device, dtype, **kwargs):
+        """Initialize transformer blocks for LTXV."""
         self.transformer_blocks = nn.ModuleList(
             [
                 BasicTransformerBlock(
                     self.inner_dim,
-                    num_attention_heads,
-                    attention_head_dim,
-                    context_dim=cross_attention_dim,
-                    # attn_precision=attn_precision,
-                    dtype=dtype, device=device, operations=operations
+                    self.num_attention_heads,
+                    self.attention_head_dim,
+                    context_dim=self.cross_attention_dim,
+                    dtype=dtype,
+                    device=device,
+                    operations=self.operations,
                 )
-                for d in range(num_layers)
+                for _ in range(self.num_layers)
             ]
         )
 
+    def _init_output_components(self, device, dtype):
+        """Initialize output components for LTXV."""
         self.scale_shift_table = nn.Parameter(torch.empty(2, self.inner_dim, dtype=dtype, device=device))
-        self.norm_out = operations.LayerNorm(self.inner_dim, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
-        self.proj_out = operations.Linear(self.inner_dim, self.out_channels, dtype=dtype, device=device)
-
-        self.patchifier = SymmetricPatchifier(1)
-
-    def forward(self, x, timestep, context, attention_mask, frame_rate=25, transformer_options={}, keyframe_idxs=None, **kwargs):
-        return comfy.patcher_extension.WrapperExecutor.new_class_executor(
-            self._forward,
-            self,
-            comfy.patcher_extension.get_all_wrappers(comfy.patcher_extension.WrappersMP.DIFFUSION_MODEL, transformer_options)
-        ).execute(x, timestep, context, attention_mask, frame_rate, transformer_options, keyframe_idxs, **kwargs)
-
-    def _forward(self, x, timestep, context, attention_mask, frame_rate=25, transformer_options={}, keyframe_idxs=None, **kwargs):
-        patches_replace = transformer_options.get("patches_replace", {})
-
-        orig_shape = list(x.shape)
+        self.norm_out = self.operations.LayerNorm(
+            self.inner_dim, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device
+        )
+        self.proj_out = self.operations.Linear(self.inner_dim, self.out_channels, dtype=dtype, device=device)
+        self.patchifier = SymmetricPatchifier(1, start_end=True)
 
+    def _process_input(self, x, keyframe_idxs, denoise_mask, **kwargs):
+        """Process input for LTXV."""
+        additional_args = {"orig_shape": list(x.shape)}
         x, latent_coords = self.patchifier.patchify(x)
         pixel_coords = latent_to_pixel_coords(
             latent_coords=latent_coords,
@@ -439,44 +880,30 @@ class LTXVModel(torch.nn.Module):
             causal_fix=self.causal_temporal_positioning,
         )
 
+        grid_mask = None
         if keyframe_idxs is not None:
-            pixel_coords[:, :, -keyframe_idxs.shape[2]:] = keyframe_idxs
+            additional_args.update({ "orig_patchified_shape": list(x.shape)})
+            denoise_mask = self.patchifier.patchify(denoise_mask)[0]
+            grid_mask = ~torch.any(denoise_mask < 0, dim=-1)[0]
+            additional_args.update({"grid_mask": grid_mask})
+            x = x[:, grid_mask, :]
+            pixel_coords = pixel_coords[:, :, grid_mask, ...]
 
-        fractional_coords = pixel_coords.to(torch.float32)
-        fractional_coords[:, 0] = fractional_coords[:, 0] * (1.0 / frame_rate)
+            kf_grid_mask = grid_mask[-keyframe_idxs.shape[2]:]
+            keyframe_idxs = keyframe_idxs[..., kf_grid_mask, :]
+            pixel_coords[:, :, -keyframe_idxs.shape[2]:, :] = keyframe_idxs
 
         x = self.patchify_proj(x)
-        timestep = timestep * 1000.0
-
-        if attention_mask is not None and not torch.is_floating_point(attention_mask):
-            attention_mask = (attention_mask - 1).to(x.dtype).reshape((attention_mask.shape[0], 1, -1, attention_mask.shape[-1])) * torch.finfo(x.dtype).max
-
-        pe = precompute_freqs_cis(fractional_coords, dim=self.inner_dim, out_dtype=x.dtype)
-
-        batch_size = x.shape[0]
-        timestep, embedded_timestep = self.adaln_single(
-            timestep.flatten(),
-            {"resolution": None, "aspect_ratio": None},
-            batch_size=batch_size,
-            hidden_dtype=x.dtype,
-        )
-        # Second dimension is 1 or number of tokens (if timestep_per_token)
-        timestep = timestep.view(batch_size, -1, timestep.shape[-1])
-        embedded_timestep = embedded_timestep.view(
-            batch_size, -1, embedded_timestep.shape[-1]
-        )
-
-        # 2. Blocks
-        if self.caption_projection is not None:
-            batch_size = x.shape[0]
-            context = self.caption_projection(context)
-            context = context.view(
-                batch_size, -1, x.shape[-1]
-            )
+        return x, pixel_coords, additional_args
 
+    def _process_transformer_blocks(self, x, context, attention_mask, timestep, pe, transformer_options={}, **kwargs):
+        """Process transformer blocks for LTXV."""
+        patches_replace = transformer_options.get("patches_replace", {})
         blocks_replace = patches_replace.get("dit", {})
+
         for i, block in enumerate(self.transformer_blocks):
             if ("double_block", i) in blocks_replace:
+
                 def block_wrap(args):
                     out = {}
                     out["img"] = block(args["img"], context=args["txt"], attention_mask=args["attention_mask"], timestep=args["vec"], pe=args["pe"], transformer_options=args["transformer_options"])
@@ -494,16 +921,28 @@ class LTXVModel(torch.nn.Module):
                     transformer_options=transformer_options,
                 )
 
-        # 3. Output
+        return x
+
+    def _process_output(self, x, embedded_timestep, keyframe_idxs, **kwargs):
+        """Process output for LTXV."""
+        # Apply scale-shift modulation
         scale_shift_values = (
             self.scale_shift_table[None, None].to(device=x.device, dtype=x.dtype) + embedded_timestep[:, :, None]
         )
         shift, scale = scale_shift_values[:, :, 0], scale_shift_values[:, :, 1]
+
         x = self.norm_out(x)
-        # Modulation
         x = x * (1 + scale) + shift
         x = self.proj_out(x)
 
+        if keyframe_idxs is not None:
+            grid_mask = kwargs["grid_mask"]
+            orig_patchified_shape = kwargs["orig_patchified_shape"]
+            full_x = torch.zeros(orig_patchified_shape, dtype=x.dtype, device=x.device)
+            full_x[:, grid_mask, :] = x
+            x = full_x
+        # Unpatchify to restore original dimensions
+        orig_shape = kwargs["orig_shape"]
         x = self.patchifier.unpatchify(
             latents=x,
             output_height=orig_shape[3],

comfy/ldm/lightricks/symmetric_patchifier.py

@@ -21,20 +21,23 @@ def latent_to_pixel_coords(
     Returns:
         Tensor: A tensor of pixel coordinates corresponding to the input latent coordinates.
     """
+    shape = [1] * latent_coords.ndim
+    shape[1] = -1
     pixel_coords = (
         latent_coords
-        * torch.tensor(scale_factors, device=latent_coords.device)[None, :, None]
+        * torch.tensor(scale_factors, device=latent_coords.device).view(*shape)
     )
     if causal_fix:
         # Fix temporal scale for first frame to 1 due to causality
-        pixel_coords[:, 0] = (pixel_coords[:, 0] + 1 - scale_factors[0]).clamp(min=0)
+        pixel_coords[:, 0, ...] = (pixel_coords[:, 0, ...] + 1 - scale_factors[0]).clamp(min=0)
     return pixel_coords
 
 
 class Patchifier(ABC):
-    def __init__(self, patch_size: int):
+    def __init__(self, patch_size: int, start_end: bool=False):
         super().__init__()
         self._patch_size = (1, patch_size, patch_size)
+        self.start_end = start_end
 
     @abstractmethod
     def patchify(
@@ -71,11 +74,23 @@ class Patchifier(ABC):
             torch.arange(0, latent_width, self._patch_size[2], device=device),
             indexing="ij",
         )
-        latent_sample_coords = torch.stack(latent_sample_coords, dim=0)
-        latent_coords = latent_sample_coords.unsqueeze(0).repeat(batch_size, 1, 1, 1, 1)
-        latent_coords = rearrange(
-            latent_coords, "b c f h w -> b c (f h w)", b=batch_size
+        latent_sample_coords_start = torch.stack(latent_sample_coords, dim=0)
+        delta = torch.tensor(self._patch_size, device=latent_sample_coords_start.device, dtype=latent_sample_coords_start.dtype)[:, None, None, None]
+        latent_sample_coords_end = latent_sample_coords_start + delta
+
+        latent_sample_coords_start = latent_sample_coords_start.unsqueeze(0).repeat(batch_size, 1, 1, 1, 1)
+        latent_sample_coords_start = rearrange(
+            latent_sample_coords_start, "b c f h w -> b c (f h w)", b=batch_size
         )
+        if self.start_end:
+            latent_sample_coords_end = latent_sample_coords_end.unsqueeze(0).repeat(batch_size, 1, 1, 1, 1)
+            latent_sample_coords_end = rearrange(
+                latent_sample_coords_end, "b c f h w -> b c (f h w)", b=batch_size
+            )
+
+            latent_coords = torch.stack((latent_sample_coords_start, latent_sample_coords_end), dim=-1)
+        else:
+            latent_coords = latent_sample_coords_start
         return latent_coords
 
 
@@ -115,3 +130,61 @@ class SymmetricPatchifier(Patchifier):
             q=self._patch_size[2],
         )
         return latents
+
+
+class AudioPatchifier(Patchifier):
+    def __init__(self, patch_size: int,
+        sample_rate=16000,
+        hop_length=160,
+        audio_latent_downsample_factor=4,
+        is_causal=True,
+        start_end=False,
+        shift = 0
+    ):
+        super().__init__(patch_size, start_end=start_end)
+        self.hop_length = hop_length
+        self.sample_rate = sample_rate
+        self.audio_latent_downsample_factor = audio_latent_downsample_factor
+        self.is_causal = is_causal
+        self.shift = shift
+
+    def copy_with_shift(self, shift):
+        return AudioPatchifier(
+            self.patch_size, self.sample_rate, self.hop_length, self.audio_latent_downsample_factor,
+            self.is_causal, self.start_end, shift
+        )
+
+    def _get_audio_latent_time_in_sec(self, start_latent, end_latent: int, dtype: torch.dtype, device=torch.device):
+        audio_latent_frame = torch.arange(start_latent, end_latent, dtype=dtype, device=device)
+        audio_mel_frame = audio_latent_frame * self.audio_latent_downsample_factor
+        if self.is_causal:
+            audio_mel_frame = (audio_mel_frame + 1 - self.audio_latent_downsample_factor).clip(min=0)
+        return audio_mel_frame * self.hop_length / self.sample_rate
+
+
+    def patchify(self, audio_latents: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
+        # audio_latents: (batch, channels, time, freq)
+        b, _, t, _ = audio_latents.shape
+        audio_latents = rearrange(
+            audio_latents,
+            "b c t f -> b t (c f)",
+        )
+
+        audio_latents_start_timings = self._get_audio_latent_time_in_sec(self.shift, t + self.shift, torch.float32, audio_latents.device)
+        audio_latents_start_timings = audio_latents_start_timings.unsqueeze(0).expand(b, -1).unsqueeze(1)
+
+        if self.start_end:
+            audio_latents_end_timings = self._get_audio_latent_time_in_sec(self.shift + 1, t + self.shift + 1, torch.float32, audio_latents.device)
+            audio_latents_end_timings = audio_latents_end_timings.unsqueeze(0).expand(b, -1).unsqueeze(1)
+
+            audio_latents_timings = torch.stack([audio_latents_start_timings, audio_latents_end_timings], dim=-1)
+        else:
+            audio_latents_timings = audio_latents_start_timings
+        return audio_latents, audio_latents_timings
+
+    def unpatchify(self, audio_latents: torch.Tensor, channels: int, freq: int) -> torch.Tensor:
+        # audio_latents: (batch, time, freq * channels)
+        audio_latents = rearrange(
+            audio_latents, "b t (c f) -> b c t f", c=channels, f=freq
+        )
+        return audio_latents

comfy/ldm/lightricks/vae/audio_vae.py

@@ -0,0 +1,279 @@
+import json
+from dataclasses import dataclass
+import math
+import torch
+import torchaudio
+
+import comfy.model_management
+import comfy.model_patcher
+import comfy.utils as utils
+from comfy.ldm.mmaudio.vae.distributions import DiagonalGaussianDistribution
+from comfy.ldm.lightricks.symmetric_patchifier import AudioPatchifier
+from comfy.ldm.lightricks.vae.causal_audio_autoencoder import (
+    CausalityAxis,
+    CausalAudioAutoencoder,
+)
+from comfy.ldm.lightricks.vocoders.vocoder import Vocoder
+
+LATENT_DOWNSAMPLE_FACTOR = 4
+
+
+@dataclass(frozen=True)
+class AudioVAEComponentConfig:
+    """Container for model component configuration extracted from metadata."""
+
+    autoencoder: dict
+    vocoder: dict
+
+    @classmethod
+    def from_metadata(cls, metadata: dict) -> "AudioVAEComponentConfig":
+        assert metadata is not None and "config" in metadata, "Metadata is required for audio VAE"
+
+        raw_config = metadata["config"]
+        if isinstance(raw_config, str):
+            parsed_config = json.loads(raw_config)
+        else:
+            parsed_config = raw_config
+
+        audio_config = parsed_config.get("audio_vae")
+        vocoder_config = parsed_config.get("vocoder")
+
+        assert audio_config is not None, "Audio VAE config is required for audio VAE"
+        assert vocoder_config is not None, "Vocoder config is required for audio VAE"
+
+        return cls(autoencoder=audio_config, vocoder=vocoder_config)
+
+
+class ModelDeviceManager:
+    """Manages device placement and GPU residency for the composed model."""
+
+    def __init__(self, module: torch.nn.Module):
+        load_device = comfy.model_management.get_torch_device()
+        offload_device = comfy.model_management.vae_offload_device()
+        self.patcher = comfy.model_patcher.ModelPatcher(module, load_device, offload_device)
+
+    def ensure_model_loaded(self) -> None:
+        comfy.model_management.free_memory(
+            self.patcher.model_size(),
+            self.patcher.load_device,
+        )
+        comfy.model_management.load_model_gpu(self.patcher)
+
+    def move_to_load_device(self, tensor: torch.Tensor) -> torch.Tensor:
+        return tensor.to(self.patcher.load_device)
+
+    @property
+    def load_device(self):
+        return self.patcher.load_device
+
+
+class AudioLatentNormalizer:
+    """Applies per-channel statistics in patch space and restores original layout."""
+
+    def __init__(self, patchfier: AudioPatchifier, statistics_processor: torch.nn.Module):
+        self.patchifier = patchfier
+        self.statistics = statistics_processor
+
+    def normalize(self, latents: torch.Tensor) -> torch.Tensor:
+        channels = latents.shape[1]
+        freq = latents.shape[3]
+        patched, _ = self.patchifier.patchify(latents)
+        normalized = self.statistics.normalize(patched)
+        return self.patchifier.unpatchify(normalized, channels=channels, freq=freq)
+
+    def denormalize(self, latents: torch.Tensor) -> torch.Tensor:
+        channels = latents.shape[1]
+        freq = latents.shape[3]
+        patched, _ = self.patchifier.patchify(latents)
+        denormalized = self.statistics.un_normalize(patched)
+        return self.patchifier.unpatchify(denormalized, channels=channels, freq=freq)
+
+
+class AudioPreprocessor:
+    """Prepares raw waveforms for the autoencoder by matching training conditions."""
+
+    def __init__(self, target_sample_rate: int, mel_bins: int, mel_hop_length: int, n_fft: int):
+        self.target_sample_rate = target_sample_rate
+        self.mel_bins = mel_bins
+        self.mel_hop_length = mel_hop_length
+        self.n_fft = n_fft
+
+    def resample(self, waveform: torch.Tensor, source_rate: int) -> torch.Tensor:
+        if source_rate == self.target_sample_rate:
+            return waveform
+        return torchaudio.functional.resample(waveform, source_rate, self.target_sample_rate)
+
+    def waveform_to_mel(
+        self, waveform: torch.Tensor, waveform_sample_rate: int, device
+    ) -> torch.Tensor:
+        waveform = self.resample(waveform, waveform_sample_rate)
+
+        mel_transform = torchaudio.transforms.MelSpectrogram(
+            sample_rate=self.target_sample_rate,
+            n_fft=self.n_fft,
+            win_length=self.n_fft,
+            hop_length=self.mel_hop_length,
+            f_min=0.0,
+            f_max=self.target_sample_rate / 2.0,
+            n_mels=self.mel_bins,
+            window_fn=torch.hann_window,
+            center=True,
+            pad_mode="reflect",
+            power=1.0,
+            mel_scale="slaney",
+            norm="slaney",
+        ).to(device)
+
+        mel = mel_transform(waveform)
+        mel = torch.log(torch.clamp(mel, min=1e-5))
+        return mel.permute(0, 1, 3, 2).contiguous()
+
+
+class AudioVAE(torch.nn.Module):
+    """High-level Audio VAE wrapper exposing encode and decode entry points."""
+
+    def __init__(self, state_dict: dict, metadata: dict):
+        super().__init__()
+
+        component_config = AudioVAEComponentConfig.from_metadata(metadata)
+
+        vae_sd = utils.state_dict_prefix_replace(state_dict, {"audio_vae.": ""}, filter_keys=True)
+        vocoder_sd = utils.state_dict_prefix_replace(state_dict, {"vocoder.": ""}, filter_keys=True)
+
+        self.autoencoder = CausalAudioAutoencoder(config=component_config.autoencoder)
+        self.vocoder = Vocoder(config=component_config.vocoder)
+
+        self.autoencoder.load_state_dict(vae_sd, strict=False)
+        self.vocoder.load_state_dict(vocoder_sd, strict=False)
+
+        autoencoder_config = self.autoencoder.get_config()
+        self.normalizer = AudioLatentNormalizer(
+            AudioPatchifier(
+                patch_size=1,
+                audio_latent_downsample_factor=LATENT_DOWNSAMPLE_FACTOR,
+                sample_rate=autoencoder_config["sampling_rate"],
+                hop_length=autoencoder_config["mel_hop_length"],
+                is_causal=autoencoder_config["is_causal"],
+            ),
+            self.autoencoder.per_channel_statistics,
+        )
+
+        self.preprocessor = AudioPreprocessor(
+            target_sample_rate=autoencoder_config["sampling_rate"],
+            mel_bins=autoencoder_config["mel_bins"],
+            mel_hop_length=autoencoder_config["mel_hop_length"],
+            n_fft=autoencoder_config["n_fft"],
+        )
+
+        self.device_manager = ModelDeviceManager(self)
+
+    def encode(self, audio: dict) -> torch.Tensor:
+        """Encode a waveform dictionary into normalized latent tensors."""
+
+        waveform = audio["waveform"]
+        waveform_sample_rate = audio["sample_rate"]
+        input_device = waveform.device
+        # Ensure that Audio VAE is loaded on the correct device.
+        self.device_manager.ensure_model_loaded()
+
+        waveform = self.device_manager.move_to_load_device(waveform)
+        expected_channels = self.autoencoder.encoder.in_channels
+        if waveform.shape[1] != expected_channels:
+            if waveform.shape[1] == 1:
+                waveform = waveform.expand(-1, expected_channels, *waveform.shape[2:])
+            else:
+                raise ValueError(
+                    f"Input audio must have {expected_channels} channels, got {waveform.shape[1]}"
+                )
+
+        mel_spec = self.preprocessor.waveform_to_mel(
+            waveform, waveform_sample_rate, device=self.device_manager.load_device
+        )
+
+        latents = self.autoencoder.encode(mel_spec)
+        posterior = DiagonalGaussianDistribution(latents)
+        latent_mode = posterior.mode()
+
+        normalized = self.normalizer.normalize(latent_mode)
+        return normalized.to(input_device)
+
+    def decode(self, latents: torch.Tensor) -> torch.Tensor:
+        """Decode normalized latent tensors into an audio waveform."""
+        original_shape = latents.shape
+
+        # Ensure that Audio VAE is loaded on the correct device.
+        self.device_manager.ensure_model_loaded()
+
+        latents = self.device_manager.move_to_load_device(latents)
+        latents = self.normalizer.denormalize(latents)
+
+        target_shape = self.target_shape_from_latents(original_shape)
+        mel_spec = self.autoencoder.decode(latents, target_shape=target_shape)
+
+        waveform = self.run_vocoder(mel_spec)
+        return self.device_manager.move_to_load_device(waveform)
+
+    def target_shape_from_latents(self, latents_shape):
+        batch, _, time, _ = latents_shape
+        target_length = time * LATENT_DOWNSAMPLE_FACTOR
+        if self.autoencoder.causality_axis != CausalityAxis.NONE:
+            target_length -= LATENT_DOWNSAMPLE_FACTOR - 1
+        return (
+            batch,
+            self.autoencoder.decoder.out_ch,
+            target_length,
+            self.autoencoder.mel_bins,
+        )
+
+    def num_of_latents_from_frames(self, frames_number: int, frame_rate: int) -> int:
+        return math.ceil((float(frames_number) / frame_rate) * self.latents_per_second)
+
+    def run_vocoder(self, mel_spec: torch.Tensor) -> torch.Tensor:
+        audio_channels = self.autoencoder.decoder.out_ch
+        vocoder_input = mel_spec.transpose(2, 3)
+
+        if audio_channels == 1:
+            vocoder_input = vocoder_input.squeeze(1)
+        elif audio_channels != 2:
+            raise ValueError(f"Unsupported audio_channels: {audio_channels}")
+
+        return self.vocoder(vocoder_input)
+
+    @property
+    def sample_rate(self) -> int:
+        return int(self.autoencoder.sampling_rate)
+
+    @property
+    def mel_hop_length(self) -> int:
+        return int(self.autoencoder.mel_hop_length)
+
+    @property
+    def mel_bins(self) -> int:
+        return int(self.autoencoder.mel_bins)
+
+    @property
+    def latent_channels(self) -> int:
+        return int(self.autoencoder.decoder.z_channels)
+
+    @property
+    def latent_frequency_bins(self) -> int:
+        return int(self.mel_bins // LATENT_DOWNSAMPLE_FACTOR)
+
+    @property
+    def latents_per_second(self) -> float:
+        return self.sample_rate / self.mel_hop_length / LATENT_DOWNSAMPLE_FACTOR
+
+    @property
+    def output_sample_rate(self) -> int:
+        output_rate = getattr(self.vocoder, "output_sample_rate", None)
+        if output_rate is not None:
+            return int(output_rate)
+        upsample_factor = getattr(self.vocoder, "upsample_factor", None)
+        if upsample_factor is None:
+            raise AttributeError(
+                "Vocoder is missing upsample_factor; cannot infer output sample rate"
+            )
+        return int(self.sample_rate * upsample_factor / self.mel_hop_length)
+
+    def memory_required(self, input_shape):
+        return self.device_manager.patcher.model_size()

comfy/ldm/lightricks/vae/causal_audio_autoencoder.py

@@ -0,0 +1,909 @@
+from __future__ import annotations
+import torch
+from torch import nn
+from torch.nn import functional as F
+from typing import Optional
+from enum import Enum
+from .pixel_norm import PixelNorm
+import comfy.ops
+import logging
+
+ops = comfy.ops.disable_weight_init
+
+
+class StringConvertibleEnum(Enum):
+    """
+    Base enum class that provides string-to-enum conversion functionality.
+
+    This mixin adds a str_to_enum() class method that handles conversion from
+    strings, None, or existing enum instances with case-insensitive matching.
+    """
+
+    @classmethod
+    def str_to_enum(cls, value):
+        """
+        Convert a string, enum instance, or None to the appropriate enum member.
+
+        Args:
+            value: Can be an enum instance of this class, a string, or None
+
+        Returns:
+            Enum member of this class
+
+        Raises:
+            ValueError: If the value cannot be converted to a valid enum member
+        """
+        # Already an enum instance of this class
+        if isinstance(value, cls):
+            return value
+
+        # None maps to NONE member if it exists
+        if value is None:
+            if hasattr(cls, "NONE"):
+                return cls.NONE
+            raise ValueError(f"{cls.__name__} does not have a NONE member to map None to")
+
+        # String conversion (case-insensitive)
+        if isinstance(value, str):
+            value_lower = value.lower()
+
+            # Try to match against enum values
+            for member in cls:
+                # Handle members with None values
+                if member.value is None:
+                    if value_lower == "none":
+                        return member
+                # Handle members with string values
+                elif isinstance(member.value, str) and member.value.lower() == value_lower:
+                    return member
+
+            # Build helpful error message with valid values
+            valid_values = []
+            for member in cls:
+                if member.value is None:
+                    valid_values.append("none")
+                elif isinstance(member.value, str):
+                    valid_values.append(member.value)
+
+            raise ValueError(f"Invalid {cls.__name__} string: '{value}'. " f"Valid values are: {valid_values}")
+
+        raise ValueError(
+            f"Cannot convert type {type(value).__name__} to {cls.__name__} enum. "
+            f"Expected string, None, or {cls.__name__} instance."
+        )
+
+
+class AttentionType(StringConvertibleEnum):
+    """Enum for specifying the attention mechanism type."""
+
+    VANILLA = "vanilla"
+    LINEAR = "linear"
+    NONE = "none"
+
+
+class CausalityAxis(StringConvertibleEnum):
+    """Enum for specifying the causality axis in causal convolutions."""
+
+    NONE = None
+    WIDTH = "width"
+    HEIGHT = "height"
+    WIDTH_COMPATIBILITY = "width-compatibility"
+
+
+def Normalize(in_channels, *, num_groups=32, normtype="group"):
+    if normtype == "group":
+        return torch.nn.GroupNorm(num_groups=num_groups, num_channels=in_channels, eps=1e-6, affine=True)
+    elif normtype == "pixel":
+        return PixelNorm(dim=1, eps=1e-6)
+    else:
+        raise ValueError(f"Invalid normalization type: {normtype}")
+
+
+class CausalConv2d(nn.Module):
+    """
+    A causal 2D convolution.
+
+    This layer ensures that the output at time `t` only depends on inputs
+    at time `t` and earlier. It achieves this by applying asymmetric padding
+    to the time dimension (width) before the convolution.
+    """
+
+    def __init__(
+        self,
+        in_channels,
+        out_channels,
+        kernel_size,
+        stride=1,
+        dilation=1,
+        groups=1,
+        bias=True,
+        causality_axis: CausalityAxis = CausalityAxis.HEIGHT,
+    ):
+        super().__init__()
+
+        self.causality_axis = causality_axis
+
+        # Ensure kernel_size and dilation are tuples
+        kernel_size = nn.modules.utils._pair(kernel_size)
+        dilation = nn.modules.utils._pair(dilation)
+
+        # Calculate padding dimensions
+        pad_h = (kernel_size[0] - 1) * dilation[0]
+        pad_w = (kernel_size[1] - 1) * dilation[1]
+
+        # The padding tuple for F.pad is (pad_left, pad_right, pad_top, pad_bottom)
+        match self.causality_axis:
+            case CausalityAxis.NONE:
+                self.padding = (pad_w // 2, pad_w - pad_w // 2, pad_h // 2, pad_h - pad_h // 2)
+            case CausalityAxis.WIDTH | CausalityAxis.WIDTH_COMPATIBILITY:
+                self.padding = (pad_w, 0, pad_h // 2, pad_h - pad_h // 2)
+            case CausalityAxis.HEIGHT:
+                self.padding = (pad_w // 2, pad_w - pad_w // 2, pad_h, 0)
+            case _:
+                raise ValueError(f"Invalid causality_axis: {causality_axis}")
+
+        # The internal convolution layer uses no padding, as we handle it manually
+        self.conv = ops.Conv2d(
+            in_channels,
+            out_channels,
+            kernel_size,
+            stride=stride,
+            padding=0,
+            dilation=dilation,
+            groups=groups,
+            bias=bias,
+        )
+
+    def forward(self, x):
+        # Apply causal padding before convolution
+        x = F.pad(x, self.padding)
+        return self.conv(x)
+
+
+def make_conv2d(
+    in_channels,
+    out_channels,
+    kernel_size,
+    stride=1,
+    padding=None,
+    dilation=1,
+    groups=1,
+    bias=True,
+    causality_axis: Optional[CausalityAxis] = None,
+):
+    """
+    Create a 2D convolution layer that can be either causal or non-causal.
+
+    Args:
+        in_channels: Number of input channels
+        out_channels: Number of output channels
+        kernel_size: Size of the convolution kernel
+        stride: Convolution stride
+        padding: Padding (if None, will be calculated based on causal flag)
+        dilation: Dilation rate
+        groups: Number of groups for grouped convolution
+        bias: Whether to use bias
+        causality_axis: Dimension along which to apply causality.
+
+    Returns:
+        Either a regular Conv2d or CausalConv2d layer
+    """
+    if causality_axis is not None:
+        # For causal convolution, padding is handled internally by CausalConv2d
+        return CausalConv2d(in_channels, out_channels, kernel_size, stride, dilation, groups, bias, causality_axis)
+    else:
+        # For non-causal convolution, use symmetric padding if not specified
+        if padding is None:
+            if isinstance(kernel_size, int):
+                padding = kernel_size // 2
+            else:
+                padding = tuple(k // 2 for k in kernel_size)
+        return ops.Conv2d(
+            in_channels,
+            out_channels,
+            kernel_size,
+            stride,
+            padding,
+            dilation,
+            groups,
+            bias,
+        )
+
+
+class Upsample(nn.Module):
+    def __init__(self, in_channels, with_conv, causality_axis: CausalityAxis = CausalityAxis.HEIGHT):
+        super().__init__()
+        self.with_conv = with_conv
+        self.causality_axis = causality_axis
+        if self.with_conv:
+            self.conv = make_conv2d(in_channels, in_channels, kernel_size=3, stride=1, causality_axis=causality_axis)
+
+    def forward(self, x):
+        x = torch.nn.functional.interpolate(x, scale_factor=2.0, mode="nearest")
+        if self.with_conv:
+            x = self.conv(x)
+            # Drop FIRST element in the causal axis to undo encoder's padding, while keeping the length 1 + 2 * n.
+            # For example, if the input is [0, 1, 2], after interpolation, the output is [0, 0, 1, 1, 2, 2].
+            # The causal convolution will pad the first element as [-, -, 0, 0, 1, 1, 2, 2],
+            # So the output elements rely on the following windows:
+            # 0: [-,-,0]
+            # 1: [-,0,0]
+            # 2: [0,0,1]
+            # 3: [0,1,1]
+            # 4: [1,1,2]
+            # 5: [1,2,2]
+            # Notice that the first and second elements in the output rely only on the first element in the input,
+            # while all other elements rely on two elements in the input.
+            # So we can drop the first element to undo the padding (rather than the last element).
+            # This is a no-op for non-causal convolutions.
+            match self.causality_axis:
+                case CausalityAxis.NONE:
+                    pass  # x remains unchanged
+                case CausalityAxis.HEIGHT:
+                    x = x[:, :, 1:, :]
+                case CausalityAxis.WIDTH:
+                    x = x[:, :, :, 1:]
+                case CausalityAxis.WIDTH_COMPATIBILITY:
+                    pass  # x remains unchanged
+                case _:
+                    raise ValueError(f"Invalid causality_axis: {self.causality_axis}")
+
+        return x
+
+
+class Downsample(nn.Module):
+    """
+    A downsampling layer that can use either a strided convolution
+    or average pooling. Supports standard and causal padding for the
+    convolutional mode.
+    """
+
+    def __init__(self, in_channels, with_conv, causality_axis: CausalityAxis = CausalityAxis.WIDTH):
+        super().__init__()
+        self.with_conv = with_conv
+        self.causality_axis = causality_axis
+
+        if self.causality_axis != CausalityAxis.NONE and not self.with_conv:
+            raise ValueError("causality is only supported when `with_conv=True`.")
+
+        if self.with_conv:
+            # Do time downsampling here
+            # no asymmetric padding in torch conv, must do it ourselves
+            self.conv = ops.Conv2d(in_channels, in_channels, kernel_size=3, stride=2, padding=0)
+
+    def forward(self, x):
+        if self.with_conv:
+            # (pad_left, pad_right, pad_top, pad_bottom)
+            match self.causality_axis:
+                case CausalityAxis.NONE:
+                    pad = (0, 1, 0, 1)
+                case CausalityAxis.WIDTH:
+                    pad = (2, 0, 0, 1)
+                case CausalityAxis.HEIGHT:
+                    pad = (0, 1, 2, 0)
+                case CausalityAxis.WIDTH_COMPATIBILITY:
+                    pad = (1, 0, 0, 1)
+                case _:
+                    raise ValueError(f"Invalid causality_axis: {self.causality_axis}")
+
+            x = torch.nn.functional.pad(x, pad, mode="constant", value=0)
+            x = self.conv(x)
+        else:
+            # This branch is only taken if with_conv=False, which implies causality_axis is NONE.
+            x = torch.nn.functional.avg_pool2d(x, kernel_size=2, stride=2)
+
+        return x
+
+
+class ResnetBlock(nn.Module):
+    def __init__(
+        self,
+        *,
+        in_channels,
+        out_channels=None,
+        conv_shortcut=False,
+        dropout,
+        temb_channels=512,
+        norm_type="group",
+        causality_axis: CausalityAxis = CausalityAxis.HEIGHT,
+    ):
+        super().__init__()
+        self.causality_axis = causality_axis
+
+        if self.causality_axis != CausalityAxis.NONE and norm_type == "group":
+            raise ValueError("Causal ResnetBlock with GroupNorm is not supported.")
+        self.in_channels = in_channels
+        out_channels = in_channels if out_channels is None else out_channels
+        self.out_channels = out_channels
+        self.use_conv_shortcut = conv_shortcut
+
+        self.norm1 = Normalize(in_channels, normtype=norm_type)
+        self.non_linearity = nn.SiLU()
+        self.conv1 = make_conv2d(in_channels, out_channels, kernel_size=3, stride=1, causality_axis=causality_axis)
+        if temb_channels > 0:
+            self.temb_proj = ops.Linear(temb_channels, out_channels)
+        self.norm2 = Normalize(out_channels, normtype=norm_type)
+        self.dropout = torch.nn.Dropout(dropout)
+        self.conv2 = make_conv2d(out_channels, out_channels, kernel_size=3, stride=1, causality_axis=causality_axis)
+        if self.in_channels != self.out_channels:
+            if self.use_conv_shortcut:
+                self.conv_shortcut = make_conv2d(
+                    in_channels, out_channels, kernel_size=3, stride=1, causality_axis=causality_axis
+                )
+            else:
+                self.nin_shortcut = make_conv2d(
+                    in_channels, out_channels, kernel_size=1, stride=1, causality_axis=causality_axis
+                )
+
+    def forward(self, x, temb):
+        h = x
+        h = self.norm1(h)
+        h = self.non_linearity(h)
+        h = self.conv1(h)
+
+        if temb is not None:
+            h = h + self.temb_proj(self.non_linearity(temb))[:, :, None, None]
+
+        h = self.norm2(h)
+        h = self.non_linearity(h)
+        h = self.dropout(h)
+        h = self.conv2(h)
+
+        if self.in_channels != self.out_channels:
+            if self.use_conv_shortcut:
+                x = self.conv_shortcut(x)
+            else:
+                x = self.nin_shortcut(x)
+
+        return x + h
+
+
+class AttnBlock(nn.Module):
+    def __init__(self, in_channels, norm_type="group"):
+        super().__init__()
+        self.in_channels = in_channels
+
+        self.norm = Normalize(in_channels, normtype=norm_type)
+        self.q = ops.Conv2d(in_channels, in_channels, kernel_size=1, stride=1, padding=0)
+        self.k = ops.Conv2d(in_channels, in_channels, kernel_size=1, stride=1, padding=0)
+        self.v = ops.Conv2d(in_channels, in_channels, kernel_size=1, stride=1, padding=0)
+        self.proj_out = ops.Conv2d(in_channels, in_channels, kernel_size=1, stride=1, padding=0)
+
+    def forward(self, x):
+        h_ = x
+        h_ = self.norm(h_)
+        q = self.q(h_)
+        k = self.k(h_)
+        v = self.v(h_)
+
+        # compute attention
+        b, c, h, w = q.shape
+        q = q.reshape(b, c, h * w).contiguous()
+        q = q.permute(0, 2, 1).contiguous()  # b,hw,c
+        k = k.reshape(b, c, h * w).contiguous()  # b,c,hw
+        w_ = torch.bmm(q, k).contiguous()  # b,hw,hw    w[b,i,j]=sum_c q[b,i,c]k[b,c,j]
+        w_ = w_ * (int(c) ** (-0.5))
+        w_ = torch.nn.functional.softmax(w_, dim=2)
+
+        # attend to values
+        v = v.reshape(b, c, h * w).contiguous()
+        w_ = w_.permute(0, 2, 1).contiguous()  # b,hw,hw (first hw of k, second of q)
+        h_ = torch.bmm(v, w_).contiguous()  # b, c,hw (hw of q) h_[b,c,j] = sum_i v[b,c,i] w_[b,i,j]
+        h_ = h_.reshape(b, c, h, w).contiguous()
+
+        h_ = self.proj_out(h_)
+
+        return x + h_
+
+
+def make_attn(in_channels, attn_type="vanilla", norm_type="group"):
+    # Convert string to enum if needed
+    attn_type = AttentionType.str_to_enum(attn_type)
+
+    if attn_type != AttentionType.NONE:
+        logging.info(f"making attention of type '{attn_type.value}' with {in_channels} in_channels")
+    else:
+        logging.info(f"making identity attention with {in_channels} in_channels")
+
+    match attn_type:
+        case AttentionType.VANILLA:
+            return AttnBlock(in_channels, norm_type=norm_type)
+        case AttentionType.NONE:
+            return nn.Identity(in_channels)
+        case AttentionType.LINEAR:
+            raise NotImplementedError(f"Attention type {attn_type.value} is not supported yet.")
+        case _:
+            raise ValueError(f"Unknown attention type: {attn_type}")
+
+
+class Encoder(nn.Module):
+    def __init__(
+        self,
+        *,
+        ch,
+        out_ch,
+        ch_mult=(1, 2, 4, 8),
+        num_res_blocks,
+        attn_resolutions,
+        dropout=0.0,
+        resamp_with_conv=True,
+        in_channels,
+        resolution,
+        z_channels,
+        double_z=True,
+        attn_type="vanilla",
+        mid_block_add_attention=True,
+        norm_type="group",
+        causality_axis=CausalityAxis.WIDTH.value,
+        **ignore_kwargs,
+    ):
+        super().__init__()
+        self.ch = ch
+        self.temb_ch = 0
+        self.num_resolutions = len(ch_mult)
+        self.num_res_blocks = num_res_blocks
+        self.resolution = resolution
+        self.in_channels = in_channels
+        self.z_channels = z_channels
+        self.double_z = double_z
+        self.norm_type = norm_type
+        # Convert string to enum if needed (for config loading)
+        causality_axis = CausalityAxis.str_to_enum(causality_axis)
+        self.attn_type = AttentionType.str_to_enum(attn_type)
+
+        # downsampling
+        self.conv_in = make_conv2d(
+            in_channels,
+            self.ch,
+            kernel_size=3,
+            stride=1,
+            causality_axis=causality_axis,
+        )
+
+        self.non_linearity = nn.SiLU()
+
+        curr_res = resolution
+        in_ch_mult = (1,) + tuple(ch_mult)
+        self.in_ch_mult = in_ch_mult
+        self.down = nn.ModuleList()
+
+        for i_level in range(self.num_resolutions):
+            block = nn.ModuleList()
+            attn = nn.ModuleList()
+            block_in = ch * in_ch_mult[i_level]
+            block_out = ch * ch_mult[i_level]
+
+            for _ in range(self.num_res_blocks):
+                block.append(
+                    ResnetBlock(
+                        in_channels=block_in,
+                        out_channels=block_out,
+                        temb_channels=self.temb_ch,
+                        dropout=dropout,
+                        norm_type=self.norm_type,
+                        causality_axis=causality_axis,
+                    )
+                )
+                block_in = block_out
+                if curr_res in attn_resolutions:
+                    attn.append(make_attn(block_in, attn_type=self.attn_type, norm_type=self.norm_type))
+
+            down = nn.Module()
+            down.block = block
+            down.attn = attn
+            if i_level != self.num_resolutions - 1:
+                down.downsample = Downsample(block_in, resamp_with_conv, causality_axis=causality_axis)
+                curr_res = curr_res // 2
+            self.down.append(down)
+
+        # middle
+        self.mid = nn.Module()
+        self.mid.block_1 = ResnetBlock(
+            in_channels=block_in,
+            out_channels=block_in,
+            temb_channels=self.temb_ch,
+            dropout=dropout,
+            norm_type=self.norm_type,
+            causality_axis=causality_axis,
+        )
+        if mid_block_add_attention:
+            self.mid.attn_1 = make_attn(block_in, attn_type=self.attn_type, norm_type=self.norm_type)
+        else:
+            self.mid.attn_1 = nn.Identity()
+        self.mid.block_2 = ResnetBlock(
+            in_channels=block_in,
+            out_channels=block_in,
+            temb_channels=self.temb_ch,
+            dropout=dropout,
+            norm_type=self.norm_type,
+            causality_axis=causality_axis,
+        )
+
+        # end
+        self.norm_out = Normalize(block_in, normtype=self.norm_type)
+        self.conv_out = make_conv2d(
+            block_in,
+            2 * z_channels if double_z else z_channels,
+            kernel_size=3,
+            stride=1,
+            causality_axis=causality_axis,
+        )
+
+    def forward(self, x):
+        """
+        Forward pass through the encoder.
+
+        Args:
+            x: Input tensor of shape [batch, channels, time, n_mels]
+
+        Returns:
+            Encoded latent representation
+        """
+        feature_maps = [self.conv_in(x)]
+
+        # Process each resolution level (from high to low resolution)
+        for resolution_level in range(self.num_resolutions):
+            # Apply residual blocks at current resolution level
+            for block_idx in range(self.num_res_blocks):
+                # Apply ResNet block with optional timestep embedding
+                current_features = self.down[resolution_level].block[block_idx](feature_maps[-1], temb=None)
+
+                # Apply attention if configured for this resolution level
+                if len(self.down[resolution_level].attn) > 0:
+                    current_features = self.down[resolution_level].attn[block_idx](current_features)
+
+                # Store processed features
+                feature_maps.append(current_features)
+
+            # Downsample spatial dimensions (except at the final resolution level)
+            if resolution_level != self.num_resolutions - 1:
+                downsampled_features = self.down[resolution_level].downsample(feature_maps[-1])
+                feature_maps.append(downsampled_features)
+
+        # === MIDDLE PROCESSING PHASE ===
+        # Take the lowest resolution features for middle processing
+        bottleneck_features = feature_maps[-1]
+
+        # Apply first middle ResNet block
+        bottleneck_features = self.mid.block_1(bottleneck_features, temb=None)
+
+        # Apply middle attention block
+        bottleneck_features = self.mid.attn_1(bottleneck_features)
+
+        # Apply second middle ResNet block
+        bottleneck_features = self.mid.block_2(bottleneck_features, temb=None)
+
+        # === OUTPUT PHASE ===
+        # Normalize the bottleneck features
+        output_features = self.norm_out(bottleneck_features)
+
+        # Apply non-linearity (SiLU activation)
+        output_features = self.non_linearity(output_features)
+
+        # Final convolution to produce latent representation
+        # [batch, channels, time, n_mels] -> [batch, 2 * z_channels if double_z else z_channels, time, n_mels]
+        return self.conv_out(output_features)
+
+
+class Decoder(nn.Module):
+    def __init__(
+        self,
+        *,
+        ch,
+        out_ch,
+        ch_mult=(1, 2, 4, 8),
+        num_res_blocks,
+        attn_resolutions,
+        dropout=0.0,
+        resamp_with_conv=True,
+        in_channels,
+        resolution,
+        z_channels,
+        give_pre_end=False,
+        tanh_out=False,
+        attn_type="vanilla",
+        mid_block_add_attention=True,
+        norm_type="group",
+        causality_axis=CausalityAxis.WIDTH.value,
+        **ignorekwargs,
+    ):
+        super().__init__()
+        self.ch = ch
+        self.temb_ch = 0
+        self.num_resolutions = len(ch_mult)
+        self.num_res_blocks = num_res_blocks
+        self.resolution = resolution
+        self.in_channels = in_channels
+        self.out_ch = out_ch
+        self.give_pre_end = give_pre_end
+        self.tanh_out = tanh_out
+        self.norm_type = norm_type
+        self.z_channels = z_channels
+        # Convert string to enum if needed (for config loading)
+        causality_axis = CausalityAxis.str_to_enum(causality_axis)
+        self.attn_type = AttentionType.str_to_enum(attn_type)
+
+        # compute block_in and curr_res at lowest res
+        block_in = ch * ch_mult[self.num_resolutions - 1]
+        curr_res = resolution // 2 ** (self.num_resolutions - 1)
+        self.z_shape = (1, z_channels, curr_res, curr_res)
+
+        # z to block_in
+        self.conv_in = make_conv2d(z_channels, block_in, kernel_size=3, stride=1, causality_axis=causality_axis)
+
+        self.non_linearity = nn.SiLU()
+
+        # middle
+        self.mid = nn.Module()
+        self.mid.block_1 = ResnetBlock(
+            in_channels=block_in,
+            out_channels=block_in,
+            temb_channels=self.temb_ch,
+            dropout=dropout,
+            norm_type=self.norm_type,
+            causality_axis=causality_axis,
+        )
+        if mid_block_add_attention:
+            self.mid.attn_1 = make_attn(block_in, attn_type=self.attn_type, norm_type=self.norm_type)
+        else:
+            self.mid.attn_1 = nn.Identity()
+        self.mid.block_2 = ResnetBlock(
+            in_channels=block_in,
+            out_channels=block_in,
+            temb_channels=self.temb_ch,
+            dropout=dropout,
+            norm_type=self.norm_type,
+            causality_axis=causality_axis,
+        )
+
+        # upsampling
+        self.up = nn.ModuleList()
+        for i_level in reversed(range(self.num_resolutions)):
+            block = nn.ModuleList()
+            attn = nn.ModuleList()
+            block_out = ch * ch_mult[i_level]
+            for _ in range(self.num_res_blocks + 1):
+                block.append(
+                    ResnetBlock(
+                        in_channels=block_in,
+                        out_channels=block_out,
+                        temb_channels=self.temb_ch,
+                        dropout=dropout,
+                        norm_type=self.norm_type,
+                        causality_axis=causality_axis,
+                    )
+                )
+                block_in = block_out
+                if curr_res in attn_resolutions:
+                    attn.append(make_attn(block_in, attn_type=self.attn_type, norm_type=self.norm_type))
+            up = nn.Module()
+            up.block = block
+            up.attn = attn
+            if i_level != 0:
+                up.upsample = Upsample(block_in, resamp_with_conv, causality_axis=causality_axis)
+                curr_res = curr_res * 2
+            self.up.insert(0, up)  # prepend to get consistent order
+
+        # end
+        self.norm_out = Normalize(block_in, normtype=self.norm_type)
+        self.conv_out = make_conv2d(block_in, out_ch, kernel_size=3, stride=1, causality_axis=causality_axis)
+
+    def _adjust_output_shape(self, decoded_output, target_shape):
+        """
+        Adjust output shape to match target dimensions for variable-length audio.
+
+        This function handles the common case where decoded audio spectrograms need to be
+        resized to match a specific target shape.
+
+        Args:
+            decoded_output: Tensor of shape (batch, channels, time, frequency)
+            target_shape: Target shape tuple (batch, channels, time, frequency)
+
+        Returns:
+            Tensor adjusted to match target_shape exactly
+        """
+        # Current output shape: (batch, channels, time, frequency)
+        _, _, current_time, current_freq = decoded_output.shape
+        _, target_channels, target_time, target_freq = target_shape
+
+        # Step 1: Crop first to avoid exceeding target dimensions
+        decoded_output = decoded_output[
+            :, :target_channels, : min(current_time, target_time), : min(current_freq, target_freq)
+        ]
+
+        # Step 2: Calculate padding needed for time and frequency dimensions
+        time_padding_needed = target_time - decoded_output.shape[2]
+        freq_padding_needed = target_freq - decoded_output.shape[3]
+
+        # Step 3: Apply padding if needed
+        if time_padding_needed > 0 or freq_padding_needed > 0:
+            # PyTorch padding format: (pad_left, pad_right, pad_top, pad_bottom)
+            # For audio: pad_left/right = frequency, pad_top/bottom = time
+            padding = (
+                0,
+                max(freq_padding_needed, 0),  # frequency padding (left, right)
+                0,
+                max(time_padding_needed, 0),  # time padding (top, bottom)
+            )
+            decoded_output = F.pad(decoded_output, padding)
+
+        # Step 4: Final safety crop to ensure exact target shape
+        decoded_output = decoded_output[:, :target_channels, :target_time, :target_freq]
+
+        return decoded_output
+
+    def get_config(self):
+        return {
+            "ch": self.ch,
+            "out_ch": self.out_ch,
+            "ch_mult": self.ch_mult,
+            "num_res_blocks": self.num_res_blocks,
+            "in_channels": self.in_channels,
+            "resolution": self.resolution,
+            "z_channels": self.z_channels,
+        }
+
+    def forward(self, latent_features, target_shape=None):
+        """
+        Decode latent features back to audio spectrograms.
+
+        Args:
+            latent_features: Encoded latent representation of shape (batch, channels, height, width)
+            target_shape: Optional target output shape (batch, channels, time, frequency)
+                         If provided, output will be cropped/padded to match this shape
+
+        Returns:
+            Reconstructed audio spectrogram of shape (batch, channels, time, frequency)
+        """
+        assert target_shape is not None, "Target shape is required for CausalAudioAutoencoder Decoder"
+
+        # Transform latent features to decoder's internal feature dimension
+        hidden_features = self.conv_in(latent_features)
+
+        # Middle processing
+        hidden_features = self.mid.block_1(hidden_features, temb=None)
+        hidden_features = self.mid.attn_1(hidden_features)
+        hidden_features = self.mid.block_2(hidden_features, temb=None)
+
+        # Upsampling
+        # Progressively increase spatial resolution from lowest to highest
+        for resolution_level in reversed(range(self.num_resolutions)):
+            # Apply residual blocks at current resolution level
+            for block_index in range(self.num_res_blocks + 1):
+                hidden_features = self.up[resolution_level].block[block_index](hidden_features, temb=None)
+
+                if len(self.up[resolution_level].attn) > 0:
+                    hidden_features = self.up[resolution_level].attn[block_index](hidden_features)
+
+            if resolution_level != 0:
+                hidden_features = self.up[resolution_level].upsample(hidden_features)
+
+        # Output
+        if self.give_pre_end:
+            # Return intermediate features before final processing (for debugging/analysis)
+            decoded_output = hidden_features
+        else:
+            # Standard output path: normalize, activate, and convert to output channels
+            # Final normalization layer
+            hidden_features = self.norm_out(hidden_features)
+
+            # Apply SiLU (Swish) activation function
+            hidden_features = self.non_linearity(hidden_features)
+
+            # Final convolution to map to output channels (typically 2 for stereo audio)
+            decoded_output = self.conv_out(hidden_features)
+
+            # Optional tanh activation to bound output values to [-1, 1] range
+            if self.tanh_out:
+                decoded_output = torch.tanh(decoded_output)
+
+        # Adjust shape for audio data
+        if target_shape is not None:
+            decoded_output = self._adjust_output_shape(decoded_output, target_shape)
+
+        return decoded_output
+
+
+class processor(nn.Module):
+    def __init__(self):
+        super().__init__()
+        self.register_buffer("std-of-means", torch.empty(128))
+        self.register_buffer("mean-of-means", torch.empty(128))
+
+    def un_normalize(self, x):
+        return (x * self.get_buffer("std-of-means").to(x)) + self.get_buffer("mean-of-means").to(x)
+
+    def normalize(self, x):
+        return (x - self.get_buffer("mean-of-means").to(x)) / self.get_buffer("std-of-means").to(x)
+
+
+class CausalAudioAutoencoder(nn.Module):
+    def __init__(self, config=None):
+        super().__init__()
+
+        if config is None:
+            config = self._guess_config()
+
+        # Extract encoder and decoder configs from the new format
+        model_config = config.get("model", {}).get("params", {})
+        variables_config = config.get("variables", {})
+
+        self.sampling_rate = variables_config.get(
+            "sampling_rate",
+            model_config.get("sampling_rate", config.get("sampling_rate", 16000)),
+        )
+        encoder_config = model_config.get("encoder", model_config.get("ddconfig", {}))
+        decoder_config = model_config.get("decoder", encoder_config)
+
+        # Load mel spectrogram parameters
+        self.mel_bins = encoder_config.get("mel_bins", 64)
+        self.mel_hop_length = model_config.get("preprocessing", {}).get("stft", {}).get("hop_length", 160)
+        self.n_fft = model_config.get("preprocessing", {}).get("stft", {}).get("filter_length", 1024)
+
+        # Store causality configuration at VAE level (not just in encoder internals)
+        causality_axis_value = encoder_config.get("causality_axis", CausalityAxis.WIDTH.value)
+        self.causality_axis = CausalityAxis.str_to_enum(causality_axis_value)
+        self.is_causal = self.causality_axis == CausalityAxis.HEIGHT
+
+        self.encoder = Encoder(**encoder_config)
+        self.decoder = Decoder(**decoder_config)
+
+        self.per_channel_statistics = processor()
+
+    def _guess_config(self):
+        encoder_config = {
+            # Required parameters - based on ltx-video-av-1679000 model metadata
+            "ch": 128,
+            "out_ch": 8,
+            "ch_mult": [1, 2, 4],  # Based on metadata: [1, 2, 4] not [1, 2, 4, 8]
+            "num_res_blocks": 2,
+            "attn_resolutions": [],  # Based on metadata: empty list, no attention
+            "dropout": 0.0,
+            "resamp_with_conv": True,
+            "in_channels": 2,  # stereo
+            "resolution": 256,
+            "z_channels": 8,
+            "double_z": True,
+            "attn_type": "vanilla",
+            "mid_block_add_attention": False,  # Based on metadata: false
+            "norm_type": "pixel",
+            "causality_axis": "height",  # Based on metadata
+            "mel_bins": 64,  # Based on metadata: mel_bins = 64
+        }
+
+        decoder_config = {
+            # Inherits encoder config, can override specific params
+            **encoder_config,
+            "out_ch": 2,  # Stereo audio output (2 channels)
+            "give_pre_end": False,
+            "tanh_out": False,
+        }
+
+        config = {
+            "_class_name": "CausalAudioAutoencoder",
+            "sampling_rate": 16000,
+            "model": {
+                "params": {
+                    "encoder": encoder_config,
+                    "decoder": decoder_config,
+                }
+            },
+        }
+
+        return config
+
+    def get_config(self):
+        return {
+            "sampling_rate": self.sampling_rate,
+            "mel_bins": self.mel_bins,
+            "mel_hop_length": self.mel_hop_length,
+            "n_fft": self.n_fft,
+            "causality_axis": self.causality_axis.value,
+            "is_causal": self.is_causal,
+        }
+
+    def encode(self, x):
+        return self.encoder(x)
+
+    def decode(self, x, target_shape=None):
+        return self.decoder(x, target_shape=target_shape)

comfy/ldm/lightricks/vae/causal_conv3d.py

@@ -1,11 +1,11 @@
 from typing import Tuple, Union
 
+import threading
 import torch
 import torch.nn as nn
 import comfy.ops
 ops = comfy.ops.disable_weight_init
 
-
 class CausalConv3d(nn.Module):
     def __init__(
         self,
@@ -42,23 +42,34 @@ class CausalConv3d(nn.Module):
             padding_mode=spatial_padding_mode,
             groups=groups,
         )
+        self.temporal_cache_state={}
 
     def forward(self, x, causal: bool = True):
-        if causal:
-            first_frame_pad = x[:, :, :1, :, :].repeat(
-                (1, 1, self.time_kernel_size - 1, 1, 1)
-            )
-            x = torch.concatenate((first_frame_pad, x), dim=2)
-        else:
-            first_frame_pad = x[:, :, :1, :, :].repeat(
-                (1, 1, (self.time_kernel_size - 1) // 2, 1, 1)
-            )
-            last_frame_pad = x[:, :, -1:, :, :].repeat(
-                (1, 1, (self.time_kernel_size - 1) // 2, 1, 1)
-            )
-            x = torch.concatenate((first_frame_pad, x, last_frame_pad), dim=2)
-        x = self.conv(x)
-        return x
+        tid = threading.get_ident()
+
+        cached, is_end = self.temporal_cache_state.get(tid, (None, False))
+        if cached is None:
+            padding_length = self.time_kernel_size - 1
+            if not causal:
+                padding_length = padding_length // 2
+            if x.shape[2] == 0:
+                return x
+            cached = x[:, :, :1, :, :].repeat((1, 1, padding_length, 1, 1))
+        pieces = [ cached, x ]
+        if is_end and not causal:
+            pieces.append(x[:, :, -1:, :, :].repeat((1, 1, (self.time_kernel_size - 1) // 2, 1, 1)))
+
+        needs_caching = not is_end
+        if needs_caching and x.shape[2] >= self.time_kernel_size - 1:
+            needs_caching = False
+            self.temporal_cache_state[tid] = (x[:, :, -(self.time_kernel_size - 1):, :, :], False)
+
+        x = torch.cat(pieces, dim=2)
+
+        if needs_caching:
+            self.temporal_cache_state[tid] = (x[:, :, -(self.time_kernel_size - 1):, :, :], False)
+
+        return self.conv(x) if x.shape[2] >= self.time_kernel_size else x[:, :, :0, :, :]
 
     @property
     def weight(self):

comfy/ldm/lightricks/vae/causal_video_autoencoder.py

@@ -1,4 +1,5 @@
 from __future__ import annotations
+import threading
 import torch
 from torch import nn
 from functools import partial
@@ -6,12 +7,35 @@ import math
 from einops import rearrange
 from typing import List, Optional, Tuple, Union
 from .conv_nd_factory import make_conv_nd, make_linear_nd
+from .causal_conv3d import CausalConv3d
 from .pixel_norm import PixelNorm
 from ..model import PixArtAlphaCombinedTimestepSizeEmbeddings
 import comfy.ops
+from comfy.ldm.modules.diffusionmodules.model import torch_cat_if_needed
 
 ops = comfy.ops.disable_weight_init
 
+def mark_conv3d_ended(module):
+    tid = threading.get_ident()
+    for _, m in module.named_modules():
+        if isinstance(m, CausalConv3d):
+            current = m.temporal_cache_state.get(tid, (None, False))
+            m.temporal_cache_state[tid] = (current[0], True)
+
+def split2(tensor, split_point, dim=2):
+    return torch.split(tensor, [split_point, tensor.shape[dim] - split_point], dim=dim)
+
+def add_exchange_cache(dest, cache_in, new_input, dim=2):
+    if dest is not None:
+        if cache_in is not None:
+            cache_to_dest = min(dest.shape[dim], cache_in.shape[dim])
+            lead_in_dest, dest = split2(dest, cache_to_dest, dim=dim)
+            lead_in_source, cache_in = split2(cache_in, cache_to_dest, dim=dim)
+            lead_in_dest.add_(lead_in_source)
+        body, new_input = split2(new_input, dest.shape[dim], dim)
+        dest.add_(body)
+    return torch_cat_if_needed([cache_in, new_input], dim=dim)
+
 class Encoder(nn.Module):
     r"""
     The `Encoder` layer of a variational autoencoder that encodes its input into a latent representation.
@@ -205,7 +229,7 @@ class Encoder(nn.Module):
 
         self.gradient_checkpointing = False
 
-    def forward(self, sample: torch.FloatTensor) -> torch.FloatTensor:
+    def forward_orig(self, sample: torch.FloatTensor) -> torch.FloatTensor:
         r"""The forward method of the `Encoder` class."""
 
         sample = patchify(sample, patch_size_hw=self.patch_size, patch_size_t=1)
@@ -254,6 +278,22 @@ class Encoder(nn.Module):
 
         return sample
 
+    def forward(self, *args, **kwargs):
+        #No encoder support so just flag the end so it doesnt use the cache.
+        mark_conv3d_ended(self)
+        try:
+            return self.forward_orig(*args, **kwargs)
+        finally:
+            tid = threading.get_ident()
+            for _, module in self.named_modules():
+                # ComfyUI doesn't thread this kind of stuff today, but just in case
+                # we key on the thread to make it thread safe.
+                tid = threading.get_ident()
+                if hasattr(module, "temporal_cache_state"):
+                    module.temporal_cache_state.pop(tid, None)
+
+
+MAX_CHUNK_SIZE=(128 * 1024 ** 2)
 
 class Decoder(nn.Module):
     r"""
@@ -341,18 +381,6 @@ class Decoder(nn.Module):
                     timestep_conditioning=timestep_conditioning,
                     spatial_padding_mode=spatial_padding_mode,
                 )
-            elif block_name == "attn_res_x":
-                block = UNetMidBlock3D(
-                    dims=dims,
-                    in_channels=input_channel,
-                    num_layers=block_params["num_layers"],
-                    resnet_groups=norm_num_groups,
-                    norm_layer=norm_layer,
-                    inject_noise=block_params.get("inject_noise", False),
-                    timestep_conditioning=timestep_conditioning,
-                    attention_head_dim=block_params["attention_head_dim"],
-                    spatial_padding_mode=spatial_padding_mode,
-                )
             elif block_name == "res_x_y":
                 output_channel = output_channel // block_params.get("multiplier", 2)
                 block = ResnetBlock3D(
@@ -428,8 +456,9 @@ class Decoder(nn.Module):
             )
             self.last_scale_shift_table = nn.Parameter(torch.empty(2, output_channel))
 
+
     # def forward(self, sample: torch.FloatTensor, target_shape) -> torch.FloatTensor:
-    def forward(
+    def forward_orig(
         self,
         sample: torch.FloatTensor,
         timestep: Optional[torch.Tensor] = None,
@@ -437,6 +466,7 @@ class Decoder(nn.Module):
         r"""The forward method of the `Decoder` class."""
         batch_size = sample.shape[0]
 
+        mark_conv3d_ended(self.conv_in)
         sample = self.conv_in(sample, causal=self.causal)
 
         checkpoint_fn = (
@@ -445,24 +475,12 @@ class Decoder(nn.Module):
             else lambda x: x
         )
 
-        scaled_timestep = None
+        timestep_shift_scale = None
         if self.timestep_conditioning:
             assert (
                 timestep is not None
             ), "should pass timestep with timestep_conditioning=True"
             scaled_timestep = timestep * self.timestep_scale_multiplier.to(dtype=sample.dtype, device=sample.device)
-
-        for up_block in self.up_blocks:
-            if self.timestep_conditioning and isinstance(up_block, UNetMidBlock3D):
-                sample = checkpoint_fn(up_block)(
-                    sample, causal=self.causal, timestep=scaled_timestep
-                )
-            else:
-                sample = checkpoint_fn(up_block)(sample, causal=self.causal)
-
-        sample = self.conv_norm_out(sample)
-
-        if self.timestep_conditioning:
             embedded_timestep = self.last_time_embedder(
                 timestep=scaled_timestep.flatten(),
                 resolution=None,
@@ -483,16 +501,62 @@ class Decoder(nn.Module):
                 embedded_timestep.shape[-2],
                 embedded_timestep.shape[-1],
             )
-            shift, scale = ada_values.unbind(dim=1)
-            sample = sample * (1 + scale) + shift
+            timestep_shift_scale = ada_values.unbind(dim=1)
 
-        sample = self.conv_act(sample)
-        sample = self.conv_out(sample, causal=self.causal)
+        output = []
+
+        def run_up(idx, sample, ended):
+            if idx >= len(self.up_blocks):
+                sample = self.conv_norm_out(sample)
+                if timestep_shift_scale is not None:
+                    shift, scale = timestep_shift_scale
+                    sample = sample * (1 + scale) + shift
+                sample = self.conv_act(sample)
+                if ended:
+                    mark_conv3d_ended(self.conv_out)
+                sample = self.conv_out(sample, causal=self.causal)
+                if sample is not None and sample.shape[2] > 0:
+                    output.append(sample)
+                return
+
+            up_block = self.up_blocks[idx]
+            if (ended):
+                mark_conv3d_ended(up_block)
+            if self.timestep_conditioning and isinstance(up_block, UNetMidBlock3D):
+                sample = checkpoint_fn(up_block)(
+                    sample, causal=self.causal, timestep=scaled_timestep
+                )
+            else:
+                sample = checkpoint_fn(up_block)(sample, causal=self.causal)
+
+            if sample is None or sample.shape[2] == 0:
+                return
+
+            total_bytes = sample.numel() * sample.element_size()
+            num_chunks = (total_bytes + MAX_CHUNK_SIZE - 1) // MAX_CHUNK_SIZE
+            samples = torch.chunk(sample, chunks=num_chunks, dim=2)
+
+            for chunk_idx, sample1 in enumerate(samples):
+                run_up(idx + 1, sample1, ended and chunk_idx == len(samples) - 1)
+
+        run_up(0, sample, True)
+        sample = torch.cat(output, dim=2)
 
         sample = unpatchify(sample, patch_size_hw=self.patch_size, patch_size_t=1)
 
         return sample
 
+    def forward(self, *args, **kwargs):
+        try:
+            return self.forward_orig(*args, **kwargs)
+        finally:
+            for _, module in self.named_modules():
+                #ComfyUI doesn't thread this kind of stuff today, but just incase
+                #we key on the thread to make it thread safe.
+                tid = threading.get_ident()
+                if hasattr(module, "temporal_cache_state"):
+                    module.temporal_cache_state.pop(tid, None)
+
 
 class UNetMidBlock3D(nn.Module):
     """
@@ -663,8 +727,22 @@ class DepthToSpaceUpsample(nn.Module):
         )
         self.residual = residual
         self.out_channels_reduction_factor = out_channels_reduction_factor
+        self.temporal_cache_state = {}
 
     def forward(self, x, causal: bool = True, timestep: Optional[torch.Tensor] = None):
+        tid = threading.get_ident()
+        cached, drop_first_conv, drop_first_res = self.temporal_cache_state.get(tid, (None, True, True))
+        y = self.conv(x, causal=causal)
+        y = rearrange(
+            y,
+            "b (c p1 p2 p3) d h w -> b c (d p1) (h p2) (w p3)",
+            p1=self.stride[0],
+            p2=self.stride[1],
+            p3=self.stride[2],
+        )
+        if self.stride[0] == 2 and y.shape[2] > 0 and drop_first_conv:
+            y = y[:, :, 1:, :, :]
+            drop_first_conv = False
         if self.residual:
             # Reshape and duplicate the input to match the output shape
             x_in = rearrange(
@@ -676,21 +754,20 @@ class DepthToSpaceUpsample(nn.Module):
             )
             num_repeat = math.prod(self.stride) // self.out_channels_reduction_factor
             x_in = x_in.repeat(1, num_repeat, 1, 1, 1)
-            if self.stride[0] == 2:
+            if self.stride[0] == 2 and x_in.shape[2] > 0 and drop_first_res:
                 x_in = x_in[:, :, 1:, :, :]
-        x = self.conv(x, causal=causal)
-        x = rearrange(
-            x,
-            "b (c p1 p2 p3) d h w -> b c (d p1) (h p2) (w p3)",
-            p1=self.stride[0],
-            p2=self.stride[1],
-            p3=self.stride[2],
-        )
-        if self.stride[0] == 2:
-            x = x[:, :, 1:, :, :]
-        if self.residual:
-            x = x + x_in
-        return x
+                drop_first_res = False
+
+            if y.shape[2] == 0:
+                y = None
+
+            cached = add_exchange_cache(y, cached, x_in, dim=2)
+            self.temporal_cache_state[tid] = (cached, drop_first_conv, drop_first_res)
+
+        else:
+            self.temporal_cache_state[tid] = (None, drop_first_conv, False)
+
+        return y
 
 class LayerNorm(nn.Module):
     def __init__(self, dim, eps, elementwise_affine=True) -> None:
@@ -807,6 +884,8 @@ class ResnetBlock3D(nn.Module):
                 torch.randn(4, in_channels) / in_channels**0.5
             )
 
+        self.temporal_cache_state={}
+
     def _feed_spatial_noise(
         self, hidden_states: torch.FloatTensor, per_channel_scale: torch.FloatTensor
     ) -> torch.FloatTensor:
@@ -880,9 +959,12 @@ class ResnetBlock3D(nn.Module):
 
         input_tensor = self.conv_shortcut(input_tensor)
 
-        output_tensor = input_tensor + hidden_states
+        tid = threading.get_ident()
+        cached = self.temporal_cache_state.get(tid, None)
+        cached = add_exchange_cache(hidden_states, cached, input_tensor, dim=2)
+        self.temporal_cache_state[tid] = cached
 
-        return output_tensor
+        return hidden_states
 
 
 def patchify(x, patch_size_hw, patch_size_t=1):

comfy/ldm/lightricks/vocoders/vocoder.py

@@ -0,0 +1,213 @@
+import torch
+import torch.nn.functional as F
+import torch.nn as nn
+import comfy.ops
+import numpy as np
+
+ops = comfy.ops.disable_weight_init
+
+LRELU_SLOPE = 0.1
+
+def get_padding(kernel_size, dilation=1):
+    return int((kernel_size * dilation - dilation) / 2)
+
+
+class ResBlock1(torch.nn.Module):
+    def __init__(self, channels, kernel_size=3, dilation=(1, 3, 5)):
+        super(ResBlock1, self).__init__()
+        self.convs1 = nn.ModuleList(
+            [
+                ops.Conv1d(
+                    channels,
+                    channels,
+                    kernel_size,
+                    1,
+                    dilation=dilation[0],
+                    padding=get_padding(kernel_size, dilation[0]),
+                ),
+                ops.Conv1d(
+                    channels,
+                    channels,
+                    kernel_size,
+                    1,
+                    dilation=dilation[1],
+                    padding=get_padding(kernel_size, dilation[1]),
+                ),
+                ops.Conv1d(
+                    channels,
+                    channels,
+                    kernel_size,
+                    1,
+                    dilation=dilation[2],
+                    padding=get_padding(kernel_size, dilation[2]),
+                ),
+            ]
+        )
+
+        self.convs2 = nn.ModuleList(
+            [
+                ops.Conv1d(
+                    channels,
+                    channels,
+                    kernel_size,
+                    1,
+                    dilation=1,
+                    padding=get_padding(kernel_size, 1),
+                ),
+                ops.Conv1d(
+                    channels,
+                    channels,
+                    kernel_size,
+                    1,
+                    dilation=1,
+                    padding=get_padding(kernel_size, 1),
+                ),
+                ops.Conv1d(
+                    channels,
+                    channels,
+                    kernel_size,
+                    1,
+                    dilation=1,
+                    padding=get_padding(kernel_size, 1),
+                ),
+            ]
+        )
+
+    def forward(self, x):
+        for c1, c2 in zip(self.convs1, self.convs2):
+            xt = F.leaky_relu(x, LRELU_SLOPE)
+            xt = c1(xt)
+            xt = F.leaky_relu(xt, LRELU_SLOPE)
+            xt = c2(xt)
+            x = xt + x
+        return x
+
+
+class ResBlock2(torch.nn.Module):
+    def __init__(self, channels, kernel_size=3, dilation=(1, 3)):
+        super(ResBlock2, self).__init__()
+        self.convs = nn.ModuleList(
+            [
+                ops.Conv1d(
+                    channels,
+                    channels,
+                    kernel_size,
+                    1,
+                    dilation=dilation[0],
+                    padding=get_padding(kernel_size, dilation[0]),
+                ),
+                ops.Conv1d(
+                    channels,
+                    channels,
+                    kernel_size,
+                    1,
+                    dilation=dilation[1],
+                    padding=get_padding(kernel_size, dilation[1]),
+                ),
+            ]
+        )
+
+    def forward(self, x):
+        for c in self.convs:
+            xt = F.leaky_relu(x, LRELU_SLOPE)
+            xt = c(xt)
+            x = xt + x
+        return x
+
+
+class Vocoder(torch.nn.Module):
+    """
+    Vocoder model for synthesizing audio from spectrograms, based on: https://github.com/jik876/hifi-gan.
+
+    """
+
+    def __init__(self, config=None):
+        super(Vocoder, self).__init__()
+
+        if config is None:
+            config = self.get_default_config()
+
+        resblock_kernel_sizes = config.get("resblock_kernel_sizes", [3, 7, 11])
+        upsample_rates = config.get("upsample_rates", [6, 5, 2, 2, 2])
+        upsample_kernel_sizes = config.get("upsample_kernel_sizes", [16, 15, 8, 4, 4])
+        resblock_dilation_sizes = config.get("resblock_dilation_sizes", [[1, 3, 5], [1, 3, 5], [1, 3, 5]])
+        upsample_initial_channel = config.get("upsample_initial_channel", 1024)
+        stereo = config.get("stereo", True)
+        resblock = config.get("resblock", "1")
+
+        self.output_sample_rate = config.get("output_sample_rate")
+        self.num_kernels = len(resblock_kernel_sizes)
+        self.num_upsamples = len(upsample_rates)
+        in_channels = 128 if stereo else 64
+        self.conv_pre = ops.Conv1d(in_channels, upsample_initial_channel, 7, 1, padding=3)
+        resblock_class = ResBlock1 if resblock == "1" else ResBlock2
+
+        self.ups = nn.ModuleList()
+        for i, (u, k) in enumerate(zip(upsample_rates, upsample_kernel_sizes)):
+            self.ups.append(
+                ops.ConvTranspose1d(
+                    upsample_initial_channel // (2**i),
+                    upsample_initial_channel // (2 ** (i + 1)),
+                    k,
+                    u,
+                    padding=(k - u) // 2,
+                )
+            )
+
+        self.resblocks = nn.ModuleList()
+        for i in range(len(self.ups)):
+            ch = upsample_initial_channel // (2 ** (i + 1))
+            for _, (k, d) in enumerate(zip(resblock_kernel_sizes, resblock_dilation_sizes)):
+                self.resblocks.append(resblock_class(ch, k, d))
+
+        out_channels = 2 if stereo else 1
+        self.conv_post = ops.Conv1d(ch, out_channels, 7, 1, padding=3)
+
+        self.upsample_factor = np.prod([self.ups[i].stride[0] for i in range(len(self.ups))])
+
+    def get_default_config(self):
+        """Generate default configuration for the vocoder."""
+
+        config = {
+            "resblock_kernel_sizes": [3, 7, 11],
+            "upsample_rates": [6, 5, 2, 2, 2],
+            "upsample_kernel_sizes": [16, 15, 8, 4, 4],
+            "resblock_dilation_sizes": [[1, 3, 5], [1, 3, 5], [1, 3, 5]],
+            "upsample_initial_channel": 1024,
+            "stereo": True,
+            "resblock": "1",
+        }
+
+        return config
+
+    def forward(self, x):
+        """
+        Forward pass of the vocoder.
+
+        Args:
+            x: Input spectrogram tensor. Can be:
+               - 3D: (batch_size, channels, time_steps) for mono
+               - 4D: (batch_size, 2, channels, time_steps) for stereo
+
+        Returns:
+            Audio tensor of shape (batch_size, out_channels, audio_length)
+        """
+        if x.dim() == 4:  # stereo
+            assert x.shape[1] == 2, "Input must have 2 channels for stereo"
+            x = torch.cat((x[:, 0, :, :], x[:, 1, :, :]), dim=1)
+        x = self.conv_pre(x)
+        for i in range(self.num_upsamples):
+            x = F.leaky_relu(x, LRELU_SLOPE)
+            x = self.ups[i](x)
+            xs = None
+            for j in range(self.num_kernels):
+                if xs is None:
+                    xs = self.resblocks[i * self.num_kernels + j](x)
+                else:
+                    xs += self.resblocks[i * self.num_kernels + j](x)
+            x = xs / self.num_kernels
+        x = F.leaky_relu(x)
+        x = self.conv_post(x)
+        x = torch.tanh(x)
+
+        return x

comfy/ldm/lumina/controlnet.py

@@ -0,0 +1,160 @@
+import torch
+from torch import nn
+
+from .model import JointTransformerBlock
+
+class ZImageControlTransformerBlock(JointTransformerBlock):
+    def __init__(
+        self,
+        layer_id: int,
+        dim: int,
+        n_heads: int,
+        n_kv_heads: int,
+        multiple_of: int,
+        ffn_dim_multiplier: float,
+        norm_eps: float,
+        qk_norm: bool,
+        modulation=True,
+        block_id=0,
+        operation_settings=None,
+    ):
+        super().__init__(layer_id, dim, n_heads, n_kv_heads, multiple_of, ffn_dim_multiplier, norm_eps, qk_norm, modulation, z_image_modulation=True, operation_settings=operation_settings)
+        self.block_id = block_id
+        if block_id == 0:
+            self.before_proj = operation_settings.get("operations").Linear(self.dim, self.dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.after_proj = operation_settings.get("operations").Linear(self.dim, self.dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+
+    def forward(self, c, x, **kwargs):
+        if self.block_id == 0:
+            c = self.before_proj(c) + x
+        c = super().forward(c, **kwargs)
+        c_skip = self.after_proj(c)
+        return c_skip, c
+
+class ZImage_Control(torch.nn.Module):
+    def __init__(
+        self,
+        dim: int = 3840,
+        n_heads: int = 30,
+        n_kv_heads: int = 30,
+        multiple_of: int = 256,
+        ffn_dim_multiplier: float = (8.0 / 3.0),
+        norm_eps: float = 1e-5,
+        qk_norm: bool = True,
+        n_control_layers=6,
+        control_in_dim=16,
+        additional_in_dim=0,
+        broken=False,
+        refiner_control=False,
+        dtype=None,
+        device=None,
+        operations=None,
+        **kwargs
+    ):
+        super().__init__()
+        operation_settings = {"operations": operations, "device": device, "dtype": dtype}
+
+        self.broken = broken
+        self.additional_in_dim = additional_in_dim
+        self.control_in_dim = control_in_dim
+        n_refiner_layers = 2
+        self.n_control_layers = n_control_layers
+        self.control_layers = nn.ModuleList(
+            [
+                ZImageControlTransformerBlock(
+                    i,
+                    dim,
+                    n_heads,
+                    n_kv_heads,
+                    multiple_of,
+                    ffn_dim_multiplier,
+                    norm_eps,
+                    qk_norm,
+                    block_id=i,
+                    operation_settings=operation_settings,
+                )
+                for i in range(self.n_control_layers)
+            ]
+        )
+
+        all_x_embedder = {}
+        patch_size = 2
+        f_patch_size = 1
+        x_embedder = operations.Linear(f_patch_size * patch_size * patch_size * (self.control_in_dim + self.additional_in_dim), dim, bias=True, device=device, dtype=dtype)
+        all_x_embedder[f"{patch_size}-{f_patch_size}"] = x_embedder
+
+        self.refiner_control = refiner_control
+
+        self.control_all_x_embedder = nn.ModuleDict(all_x_embedder)
+        if self.refiner_control:
+            self.control_noise_refiner = nn.ModuleList(
+                [
+                    ZImageControlTransformerBlock(
+                        layer_id,
+                        dim,
+                        n_heads,
+                        n_kv_heads,
+                        multiple_of,
+                        ffn_dim_multiplier,
+                        norm_eps,
+                        qk_norm,
+                        block_id=layer_id,
+                        operation_settings=operation_settings,
+                    )
+                    for layer_id in range(n_refiner_layers)
+                ]
+            )
+        else:
+            self.control_noise_refiner = nn.ModuleList(
+                [
+                    JointTransformerBlock(
+                        layer_id,
+                        dim,
+                        n_heads,
+                        n_kv_heads,
+                        multiple_of,
+                        ffn_dim_multiplier,
+                        norm_eps,
+                        qk_norm,
+                        modulation=True,
+                        z_image_modulation=True,
+                        operation_settings=operation_settings,
+                    )
+                    for layer_id in range(n_refiner_layers)
+                ]
+            )
+
+    def forward(self, cap_feats, control_context, x_freqs_cis, adaln_input):
+        patch_size = 2
+        f_patch_size = 1
+        pH = pW = patch_size
+        B, C, H, W = control_context.shape
+        control_context = self.control_all_x_embedder[f"{patch_size}-{f_patch_size}"](control_context.view(B, C, H // pH, pH, W // pW, pW).permute(0, 2, 4, 3, 5, 1).flatten(3).flatten(1, 2))
+
+        x_attn_mask = None
+        if not self.refiner_control:
+            for layer in self.control_noise_refiner:
+                control_context = layer(control_context, x_attn_mask, x_freqs_cis[:control_context.shape[0], :control_context.shape[1]], adaln_input)
+
+        return control_context
+
+    def forward_noise_refiner_block(self, layer_id, control_context, x, x_attn_mask, x_freqs_cis, adaln_input):
+        if self.refiner_control:
+            if self.broken:
+                if layer_id == 0:
+                    return self.control_layers[layer_id](control_context, x, x_mask=x_attn_mask, freqs_cis=x_freqs_cis[:control_context.shape[0], :control_context.shape[1]], adaln_input=adaln_input)
+                if layer_id > 0:
+                    out = None
+                    for i in range(1, len(self.control_layers)):
+                        o, control_context = self.control_layers[i](control_context, x, x_mask=x_attn_mask, freqs_cis=x_freqs_cis[:control_context.shape[0], :control_context.shape[1]], adaln_input=adaln_input)
+                        if out is None:
+                            out = o
+
+                    return (out, control_context)
+            else:
+                return self.control_noise_refiner[layer_id](control_context, x, x_mask=x_attn_mask, freqs_cis=x_freqs_cis[:control_context.shape[0], :control_context.shape[1]], adaln_input=adaln_input)
+        else:
+            return (None, control_context)
+
+    def forward_control_block(self, layer_id, control_context, x, x_attn_mask, x_freqs_cis, adaln_input):
+        return self.control_layers[layer_id](control_context, x, x_mask=x_attn_mask, freqs_cis=x_freqs_cis[:control_context.shape[0], :control_context.shape[1]], adaln_input=adaln_input)

comfy/ldm/lumina/model.py

@@ -11,16 +11,64 @@ import comfy.ldm.common_dit
 from comfy.ldm.modules.diffusionmodules.mmdit import TimestepEmbedder
 from comfy.ldm.modules.attention import optimized_attention_masked
 from comfy.ldm.flux.layers import EmbedND
+from comfy.ldm.flux.math import apply_rope
 import comfy.patcher_extension
+import comfy.utils
 
 
-def modulate(x, scale):
-    return x * (1 + scale.unsqueeze(1))
+def invert_slices(slices, length):
+    sorted_slices = sorted(slices)
+    result = []
+    current = 0
+
+    for start, end in sorted_slices:
+        if current < start:
+            result.append((current, start))
+        current = max(current, end)
+
+    if current < length:
+        result.append((current, length))
+
+    return result
+
+
+def modulate(x, scale, timestep_zero_index=None):
+    if timestep_zero_index is None:
+        return x * (1 + scale.unsqueeze(1))
+    else:
+        scale = (1 + scale.unsqueeze(1))
+        actual_batch = scale.size(0) // 2
+        slices = timestep_zero_index
+        invert = invert_slices(timestep_zero_index, x.shape[1])
+        for s in slices:
+            x[:, s[0]:s[1]] *= scale[actual_batch:]
+        for s in invert:
+            x[:, s[0]:s[1]] *= scale[:actual_batch]
+        return x
+
+
+def apply_gate(gate, x, timestep_zero_index=None):
+    if timestep_zero_index is None:
+        return gate * x
+    else:
+        actual_batch = gate.size(0) // 2
+
+        slices = timestep_zero_index
+        invert = invert_slices(timestep_zero_index, x.shape[1])
+        for s in slices:
+            x[:, s[0]:s[1]] *= gate[actual_batch:]
+        for s in invert:
+            x[:, s[0]:s[1]] *= gate[:actual_batch]
+        return x
 
 #############################################################################
 #                               Core NextDiT Model                              #
 #############################################################################
 
+def clamp_fp16(x):
+    if x.dtype == torch.float16:
+        return torch.nan_to_num(x, nan=0.0, posinf=65504, neginf=-65504)
+    return x
 
 class JointAttention(nn.Module):
     """Multi-head attention module."""
@@ -31,6 +79,7 @@ class JointAttention(nn.Module):
         n_heads: int,
         n_kv_heads: Optional[int],
         qk_norm: bool,
+        out_bias: bool = False,
         operation_settings={},
     ):
         """
@@ -59,7 +108,7 @@ class JointAttention(nn.Module):
         self.out = operation_settings.get("operations").Linear(
             n_heads * self.head_dim,
             dim,
-            bias=False,
+            bias=out_bias,
             device=operation_settings.get("device"),
             dtype=operation_settings.get("dtype"),
         )
@@ -70,35 +119,6 @@ class JointAttention(nn.Module):
         else:
             self.q_norm = self.k_norm = nn.Identity()
 
-    @staticmethod
-    def apply_rotary_emb(
-        x_in: torch.Tensor,
-        freqs_cis: torch.Tensor,
-    ) -> torch.Tensor:
-        """
-        Apply rotary embeddings to input tensors using the given frequency
-        tensor.
-
-        This function applies rotary embeddings to the given query 'xq' and
-        key 'xk' tensors using the provided frequency tensor 'freqs_cis'. The
-        input tensors are reshaped as complex numbers, and the frequency tensor
-        is reshaped for broadcasting compatibility. The resulting tensors
-        contain rotary embeddings and are returned as real tensors.
-
-        Args:
-            x_in (torch.Tensor): Query or Key tensor to apply rotary embeddings.
-            freqs_cis (torch.Tensor): Precomputed frequency tensor for complex
-                exponentials.
-
-        Returns:
-            Tuple[torch.Tensor, torch.Tensor]: Tuple of modified query tensor
-                and key tensor with rotary embeddings.
-        """
-
-        t_ = x_in.reshape(*x_in.shape[:-1], -1, 1, 2)
-        t_out = freqs_cis[..., 0] * t_[..., 0] + freqs_cis[..., 1] * t_[..., 1]
-        return t_out.reshape(*x_in.shape)
-
     def forward(
         self,
         x: torch.Tensor,
@@ -134,8 +154,7 @@ class JointAttention(nn.Module):
         xq = self.q_norm(xq)
         xk = self.k_norm(xk)
 
-        xq = JointAttention.apply_rotary_emb(xq, freqs_cis=freqs_cis)
-        xk = JointAttention.apply_rotary_emb(xk, freqs_cis=freqs_cis)
+        xq, xk = apply_rope(xq, xk, freqs_cis)
 
         n_rep = self.n_local_heads // self.n_local_kv_heads
         if n_rep >= 1:
@@ -197,7 +216,7 @@ class FeedForward(nn.Module):
 
     # @torch.compile
     def _forward_silu_gating(self, x1, x3):
-        return F.silu(x1) * x3
+        return clamp_fp16(F.silu(x1) * x3)
 
     def forward(self, x):
         return self.w2(self._forward_silu_gating(self.w1(x), self.w3(x)))
@@ -215,6 +234,8 @@ class JointTransformerBlock(nn.Module):
         norm_eps: float,
         qk_norm: bool,
         modulation=True,
+        z_image_modulation=False,
+        attn_out_bias=False,
         operation_settings={},
     ) -> None:
         """
@@ -235,10 +256,10 @@ class JointTransformerBlock(nn.Module):
         super().__init__()
         self.dim = dim
         self.head_dim = dim // n_heads
-        self.attention = JointAttention(dim, n_heads, n_kv_heads, qk_norm, operation_settings=operation_settings)
+        self.attention = JointAttention(dim, n_heads, n_kv_heads, qk_norm, out_bias=attn_out_bias, operation_settings=operation_settings)
         self.feed_forward = FeedForward(
             dim=dim,
-            hidden_dim=4 * dim,
+            hidden_dim=dim,
             multiple_of=multiple_of,
             ffn_dim_multiplier=ffn_dim_multiplier,
             operation_settings=operation_settings,
@@ -252,16 +273,27 @@ class JointTransformerBlock(nn.Module):
 
         self.modulation = modulation
         if modulation:
-            self.adaLN_modulation = nn.Sequential(
-                nn.SiLU(),
-                operation_settings.get("operations").Linear(
-                    min(dim, 1024),
-                    4 * dim,
-                    bias=True,
-                    device=operation_settings.get("device"),
-                    dtype=operation_settings.get("dtype"),
-                ),
-            )
+            if z_image_modulation:
+                self.adaLN_modulation = nn.Sequential(
+                    operation_settings.get("operations").Linear(
+                        min(dim, 256),
+                        4 * dim,
+                        bias=True,
+                        device=operation_settings.get("device"),
+                        dtype=operation_settings.get("dtype"),
+                    ),
+                )
+            else:
+                self.adaLN_modulation = nn.Sequential(
+                    nn.SiLU(),
+                    operation_settings.get("operations").Linear(
+                        min(dim, 1024),
+                        4 * dim,
+                        bias=True,
+                        device=operation_settings.get("device"),
+                        dtype=operation_settings.get("dtype"),
+                    ),
+                )
 
     def forward(
         self,
@@ -269,6 +301,7 @@ class JointTransformerBlock(nn.Module):
         x_mask: torch.Tensor,
         freqs_cis: torch.Tensor,
         adaln_input: Optional[torch.Tensor]=None,
+        timestep_zero_index=None,
         transformer_options={},
     ):
         """
@@ -287,28 +320,28 @@ class JointTransformerBlock(nn.Module):
             assert adaln_input is not None
             scale_msa, gate_msa, scale_mlp, gate_mlp = self.adaLN_modulation(adaln_input).chunk(4, dim=1)
 
-            x = x + gate_msa.unsqueeze(1).tanh() * self.attention_norm2(
-                self.attention(
-                    modulate(self.attention_norm1(x), scale_msa),
+            x = x + apply_gate(gate_msa.unsqueeze(1).tanh(), self.attention_norm2(
+                clamp_fp16(self.attention(
+                    modulate(self.attention_norm1(x), scale_msa, timestep_zero_index=timestep_zero_index),
                     x_mask,
                     freqs_cis,
                     transformer_options=transformer_options,
-                )
+                ))), timestep_zero_index=timestep_zero_index
             )
-            x = x + gate_mlp.unsqueeze(1).tanh() * self.ffn_norm2(
-                self.feed_forward(
-                    modulate(self.ffn_norm1(x), scale_mlp),
-                )
+            x = x + apply_gate(gate_mlp.unsqueeze(1).tanh(), self.ffn_norm2(
+                clamp_fp16(self.feed_forward(
+                    modulate(self.ffn_norm1(x), scale_mlp, timestep_zero_index=timestep_zero_index),
+                ))), timestep_zero_index=timestep_zero_index
             )
         else:
             assert adaln_input is None
             x = x + self.attention_norm2(
-                self.attention(
+                clamp_fp16(self.attention(
                     self.attention_norm1(x),
                     x_mask,
                     freqs_cis,
                     transformer_options=transformer_options,
-                )
+                ))
             )
             x = x + self.ffn_norm2(
                 self.feed_forward(
@@ -323,7 +356,7 @@ class FinalLayer(nn.Module):
     The final layer of NextDiT.
     """
 
-    def __init__(self, hidden_size, patch_size, out_channels, operation_settings={}):
+    def __init__(self, hidden_size, patch_size, out_channels, z_image_modulation=False, operation_settings={}):
         super().__init__()
         self.norm_final = operation_settings.get("operations").LayerNorm(
             hidden_size,
@@ -340,10 +373,15 @@ class FinalLayer(nn.Module):
             dtype=operation_settings.get("dtype"),
         )
 
+        if z_image_modulation:
+            min_mod = 256
+        else:
+            min_mod = 1024
+
         self.adaLN_modulation = nn.Sequential(
             nn.SiLU(),
             operation_settings.get("operations").Linear(
-                min(hidden_size, 1024),
+                min(hidden_size, min_mod),
                 hidden_size,
                 bias=True,
                 device=operation_settings.get("device"),
@@ -351,13 +389,37 @@ class FinalLayer(nn.Module):
             ),
         )
 
-    def forward(self, x, c):
+    def forward(self, x, c, timestep_zero_index=None):
         scale = self.adaLN_modulation(c)
-        x = modulate(self.norm_final(x), scale)
+        x = modulate(self.norm_final(x), scale, timestep_zero_index=timestep_zero_index)
         x = self.linear(x)
         return x
 
 
+def pad_zimage(feats, pad_token, pad_tokens_multiple):
+    pad_extra = (-feats.shape[1]) % pad_tokens_multiple
+    return torch.cat((feats, pad_token.to(device=feats.device, dtype=feats.dtype, copy=True).unsqueeze(0).repeat(feats.shape[0], pad_extra, 1)), dim=1), pad_extra
+
+
+def pos_ids_x(start_t, H_tokens, W_tokens, batch_size, device, transformer_options={}):
+    rope_options = transformer_options.get("rope_options", None)
+    h_scale = 1.0
+    w_scale = 1.0
+    h_start = 0
+    w_start = 0
+    if rope_options is not None:
+        h_scale = rope_options.get("scale_y", 1.0)
+        w_scale = rope_options.get("scale_x", 1.0)
+
+        h_start = rope_options.get("shift_y", 0.0)
+        w_start = rope_options.get("shift_x", 0.0)
+    x_pos_ids = torch.zeros((batch_size, H_tokens * W_tokens, 3), dtype=torch.float32, device=device)
+    x_pos_ids[:, :, 0] = start_t
+    x_pos_ids[:, :, 1] = (torch.arange(H_tokens, dtype=torch.float32, device=device) * h_scale + h_start).view(-1, 1).repeat(1, W_tokens).flatten()
+    x_pos_ids[:, :, 2] = (torch.arange(W_tokens, dtype=torch.float32, device=device) * w_scale + w_start).view(1, -1).repeat(H_tokens, 1).flatten()
+    return x_pos_ids
+
+
 class NextDiT(nn.Module):
     """
     Diffusion model with a Transformer backbone.
@@ -373,16 +435,23 @@ class NextDiT(nn.Module):
         n_heads: int = 32,
         n_kv_heads: Optional[int] = None,
         multiple_of: int = 256,
-        ffn_dim_multiplier: Optional[float] = None,
+        ffn_dim_multiplier: float = 4.0,
         norm_eps: float = 1e-5,
         qk_norm: bool = False,
         cap_feat_dim: int = 5120,
         axes_dims: List[int] = (16, 56, 56),
         axes_lens: List[int] = (1, 512, 512),
+        rope_theta=10000.0,
+        z_image_modulation=False,
+        time_scale=1.0,
+        pad_tokens_multiple=None,
+        clip_text_dim=None,
+        siglip_feat_dim=None,
         image_model=None,
         device=None,
         dtype=None,
         operations=None,
+        **kwargs,
     ) -> None:
         super().__init__()
         self.dtype = dtype
@@ -390,6 +459,8 @@ class NextDiT(nn.Module):
         self.in_channels = in_channels
         self.out_channels = in_channels
         self.patch_size = patch_size
+        self.time_scale = time_scale
+        self.pad_tokens_multiple = pad_tokens_multiple
 
         self.x_embedder = operation_settings.get("operations").Linear(
             in_features=patch_size * patch_size * in_channels,
@@ -411,6 +482,7 @@ class NextDiT(nn.Module):
                     norm_eps,
                     qk_norm,
                     modulation=True,
+                    z_image_modulation=z_image_modulation,
                     operation_settings=operation_settings,
                 )
                 for layer_id in range(n_refiner_layers)
@@ -434,7 +506,7 @@ class NextDiT(nn.Module):
             ]
         )
 
-        self.t_embedder = TimestepEmbedder(min(dim, 1024), **operation_settings)
+        self.t_embedder = TimestepEmbedder(min(dim, 1024), output_size=256 if z_image_modulation else None, **operation_settings)
         self.cap_embedder = nn.Sequential(
             operation_settings.get("operations").RMSNorm(cap_feat_dim, eps=norm_eps, elementwise_affine=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")),
             operation_settings.get("operations").Linear(
@@ -446,6 +518,31 @@ class NextDiT(nn.Module):
             ),
         )
 
+        self.clip_text_pooled_proj = None
+
+        if clip_text_dim is not None:
+            self.clip_text_dim = clip_text_dim
+            self.clip_text_pooled_proj = nn.Sequential(
+                operation_settings.get("operations").RMSNorm(clip_text_dim, eps=norm_eps, elementwise_affine=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")),
+                operation_settings.get("operations").Linear(
+                    clip_text_dim,
+                    clip_text_dim,
+                    bias=True,
+                    device=operation_settings.get("device"),
+                    dtype=operation_settings.get("dtype"),
+                ),
+            )
+            self.time_text_embed = nn.Sequential(
+                nn.SiLU(),
+                operation_settings.get("operations").Linear(
+                    min(dim, 1024) + clip_text_dim,
+                    min(dim, 1024),
+                    bias=True,
+                    device=operation_settings.get("device"),
+                    dtype=operation_settings.get("dtype"),
+                ),
+            )
+
         self.layers = nn.ModuleList(
             [
                 JointTransformerBlock(
@@ -457,18 +554,60 @@ class NextDiT(nn.Module):
                     ffn_dim_multiplier,
                     norm_eps,
                     qk_norm,
+                    z_image_modulation=z_image_modulation,
+                    attn_out_bias=False,
                     operation_settings=operation_settings,
                 )
                 for layer_id in range(n_layers)
             ]
         )
-        self.norm_final = operation_settings.get("operations").RMSNorm(dim, eps=norm_eps, elementwise_affine=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
-        self.final_layer = FinalLayer(dim, patch_size, self.out_channels, operation_settings=operation_settings)
+
+        if siglip_feat_dim is not None:
+            self.siglip_embedder = nn.Sequential(
+                operation_settings.get("operations").RMSNorm(siglip_feat_dim, eps=norm_eps, elementwise_affine=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")),
+                operation_settings.get("operations").Linear(
+                    siglip_feat_dim,
+                    dim,
+                    bias=True,
+                    device=operation_settings.get("device"),
+                    dtype=operation_settings.get("dtype"),
+                ),
+            )
+            self.siglip_refiner = nn.ModuleList(
+                [
+                    JointTransformerBlock(
+                        layer_id,
+                        dim,
+                        n_heads,
+                        n_kv_heads,
+                        multiple_of,
+                        ffn_dim_multiplier,
+                        norm_eps,
+                        qk_norm,
+                        modulation=False,
+                        operation_settings=operation_settings,
+                    )
+                    for layer_id in range(n_refiner_layers)
+                ]
+            )
+            self.siglip_pad_token = nn.Parameter(torch.empty((1, dim), device=device, dtype=dtype))
+        else:
+            self.siglip_embedder = None
+            self.siglip_refiner = None
+            self.siglip_pad_token = None
+
+        # This norm final is in the lumina 2.0 code but isn't actually used for anything.
+        # self.norm_final = operation_settings.get("operations").RMSNorm(dim, eps=norm_eps, elementwise_affine=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.final_layer = FinalLayer(dim, patch_size, self.out_channels, z_image_modulation=z_image_modulation, operation_settings=operation_settings)
+
+        if self.pad_tokens_multiple is not None:
+            self.x_pad_token = nn.Parameter(torch.empty((1, dim), device=device, dtype=dtype))
+            self.cap_pad_token = nn.Parameter(torch.empty((1, dim), device=device, dtype=dtype))
 
         assert (dim // n_heads) == sum(axes_dims)
         self.axes_dims = axes_dims
         self.axes_lens = axes_lens
-        self.rope_embedder = EmbedND(dim=dim // n_heads, theta=10000.0, axes_dim=axes_dims)
+        self.rope_embedder = EmbedND(dim=dim // n_heads, theta=rope_theta, axes_dim=axes_dims)
         self.dim = dim
         self.n_heads = n_heads
 
@@ -497,115 +636,168 @@ class NextDiT(nn.Module):
             imgs = torch.stack(imgs, dim=0)
         return imgs
 
-    def patchify_and_embed(
-        self, x: List[torch.Tensor] | torch.Tensor, cap_feats: torch.Tensor, cap_mask: torch.Tensor, t: torch.Tensor, num_tokens, transformer_options={}
-    ) -> Tuple[torch.Tensor, torch.Tensor, List[Tuple[int, int]], List[int], torch.Tensor]:
-        bsz = len(x)
-        pH = pW = self.patch_size
-        device = x[0].device
-        dtype = x[0].dtype
-
-        if cap_mask is not None:
-            l_effective_cap_len = cap_mask.sum(dim=1).tolist()
+    def embed_cap(self, cap_feats=None, offset=0, bsz=1, device=None, dtype=None):
+        if cap_feats is not None:
+            cap_feats = self.cap_embedder(cap_feats)
+            cap_feats_len = cap_feats.shape[1]
+            if self.pad_tokens_multiple is not None:
+                cap_feats, _ = pad_zimage(cap_feats, self.cap_pad_token, self.pad_tokens_multiple)
         else:
-            l_effective_cap_len = [num_tokens] * bsz
+            cap_feats_len = 0
+            cap_feats = self.cap_pad_token.to(device=device, dtype=dtype, copy=True).unsqueeze(0).repeat(bsz, self.pad_tokens_multiple, 1)
 
-        if cap_mask is not None and not torch.is_floating_point(cap_mask):
-            cap_mask = (cap_mask - 1).to(dtype) * torch.finfo(dtype).max
+        cap_pos_ids = torch.zeros(bsz, cap_feats.shape[1], 3, dtype=torch.float32, device=device)
+        cap_pos_ids[:, :, 0] = torch.arange(cap_feats.shape[1], dtype=torch.float32, device=device) + 1.0 + offset
+        embeds = (cap_feats,)
+        freqs_cis = (self.rope_embedder(cap_pos_ids).movedim(1, 2),)
+        return embeds, freqs_cis, cap_feats_len
 
-        img_sizes = [(img.size(1), img.size(2)) for img in x]
-        l_effective_img_len = [(H // pH) * (W // pW) for (H, W) in img_sizes]
+    def embed_all(self, x, cap_feats=None, siglip_feats=None, offset=0, omni=False, transformer_options={}):
+        bsz = 1
+        pH = pW = self.patch_size
+        device = x.device
+        embeds, freqs_cis, cap_feats_len = self.embed_cap(cap_feats, offset=offset, bsz=bsz, device=device, dtype=x.dtype)
 
-        max_seq_len = max(
-            (cap_len+img_len for cap_len, img_len in zip(l_effective_cap_len, l_effective_img_len))
-        )
-        max_cap_len = max(l_effective_cap_len)
-        max_img_len = max(l_effective_img_len)
+        if (not omni) or self.siglip_embedder is None:
+            cap_feats_len = embeds[0].shape[1] + offset
+            embeds += (None,)
+            freqs_cis += (None,)
+        else:
+            cap_feats_len += offset
+            if siglip_feats is not None:
+                b, h, w, c = siglip_feats.shape
+                siglip_feats = siglip_feats.permute(0, 3, 1, 2).reshape(b, h * w, c)
+                siglip_feats = self.siglip_embedder(siglip_feats)
+                siglip_pos_ids = torch.zeros((bsz, siglip_feats.shape[1], 3), dtype=torch.float32, device=device)
+                siglip_pos_ids[:, :, 0] = cap_feats_len + 2
+                siglip_pos_ids[:, :, 1] = (torch.linspace(0, h * 8 - 1, steps=h, dtype=torch.float32, device=device).floor()).view(-1, 1).repeat(1, w).flatten()
+                siglip_pos_ids[:, :, 2] = (torch.linspace(0, w * 8 - 1, steps=w, dtype=torch.float32, device=device).floor()).view(1, -1).repeat(h, 1).flatten()
+                if self.siglip_pad_token is not None:
+                    siglip_feats, pad_extra = pad_zimage(siglip_feats, self.siglip_pad_token, self.pad_tokens_multiple)  # TODO: double check
+                    siglip_pos_ids = torch.nn.functional.pad(siglip_pos_ids, (0, 0, 0, pad_extra))
+            else:
+                if self.siglip_pad_token is not None:
+                    siglip_feats = self.siglip_pad_token.to(device=device, dtype=x.dtype, copy=True).unsqueeze(0).repeat(bsz, self.pad_tokens_multiple, 1)
+                    siglip_pos_ids = torch.zeros((bsz, siglip_feats.shape[1], 3), dtype=torch.float32, device=device)
 
-        position_ids = torch.zeros(bsz, max_seq_len, 3, dtype=torch.float32, device=device)
+            if siglip_feats is None:
+                embeds += (None,)
+                freqs_cis += (None,)
+            else:
+                embeds += (siglip_feats,)
+                freqs_cis += (self.rope_embedder(siglip_pos_ids).movedim(1, 2),)
 
-        for i in range(bsz):
-            cap_len = l_effective_cap_len[i]
-            img_len = l_effective_img_len[i]
-            H, W = img_sizes[i]
-            H_tokens, W_tokens = H // pH, W // pW
-            assert H_tokens * W_tokens == img_len
+        B, C, H, W = x.shape
+        x = self.x_embedder(x.view(B, C, H // pH, pH, W // pW, pW).permute(0, 2, 4, 3, 5, 1).flatten(3).flatten(1, 2))
+        x_pos_ids = pos_ids_x(cap_feats_len + 1, H // pH, W // pW, bsz, device, transformer_options=transformer_options)
+        if self.pad_tokens_multiple is not None:
+            x, pad_extra = pad_zimage(x, self.x_pad_token, self.pad_tokens_multiple)
+            x_pos_ids = torch.nn.functional.pad(x_pos_ids, (0, 0, 0, pad_extra))
 
-            rope_options = transformer_options.get("rope_options", None)
-            h_scale = 1.0
-            w_scale = 1.0
-            h_start = 0
-            w_start = 0
-            if rope_options is not None:
-                h_scale = rope_options.get("scale_y", 1.0)
-                w_scale = rope_options.get("scale_x", 1.0)
+        embeds += (x,)
+        freqs_cis += (self.rope_embedder(x_pos_ids).movedim(1, 2),)
+        return embeds, freqs_cis, cap_feats_len + len(freqs_cis) - 1
 
-                h_start = rope_options.get("shift_y", 0.0)
-                w_start = rope_options.get("shift_x", 0.0)
 
-            position_ids[i, :cap_len, 0] = torch.arange(cap_len, dtype=torch.float32, device=device)
-            position_ids[i, cap_len:cap_len+img_len, 0] = cap_len
-            row_ids = (torch.arange(H_tokens, dtype=torch.float32, device=device) * h_scale + h_start).view(-1, 1).repeat(1, W_tokens).flatten()
-            col_ids = (torch.arange(W_tokens, dtype=torch.float32, device=device) * w_scale + w_start).view(1, -1).repeat(H_tokens, 1).flatten()
-            position_ids[i, cap_len:cap_len+img_len, 1] = row_ids
-            position_ids[i, cap_len:cap_len+img_len, 2] = col_ids
+    def patchify_and_embed(
+        self, x: torch.Tensor, cap_feats: torch.Tensor, cap_mask: torch.Tensor, t: torch.Tensor, num_tokens, ref_latents=[], ref_contexts=[], siglip_feats=[], transformer_options={}
+    ) -> Tuple[torch.Tensor, torch.Tensor, List[Tuple[int, int]], List[int], torch.Tensor]:
+        bsz = x.shape[0]
+        cap_mask = None  # TODO?
+        main_siglip = None
+        orig_x = x
 
-        freqs_cis = self.rope_embedder(position_ids).movedim(1, 2).to(dtype)
+        embeds = ([], [], [])
+        freqs_cis = ([], [], [])
+        leftover_cap = []
 
-        # build freqs_cis for cap and image individually
-        cap_freqs_cis_shape = list(freqs_cis.shape)
-        # cap_freqs_cis_shape[1] = max_cap_len
-        cap_freqs_cis_shape[1] = cap_feats.shape[1]
-        cap_freqs_cis = torch.zeros(*cap_freqs_cis_shape, device=device, dtype=freqs_cis.dtype)
+        start_t = 0
+        omni = len(ref_latents) > 0
+        if omni:
+            for i, ref in enumerate(ref_latents):
+                if i < len(ref_contexts):
+                    ref_con = ref_contexts[i]
+                else:
+                    ref_con = None
+                if i < len(siglip_feats):
+                    sig_feat = siglip_feats[i]
+                else:
+                    sig_feat = None
 
-        img_freqs_cis_shape = list(freqs_cis.shape)
-        img_freqs_cis_shape[1] = max_img_len
-        img_freqs_cis = torch.zeros(*img_freqs_cis_shape, device=device, dtype=freqs_cis.dtype)
+                out = self.embed_all(ref, ref_con, sig_feat, offset=start_t, omni=omni, transformer_options=transformer_options)
+                for i, e in enumerate(out[0]):
+                    if e is not None:
+                        embeds[i].append(comfy.utils.repeat_to_batch_size(e, bsz))
+                        freqs_cis[i].append(out[1][i])
+                start_t = out[2]
+            leftover_cap = ref_contexts[len(ref_latents):]
 
-        for i in range(bsz):
-            cap_len = l_effective_cap_len[i]
-            img_len = l_effective_img_len[i]
-            cap_freqs_cis[i, :cap_len] = freqs_cis[i, :cap_len]
-            img_freqs_cis[i, :img_len] = freqs_cis[i, cap_len:cap_len+img_len]
+        H, W = x.shape[-2], x.shape[-1]
+        img_sizes = [(H, W)] * bsz
+        out = self.embed_all(x, cap_feats, main_siglip, offset=start_t, omni=omni, transformer_options=transformer_options)
+        img_len = out[0][-1].shape[1]
+        cap_len = out[0][0].shape[1]
+        for i, e in enumerate(out[0]):
+            if e is not None:
+                e = comfy.utils.repeat_to_batch_size(e, bsz)
+                embeds[i].append(e)
+                freqs_cis[i].append(out[1][i])
+        start_t = out[2]
+
+        for cap in leftover_cap:
+            out = self.embed_cap(cap, offset=start_t, bsz=bsz, device=x.device, dtype=x.dtype)
+            cap_len += out[0][0].shape[1]
+            embeds[0].append(comfy.utils.repeat_to_batch_size(out[0][0], bsz))
+            freqs_cis[0].append(out[1][0])
+            start_t += out[2]
+
+        patches = transformer_options.get("patches", {})
 
         # refine context
+        cap_feats = torch.cat(embeds[0], dim=1)
+        cap_freqs_cis = torch.cat(freqs_cis[0], dim=1)
         for layer in self.context_refiner:
             cap_feats = layer(cap_feats, cap_mask, cap_freqs_cis, transformer_options=transformer_options)
 
-        # refine image
-        flat_x = []
-        for i in range(bsz):
-            img = x[i]
-            C, H, W = img.size()
-            img = img.view(C, H // pH, pH, W // pW, pW).permute(1, 3, 2, 4, 0).flatten(2).flatten(0, 1)
-            flat_x.append(img)
-        x = flat_x
-        padded_img_embed = torch.zeros(bsz, max_img_len, x[0].shape[-1], device=device, dtype=x[0].dtype)
-        padded_img_mask = torch.zeros(bsz, max_img_len, dtype=dtype, device=device)
-        for i in range(bsz):
-            padded_img_embed[i, :l_effective_img_len[i]] = x[i]
-            padded_img_mask[i, l_effective_img_len[i]:] = -torch.finfo(dtype).max
+        feats = (cap_feats,)
+        fc = (cap_freqs_cis,)
 
-        padded_img_embed = self.x_embedder(padded_img_embed)
-        padded_img_mask = padded_img_mask.unsqueeze(1)
-        for layer in self.noise_refiner:
-            padded_img_embed = layer(padded_img_embed, padded_img_mask, img_freqs_cis, t, transformer_options=transformer_options)
+        if omni and len(embeds[1]) > 0:
+            siglip_mask = None
+            siglip_feats_combined = torch.cat(embeds[1], dim=1)
+            siglip_feats_freqs_cis = torch.cat(freqs_cis[1], dim=1)
+            if self.siglip_refiner is not None:
+                for layer in self.siglip_refiner:
+                    siglip_feats_combined = layer(siglip_feats_combined, siglip_mask, siglip_feats_freqs_cis, transformer_options=transformer_options)
+            feats += (siglip_feats_combined,)
+            fc += (siglip_feats_freqs_cis,)
 
-        if cap_mask is not None:
-            mask = torch.zeros(bsz, max_seq_len, dtype=dtype, device=device)
-            mask[:, :max_cap_len] = cap_mask[:, :max_cap_len]
+        padded_img_mask = None
+        x = torch.cat(embeds[-1], dim=1)
+        fc_x = torch.cat(freqs_cis[-1], dim=1)
+        if omni:
+            timestep_zero_index = [(x.shape[1] - img_len, x.shape[1])]
         else:
-            mask = None
+            timestep_zero_index = None
 
-        padded_full_embed = torch.zeros(bsz, max_seq_len, self.dim, device=device, dtype=x[0].dtype)
-        for i in range(bsz):
-            cap_len = l_effective_cap_len[i]
-            img_len = l_effective_img_len[i]
+        x_input = x
+        for i, layer in enumerate(self.noise_refiner):
+            x = layer(x, padded_img_mask, fc_x, t, timestep_zero_index=timestep_zero_index, transformer_options=transformer_options)
+            if "noise_refiner" in patches:
+                for p in patches["noise_refiner"]:
+                    out = p({"img": x, "img_input": x_input, "txt": cap_feats, "pe": fc_x, "vec": t, "x": orig_x, "block_index": i, "transformer_options": transformer_options, "block_type": "noise_refiner"})
+                    if "img" in out:
+                        x = out["img"]
 
-            padded_full_embed[i, :cap_len] = cap_feats[i, :cap_len]
-            padded_full_embed[i, cap_len:cap_len+img_len] = padded_img_embed[i, :img_len]
+        padded_full_embed = torch.cat(feats + (x,), dim=1)
+        if timestep_zero_index is not None:
+            ind = padded_full_embed.shape[1] - x.shape[1]
+            timestep_zero_index = [(ind + x.shape[1] - img_len, ind + x.shape[1])]
+            timestep_zero_index.append((feats[0].shape[1] - cap_len, feats[0].shape[1]))
 
-        return padded_full_embed, mask, img_sizes, l_effective_cap_len, freqs_cis
+        mask = None
+        l_effective_cap_len = [padded_full_embed.shape[1] - img_len] * bsz
+        return padded_full_embed, mask, img_sizes, l_effective_cap_len, torch.cat(fc + (fc_x,), dim=1), timestep_zero_index
 
     def forward(self, x, timesteps, context, num_tokens, attention_mask=None, **kwargs):
         return comfy.patcher_extension.WrapperExecutor.new_class_executor(
@@ -615,7 +807,11 @@ class NextDiT(nn.Module):
         ).execute(x, timesteps, context, num_tokens, attention_mask, **kwargs)
 
     # def forward(self, x, t, cap_feats, cap_mask):
-    def _forward(self, x, timesteps, context, num_tokens, attention_mask=None, **kwargs):
+    def _forward(self, x, timesteps, context, num_tokens, attention_mask=None, ref_latents=[], ref_contexts=[], siglip_feats=[], transformer_options={}, **kwargs):
+        omni = len(ref_latents) > 0
+        if omni:
+            timesteps = torch.cat([timesteps * 0, timesteps], dim=0)
+
         t = 1.0 - timesteps
         cap_feats = context
         cap_mask = attention_mask
@@ -627,21 +823,38 @@ class NextDiT(nn.Module):
         y: (N,) tensor of text tokens/features
         """
 
-        t = self.t_embedder(t, dtype=x.dtype)  # (N, D)
+        t = self.t_embedder(t * self.time_scale, dtype=x.dtype)  # (N, D)
         adaln_input = t
 
-        cap_feats = self.cap_embedder(cap_feats)  # (N, L, D)  # todo check if able to batchify w.o. redundant compute
+        if self.clip_text_pooled_proj is not None:
+            pooled = kwargs.get("clip_text_pooled", None)
+            if pooled is not None:
+                pooled = self.clip_text_pooled_proj(pooled)
+            else:
+                pooled = torch.zeros((x.shape[0], self.clip_text_dim), device=x.device, dtype=x.dtype)
 
-        transformer_options = kwargs.get("transformer_options", {})
+            adaln_input = self.time_text_embed(torch.cat((t, pooled), dim=-1))
+
+        patches = transformer_options.get("patches", {})
         x_is_tensor = isinstance(x, torch.Tensor)
-        x, mask, img_size, cap_size, freqs_cis = self.patchify_and_embed(x, cap_feats, cap_mask, t, num_tokens, transformer_options=transformer_options)
-        freqs_cis = freqs_cis.to(x.device)
+        img, mask, img_size, cap_size, freqs_cis, timestep_zero_index = self.patchify_and_embed(x, cap_feats, cap_mask, adaln_input, num_tokens, ref_latents=ref_latents, ref_contexts=ref_contexts, siglip_feats=siglip_feats, transformer_options=transformer_options)
+        freqs_cis = freqs_cis.to(img.device)
 
-        for layer in self.layers:
-            x = layer(x, mask, freqs_cis, adaln_input, transformer_options=transformer_options)
+        transformer_options["total_blocks"] = len(self.layers)
+        transformer_options["block_type"] = "double"
+        img_input = img
+        for i, layer in enumerate(self.layers):
+            transformer_options["block_index"] = i
+            img = layer(img, mask, freqs_cis, adaln_input, timestep_zero_index=timestep_zero_index, transformer_options=transformer_options)
+            if "double_block" in patches:
+                for p in patches["double_block"]:
+                    out = p({"img": img[:, cap_size[0]:], "img_input": img_input[:, cap_size[0]:], "txt": img[:, :cap_size[0]], "pe": freqs_cis[:, cap_size[0]:], "vec": adaln_input, "x": x, "block_index": i, "transformer_options": transformer_options})
+                    if "img" in out:
+                        img[:, cap_size[0]:] = out["img"]
+                    if "txt" in out:
+                        img[:, :cap_size[0]] = out["txt"]
 
-        x = self.final_layer(x, adaln_input)
-        x = self.unpatchify(x, img_size, cap_size, return_tensor=x_is_tensor)[:,:,:h,:w]
-
-        return -x
+        img = self.final_layer(img, adaln_input, timestep_zero_index=timestep_zero_index)
+        img = self.unpatchify(img, img_size, cap_size, return_tensor=x_is_tensor)[:, :, :h, :w]
+        return -img
 

comfy/ldm/models/autoencoder.py

@@ -9,6 +9,8 @@ from comfy.ldm.modules.distributions.distributions import DiagonalGaussianDistri
 from comfy.ldm.util import get_obj_from_str, instantiate_from_config
 from comfy.ldm.modules.ema import LitEma
 import comfy.ops
+from einops import rearrange
+import comfy.model_management
 
 class DiagonalGaussianRegularizer(torch.nn.Module):
     def __init__(self, sample: bool = False):
@@ -179,6 +181,21 @@ class AutoencodingEngineLegacy(AutoencodingEngine):
         self.post_quant_conv = conv_op(embed_dim, ddconfig["z_channels"], 1)
         self.embed_dim = embed_dim
 
+        if ddconfig.get("batch_norm_latent", False):
+            self.bn_eps = 1e-4
+            self.bn_momentum = 0.1
+            self.ps = [2, 2]
+            self.bn = torch.nn.BatchNorm2d(math.prod(self.ps) * ddconfig["z_channels"],
+                                           eps=self.bn_eps,
+                                           momentum=self.bn_momentum,
+                                           affine=False,
+                                           track_running_stats=True,
+                                           )
+            self.bn.eval()
+        else:
+            self.bn = None
+
+
     def get_autoencoder_params(self) -> list:
         params = super().get_autoencoder_params()
         return params
@@ -201,11 +218,36 @@ class AutoencodingEngineLegacy(AutoencodingEngine):
             z = torch.cat(z, 0)
 
         z, reg_log = self.regularization(z)
+
+        if self.bn is not None:
+            z = rearrange(z,
+                          "... c (i pi) (j pj)  -> ... (c pi pj) i j",
+                          pi=self.ps[0],
+                          pj=self.ps[1],
+                          )
+
+            z = torch.nn.functional.batch_norm(z,
+                                               comfy.model_management.cast_to(self.bn.running_mean, dtype=z.dtype, device=z.device),
+                                               comfy.model_management.cast_to(self.bn.running_var, dtype=z.dtype, device=z.device),
+                                               momentum=self.bn_momentum,
+                                               eps=self.bn_eps)
+
         if return_reg_log:
             return z, reg_log
         return z
 
     def decode(self, z: torch.Tensor, **decoder_kwargs) -> torch.Tensor:
+        if self.bn is not None:
+            s = torch.sqrt(comfy.model_management.cast_to(self.bn.running_var.view(1, -1, 1, 1), dtype=z.dtype, device=z.device) + self.bn_eps)
+            m = comfy.model_management.cast_to(self.bn.running_mean.view(1, -1, 1, 1), dtype=z.dtype, device=z.device)
+            z = z * s + m
+            z = rearrange(
+                z,
+                "... (c pi pj) i j -> ... c (i pi) (j pj)",
+                pi=self.ps[0],
+                pj=self.ps[1],
+            )
+
         if self.max_batch_size is None:
             dec = self.post_quant_conv(z)
             dec = self.decoder(dec, **decoder_kwargs)

comfy/ldm/modules/attention.py

@@ -30,6 +30,13 @@ except ImportError as e:
             raise e
         exit(-1)
 
+SAGE_ATTENTION3_IS_AVAILABLE = False
+try:
+    from sageattn3 import sageattn3_blackwell
+    SAGE_ATTENTION3_IS_AVAILABLE = True
+except ImportError:
+    pass
+
 FLASH_ATTENTION_IS_AVAILABLE = False
 try:
     from flash_attn import flash_attn_func
@@ -517,6 +524,10 @@ def attention_pytorch(q, k, v, heads, mask=None, attn_precision=None, skip_resha
 
 @wrap_attn
 def attention_sage(q, k, v, heads, mask=None, attn_precision=None, skip_reshape=False, skip_output_reshape=False, **kwargs):
+    if kwargs.get("low_precision_attention", True) is False:
+        return attention_pytorch(q, k, v, heads, mask=mask, skip_reshape=skip_reshape, skip_output_reshape=skip_output_reshape, **kwargs)
+
+    exception_fallback = False
     if skip_reshape:
         b, _, _, dim_head = q.shape
         tensor_layout = "HND"
@@ -541,6 +552,8 @@ def attention_sage(q, k, v, heads, mask=None, attn_precision=None, skip_reshape=
         out = sageattn(q, k, v, attn_mask=mask, is_causal=False, tensor_layout=tensor_layout)
     except Exception as e:
         logging.error("Error running sage attention: {}, using pytorch attention instead.".format(e))
+        exception_fallback = True
+    if exception_fallback:
         if tensor_layout == "NHD":
             q, k, v = map(
                 lambda t: t.transpose(1, 2),
@@ -560,6 +573,93 @@ def attention_sage(q, k, v, heads, mask=None, attn_precision=None, skip_reshape=
             out = out.reshape(b, -1, heads * dim_head)
     return out
 
+@wrap_attn
+def attention3_sage(q, k, v, heads, mask=None, attn_precision=None, skip_reshape=False, skip_output_reshape=False, **kwargs):
+    exception_fallback = False
+    if (q.device.type != "cuda" or
+        q.dtype not in (torch.float16, torch.bfloat16) or
+        mask is not None):
+        return attention_pytorch(
+            q, k, v, heads,
+            mask=mask,
+            attn_precision=attn_precision,
+            skip_reshape=skip_reshape,
+            skip_output_reshape=skip_output_reshape,
+            **kwargs
+        )
+
+    if skip_reshape:
+        B, H, L, D = q.shape
+        if H != heads:
+            return attention_pytorch(
+                q, k, v, heads,
+                mask=mask,
+                attn_precision=attn_precision,
+                skip_reshape=True,
+                skip_output_reshape=skip_output_reshape,
+                **kwargs
+            )
+        q_s, k_s, v_s = q, k, v
+        N = q.shape[2]
+        dim_head = D
+    else:
+        B, N, inner_dim = q.shape
+        if inner_dim % heads != 0:
+            return attention_pytorch(
+                q, k, v, heads,
+                mask=mask,
+                attn_precision=attn_precision,
+                skip_reshape=False,
+                skip_output_reshape=skip_output_reshape,
+                **kwargs
+            )
+        dim_head = inner_dim // heads
+
+    if dim_head >= 256 or N <= 1024:
+        return attention_pytorch(
+                q, k, v, heads,
+                mask=mask,
+                attn_precision=attn_precision,
+                skip_reshape=skip_reshape,
+                skip_output_reshape=skip_output_reshape,
+                **kwargs
+            )
+
+    if not skip_reshape:
+        q_s, k_s, v_s = map(
+            lambda t: t.view(B, -1, heads, dim_head).permute(0, 2, 1, 3).contiguous(),
+            (q, k, v),
+        )
+        B, H, L, D = q_s.shape
+
+    try:
+        out = sageattn3_blackwell(q_s, k_s, v_s, is_causal=False)
+    except Exception as e:
+        exception_fallback = True
+        logging.error("Error running SageAttention3: %s, falling back to pytorch attention.", e)
+
+    if exception_fallback:
+        if not skip_reshape:
+            del q_s, k_s, v_s
+        return attention_pytorch(
+                q, k, v, heads,
+                mask=mask,
+                attn_precision=attn_precision,
+                skip_reshape=False,
+                skip_output_reshape=skip_output_reshape,
+                **kwargs
+            )
+
+    if skip_reshape:
+        if not skip_output_reshape:
+            out = out.permute(0, 2, 1, 3).reshape(B, L, H * D)
+    else:
+        if skip_output_reshape:
+            pass
+        else:
+            out = out.permute(0, 2, 1, 3).reshape(B, L, H * D)
+
+    return out
 
 try:
     @torch.library.custom_op("flash_attention::flash_attn", mutates_args=())
@@ -647,6 +747,8 @@ optimized_attention_masked = optimized_attention
 # register core-supported attention functions
 if SAGE_ATTENTION_IS_AVAILABLE:
     register_attention_function("sage", attention_sage)
+if SAGE_ATTENTION3_IS_AVAILABLE:
+    register_attention_function("sage3", attention3_sage)
 if FLASH_ATTENTION_IS_AVAILABLE:
     register_attention_function("flash", attention_flash)
 if model_management.xformers_enabled():

comfy/ldm/modules/diffusionmodules/mmdit.py

@@ -211,12 +211,14 @@ class TimestepEmbedder(nn.Module):
     Embeds scalar timesteps into vector representations.
     """
 
-    def __init__(self, hidden_size, frequency_embedding_size=256, dtype=None, device=None, operations=None):
+    def __init__(self, hidden_size, frequency_embedding_size=256, output_size=None, dtype=None, device=None, operations=None):
         super().__init__()
+        if output_size is None:
+            output_size = hidden_size
         self.mlp = nn.Sequential(
             operations.Linear(frequency_embedding_size, hidden_size, bias=True, dtype=dtype, device=device),
             nn.SiLU(),
-            operations.Linear(hidden_size, hidden_size, bias=True, dtype=dtype, device=device),
+            operations.Linear(hidden_size, output_size, bias=True, dtype=dtype, device=device),
         )
         self.frequency_embedding_size = frequency_embedding_size
 

comfy/ldm/modules/diffusionmodules/model.py

@@ -13,6 +13,15 @@ if model_management.xformers_enabled_vae():
     import xformers
     import xformers.ops
 
+def torch_cat_if_needed(xl, dim):
+    xl = [x for x in xl if x is not None and x.shape[dim] > 0]
+    if len(xl) > 1:
+        return torch.cat(xl, dim)
+    elif len(xl) == 1:
+        return xl[0]
+    else:
+        return None
+
 def get_timestep_embedding(timesteps, embedding_dim):
     """
     This matches the implementation in Denoising Diffusion Probabilistic Models:
@@ -43,6 +52,37 @@ def Normalize(in_channels, num_groups=32):
     return ops.GroupNorm(num_groups=num_groups, num_channels=in_channels, eps=1e-6, affine=True)
 
 
+class CarriedConv3d(nn.Module):
+    def __init__(self, n_channels, out_channels, kernel_size, stride=1, dilation=1, padding=0, **kwargs):
+        super().__init__()
+        self.conv = ops.Conv3d(n_channels, out_channels, kernel_size, stride=stride, dilation=dilation, **kwargs)
+
+    def forward(self, x):
+        return self.conv(x)
+
+
+def conv_carry_causal_3d(xl, op, conv_carry_in=None, conv_carry_out=None):
+
+    x = xl[0]
+    xl.clear()
+
+    if isinstance(op, CarriedConv3d):
+        if conv_carry_in is None:
+            x = torch.nn.functional.pad(x, (1, 1, 1, 1, 2, 0), mode = 'replicate')
+        else:
+            carry_len = conv_carry_in[0].shape[2]
+            x = torch.nn.functional.pad(x, (1, 1, 1, 1, 2 - carry_len, 0), mode = 'replicate')
+            x = torch.cat([conv_carry_in.pop(0), x], dim=2)
+
+        if conv_carry_out is not None:
+            to_push = x[:, :, -2:, :, :].clone()
+            conv_carry_out.append(to_push)
+
+    out = op(x)
+
+    return out
+
+
 class VideoConv3d(nn.Module):
     def __init__(self, n_channels, out_channels, kernel_size, stride=1, dilation=1, padding_mode='replicate', padding=1, **kwargs):
         super().__init__()
@@ -89,29 +129,24 @@ class Upsample(nn.Module):
                                         stride=1,
                                         padding=1)
 
-    def forward(self, x):
+    def forward(self, x, conv_carry_in=None, conv_carry_out=None):
         scale_factor = self.scale_factor
         if isinstance(scale_factor, (int, float)):
             scale_factor = (scale_factor,) * (x.ndim - 2)
 
         if x.ndim == 5 and scale_factor[0] > 1.0:
-            t = x.shape[2]
-            if t > 1:
-                a, b = x.split((1, t - 1), dim=2)
-                del x
-                b = interpolate_up(b, scale_factor)
-            else:
-                a = x
-
-            a = interpolate_up(a.squeeze(2), scale_factor=scale_factor[1:]).unsqueeze(2)
-            if t > 1:
-                x = torch.cat((a, b), dim=2)
-            else:
-                x = a
+            results = []
+            if conv_carry_in is None:
+                first = x[:, :, :1, :, :]
+                results.append(interpolate_up(first.squeeze(2), scale_factor=scale_factor[1:]).unsqueeze(2))
+                x = x[:, :, 1:, :, :]
+            if x.shape[2] > 0:
+                results.append(interpolate_up(x, scale_factor))
+            x = torch_cat_if_needed(results, dim=2)
         else:
             x = interpolate_up(x, scale_factor)
         if self.with_conv:
-            x = self.conv(x)
+            x = conv_carry_causal_3d([x], self.conv, conv_carry_in, conv_carry_out)
         return x
 
 
@@ -127,17 +162,20 @@ class Downsample(nn.Module):
                                         stride=stride,
                                         padding=0)
 
-    def forward(self, x):
+    def forward(self, x, conv_carry_in=None, conv_carry_out=None):
         if self.with_conv:
-            if x.ndim == 4:
+            if isinstance(self.conv, CarriedConv3d):
+                x = conv_carry_causal_3d([x], self.conv, conv_carry_in, conv_carry_out)
+            elif x.ndim == 4:
                 pad = (0, 1, 0, 1)
                 mode = "constant"
                 x = torch.nn.functional.pad(x, pad, mode=mode, value=0)
+                x = self.conv(x)
             elif x.ndim == 5:
                 pad = (1, 1, 1, 1, 2, 0)
                 mode = "replicate"
                 x = torch.nn.functional.pad(x, pad, mode=mode)
-            x = self.conv(x)
+                x = self.conv(x)
         else:
             x = torch.nn.functional.avg_pool2d(x, kernel_size=2, stride=2)
         return x
@@ -183,23 +221,23 @@ class ResnetBlock(nn.Module):
                                                     stride=1,
                                                     padding=0)
 
-    def forward(self, x, temb=None):
+    def forward(self, x, temb=None, conv_carry_in=None, conv_carry_out=None):
         h = x
         h = self.norm1(h)
-        h = self.swish(h)
-        h = self.conv1(h)
+        h = [ self.swish(h) ]
+        h = conv_carry_causal_3d(h, self.conv1, conv_carry_in=conv_carry_in, conv_carry_out=conv_carry_out)
 
         if temb is not None:
             h = h + self.temb_proj(self.swish(temb))[:,:,None,None]
 
         h = self.norm2(h)
         h = self.swish(h)
-        h = self.dropout(h)
-        h = self.conv2(h)
+        h = [ self.dropout(h) ]
+        h = conv_carry_causal_3d(h, self.conv2, conv_carry_in=conv_carry_in, conv_carry_out=conv_carry_out)
 
         if self.in_channels != self.out_channels:
             if self.use_conv_shortcut:
-                x = self.conv_shortcut(x)
+                x = conv_carry_causal_3d([x], self.conv_shortcut, conv_carry_in=conv_carry_in, conv_carry_out=conv_carry_out)
             else:
                 x = self.nin_shortcut(x)
 
@@ -279,6 +317,7 @@ def pytorch_attention(q, k, v):
     orig_shape = q.shape
     B = orig_shape[0]
     C = orig_shape[1]
+    oom_fallback = False
     q, k, v = map(
         lambda t: t.view(B, 1, C, -1).transpose(2, 3).contiguous(),
         (q, k, v),
@@ -289,6 +328,8 @@ def pytorch_attention(q, k, v):
         out = out.transpose(2, 3).reshape(orig_shape)
     except model_management.OOM_EXCEPTION:
         logging.warning("scaled_dot_product_attention OOMed: switched to slice attention")
+        oom_fallback = True
+    if oom_fallback:
         out = slice_attention(q.view(B, -1, C), k.view(B, -1, C).transpose(1, 2), v.view(B, -1, C).transpose(1, 2)).reshape(orig_shape)
     return out
 
@@ -356,7 +397,8 @@ class Model(nn.Module):
                  attn_resolutions, dropout=0.0, resamp_with_conv=True, in_channels,
                  resolution, use_timestep=True, use_linear_attn=False, attn_type="vanilla"):
         super().__init__()
-        if use_linear_attn: attn_type = "linear"
+        if use_linear_attn:
+            attn_type = "linear"
         self.ch = ch
         self.temb_ch = self.ch*4
         self.num_resolutions = len(ch_mult)
@@ -510,16 +552,22 @@ class Encoder(nn.Module):
                  conv3d=False, time_compress=None,
                  **ignore_kwargs):
         super().__init__()
-        if use_linear_attn: attn_type = "linear"
+        if use_linear_attn:
+            attn_type = "linear"
         self.ch = ch
         self.temb_ch = 0
         self.num_resolutions = len(ch_mult)
         self.num_res_blocks = num_res_blocks
         self.resolution = resolution
         self.in_channels = in_channels
+        self.carried = False
 
         if conv3d:
-            conv_op = VideoConv3d
+            if not attn_resolutions:
+                conv_op = CarriedConv3d
+                self.carried = True
+            else:
+                conv_op = VideoConv3d
             mid_attn_conv_op = ops.Conv3d
         else:
             conv_op = ops.Conv2d
@@ -532,6 +580,7 @@ class Encoder(nn.Module):
                                        stride=1,
                                        padding=1)
 
+        self.time_compress = 1
         curr_res = resolution
         in_ch_mult = (1,)+tuple(ch_mult)
         self.in_ch_mult = in_ch_mult
@@ -558,10 +607,15 @@ class Encoder(nn.Module):
                 if time_compress is not None:
                     if (self.num_resolutions - 1 - i_level) > math.log2(time_compress):
                         stride = (1, 2, 2)
+                else:
+                    self.time_compress *= 2
                 down.downsample = Downsample(block_in, resamp_with_conv, stride=stride, conv_op=conv_op)
                 curr_res = curr_res // 2
             self.down.append(down)
 
+        if time_compress is not None:
+            self.time_compress = time_compress
+
         # middle
         self.mid = nn.Module()
         self.mid.block_1 = ResnetBlock(in_channels=block_in,
@@ -587,15 +641,42 @@ class Encoder(nn.Module):
     def forward(self, x):
         # timestep embedding
         temb = None
-        # downsampling
-        h = self.conv_in(x)
-        for i_level in range(self.num_resolutions):
-            for i_block in range(self.num_res_blocks):
-                h = self.down[i_level].block[i_block](h, temb)
-                if len(self.down[i_level].attn) > 0:
-                    h = self.down[i_level].attn[i_block](h)
-            if i_level != self.num_resolutions-1:
-                h = self.down[i_level].downsample(h)
+
+        if self.carried:
+            xl = [x[:, :, :1, :, :]]
+            if x.shape[2] > self.time_compress:
+                tc = self.time_compress
+                xl += torch.split(x[:, :, 1: 1 + ((x.shape[2] - 1) // tc) * tc, :, :], tc * 2, dim = 2)
+            x = xl
+        else:
+            x = [x]
+        out = []
+
+        conv_carry_in = None
+
+        for i, x1 in enumerate(x):
+            conv_carry_out = []
+            if i == len(x) - 1:
+                conv_carry_out = None
+
+            # downsampling
+            x1 = [ x1 ]
+            h1 = conv_carry_causal_3d(x1, self.conv_in, conv_carry_in, conv_carry_out)
+
+            for i_level in range(self.num_resolutions):
+                for i_block in range(self.num_res_blocks):
+                    h1 = self.down[i_level].block[i_block](h1, temb, conv_carry_in, conv_carry_out)
+                    if len(self.down[i_level].attn) > 0:
+                        assert i == 0 #carried should not happen if attn exists
+                        h1 = self.down[i_level].attn[i_block](h1)
+                if i_level != self.num_resolutions-1:
+                    h1 = self.down[i_level].downsample(h1, conv_carry_in, conv_carry_out)
+
+            out.append(h1)
+            conv_carry_in = conv_carry_out
+
+        h = torch_cat_if_needed(out, dim=2)
+        del out
 
         # middle
         h = self.mid.block_1(h, temb)
@@ -604,15 +685,15 @@ class Encoder(nn.Module):
 
         # end
         h = self.norm_out(h)
-        h = nonlinearity(h)
-        h = self.conv_out(h)
+        h = [ nonlinearity(h) ]
+        h = conv_carry_causal_3d(h, self.conv_out)
         return h
 
 
 class Decoder(nn.Module):
     def __init__(self, *, ch, out_ch, ch_mult=(1,2,4,8), num_res_blocks,
                  attn_resolutions, dropout=0.0, resamp_with_conv=True, in_channels,
-                 resolution, z_channels, give_pre_end=False, tanh_out=False, use_linear_attn=False,
+                 resolution, z_channels, tanh_out=False, use_linear_attn=False,
                  conv_out_op=ops.Conv2d,
                  resnet_op=ResnetBlock,
                  attn_op=AttnBlock,
@@ -626,12 +707,18 @@ class Decoder(nn.Module):
         self.num_res_blocks = num_res_blocks
         self.resolution = resolution
         self.in_channels = in_channels
-        self.give_pre_end = give_pre_end
         self.tanh_out = tanh_out
+        self.carried = False
 
         if conv3d:
-            conv_op = VideoConv3d
-            conv_out_op = VideoConv3d
+            if not attn_resolutions and resnet_op == ResnetBlock:
+                conv_op = CarriedConv3d
+                conv_out_op = CarriedConv3d
+                self.carried = True
+            else:
+                conv_op = VideoConv3d
+                conv_out_op = VideoConv3d
+
             mid_attn_conv_op = ops.Conv3d
         else:
             conv_op = ops.Conv2d
@@ -706,29 +793,43 @@ class Decoder(nn.Module):
         temb = None
 
         # z to block_in
-        h = self.conv_in(z)
+        h = conv_carry_causal_3d([z], self.conv_in)
 
         # middle
         h = self.mid.block_1(h, temb, **kwargs)
         h = self.mid.attn_1(h, **kwargs)
         h = self.mid.block_2(h, temb, **kwargs)
 
+        if self.carried:
+            h = torch.split(h, 2, dim=2)
+        else:
+            h = [ h ]
+        out = []
+
+        conv_carry_in = None
+
         # upsampling
-        for i_level in reversed(range(self.num_resolutions)):
-            for i_block in range(self.num_res_blocks+1):
-                h = self.up[i_level].block[i_block](h, temb, **kwargs)
-                if len(self.up[i_level].attn) > 0:
-                    h = self.up[i_level].attn[i_block](h, **kwargs)
-            if i_level != 0:
-                h = self.up[i_level].upsample(h)
+        for i, h1 in enumerate(h):
+            conv_carry_out = []
+            if i == len(h) - 1:
+                conv_carry_out = None
+            for i_level in reversed(range(self.num_resolutions)):
+                for i_block in range(self.num_res_blocks+1):
+                    h1 = self.up[i_level].block[i_block](h1, temb, conv_carry_in, conv_carry_out, **kwargs)
+                    if len(self.up[i_level].attn) > 0:
+                        assert i == 0 #carried should not happen if attn exists
+                        h1 = self.up[i_level].attn[i_block](h1, **kwargs)
+                if i_level != 0:
+                    h1 = self.up[i_level].upsample(h1, conv_carry_in, conv_carry_out)
 
-        # end
-        if self.give_pre_end:
-            return h
+            h1 = self.norm_out(h1)
+            h1 = [ nonlinearity(h1) ]
+            h1 = conv_carry_causal_3d(h1, self.conv_out, conv_carry_in, conv_carry_out)
+            if self.tanh_out:
+                h1 = torch.tanh(h1)
+            out.append(h1)
+            conv_carry_in = conv_carry_out
 
-        h = self.norm_out(h)
-        h = nonlinearity(h)
-        h = self.conv_out(h, **kwargs)
-        if self.tanh_out:
-            h = torch.tanh(h)
-        return h
+        out = torch_cat_if_needed(out, dim=2)
+
+        return out

comfy/ldm/modules/ema.py

@@ -45,7 +45,7 @@ class LitEma(nn.Module):
                     shadow_params[sname] = shadow_params[sname].type_as(m_param[key])
                     shadow_params[sname].sub_(one_minus_decay * (shadow_params[sname] - m_param[key]))
                 else:
-                    assert not key in self.m_name2s_name
+                    assert key not in self.m_name2s_name
 
     def copy_to(self, model):
         m_param = dict(model.named_parameters())
@@ -54,7 +54,7 @@ class LitEma(nn.Module):
             if m_param[key].requires_grad:
                 m_param[key].data.copy_(shadow_params[self.m_name2s_name[key]].data)
             else:
-                assert not key in self.m_name2s_name
+                assert key not in self.m_name2s_name
 
     def store(self, parameters):
         """

comfy/ldm/qwen_image/controlnet.py

@@ -44,7 +44,7 @@ class QwenImageControlNetModel(QwenImageTransformer2DModel):
         txt_start = round(max(((x.shape[-1] + (self.patch_size // 2)) // self.patch_size) // 2, ((x.shape[-2] + (self.patch_size // 2)) // self.patch_size) // 2))
         txt_ids = torch.arange(txt_start, txt_start + context.shape[1], device=x.device).reshape(1, -1, 1).repeat(x.shape[0], 1, 3)
         ids = torch.cat((txt_ids, img_ids), dim=1)
-        image_rotary_emb = self.pe_embedder(ids).squeeze(1).unsqueeze(2).to(x.dtype)
+        image_rotary_emb = self.pe_embedder(ids).to(x.dtype).contiguous()
         del ids, txt_ids, img_ids
 
         hidden_states = self.img_in(hidden_states) + self.controlnet_x_embedder(hint)

comfy/ldm/qwen_image/model.py

@@ -10,6 +10,7 @@ from comfy.ldm.modules.attention import optimized_attention_masked
 from comfy.ldm.flux.layers import EmbedND
 import comfy.ldm.common_dit
 import comfy.patcher_extension
+from comfy.ldm.flux.math import apply_rope1
 
 class GELU(nn.Module):
     def __init__(self, dim_in: int, dim_out: int, approximate: str = "none", bias: bool = True, dtype=None, device=None, operations=None):
@@ -60,7 +61,7 @@ def apply_rotary_emb(x, freqs_cis):
 
 
 class QwenTimestepProjEmbeddings(nn.Module):
-    def __init__(self, embedding_dim, pooled_projection_dim, dtype=None, device=None, operations=None):
+    def __init__(self, embedding_dim, pooled_projection_dim, use_additional_t_cond=False, dtype=None, device=None, operations=None):
         super().__init__()
         self.time_proj = Timesteps(num_channels=256, flip_sin_to_cos=True, downscale_freq_shift=0, scale=1000)
         self.timestep_embedder = TimestepEmbedding(
@@ -71,9 +72,19 @@ class QwenTimestepProjEmbeddings(nn.Module):
             operations=operations
         )
 
-    def forward(self, timestep, hidden_states):
+        self.use_additional_t_cond = use_additional_t_cond
+        if self.use_additional_t_cond:
+            self.addition_t_embedding = operations.Embedding(2, embedding_dim, device=device, dtype=dtype)
+
+    def forward(self, timestep, hidden_states, addition_t_cond=None):
         timesteps_proj = self.time_proj(timestep)
         timesteps_emb = self.timestep_embedder(timesteps_proj.to(dtype=hidden_states.dtype))
+
+        if self.use_additional_t_cond:
+            if addition_t_cond is None:
+                addition_t_cond = torch.zeros((timesteps_emb.shape[0]), device=timesteps_emb.device, dtype=torch.long)
+            timesteps_emb += self.addition_t_embedding(addition_t_cond, out_dtype=timesteps_emb.dtype)
+
         return timesteps_emb
 
 
@@ -134,33 +145,40 @@ class Attention(nn.Module):
         image_rotary_emb: Optional[torch.Tensor] = None,
         transformer_options={},
     ) -> Tuple[torch.Tensor, torch.Tensor]:
+        batch_size = hidden_states.shape[0]
+        seq_img = hidden_states.shape[1]
         seq_txt = encoder_hidden_states.shape[1]
 
-        img_query = self.to_q(hidden_states).unflatten(-1, (self.heads, -1))
-        img_key = self.to_k(hidden_states).unflatten(-1, (self.heads, -1))
-        img_value = self.to_v(hidden_states).unflatten(-1, (self.heads, -1))
+        # Project and reshape to BHND format (batch, heads, seq, dim)
+        img_query = self.to_q(hidden_states).view(batch_size, seq_img, self.heads, -1).transpose(1, 2).contiguous()
+        img_key = self.to_k(hidden_states).view(batch_size, seq_img, self.heads, -1).transpose(1, 2).contiguous()
+        img_value = self.to_v(hidden_states).view(batch_size, seq_img, self.heads, -1).transpose(1, 2)
 
-        txt_query = self.add_q_proj(encoder_hidden_states).unflatten(-1, (self.heads, -1))
-        txt_key = self.add_k_proj(encoder_hidden_states).unflatten(-1, (self.heads, -1))
-        txt_value = self.add_v_proj(encoder_hidden_states).unflatten(-1, (self.heads, -1))
+        txt_query = self.add_q_proj(encoder_hidden_states).view(batch_size, seq_txt, self.heads, -1).transpose(1, 2).contiguous()
+        txt_key = self.add_k_proj(encoder_hidden_states).view(batch_size, seq_txt, self.heads, -1).transpose(1, 2).contiguous()
+        txt_value = self.add_v_proj(encoder_hidden_states).view(batch_size, seq_txt, self.heads, -1).transpose(1, 2)
 
         img_query = self.norm_q(img_query)
         img_key = self.norm_k(img_key)
         txt_query = self.norm_added_q(txt_query)
         txt_key = self.norm_added_k(txt_key)
 
-        joint_query = torch.cat([txt_query, img_query], dim=1)
-        joint_key = torch.cat([txt_key, img_key], dim=1)
-        joint_value = torch.cat([txt_value, img_value], dim=1)
+        joint_query = torch.cat([txt_query, img_query], dim=2)
+        joint_key = torch.cat([txt_key, img_key], dim=2)
+        joint_value = torch.cat([txt_value, img_value], dim=2)
 
-        joint_query = apply_rotary_emb(joint_query, image_rotary_emb)
-        joint_key = apply_rotary_emb(joint_key, image_rotary_emb)
+        joint_query = apply_rope1(joint_query, image_rotary_emb)
+        joint_key = apply_rope1(joint_key, image_rotary_emb)
 
-        joint_query = joint_query.flatten(start_dim=2)
-        joint_key = joint_key.flatten(start_dim=2)
-        joint_value = joint_value.flatten(start_dim=2)
+        if encoder_hidden_states_mask is not None:
+            attn_mask = torch.zeros((batch_size, 1, seq_txt + seq_img), dtype=hidden_states.dtype, device=hidden_states.device)
+            attn_mask[:, 0, :seq_txt] = encoder_hidden_states_mask
+        else:
+            attn_mask = None
 
-        joint_hidden_states = optimized_attention_masked(joint_query, joint_key, joint_value, self.heads, attention_mask, transformer_options=transformer_options)
+        joint_hidden_states = optimized_attention_masked(joint_query, joint_key, joint_value, self.heads,
+                                                         attn_mask, transformer_options=transformer_options,
+                                                         skip_reshape=True)
 
         txt_attn_output = joint_hidden_states[:, :seq_txt, :]
         img_attn_output = joint_hidden_states[:, seq_txt:, :]
@@ -216,9 +234,24 @@ class QwenImageTransformerBlock(nn.Module):
             operations=operations,
         )
 
-    def _modulate(self, x: torch.Tensor, mod_params: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
+    def _apply_gate(self, x, y, gate, timestep_zero_index=None):
+        if timestep_zero_index is not None:
+            return y + torch.cat((x[:, :timestep_zero_index] * gate[0], x[:, timestep_zero_index:] * gate[1]), dim=1)
+        else:
+            return torch.addcmul(y, gate, x)
+
+    def _modulate(self, x: torch.Tensor, mod_params: torch.Tensor, timestep_zero_index=None) -> Tuple[torch.Tensor, torch.Tensor]:
         shift, scale, gate = torch.chunk(mod_params, 3, dim=-1)
-        return torch.addcmul(shift.unsqueeze(1), x, 1 + scale.unsqueeze(1)), gate.unsqueeze(1)
+        if timestep_zero_index is not None:
+            actual_batch = shift.size(0) // 2
+            shift, shift_0 = shift[:actual_batch], shift[actual_batch:]
+            scale, scale_0 = scale[:actual_batch], scale[actual_batch:]
+            gate, gate_0 = gate[:actual_batch], gate[actual_batch:]
+            reg = torch.addcmul(shift.unsqueeze(1), x[:, :timestep_zero_index], 1 + scale.unsqueeze(1))
+            zero = torch.addcmul(shift_0.unsqueeze(1), x[:, timestep_zero_index:], 1 + scale_0.unsqueeze(1))
+            return torch.cat((reg, zero), dim=1), (gate.unsqueeze(1), gate_0.unsqueeze(1))
+        else:
+            return torch.addcmul(shift.unsqueeze(1), x, 1 + scale.unsqueeze(1)), gate.unsqueeze(1)
 
     def forward(
         self,
@@ -227,17 +260,22 @@ class QwenImageTransformerBlock(nn.Module):
         encoder_hidden_states_mask: torch.Tensor,
         temb: torch.Tensor,
         image_rotary_emb: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+        timestep_zero_index=None,
         transformer_options={},
     ) -> Tuple[torch.Tensor, torch.Tensor]:
         img_mod_params = self.img_mod(temb)
+
+        if timestep_zero_index is not None:
+            temb = temb.chunk(2, dim=0)[0]
+
         txt_mod_params = self.txt_mod(temb)
         img_mod1, img_mod2 = img_mod_params.chunk(2, dim=-1)
         txt_mod1, txt_mod2 = txt_mod_params.chunk(2, dim=-1)
 
-        img_normed = self.img_norm1(hidden_states)
-        img_modulated, img_gate1 = self._modulate(img_normed, img_mod1)
-        txt_normed = self.txt_norm1(encoder_hidden_states)
-        txt_modulated, txt_gate1 = self._modulate(txt_normed, txt_mod1)
+        img_modulated, img_gate1 = self._modulate(self.img_norm1(hidden_states), img_mod1, timestep_zero_index)
+        del img_mod1
+        txt_modulated, txt_gate1 = self._modulate(self.txt_norm1(encoder_hidden_states), txt_mod1)
+        del txt_mod1
 
         img_attn_output, txt_attn_output = self.attn(
             hidden_states=img_modulated,
@@ -246,16 +284,20 @@ class QwenImageTransformerBlock(nn.Module):
             image_rotary_emb=image_rotary_emb,
             transformer_options=transformer_options,
         )
+        del img_modulated
+        del txt_modulated
 
-        hidden_states = hidden_states + img_gate1 * img_attn_output
+        hidden_states = self._apply_gate(img_attn_output, hidden_states, img_gate1, timestep_zero_index)
         encoder_hidden_states = encoder_hidden_states + txt_gate1 * txt_attn_output
+        del img_attn_output
+        del txt_attn_output
+        del img_gate1
+        del txt_gate1
 
-        img_normed2 = self.img_norm2(hidden_states)
-        img_modulated2, img_gate2 = self._modulate(img_normed2, img_mod2)
-        hidden_states = torch.addcmul(hidden_states, img_gate2, self.img_mlp(img_modulated2))
+        img_modulated2, img_gate2 = self._modulate(self.img_norm2(hidden_states), img_mod2, timestep_zero_index)
+        hidden_states = self._apply_gate(self.img_mlp(img_modulated2), hidden_states, img_gate2, timestep_zero_index)
 
-        txt_normed2 = self.txt_norm2(encoder_hidden_states)
-        txt_modulated2, txt_gate2 = self._modulate(txt_normed2, txt_mod2)
+        txt_modulated2, txt_gate2 = self._modulate(self.txt_norm2(encoder_hidden_states), txt_mod2)
         encoder_hidden_states = torch.addcmul(encoder_hidden_states, txt_gate2, self.txt_mlp(txt_modulated2))
 
         return encoder_hidden_states, hidden_states
@@ -294,10 +336,11 @@ class QwenImageTransformer2DModel(nn.Module):
         num_attention_heads: int = 24,
         joint_attention_dim: int = 3584,
         pooled_projection_dim: int = 768,
-        guidance_embeds: bool = False,
         axes_dims_rope: Tuple[int, int, int] = (16, 56, 56),
+        default_ref_method="index",
         image_model=None,
         final_layer=True,
+        use_additional_t_cond=False,
         dtype=None,
         device=None,
         operations=None,
@@ -308,12 +351,14 @@ class QwenImageTransformer2DModel(nn.Module):
         self.in_channels = in_channels
         self.out_channels = out_channels or in_channels
         self.inner_dim = num_attention_heads * attention_head_dim
+        self.default_ref_method = default_ref_method
 
         self.pe_embedder = EmbedND(dim=attention_head_dim, theta=10000, axes_dim=list(axes_dims_rope))
 
         self.time_text_embed = QwenTimestepProjEmbeddings(
             embedding_dim=self.inner_dim,
             pooled_projection_dim=pooled_projection_dim,
+            use_additional_t_cond=use_additional_t_cond,
             dtype=dtype,
             device=device,
             operations=operations
@@ -335,6 +380,9 @@ class QwenImageTransformer2DModel(nn.Module):
             for _ in range(num_layers)
         ])
 
+        if self.default_ref_method == "index_timestep_zero":
+            self.register_buffer("__index_timestep_zero__", torch.tensor([]))
+
         if final_layer:
             self.norm_out = LastLayer(self.inner_dim, self.inner_dim, dtype=dtype, device=device, operations=operations)
             self.proj_out = operations.Linear(self.inner_dim, patch_size * patch_size * self.out_channels, bias=True, dtype=dtype, device=device)
@@ -344,27 +392,33 @@ class QwenImageTransformer2DModel(nn.Module):
         patch_size = self.patch_size
         hidden_states = comfy.ldm.common_dit.pad_to_patch_size(x, (1, self.patch_size, self.patch_size))
         orig_shape = hidden_states.shape
-        hidden_states = hidden_states.view(orig_shape[0], orig_shape[1], orig_shape[-2] // 2, 2, orig_shape[-1] // 2, 2)
-        hidden_states = hidden_states.permute(0, 2, 4, 1, 3, 5)
-        hidden_states = hidden_states.reshape(orig_shape[0], (orig_shape[-2] // 2) * (orig_shape[-1] // 2), orig_shape[1] * 4)
+        hidden_states = hidden_states.view(orig_shape[0], orig_shape[1], orig_shape[-3], orig_shape[-2] // 2, 2, orig_shape[-1] // 2, 2)
+        hidden_states = hidden_states.permute(0, 2, 3, 5, 1, 4, 6)
+        hidden_states = hidden_states.reshape(orig_shape[0], orig_shape[-3] * (orig_shape[-2] // 2) * (orig_shape[-1] // 2), orig_shape[1] * 4)
+        t_len = t
         h_len = ((h + (patch_size // 2)) // patch_size)
         w_len = ((w + (patch_size // 2)) // patch_size)
 
         h_offset = ((h_offset + (patch_size // 2)) // patch_size)
         w_offset = ((w_offset + (patch_size // 2)) // patch_size)
 
-        img_ids = torch.zeros((h_len, w_len, 3), device=x.device)
-        img_ids[:, :, 0] = img_ids[:, :, 1] + index
-        img_ids[:, :, 1] = img_ids[:, :, 1] + torch.linspace(h_offset, h_len - 1 + h_offset, steps=h_len, device=x.device, dtype=x.dtype).unsqueeze(1) - (h_len // 2)
-        img_ids[:, :, 2] = img_ids[:, :, 2] + torch.linspace(w_offset, w_len - 1 + w_offset, steps=w_len, device=x.device, dtype=x.dtype).unsqueeze(0) - (w_len // 2)
-        return hidden_states, repeat(img_ids, "h w c -> b (h w) c", b=bs), orig_shape
+        img_ids = torch.zeros((t_len, h_len, w_len, 3), device=x.device)
 
-    def forward(self, x, timestep, context, attention_mask=None, guidance=None, ref_latents=None, transformer_options={}, **kwargs):
+        if t_len > 1:
+            img_ids[:, :, :, 0] = img_ids[:, :, :, 0] + torch.linspace(0, t_len - 1, steps=t_len, device=x.device, dtype=x.dtype).unsqueeze(1).unsqueeze(1)
+        else:
+            img_ids[:, :, :, 0] = img_ids[:, :, :, 0] + index
+
+        img_ids[:, :, :, 1] = img_ids[:, :, :, 1] + torch.linspace(h_offset, h_len - 1 + h_offset, steps=h_len, device=x.device, dtype=x.dtype).unsqueeze(1).unsqueeze(0) - (h_len // 2)
+        img_ids[:, :, :, 2] = img_ids[:, :, :, 2] + torch.linspace(w_offset, w_len - 1 + w_offset, steps=w_len, device=x.device, dtype=x.dtype).unsqueeze(0).unsqueeze(0) - (w_len // 2)
+        return hidden_states, repeat(img_ids, "t h w c -> b (t h w) c", b=bs), orig_shape
+
+    def forward(self, x, timestep, context, attention_mask=None, ref_latents=None, additional_t_cond=None, transformer_options={}, **kwargs):
         return comfy.patcher_extension.WrapperExecutor.new_class_executor(
             self._forward,
             self,
             comfy.patcher_extension.get_all_wrappers(comfy.patcher_extension.WrappersMP.DIFFUSION_MODEL, transformer_options)
-        ).execute(x, timestep, context, attention_mask, guidance, ref_latents, transformer_options, **kwargs)
+        ).execute(x, timestep, context, attention_mask, ref_latents, additional_t_cond, transformer_options, **kwargs)
 
     def _forward(
         self,
@@ -372,8 +426,8 @@ class QwenImageTransformer2DModel(nn.Module):
         timesteps,
         context,
         attention_mask=None,
-        guidance: torch.Tensor = None,
         ref_latents=None,
+        additional_t_cond=None,
         transformer_options={},
         control=None,
         **kwargs
@@ -382,19 +436,30 @@ class QwenImageTransformer2DModel(nn.Module):
         encoder_hidden_states = context
         encoder_hidden_states_mask = attention_mask
 
+        if encoder_hidden_states_mask is not None and not torch.is_floating_point(encoder_hidden_states_mask):
+            encoder_hidden_states_mask = (encoder_hidden_states_mask - 1).to(x.dtype) * torch.finfo(x.dtype).max
+
         hidden_states, img_ids, orig_shape = self.process_img(x)
         num_embeds = hidden_states.shape[1]
 
+        timestep_zero_index = None
         if ref_latents is not None:
             h = 0
             w = 0
             index = 0
-            index_ref_method = kwargs.get("ref_latents_method", "index") == "index"
+            ref_method = kwargs.get("ref_latents_method", self.default_ref_method)
+            index_ref_method = (ref_method == "index") or (ref_method == "index_timestep_zero")
+            negative_ref_method = ref_method == "negative_index"
+            timestep_zero = ref_method == "index_timestep_zero"
             for ref in ref_latents:
                 if index_ref_method:
                     index += 1
                     h_offset = 0
                     w_offset = 0
+                elif negative_ref_method:
+                    index -= 1
+                    h_offset = 0
+                    w_offset = 0
                 else:
                     index = 1
                     h_offset = 0
@@ -409,35 +474,35 @@ class QwenImageTransformer2DModel(nn.Module):
                 kontext, kontext_ids, _ = self.process_img(ref, index=index, h_offset=h_offset, w_offset=w_offset)
                 hidden_states = torch.cat([hidden_states, kontext], dim=1)
                 img_ids = torch.cat([img_ids, kontext_ids], dim=1)
+            if timestep_zero:
+                if index > 0:
+                    timestep = torch.cat([timestep, timestep * 0], dim=0)
+                    timestep_zero_index = num_embeds
 
         txt_start = round(max(((x.shape[-1] + (self.patch_size // 2)) // self.patch_size) // 2, ((x.shape[-2] + (self.patch_size // 2)) // self.patch_size) // 2))
         txt_ids = torch.arange(txt_start, txt_start + context.shape[1], device=x.device).reshape(1, -1, 1).repeat(x.shape[0], 1, 3)
         ids = torch.cat((txt_ids, img_ids), dim=1)
-        image_rotary_emb = self.pe_embedder(ids).squeeze(1).unsqueeze(2).to(x.dtype)
+        image_rotary_emb = self.pe_embedder(ids).to(x.dtype).contiguous()
         del ids, txt_ids, img_ids
 
         hidden_states = self.img_in(hidden_states)
         encoder_hidden_states = self.txt_norm(encoder_hidden_states)
         encoder_hidden_states = self.txt_in(encoder_hidden_states)
 
-        if guidance is not None:
-            guidance = guidance * 1000
-
-        temb = (
-            self.time_text_embed(timestep, hidden_states)
-            if guidance is None
-            else self.time_text_embed(timestep, guidance, hidden_states)
-        )
+        temb = self.time_text_embed(timestep, hidden_states, additional_t_cond)
 
         patches_replace = transformer_options.get("patches_replace", {})
         patches = transformer_options.get("patches", {})
         blocks_replace = patches_replace.get("dit", {})
 
+        transformer_options["total_blocks"] = len(self.transformer_blocks)
+        transformer_options["block_type"] = "double"
         for i, block in enumerate(self.transformer_blocks):
+            transformer_options["block_index"] = i
             if ("double_block", i) in blocks_replace:
                 def block_wrap(args):
                     out = {}
-                    out["txt"], out["img"] = block(hidden_states=args["img"], encoder_hidden_states=args["txt"], encoder_hidden_states_mask=encoder_hidden_states_mask, temb=args["vec"], image_rotary_emb=args["pe"], transformer_options=args["transformer_options"])
+                    out["txt"], out["img"] = block(hidden_states=args["img"], encoder_hidden_states=args["txt"], encoder_hidden_states_mask=encoder_hidden_states_mask, temb=args["vec"], image_rotary_emb=args["pe"], timestep_zero_index=timestep_zero_index, transformer_options=args["transformer_options"])
                     return out
                 out = blocks_replace[("double_block", i)]({"img": hidden_states, "txt": encoder_hidden_states, "vec": temb, "pe": image_rotary_emb, "transformer_options": transformer_options}, {"original_block": block_wrap})
                 hidden_states = out["img"]
@@ -449,6 +514,7 @@ class QwenImageTransformer2DModel(nn.Module):
                     encoder_hidden_states_mask=encoder_hidden_states_mask,
                     temb=temb,
                     image_rotary_emb=image_rotary_emb,
+                    timestep_zero_index=timestep_zero_index,
                     transformer_options=transformer_options,
                 )
 
@@ -465,9 +531,12 @@ class QwenImageTransformer2DModel(nn.Module):
                     if add is not None:
                         hidden_states[:, :add.shape[1]] += add
 
+        if timestep_zero_index is not None:
+            temb = temb.chunk(2, dim=0)[0]
+
         hidden_states = self.norm_out(hidden_states, temb)
         hidden_states = self.proj_out(hidden_states)
 
-        hidden_states = hidden_states[:, :num_embeds].view(orig_shape[0], orig_shape[-2] // 2, orig_shape[-1] // 2, orig_shape[1], 2, 2)
-        hidden_states = hidden_states.permute(0, 3, 1, 4, 2, 5)
+        hidden_states = hidden_states[:, :num_embeds].view(orig_shape[0], orig_shape[-3], orig_shape[-2] // 2, orig_shape[-1] // 2, orig_shape[1], 2, 2)
+        hidden_states = hidden_states.permute(0, 4, 1, 2, 5, 3, 6)
         return hidden_states.reshape(orig_shape)[:, :, :, :x.shape[-2], :x.shape[-1]]

comfy/ldm/util.py

@@ -71,7 +71,7 @@ def count_params(model, verbose=False):
 
 
 def instantiate_from_config(config):
-    if not "target" in config:
+    if "target" not in config:
         if config == '__is_first_stage__':
             return None
         elif config == "__is_unconditional__":

comfy/ldm/wan/model.py

@@ -62,6 +62,8 @@ class WanSelfAttention(nn.Module):
             x(Tensor): Shape [B, L, num_heads, C / num_heads]
             freqs(Tensor): Rope freqs, shape [1024, C / num_heads / 2]
         """
+        patches = transformer_options.get("patches", {})
+
         b, s, n, d = *x.shape[:2], self.num_heads, self.head_dim
 
         def qkv_fn_q(x):
@@ -86,6 +88,10 @@ class WanSelfAttention(nn.Module):
             transformer_options=transformer_options,
         )
 
+        if "attn1_patch" in patches:
+            for p in patches["attn1_patch"]:
+                x = p({"x": x, "q": q, "k": k, "transformer_options": transformer_options})
+
         x = self.o(x)
         return x
 
@@ -225,6 +231,8 @@ class WanAttentionBlock(nn.Module):
         """
         # assert e.dtype == torch.float32
 
+        patches = transformer_options.get("patches", {})
+
         if e.ndim < 4:
             e = (comfy.model_management.cast_to(self.modulation, dtype=x.dtype, device=x.device) + e).chunk(6, dim=1)
         else:
@@ -232,6 +240,7 @@ class WanAttentionBlock(nn.Module):
         # assert e[0].dtype == torch.float32
 
         # self-attention
+        x = x.contiguous() # otherwise implicit in LayerNorm
         y = self.self_attn(
             torch.addcmul(repeat_e(e[0], x), self.norm1(x), 1 + repeat_e(e[1], x)),
             freqs, transformer_options=transformer_options)
@@ -241,6 +250,11 @@ class WanAttentionBlock(nn.Module):
 
         # cross-attention & ffn
         x = x + self.cross_attn(self.norm3(x), context, context_img_len=context_img_len, transformer_options=transformer_options)
+
+        if "attn2_patch" in patches:
+            for p in patches["attn2_patch"]:
+                x = p({"x": x, "transformer_options": transformer_options})
+
         y = self.ffn(torch.addcmul(repeat_e(e[3], x), self.norm2(x), 1 + repeat_e(e[4], x)))
         x = torch.addcmul(x, y, repeat_e(e[5], x))
         return x
@@ -487,7 +501,7 @@ class WanModel(torch.nn.Module):
         self.blocks = nn.ModuleList([
             wan_attn_block_class(cross_attn_type, dim, ffn_dim, num_heads,
                                  window_size, qk_norm, cross_attn_norm, eps, operation_settings=operation_settings)
-            for _ in range(num_layers)
+            for i in range(num_layers)
         ])
 
         # head
@@ -540,6 +554,7 @@ class WanModel(torch.nn.Module):
         # embeddings
         x = self.patch_embedding(x.float()).to(x.dtype)
         grid_sizes = x.shape[2:]
+        transformer_options["grid_sizes"] = grid_sizes
         x = x.flatten(2).transpose(1, 2)
 
         # time embeddings
@@ -567,7 +582,10 @@ class WanModel(torch.nn.Module):
 
         patches_replace = transformer_options.get("patches_replace", {})
         blocks_replace = patches_replace.get("dit", {})
+        transformer_options["total_blocks"] = len(self.blocks)
+        transformer_options["block_type"] = "double"
         for i, block in enumerate(self.blocks):
+            transformer_options["block_index"] = i
             if ("double_block", i) in blocks_replace:
                 def block_wrap(args):
                     out = {}
@@ -734,6 +752,7 @@ class VaceWanModel(WanModel):
         # embeddings
         x = self.patch_embedding(x.float()).to(x.dtype)
         grid_sizes = x.shape[2:]
+        transformer_options["grid_sizes"] = grid_sizes
         x = x.flatten(2).transpose(1, 2)
 
         # time embeddings
@@ -762,7 +781,10 @@ class VaceWanModel(WanModel):
 
         patches_replace = transformer_options.get("patches_replace", {})
         blocks_replace = patches_replace.get("dit", {})
+        transformer_options["total_blocks"] = len(self.blocks)
+        transformer_options["block_type"] = "double"
         for i, block in enumerate(self.blocks):
+            transformer_options["block_index"] = i
             if ("double_block", i) in blocks_replace:
                 def block_wrap(args):
                     out = {}
@@ -861,7 +883,10 @@ class CameraWanModel(WanModel):
 
         patches_replace = transformer_options.get("patches_replace", {})
         blocks_replace = patches_replace.get("dit", {})
+        transformer_options["total_blocks"] = len(self.blocks)
+        transformer_options["block_type"] = "double"
         for i, block in enumerate(self.blocks):
+            transformer_options["block_index"] = i
             if ("double_block", i) in blocks_replace:
                 def block_wrap(args):
                     out = {}
@@ -1325,16 +1350,19 @@ class WanModel_S2V(WanModel):
 
         patches_replace = transformer_options.get("patches_replace", {})
         blocks_replace = patches_replace.get("dit", {})
+        transformer_options["total_blocks"] = len(self.blocks)
+        transformer_options["block_type"] = "double"
         for i, block in enumerate(self.blocks):
+            transformer_options["block_index"] = i
             if ("double_block", i) in blocks_replace:
                 def block_wrap(args):
                     out = {}
-                    out["img"] = block(args["img"], context=args["txt"], e=args["vec"], freqs=args["pe"])
+                    out["img"] = block(args["img"], context=args["txt"], e=args["vec"], freqs=args["pe"], transformer_options=args["transformer_options"])
                     return out
-                out = blocks_replace[("double_block", i)]({"img": x, "txt": context, "vec": e0, "pe": freqs}, {"original_block": block_wrap})
+                out = blocks_replace[("double_block", i)]({"img": x, "txt": context, "vec": e0, "pe": freqs, "transformer_options": transformer_options}, {"original_block": block_wrap})
                 x = out["img"]
             else:
-                x = block(x, e=e0, freqs=freqs, context=context)
+                x = block(x, e=e0, freqs=freqs, context=context, transformer_options=transformer_options)
             if audio_emb is not None:
                 x = self.audio_injector(x, i, audio_emb, audio_emb_global, seq_len)
         # head
@@ -1573,7 +1601,10 @@ class HumoWanModel(WanModel):
 
         patches_replace = transformer_options.get("patches_replace", {})
         blocks_replace = patches_replace.get("dit", {})
+        transformer_options["total_blocks"] = len(self.blocks)
+        transformer_options["block_type"] = "double"
         for i, block in enumerate(self.blocks):
+            transformer_options["block_index"] = i
             if ("double_block", i) in blocks_replace:
                 def block_wrap(args):
                     out = {}

comfy/ldm/wan/model_animate.py

@@ -523,7 +523,10 @@ class AnimateWanModel(WanModel):
 
         patches_replace = transformer_options.get("patches_replace", {})
         blocks_replace = patches_replace.get("dit", {})
+        transformer_options["total_blocks"] = len(self.blocks)
+        transformer_options["block_type"] = "double"
         for i, block in enumerate(self.blocks):
+            transformer_options["block_index"] = i
             if ("double_block", i) in blocks_replace:
                 def block_wrap(args):
                     out = {}

comfy/ldm/wan/model_multitalk.py

@@ -0,0 +1,500 @@
+import torch
+from einops import rearrange, repeat
+import comfy
+from comfy.ldm.modules.attention import optimized_attention
+
+
+def calculate_x_ref_attn_map(visual_q, ref_k, ref_target_masks, split_num=8):
+    scale = 1.0 / visual_q.shape[-1] ** 0.5
+    visual_q = visual_q.transpose(1, 2) * scale
+
+    B, H, x_seqlens, K = visual_q.shape
+
+    x_ref_attn_maps = []
+    for class_idx, ref_target_mask in enumerate(ref_target_masks):
+        ref_target_mask = ref_target_mask.view(1, 1, 1, -1)
+
+        x_ref_attnmap = torch.zeros(B, H, x_seqlens, device=visual_q.device, dtype=visual_q.dtype)
+        chunk_size = min(max(x_seqlens // split_num, 1), x_seqlens)
+
+        for i in range(0, x_seqlens, chunk_size):
+            end_i = min(i + chunk_size, x_seqlens)
+
+            attn_chunk = visual_q[:, :, i:end_i] @ ref_k.permute(0, 2, 3, 1)  # B, H, chunk, ref_seqlens
+
+            # Apply softmax
+            attn_max = attn_chunk.max(dim=-1, keepdim=True).values
+            attn_chunk = (attn_chunk - attn_max).exp()
+            attn_sum = attn_chunk.sum(dim=-1, keepdim=True)
+            attn_chunk = attn_chunk / (attn_sum + 1e-8)
+
+            # Apply mask and sum
+            masked_attn = attn_chunk * ref_target_mask
+            x_ref_attnmap[:, :, i:end_i] = masked_attn.sum(-1) / (ref_target_mask.sum() + 1e-8)
+
+            del attn_chunk, masked_attn
+
+        # Average across heads
+        x_ref_attnmap = x_ref_attnmap.mean(dim=1)  # B, x_seqlens
+        x_ref_attn_maps.append(x_ref_attnmap)
+
+    del visual_q, ref_k
+
+    return torch.cat(x_ref_attn_maps, dim=0)
+
+def get_attn_map_with_target(visual_q, ref_k, shape, ref_target_masks=None, split_num=2):
+    """Args:
+        query (torch.tensor): B M H K
+        key (torch.tensor): B M H K
+        shape (tuple): (N_t, N_h, N_w)
+        ref_target_masks: [B, N_h * N_w]
+    """
+
+    N_t, N_h, N_w = shape
+
+    x_seqlens = N_h * N_w
+    ref_k     = ref_k[:, :x_seqlens]
+    _, seq_lens, heads, _ = visual_q.shape
+    class_num, _ = ref_target_masks.shape
+    x_ref_attn_maps = torch.zeros(class_num, seq_lens).to(visual_q)
+
+    split_chunk = heads // split_num
+
+    for i in range(split_num):
+        x_ref_attn_maps_perhead = calculate_x_ref_attn_map(
+            visual_q[:, :, i*split_chunk:(i+1)*split_chunk, :],
+            ref_k[:, :, i*split_chunk:(i+1)*split_chunk, :],
+            ref_target_masks
+            )
+        x_ref_attn_maps += x_ref_attn_maps_perhead
+
+    return x_ref_attn_maps / split_num
+
+
+def normalize_and_scale(column, source_range, target_range, epsilon=1e-8):
+    source_min, source_max = source_range
+    new_min, new_max = target_range
+    normalized = (column - source_min) / (source_max - source_min + epsilon)
+    scaled = normalized * (new_max - new_min) + new_min
+    return scaled
+
+
+def rotate_half(x):
+    x = rearrange(x, "... (d r) -> ... d r", r=2)
+    x1, x2 = x.unbind(dim=-1)
+    x = torch.stack((-x2, x1), dim=-1)
+    return rearrange(x, "... d r -> ... (d r)")
+
+
+def get_audio_embeds(encoded_audio, audio_start, audio_end):
+    audio_embs = []
+    human_num = len(encoded_audio)
+    audio_frames = encoded_audio[0].shape[0]
+
+    indices = (torch.arange(4 + 1) - 2) * 1
+
+    for human_idx in range(human_num):
+        if audio_end > audio_frames: # in case of not enough audio for current window, pad with first audio frame as that's most likely silence
+            pad_len = audio_end - audio_frames
+            pad_shape = list(encoded_audio[human_idx].shape)
+            pad_shape[0] = pad_len
+            pad_tensor = encoded_audio[human_idx][:1].repeat(pad_len, *([1] * (encoded_audio[human_idx].dim() - 1)))
+            encoded_audio_in = torch.cat([encoded_audio[human_idx], pad_tensor], dim=0)
+        else:
+            encoded_audio_in = encoded_audio[human_idx]
+        center_indices = torch.arange(audio_start, audio_end, 1).unsqueeze(1) + indices.unsqueeze(0)
+        center_indices = torch.clamp(center_indices, min=0, max=encoded_audio_in.shape[0] - 1)
+        audio_emb = encoded_audio_in[center_indices].unsqueeze(0)
+        audio_embs.append(audio_emb)
+
+    return torch.cat(audio_embs, dim=0)
+
+
+def project_audio_features(audio_proj, encoded_audio, audio_start, audio_end):
+    audio_embs = get_audio_embeds(encoded_audio, audio_start, audio_end)
+
+    first_frame_audio_emb_s = audio_embs[:, :1, ...]
+    latter_frame_audio_emb = audio_embs[:, 1:, ...]
+    latter_frame_audio_emb = rearrange(latter_frame_audio_emb, "b (n_t n) w s c -> b n_t n w s c", n=4)
+
+    middle_index = audio_proj.seq_len // 2
+
+    latter_first_frame_audio_emb = latter_frame_audio_emb[:, :, :1, :middle_index+1, ...]
+    latter_first_frame_audio_emb = rearrange(latter_first_frame_audio_emb, "b n_t n w s c -> b n_t (n w) s c")
+    latter_last_frame_audio_emb = latter_frame_audio_emb[:, :, -1:, middle_index:, ...]
+    latter_last_frame_audio_emb = rearrange(latter_last_frame_audio_emb, "b n_t n w s c -> b n_t (n w) s c")
+    latter_middle_frame_audio_emb = latter_frame_audio_emb[:, :, 1:-1, middle_index:middle_index+1, ...]
+    latter_middle_frame_audio_emb = rearrange(latter_middle_frame_audio_emb, "b n_t n w s c -> b n_t (n w) s c")
+    latter_frame_audio_emb_s = torch.cat([latter_first_frame_audio_emb, latter_middle_frame_audio_emb, latter_last_frame_audio_emb], dim=2)
+
+    audio_emb = audio_proj(first_frame_audio_emb_s, latter_frame_audio_emb_s)
+    audio_emb = torch.cat(audio_emb.split(1), dim=2)
+
+    return audio_emb
+
+
+class RotaryPositionalEmbedding1D(torch.nn.Module):
+    def __init__(self,
+                 head_dim,
+                 ):
+        super().__init__()
+        self.head_dim = head_dim
+        self.base = 10000
+
+    def precompute_freqs_cis_1d(self, pos_indices):
+        freqs = 1.0 / (self.base ** (torch.arange(0, self.head_dim, 2)[: (self.head_dim // 2)].float() / self.head_dim))
+        freqs = freqs.to(pos_indices.device)
+        freqs = torch.einsum("..., f -> ... f", pos_indices.float(), freqs)
+        freqs = repeat(freqs, "... n -> ... (n r)", r=2)
+        return freqs
+
+    def forward(self, x, pos_indices):
+        freqs_cis = self.precompute_freqs_cis_1d(pos_indices)
+
+        x_ = x.float()
+
+        freqs_cis = freqs_cis.float().to(x.device)
+        cos, sin = freqs_cis.cos(), freqs_cis.sin()
+        cos, sin = rearrange(cos, 'n d -> 1 1 n d'), rearrange(sin, 'n d -> 1 1 n d')
+        x_ = (x_ * cos) + (rotate_half(x_) * sin)
+
+        return x_.type_as(x)
+
+class SingleStreamAttention(torch.nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        encoder_hidden_states_dim: int,
+        num_heads: int,
+        qkv_bias: bool,
+        device=None, dtype=None, operations=None
+    ) -> None:
+        super().__init__()
+        self.dim = dim
+        self.encoder_hidden_states_dim = encoder_hidden_states_dim
+        self.num_heads = num_heads
+        self.head_dim = dim // num_heads
+
+        self.q_linear = operations.Linear(dim, dim, bias=qkv_bias, device=device, dtype=dtype)
+        self.proj = operations.Linear(dim, dim, device=device, dtype=dtype)
+        self.kv_linear = operations.Linear(encoder_hidden_states_dim, dim * 2, bias=qkv_bias, device=device, dtype=dtype)
+
+    def forward(self, x: torch.Tensor, encoder_hidden_states: torch.Tensor, shape=None) -> torch.Tensor:
+        N_t, N_h, N_w = shape
+
+        expected_tokens = N_t * N_h * N_w
+        actual_tokens = x.shape[1]
+        x_extra = None
+
+        if actual_tokens != expected_tokens:
+            x_extra = x[:, -N_h * N_w:, :]
+            x = x[:, :-N_h * N_w, :]
+            N_t = N_t - 1
+
+        B = x.shape[0]
+        S = N_h * N_w
+        x = x.view(B * N_t, S, self.dim)
+
+        # get q for hidden_state
+        q = self.q_linear(x).view(B * N_t, S, self.num_heads, self.head_dim)
+
+        # get kv from encoder_hidden_states # shape: (B, N, num_heads, head_dim)
+        kv = self.kv_linear(encoder_hidden_states)
+        encoder_k, encoder_v = kv.view(B * N_t, encoder_hidden_states.shape[1], 2, self.num_heads, self.head_dim).unbind(2)
+
+        #print("q.shape", q.shape) #torch.Size([21, 1024, 40, 128])
+        x = optimized_attention(
+            q.transpose(1, 2),
+            encoder_k.transpose(1, 2),
+            encoder_v.transpose(1, 2),
+            heads=self.num_heads, skip_reshape=True, skip_output_reshape=True).transpose(1, 2)
+
+        # linear transform
+        x = self.proj(x.reshape(B * N_t, S, self.dim))
+        x = x.view(B, N_t * S, self.dim)
+
+        if x_extra is not None:
+            x = torch.cat([x, torch.zeros_like(x_extra)], dim=1)
+
+        return x
+
+class SingleStreamMultiAttention(SingleStreamAttention):
+    def __init__(
+        self,
+        dim: int,
+        encoder_hidden_states_dim: int,
+        num_heads: int,
+        qkv_bias: bool,
+        class_range: int = 24,
+        class_interval: int = 4,
+        device=None, dtype=None, operations=None
+    ) -> None:
+        super().__init__(
+            dim=dim,
+            encoder_hidden_states_dim=encoder_hidden_states_dim,
+            num_heads=num_heads,
+            qkv_bias=qkv_bias,
+            device=device,
+            dtype=dtype,
+            operations=operations
+        )
+
+        # Rotary-embedding layout parameters
+        self.class_interval = class_interval
+        self.class_range = class_range
+        self.max_humans = self.class_range // self.class_interval
+
+        # Constant bucket used for background tokens
+        self.rope_bak = int(self.class_range // 2)
+
+        self.rope_1d = RotaryPositionalEmbedding1D(self.head_dim)
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        encoder_hidden_states: torch.Tensor,
+        shape=None,
+        x_ref_attn_map=None
+    ) -> torch.Tensor:
+        encoder_hidden_states = encoder_hidden_states.squeeze(0).to(x.device)
+        human_num = x_ref_attn_map.shape[0] if x_ref_attn_map is not None else 1
+        # Single-speaker fall-through
+        if human_num <= 1:
+            return super().forward(x, encoder_hidden_states, shape)
+
+        N_t, N_h, N_w = shape
+
+        x_extra = None
+        if x.shape[0] * N_t != encoder_hidden_states.shape[0]:
+            x_extra = x[:, -N_h * N_w:, :]
+            x = x[:, :-N_h * N_w, :]
+            N_t = N_t - 1
+        x = rearrange(x, "B (N_t S) C -> (B N_t) S C", N_t=N_t)
+
+        # Query projection
+        B, N, C = x.shape
+        q = self.q_linear(x)
+        q = q.view(B, N, self.num_heads, self.head_dim).permute(0, 2, 1, 3)
+
+        # Use `class_range` logic for 2 speakers
+        rope_h1 = (0, self.class_interval)
+        rope_h2 = (self.class_range - self.class_interval, self.class_range)
+        rope_bak = int(self.class_range // 2)
+
+        # Normalize and scale attention maps for each speaker
+        max_values = x_ref_attn_map.max(1).values[:, None, None]
+        min_values = x_ref_attn_map.min(1).values[:, None, None]
+        max_min_values = torch.cat([max_values, min_values], dim=2)
+
+        human1_max_value, human1_min_value = max_min_values[0, :, 0].max(), max_min_values[0, :, 1].min()
+        human2_max_value, human2_min_value = max_min_values[1, :, 0].max(), max_min_values[1, :, 1].min()
+
+        human1 = normalize_and_scale(x_ref_attn_map[0], (human1_min_value, human1_max_value), rope_h1)
+        human2 = normalize_and_scale(x_ref_attn_map[1], (human2_min_value, human2_max_value), rope_h2)
+        back = torch.full((x_ref_attn_map.size(1),), rope_bak, dtype=human1.dtype, device=human1.device)
+
+        # Token-wise speaker dominance
+        max_indices = x_ref_attn_map.argmax(dim=0)
+        normalized_map = torch.stack([human1, human2, back], dim=1)
+        normalized_pos = normalized_map[torch.arange(x_ref_attn_map.size(1)), max_indices]
+
+        # Apply rotary to Q
+        q = rearrange(q, "(B N_t) H S C -> B H (N_t S) C", N_t=N_t)
+        q = self.rope_1d(q, normalized_pos)
+        q = rearrange(q, "B H (N_t S) C -> (B N_t) H S C", N_t=N_t)
+
+        # Keys / Values
+        _, N_a, _ = encoder_hidden_states.shape
+        encoder_kv = self.kv_linear(encoder_hidden_states)
+        encoder_kv = encoder_kv.view(B, N_a, 2, self.num_heads, self.head_dim).permute(2, 0, 3, 1, 4)
+        encoder_k, encoder_v = encoder_kv.unbind(0)
+
+        # Rotary for keys – assign centre of each speaker bucket to its context tokens
+        per_frame = torch.zeros(N_a, dtype=encoder_k.dtype, device=encoder_k.device)
+        per_frame[: per_frame.size(0) // 2] = (rope_h1[0] + rope_h1[1]) / 2
+        per_frame[per_frame.size(0) // 2 :] = (rope_h2[0] + rope_h2[1]) / 2
+        encoder_pos = torch.cat([per_frame] * N_t, dim=0)
+
+        encoder_k = rearrange(encoder_k, "(B N_t) H S C -> B H (N_t S) C", N_t=N_t)
+        encoder_k = self.rope_1d(encoder_k, encoder_pos)
+        encoder_k = rearrange(encoder_k, "B H (N_t S) C -> (B N_t) H S C", N_t=N_t)
+
+        # Final attention
+        q = rearrange(q, "B H M K -> B M H K")
+        encoder_k = rearrange(encoder_k, "B H M K -> B M H K")
+        encoder_v = rearrange(encoder_v, "B H M K -> B M H K")
+
+        x = optimized_attention(
+            q.transpose(1, 2),
+            encoder_k.transpose(1, 2),
+            encoder_v.transpose(1, 2),
+            heads=self.num_heads, skip_reshape=True, skip_output_reshape=True).transpose(1, 2)
+
+        # Linear projection
+        x = x.reshape(B, N, C)
+        x = self.proj(x)
+
+        # Restore original layout
+        x = rearrange(x, "(B N_t) S C -> B (N_t S) C", N_t=N_t)
+        if x_extra is not None:
+            x = torch.cat([x, torch.zeros_like(x_extra)], dim=1)
+
+        return x
+
+
+class MultiTalkAudioProjModel(torch.nn.Module):
+    def __init__(
+        self,
+        seq_len: int = 5,
+        seq_len_vf: int = 12,
+        blocks: int = 12,
+        channels: int = 768,
+        intermediate_dim: int = 512,
+        out_dim: int = 768,
+        context_tokens: int = 32,
+        device=None, dtype=None, operations=None
+    ):
+        super().__init__()
+
+        self.seq_len = seq_len
+        self.blocks = blocks
+        self.channels = channels
+        self.input_dim = seq_len * blocks * channels
+        self.input_dim_vf = seq_len_vf * blocks * channels
+        self.intermediate_dim = intermediate_dim
+        self.context_tokens = context_tokens
+        self.out_dim = out_dim
+
+        # define multiple linear layers
+        self.proj1 = operations.Linear(self.input_dim, intermediate_dim, device=device, dtype=dtype)
+        self.proj1_vf = operations.Linear(self.input_dim_vf, intermediate_dim, device=device, dtype=dtype)
+        self.proj2 = operations.Linear(intermediate_dim, intermediate_dim, device=device, dtype=dtype)
+        self.proj3 = operations.Linear(intermediate_dim, context_tokens * out_dim, device=device, dtype=dtype)
+        self.norm = operations.LayerNorm(out_dim, device=device, dtype=dtype)
+
+    def forward(self, audio_embeds, audio_embeds_vf):
+        video_length = audio_embeds.shape[1] + audio_embeds_vf.shape[1]
+        B, _, _, S, C = audio_embeds.shape
+
+        # process audio of first frame
+        audio_embeds = rearrange(audio_embeds, "bz f w b c -> (bz f) w b c")
+        batch_size, window_size, blocks, channels = audio_embeds.shape
+        audio_embeds = audio_embeds.view(batch_size, window_size * blocks * channels)
+
+        # process audio of latter frame
+        audio_embeds_vf = rearrange(audio_embeds_vf, "bz f w b c -> (bz f) w b c")
+        batch_size_vf, window_size_vf, blocks_vf, channels_vf = audio_embeds_vf.shape
+        audio_embeds_vf = audio_embeds_vf.view(batch_size_vf, window_size_vf * blocks_vf * channels_vf)
+
+        # first projection
+        audio_embeds = torch.relu(self.proj1(audio_embeds))
+        audio_embeds_vf = torch.relu(self.proj1_vf(audio_embeds_vf))
+        audio_embeds = rearrange(audio_embeds, "(bz f) c -> bz f c", bz=B)
+        audio_embeds_vf = rearrange(audio_embeds_vf, "(bz f) c -> bz f c", bz=B)
+        audio_embeds_c = torch.concat([audio_embeds, audio_embeds_vf], dim=1)
+        batch_size_c, N_t, C_a = audio_embeds_c.shape
+        audio_embeds_c = audio_embeds_c.view(batch_size_c*N_t, C_a)
+
+        # second projection
+        audio_embeds_c = torch.relu(self.proj2(audio_embeds_c))
+
+        context_tokens = self.proj3(audio_embeds_c).reshape(batch_size_c*N_t, self.context_tokens, self.out_dim)
+
+        # normalization and reshape
+        context_tokens = self.norm(context_tokens)
+        context_tokens = rearrange(context_tokens, "(bz f) m c -> bz f m c", f=video_length)
+
+        return context_tokens
+
+
+class WanMultiTalkAttentionBlock(torch.nn.Module):
+    def __init__(self, in_dim=5120, out_dim=768, device=None, dtype=None, operations=None):
+        super().__init__()
+        self.audio_cross_attn = SingleStreamMultiAttention(in_dim, out_dim, num_heads=40, qkv_bias=True, device=device, dtype=dtype, operations=operations)
+        self.norm_x = operations.LayerNorm(in_dim, device=device, dtype=dtype, elementwise_affine=True)
+
+
+class MultiTalkGetAttnMapPatch:
+    def __init__(self, ref_target_masks=None):
+        self.ref_target_masks = ref_target_masks
+
+    def __call__(self, kwargs):
+        transformer_options = kwargs.get("transformer_options", {})
+        x = kwargs["x"]
+
+        if self.ref_target_masks is not None:
+            x_ref_attn_map = get_attn_map_with_target(kwargs["q"], kwargs["k"], transformer_options["grid_sizes"], ref_target_masks=self.ref_target_masks.to(x.device))
+            transformer_options["x_ref_attn_map"] = x_ref_attn_map
+        return x
+
+
+class MultiTalkCrossAttnPatch:
+    def __init__(self, model_patch, audio_scale=1.0, ref_target_masks=None):
+        self.model_patch = model_patch
+        self.audio_scale = audio_scale
+        self.ref_target_masks = ref_target_masks
+
+    def __call__(self, kwargs):
+        transformer_options = kwargs.get("transformer_options", {})
+        block_idx = transformer_options.get("block_index", None)
+        x = kwargs["x"]
+        if block_idx is None:
+            return torch.zeros_like(x)
+
+        audio_embeds = transformer_options.get("audio_embeds")
+        x_ref_attn_map = transformer_options.pop("x_ref_attn_map", None)
+
+        norm_x = self.model_patch.model.blocks[block_idx].norm_x(x)
+        x_audio = self.model_patch.model.blocks[block_idx].audio_cross_attn(
+            norm_x, audio_embeds.to(x.dtype),
+            shape=transformer_options["grid_sizes"],
+            x_ref_attn_map=x_ref_attn_map
+        )
+        x = x + x_audio * self.audio_scale
+        return x
+
+    def models(self):
+        return [self.model_patch]
+
+class MultiTalkApplyModelWrapper:
+    def __init__(self, init_latents):
+        self.init_latents = init_latents
+
+    def __call__(self, executor, x, *args, **kwargs):
+        x[:, :, :self.init_latents.shape[2]] = self.init_latents.to(x)
+        samples = executor(x, *args, **kwargs)
+        return samples
+
+
+class InfiniteTalkOuterSampleWrapper:
+    def __init__(self, motion_frames_latent, model_patch, is_extend=False):
+        self.motion_frames_latent = motion_frames_latent
+        self.model_patch = model_patch
+        self.is_extend = is_extend
+
+    def __call__(self, executor, *args, **kwargs):
+        model_patcher = executor.class_obj.model_patcher
+        model_options = executor.class_obj.model_options
+        process_latent_in = model_patcher.model.process_latent_in
+
+        # for InfiniteTalk, model input first latent(s) need to always be replaced on every step
+        if self.motion_frames_latent is not None:
+            wrappers = model_options["transformer_options"]["wrappers"]
+            w = wrappers.setdefault(comfy.patcher_extension.WrappersMP.APPLY_MODEL, {})
+            w["MultiTalk_apply_model"] = [MultiTalkApplyModelWrapper(process_latent_in(self.motion_frames_latent))]
+
+        # run the sampling process
+        result = executor(*args, **kwargs)
+
+        # insert motion frames before decoding
+        if self.is_extend:
+            overlap = self.motion_frames_latent.shape[2]
+            result = torch.cat([self.motion_frames_latent.to(result), result[:, :, overlap:]], dim=2)
+
+        return result
+
+    def to(self, device_or_dtype):
+        if isinstance(device_or_dtype, torch.device):
+            if self.motion_frames_latent is not None:
+                self.motion_frames_latent = self.motion_frames_latent.to(device_or_dtype)
+        return self

comfy/ldm/wan/vae.py

@@ -5,7 +5,7 @@ import torch
 import torch.nn as nn
 import torch.nn.functional as F
 from einops import rearrange
-from comfy.ldm.modules.diffusionmodules.model import vae_attention
+from comfy.ldm.modules.diffusionmodules.model import vae_attention, torch_cat_if_needed
 
 import comfy.ops
 ops = comfy.ops.disable_weight_init
@@ -20,22 +20,29 @@ class CausalConv3d(ops.Conv3d):
 
     def __init__(self, *args, **kwargs):
         super().__init__(*args, **kwargs)
-        self._padding = (self.padding[2], self.padding[2], self.padding[1],
-                         self.padding[1], 2 * self.padding[0], 0)
-        self.padding = (0, 0, 0)
+        self._padding = 2 * self.padding[0]
+        self.padding = (0, self.padding[1], self.padding[2])
 
     def forward(self, x, cache_x=None, cache_list=None, cache_idx=None):
         if cache_list is not None:
             cache_x = cache_list[cache_idx]
             cache_list[cache_idx] = None
 
-        padding = list(self._padding)
-        if cache_x is not None and self._padding[4] > 0:
-            cache_x = cache_x.to(x.device)
-            x = torch.cat([cache_x, x], dim=2)
-            padding[4] -= cache_x.shape[2]
+        if cache_x is None and x.shape[2] == 1:
+            #Fast path - the op will pad for use by truncating the weight
+            #and save math on a pile of zeros.
+            return super().forward(x, autopad="causal_zero")
+
+        if self._padding > 0:
+            padding_needed = self._padding
+            if cache_x is not None:
+                cache_x = cache_x.to(x.device)
+                padding_needed = max(0, padding_needed - cache_x.shape[2])
+            padding_shape = list(x.shape)
+            padding_shape[2] = padding_needed
+            padding = torch.zeros(padding_shape, device=x.device, dtype=x.dtype)
+            x = torch_cat_if_needed([padding, cache_x, x], dim=2)
             del cache_x
-        x = F.pad(x, padding)
 
         return super().forward(x)
 
@@ -227,6 +234,7 @@ class Encoder3d(nn.Module):
     def __init__(self,
                  dim=128,
                  z_dim=4,
+                 input_channels=3,
                  dim_mult=[1, 2, 4, 4],
                  num_res_blocks=2,
                  attn_scales=[],
@@ -245,7 +253,7 @@ class Encoder3d(nn.Module):
         scale = 1.0
 
         # init block
-        self.conv1 = CausalConv3d(3, dims[0], 3, padding=1)
+        self.conv1 = CausalConv3d(input_channels, dims[0], 3, padding=1)
 
         # downsample blocks
         downsamples = []
@@ -331,6 +339,7 @@ class Decoder3d(nn.Module):
     def __init__(self,
                  dim=128,
                  z_dim=4,
+                 output_channels=3,
                  dim_mult=[1, 2, 4, 4],
                  num_res_blocks=2,
                  attn_scales=[],
@@ -378,7 +387,7 @@ class Decoder3d(nn.Module):
         # output blocks
         self.head = nn.Sequential(
             RMS_norm(out_dim, images=False), nn.SiLU(),
-            CausalConv3d(out_dim, 3, 3, padding=1))
+            CausalConv3d(out_dim, output_channels, 3, padding=1))
 
     def forward(self, x, feat_cache=None, feat_idx=[0]):
         ## conv1
@@ -449,6 +458,7 @@ class WanVAE(nn.Module):
                  num_res_blocks=2,
                  attn_scales=[],
                  temperal_downsample=[True, True, False],
+                 image_channels=3,
                  dropout=0.0):
         super().__init__()
         self.dim = dim
@@ -460,19 +470,21 @@ class WanVAE(nn.Module):
         self.temperal_upsample = temperal_downsample[::-1]
 
         # modules
-        self.encoder = Encoder3d(dim, z_dim * 2, dim_mult, num_res_blocks,
+        self.encoder = Encoder3d(dim, z_dim * 2, image_channels, dim_mult, num_res_blocks,
                                  attn_scales, self.temperal_downsample, dropout)
         self.conv1 = CausalConv3d(z_dim * 2, z_dim * 2, 1)
         self.conv2 = CausalConv3d(z_dim, z_dim, 1)
-        self.decoder = Decoder3d(dim, z_dim, dim_mult, num_res_blocks,
+        self.decoder = Decoder3d(dim, z_dim, image_channels, dim_mult, num_res_blocks,
                                  attn_scales, self.temperal_upsample, dropout)
 
     def encode(self, x):
         conv_idx = [0]
-        feat_map = [None] * count_conv3d(self.decoder)
         ## cache
         t = x.shape[2]
         iter_ = 1 + (t - 1) // 4
+        feat_map = None
+        if iter_ > 1:
+            feat_map = [None] * count_conv3d(self.decoder)
         ## 对encode输入的x，按时间拆分为1、4、4、4....
         for i in range(iter_):
             conv_idx = [0]
@@ -492,10 +504,11 @@ class WanVAE(nn.Module):
 
     def decode(self, z):
         conv_idx = [0]
-        feat_map = [None] * count_conv3d(self.decoder)
         # z: [b,c,t,h,w]
-
         iter_ = z.shape[2]
+        feat_map = None
+        if iter_ > 1:
+            feat_map = [None] * count_conv3d(self.decoder)
         x = self.conv2(z)
         for i in range(iter_):
             conv_idx = [0]

comfy/lora.py

@@ -260,6 +260,7 @@ def model_lora_keys_unet(model, key_map={}):
                 key_map["transformer.{}".format(k[:-len(".weight")])] = to #simpletrainer and probably regular diffusers flux lora format
                 key_map["lycoris_{}".format(k[:-len(".weight")].replace(".", "_"))] = to #simpletrainer lycoris
                 key_map["lora_transformer_{}".format(k[:-len(".weight")].replace(".", "_"))] = to #onetrainer
+                key_map[k[:-len(".weight")]] = to #DiffSynth lora format
         for k in sdk:
             hidden_size = model.model_config.unet_config.get("hidden_size", 0)
             if k.endswith(".weight") and ".linear1." in k:
@@ -313,6 +314,30 @@ def model_lora_keys_unet(model, key_map={}):
                 key_map["transformer.{}".format(key_lora)] = k
                 key_map["lycoris_{}".format(key_lora.replace(".", "_"))] = k #SimpleTuner lycoris format
 
+    if isinstance(model, comfy.model_base.Lumina2):
+        diffusers_keys = comfy.utils.z_image_to_diffusers(model.model_config.unet_config, output_prefix="diffusion_model.")
+        for k in diffusers_keys:
+            if k.endswith(".weight"):
+                to = diffusers_keys[k]
+                key_lora = k[:-len(".weight")]
+                key_map["diffusion_model.{}".format(key_lora)] = to
+                key_map["transformer.{}".format(key_lora)] = to
+                key_map["lycoris_{}".format(key_lora.replace(".", "_"))] = to
+                key_map[key_lora] = to
+
+    if isinstance(model, comfy.model_base.Kandinsky5):
+        for k in sdk:
+            if k.startswith("diffusion_model.") and k.endswith(".weight"):
+                key_lora = k[len("diffusion_model."):-len(".weight")]
+                key_map["{}".format(key_lora)] = k
+                key_map["transformer.{}".format(key_lora)] = k
+
+    if isinstance(model, comfy.model_base.ACEStep15):
+        for k in sdk:
+            if k.startswith("diffusion_model.decoder.") and k.endswith(".weight"):
+                key_lora = k[len("diffusion_model.decoder."):-len(".weight")]
+                key_map["base_model.model.{}".format(key_lora)] = k  # Official base model loras
+
     return key_map
 
 

comfy/memory_management.py

@@ -0,0 +1,81 @@
+import math
+import torch
+from typing import NamedTuple
+
+from comfy.quant_ops import QuantizedTensor
+
+class TensorGeometry(NamedTuple):
+    shape: any
+    dtype: torch.dtype
+
+    def element_size(self):
+        info = torch.finfo(self.dtype) if self.dtype.is_floating_point else torch.iinfo(self.dtype)
+        return info.bits // 8
+
+    def numel(self):
+        return math.prod(self.shape)
+
+def tensors_to_geometries(tensors, dtype=None):
+    geometries = []
+    for t in tensors:
+        if t is None or isinstance(t, QuantizedTensor):
+            geometries.append(t)
+            continue
+        tdtype = t.dtype
+        if hasattr(t, "_model_dtype"):
+            tdtype = t._model_dtype
+        if dtype is not None:
+            tdtype = dtype
+        geometries.append(TensorGeometry(shape=t.shape, dtype=tdtype))
+    return geometries
+
+def vram_aligned_size(tensor):
+    if isinstance(tensor, list):
+        return sum([vram_aligned_size(t) for t in tensor])
+
+    if isinstance(tensor, QuantizedTensor):
+        inner_tensors, _ = tensor.__tensor_flatten__()
+        return vram_aligned_size([ getattr(tensor, attr) for attr in inner_tensors ])
+
+    if tensor is None:
+        return 0
+
+    size = tensor.numel() * tensor.element_size()
+    aligment_req = 1024
+    return (size + aligment_req - 1) // aligment_req * aligment_req
+
+def interpret_gathered_like(tensors, gathered):
+    offset = 0
+    dest_views = []
+
+    if gathered.dim() != 1 or gathered.element_size() != 1:
+        raise ValueError(f"Buffer must be 1D and single-byte (got {gathered.dim()}D {gathered.dtype})")
+
+    for tensor in tensors:
+
+        if tensor is None:
+            dest_views.append(None)
+            continue
+
+        if isinstance(tensor, QuantizedTensor):
+            inner_tensors, qt_ctx = tensor.__tensor_flatten__()
+            templates = { attr: getattr(tensor, attr) for attr in inner_tensors }
+        else:
+            templates = { "data": tensor }
+
+        actuals = {}
+        for attr, template in templates.items():
+            size = template.numel() * template.element_size()
+            if offset + size > gathered.numel():
+                raise ValueError(f"Buffer too small: needs {offset + size} bytes, but only has {gathered.numel()}. ")
+            actuals[attr] = gathered[offset:offset+size].view(dtype=template.dtype).view(template.shape)
+            offset += vram_aligned_size(template)
+
+        if isinstance(tensor, QuantizedTensor):
+            dest_views.append(QuantizedTensor.__tensor_unflatten__(actuals, qt_ctx, 0, 0))
+        else:
+            dest_views.append(actuals["data"])
+
+    return dest_views
+
+aimdo_allocator = None

comfy/model_base.py

@@ -20,6 +20,7 @@ import comfy.ldm.hunyuan3dv2_1
 import comfy.ldm.hunyuan3dv2_1.hunyuandit
 import torch
 import logging
+import comfy.ldm.lightricks.av_model
 from comfy.ldm.modules.diffusionmodules.openaimodel import UNetModel, Timestep
 from comfy.ldm.cascade.stage_c import StageC
 from comfy.ldm.cascade.stage_b import StageB
@@ -47,6 +48,9 @@ import comfy.ldm.chroma_radiance.model
 import comfy.ldm.ace.model
 import comfy.ldm.omnigen.omnigen2
 import comfy.ldm.qwen_image.model
+import comfy.ldm.kandinsky5.model
+import comfy.ldm.anima.model
+import comfy.ldm.ace.ace_step15
 
 import comfy.model_management
 import comfy.patcher_extension
@@ -134,7 +138,7 @@ class BaseModel(torch.nn.Module):
         if not unet_config.get("disable_unet_model_creation", False):
             if model_config.custom_operations is None:
                 fp8 = model_config.optimizations.get("fp8", False)
-                operations = comfy.ops.pick_operations(unet_config.get("dtype", None), self.manual_cast_dtype, fp8_optimizations=fp8, scaled_fp8=model_config.scaled_fp8, model_config=model_config)
+                operations = comfy.ops.pick_operations(unet_config.get("dtype", None), self.manual_cast_dtype, fp8_optimizations=fp8, model_config=model_config)
             else:
                 operations = model_config.custom_operations
             self.diffusion_model = unet_model(**unet_config, device=device, operations=operations)
@@ -143,6 +147,8 @@ class BaseModel(torch.nn.Module):
                 self.diffusion_model.to(memory_format=torch.channels_last)
                 logging.debug("using channels last mode for diffusion model")
             logging.info("model weight dtype {}, manual cast: {}".format(self.get_dtype(), self.manual_cast_dtype))
+            comfy.model_management.archive_model_dtypes(self.diffusion_model)
+
         self.model_type = model_type
         self.model_sampling = model_sampling(model_config, model_type)
 
@@ -296,7 +302,7 @@ class BaseModel(torch.nn.Module):
 
         return out
 
-    def load_model_weights(self, sd, unet_prefix=""):
+    def load_model_weights(self, sd, unet_prefix="", assign=False):
         to_load = {}
         keys = list(sd.keys())
         for k in keys:
@@ -304,7 +310,7 @@ class BaseModel(torch.nn.Module):
                 to_load[k[len(unet_prefix):]] = sd.pop(k)
 
         to_load = self.model_config.process_unet_state_dict(to_load)
-        m, u = self.diffusion_model.load_state_dict(to_load, strict=False)
+        m, u = self.diffusion_model.load_state_dict(to_load, strict=False, assign=assign)
         if len(m) > 0:
             logging.warning("unet missing: {}".format(m))
 
@@ -319,7 +325,7 @@ class BaseModel(torch.nn.Module):
     def process_latent_out(self, latent):
         return self.latent_format.process_out(latent)
 
-    def state_dict_for_saving(self, clip_state_dict=None, vae_state_dict=None, clip_vision_state_dict=None):
+    def state_dict_for_saving(self, unet_state_dict, clip_state_dict=None, vae_state_dict=None, clip_vision_state_dict=None):
         extra_sds = []
         if clip_state_dict is not None:
             extra_sds.append(self.model_config.process_clip_state_dict_for_saving(clip_state_dict))
@@ -327,22 +333,7 @@ class BaseModel(torch.nn.Module):
             extra_sds.append(self.model_config.process_vae_state_dict_for_saving(vae_state_dict))
         if clip_vision_state_dict is not None:
             extra_sds.append(self.model_config.process_clip_vision_state_dict_for_saving(clip_vision_state_dict))
-
-        unet_state_dict = self.diffusion_model.state_dict()
-
-        if self.model_config.scaled_fp8 is not None:
-            unet_state_dict["scaled_fp8"] = torch.tensor([], dtype=self.model_config.scaled_fp8)
-
-        # Save mixed precision metadata
-        if hasattr(self.model_config, 'layer_quant_config') and self.model_config.layer_quant_config:
-            metadata = {
-                "format_version": "1.0",
-                "layers": self.model_config.layer_quant_config
-            }
-            unet_state_dict["_quantization_metadata"] = metadata
-
         unet_state_dict = self.model_config.process_unet_state_dict_for_saving(unet_state_dict)
-
         if self.model_type == ModelType.V_PREDICTION:
             unet_state_dict["v_pred"] = torch.tensor([])
 
@@ -785,8 +776,8 @@ class StableAudio1(BaseModel):
             out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn)
         return out
 
-    def state_dict_for_saving(self, clip_state_dict=None, vae_state_dict=None, clip_vision_state_dict=None):
-        sd = super().state_dict_for_saving(clip_state_dict=clip_state_dict, vae_state_dict=vae_state_dict, clip_vision_state_dict=clip_vision_state_dict)
+    def state_dict_for_saving(self, unet_state_dict, clip_state_dict=None, vae_state_dict=None, clip_vision_state_dict=None):
+        sd = super().state_dict_for_saving(unet_state_dict, clip_state_dict=clip_state_dict, vae_state_dict=vae_state_dict, clip_vision_state_dict=clip_vision_state_dict)
         d = {"conditioner.conditioners.seconds_start.": self.seconds_start_embedder.state_dict(), "conditioner.conditioners.seconds_total.": self.seconds_total_embedder.state_dict()}
         for k in d:
             s = d[k]
@@ -898,12 +889,13 @@ class Flux(BaseModel):
         attention_mask = kwargs.get("attention_mask", None)
         if attention_mask is not None:
             shape = kwargs["noise"].shape
-            mask_ref_size = kwargs["attention_mask_img_shape"]
-            # the model will pad to the patch size, and then divide
-            # essentially dividing and rounding up
-            (h_tok, w_tok) = (math.ceil(shape[2] / self.diffusion_model.patch_size), math.ceil(shape[3] / self.diffusion_model.patch_size))
-            attention_mask = utils.upscale_dit_mask(attention_mask, mask_ref_size, (h_tok, w_tok))
-            out['attention_mask'] = comfy.conds.CONDRegular(attention_mask)
+            mask_ref_size = kwargs.get("attention_mask_img_shape", None)
+            if mask_ref_size is not None:
+                # the model will pad to the patch size, and then divide
+                # essentially dividing and rounding up
+                (h_tok, w_tok) = (math.ceil(shape[2] / self.diffusion_model.patch_size), math.ceil(shape[3] / self.diffusion_model.patch_size))
+                attention_mask = utils.upscale_dit_mask(attention_mask, mask_ref_size, (h_tok, w_tok))
+                out['attention_mask'] = comfy.conds.CONDRegular(attention_mask)
 
         guidance = kwargs.get("guidance", 3.5)
         if guidance is not None:
@@ -925,9 +917,19 @@ class Flux(BaseModel):
         out = {}
         ref_latents = kwargs.get("reference_latents", None)
         if ref_latents is not None:
-            out['ref_latents'] = list([1, 16, sum(map(lambda a: math.prod(a.size()), ref_latents)) // 16])
+            out['ref_latents'] = list([1, 16, sum(map(lambda a: math.prod(a.size()[2:]), ref_latents))])
         return out
 
+class Flux2(Flux):
+    def extra_conds(self, **kwargs):
+        out = super().extra_conds(**kwargs)
+        cross_attn = kwargs.get("cross_attn", None)
+        if cross_attn is not None:
+            target_text_len = 512
+            if cross_attn.shape[1] < target_text_len:
+                cross_attn = torch.nn.functional.pad(cross_attn, (0, 0, target_text_len - cross_attn.shape[1], 0))
+            out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn)
+        return out
 
 class GenmoMochi(BaseModel):
     def __init__(self, model_config, model_type=ModelType.FLOW, device=None):
@@ -946,7 +948,7 @@ class GenmoMochi(BaseModel):
 
 class LTXV(BaseModel):
     def __init__(self, model_config, model_type=ModelType.FLUX, device=None):
-        super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.lightricks.model.LTXVModel) #TODO
+        super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.lightricks.model.LTXVModel)
 
     def extra_conds(self, **kwargs):
         out = super().extra_conds(**kwargs)
@@ -977,6 +979,60 @@ class LTXV(BaseModel):
     def scale_latent_inpaint(self, sigma, noise, latent_image, **kwargs):
         return latent_image
 
+class LTXAV(BaseModel):
+    def __init__(self, model_config, model_type=ModelType.FLUX, device=None):
+        super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.lightricks.av_model.LTXAVModel) #TODO
+
+    def extra_conds(self, **kwargs):
+        out = super().extra_conds(**kwargs)
+        attention_mask = kwargs.get("attention_mask", None)
+        if attention_mask is not None:
+            out['attention_mask'] = comfy.conds.CONDRegular(attention_mask)
+        cross_attn = kwargs.get("cross_attn", None)
+        if cross_attn is not None:
+            out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn)
+
+        out['frame_rate'] = comfy.conds.CONDConstant(kwargs.get("frame_rate", 25))
+
+        denoise_mask = kwargs.get("concat_mask", kwargs.get("denoise_mask", None))
+
+        audio_denoise_mask = None
+        if denoise_mask is not None and "latent_shapes" in kwargs:
+            denoise_mask = utils.unpack_latents(denoise_mask, kwargs["latent_shapes"])
+            if len(denoise_mask) > 1:
+                audio_denoise_mask = denoise_mask[1]
+            denoise_mask = denoise_mask[0]
+
+        if denoise_mask is not None:
+            out["denoise_mask"] = comfy.conds.CONDRegular(denoise_mask)
+
+        if audio_denoise_mask is not None:
+            out["audio_denoise_mask"] = comfy.conds.CONDRegular(audio_denoise_mask)
+
+        keyframe_idxs = kwargs.get("keyframe_idxs", None)
+        if keyframe_idxs is not None:
+            out['keyframe_idxs'] = comfy.conds.CONDRegular(keyframe_idxs)
+
+        latent_shapes = kwargs.get("latent_shapes", None)
+        if latent_shapes is not None:
+            out['latent_shapes'] = comfy.conds.CONDConstant(latent_shapes)
+
+        return out
+
+    def process_timestep(self, timestep, x, denoise_mask=None, audio_denoise_mask=None, **kwargs):
+        v_timestep = timestep
+        a_timestep = timestep
+
+        if denoise_mask is not None:
+            v_timestep = self.diffusion_model.patchifier.patchify(((denoise_mask) * timestep.view([timestep.shape[0]] + [1] * (denoise_mask.ndim - 1)))[:, :1])[0]
+        if audio_denoise_mask is not None:
+            a_timestep = self.diffusion_model.a_patchifier.patchify(((audio_denoise_mask) * timestep.view([timestep.shape[0]] + [1] * (audio_denoise_mask.ndim - 1)))[:, :1, :, :1])[0]
+
+        return v_timestep, a_timestep
+
+    def scale_latent_inpaint(self, sigma, noise, latent_image, **kwargs):
+        return latent_image
+
 class HunyuanVideo(BaseModel):
     def __init__(self, model_config, model_type=ModelType.FLOW, device=None):
         super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.hunyuan_video.model.HunyuanVideo)
@@ -1092,9 +1148,35 @@ class CosmosPredict2(BaseModel):
         sigma = (sigma / (sigma + 1))
         return latent_image / (1.0 - sigma)
 
+class Anima(BaseModel):
+    def __init__(self, model_config, model_type=ModelType.FLOW, device=None):
+        super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.anima.model.Anima)
+
+    def extra_conds(self, **kwargs):
+        out = super().extra_conds(**kwargs)
+        cross_attn = kwargs.get("cross_attn", None)
+        t5xxl_ids = kwargs.get("t5xxl_ids", None)
+        t5xxl_weights = kwargs.get("t5xxl_weights", None)
+        device = kwargs["device"]
+        if cross_attn is not None:
+            if t5xxl_ids is not None:
+                if t5xxl_weights is not None:
+                    t5xxl_weights = t5xxl_weights.unsqueeze(0).unsqueeze(-1).to(cross_attn)
+                t5xxl_ids = t5xxl_ids.unsqueeze(0)
+
+                if torch.is_inference_mode_enabled():  # if not we are training
+                    cross_attn = self.diffusion_model.preprocess_text_embeds(cross_attn.to(device=device, dtype=self.get_dtype()), t5xxl_ids.to(device=device), t5xxl_weights=t5xxl_weights.to(device=device, dtype=self.get_dtype()))
+                else:
+                    out['t5xxl_ids'] = comfy.conds.CONDRegular(t5xxl_ids)
+                    out['t5xxl_weights'] = comfy.conds.CONDRegular(t5xxl_weights)
+
+            out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn)
+        return out
+
 class Lumina2(BaseModel):
     def __init__(self, model_config, model_type=ModelType.FLOW, device=None):
         super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.lumina.model.NextDiT)
+        self.memory_usage_factor_conds = ("ref_latents",)
 
     def extra_conds(self, **kwargs):
         out = super().extra_conds(**kwargs)
@@ -1103,9 +1185,46 @@ class Lumina2(BaseModel):
             if torch.numel(attention_mask) != attention_mask.sum():
                 out['attention_mask'] = comfy.conds.CONDRegular(attention_mask)
             out['num_tokens'] = comfy.conds.CONDConstant(max(1, torch.sum(attention_mask).item()))
+
         cross_attn = kwargs.get("cross_attn", None)
         if cross_attn is not None:
             out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn)
+            if 'num_tokens' not in out:
+                out['num_tokens'] = comfy.conds.CONDConstant(cross_attn.shape[1])
+
+        clip_text_pooled = kwargs.get("pooled_output", None)  # NewBie
+        if clip_text_pooled is not None:
+            out['clip_text_pooled'] = comfy.conds.CONDRegular(clip_text_pooled)
+
+        clip_vision_outputs = kwargs.get("clip_vision_outputs", list(map(lambda a: a.get("clip_vision_output"), kwargs.get("unclip_conditioning", [{}]))))  # Z Image omni
+        if clip_vision_outputs is not None and len(clip_vision_outputs) > 0:
+            sigfeats = []
+            for clip_vision_output in clip_vision_outputs:
+                if clip_vision_output is not None:
+                    image_size = clip_vision_output.image_sizes[0]
+                    shape = clip_vision_output.last_hidden_state.shape
+                    sigfeats.append(clip_vision_output.last_hidden_state.reshape(shape[0], image_size[1] // 16, image_size[2] // 16, shape[-1]))
+            if len(sigfeats) > 0:
+                out['siglip_feats'] = comfy.conds.CONDList(sigfeats)
+
+        ref_latents = kwargs.get("reference_latents", None)
+        if ref_latents is not None:
+            latents = []
+            for lat in ref_latents:
+                latents.append(self.process_latent_in(lat))
+            out['ref_latents'] = comfy.conds.CONDList(latents)
+
+        ref_contexts = kwargs.get("reference_latents_text_embeds", None)
+        if ref_contexts is not None:
+            out['ref_contexts'] = comfy.conds.CONDList(ref_contexts)
+
+        return out
+
+    def extra_conds_shapes(self, **kwargs):
+        out = {}
+        ref_latents = kwargs.get("reference_latents", None)
+        if ref_latents is not None:
+            out['ref_latents'] = list([1, 16, sum(map(lambda a: math.prod(a.size()[2:]), ref_latents))])
         return out
 
 class WAN21(BaseModel):
@@ -1426,6 +1545,49 @@ class ACEStep(BaseModel):
         out['lyrics_strength'] = comfy.conds.CONDConstant(kwargs.get("lyrics_strength", 1.0))
         return out
 
+class ACEStep15(BaseModel):
+    def __init__(self, model_config, model_type=ModelType.FLOW, device=None):
+        super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.ace.ace_step15.AceStepConditionGenerationModel)
+
+    def extra_conds(self, **kwargs):
+        out = super().extra_conds(**kwargs)
+        device = kwargs["device"]
+        noise = kwargs["noise"]
+
+        cross_attn = kwargs.get("cross_attn", None)
+        if cross_attn is not None:
+            if torch.count_nonzero(cross_attn) == 0:
+                out['replace_with_null_embeds'] = comfy.conds.CONDConstant(True)
+            out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn)
+
+        conditioning_lyrics = kwargs.get("conditioning_lyrics", None)
+        if cross_attn is not None:
+            out['lyric_embed'] = comfy.conds.CONDRegular(conditioning_lyrics)
+
+        refer_audio = kwargs.get("reference_audio_timbre_latents", None)
+        if refer_audio is None or len(refer_audio) == 0:
+            refer_audio = comfy.ldm.ace.ace_step15.get_silence_latent(noise.shape[2], device)
+            pass_audio_codes = True
+        else:
+            refer_audio = refer_audio[-1][:, :, :noise.shape[2]]
+            out['is_covers'] = comfy.conds.CONDConstant(True)
+            pass_audio_codes = False
+
+        if pass_audio_codes:
+            audio_codes = kwargs.get("audio_codes", None)
+            if audio_codes is not None:
+                out['audio_codes'] = comfy.conds.CONDRegular(torch.tensor(audio_codes, device=device))
+                refer_audio = refer_audio[:, :, :750]
+            else:
+                out['is_covers'] = comfy.conds.CONDConstant(False)
+
+        if refer_audio.shape[2] < noise.shape[2]:
+            pad = comfy.ldm.ace.ace_step15.get_silence_latent(noise.shape[2], device)
+            refer_audio = torch.cat([refer_audio.to(pad), pad[:, :, refer_audio.shape[2]:]], dim=2)
+
+        out['refer_audio'] = comfy.conds.CONDRegular(refer_audio)
+        return out
+
 class Omnigen2(BaseModel):
     def __init__(self, model_config, model_type=ModelType.FLOW, device=None):
         super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.omnigen.omnigen2.OmniGen2Transformer2DModel)
@@ -1463,6 +1625,9 @@ class QwenImage(BaseModel):
 
     def extra_conds(self, **kwargs):
         out = super().extra_conds(**kwargs)
+        attention_mask = kwargs.get("attention_mask", None)
+        if attention_mask is not None:
+            out['attention_mask'] = comfy.conds.CONDRegular(attention_mask)
         cross_attn = kwargs.get("cross_attn", None)
         if cross_attn is not None:
             out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn)
@@ -1536,3 +1701,140 @@ class HunyuanImage21Refiner(HunyuanImage21):
         out = super().extra_conds(**kwargs)
         out['disable_time_r'] = comfy.conds.CONDConstant(True)
         return out
+
+class HunyuanVideo15(HunyuanVideo):
+    def __init__(self, model_config, model_type=ModelType.FLOW, device=None):
+        super().__init__(model_config, model_type, device=device)
+
+    def concat_cond(self, **kwargs):
+        noise = kwargs.get("noise", None)
+        extra_channels = self.diffusion_model.img_in.proj.weight.shape[1] - noise.shape[1] - 1 #noise 32 img cond 32 + mask 1
+        if extra_channels == 0:
+            return None
+
+        image = kwargs.get("concat_latent_image", None)
+        device = kwargs["device"]
+
+        if image is None:
+            shape_image = list(noise.shape)
+            shape_image[1] = extra_channels
+            image = torch.zeros(shape_image, dtype=noise.dtype, layout=noise.layout, device=noise.device)
+        else:
+            latent_dim = self.latent_format.latent_channels
+            image = utils.common_upscale(image.to(device), noise.shape[-1], noise.shape[-2], "bilinear", "center")
+            for i in range(0, image.shape[1], latent_dim):
+                image[:, i: i + latent_dim] = self.process_latent_in(image[:, i: i + latent_dim])
+            image = utils.resize_to_batch_size(image, noise.shape[0])
+
+        mask = kwargs.get("concat_mask", kwargs.get("denoise_mask", None))
+        if mask is None:
+            mask = torch.zeros_like(noise)[:, :1]
+        else:
+            mask = 1.0 - mask
+            mask = utils.common_upscale(mask.to(device), noise.shape[-1], noise.shape[-2], "bilinear", "center")
+            if mask.shape[-3] < noise.shape[-3]:
+                mask = torch.nn.functional.pad(mask, (0, 0, 0, 0, 0, noise.shape[-3] - mask.shape[-3]), mode='constant', value=0)
+            mask = utils.resize_to_batch_size(mask, noise.shape[0])
+
+        return torch.cat((image, mask), dim=1)
+
+    def extra_conds(self, **kwargs):
+        out = super().extra_conds(**kwargs)
+        attention_mask = kwargs.get("attention_mask", None)
+        if attention_mask is not None:
+            if torch.numel(attention_mask) != attention_mask.sum():
+                out['attention_mask'] = comfy.conds.CONDRegular(attention_mask)
+        cross_attn = kwargs.get("cross_attn", None)
+        if cross_attn is not None:
+            out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn)
+
+        conditioning_byt5small = kwargs.get("conditioning_byt5small", None)
+        if conditioning_byt5small is not None:
+            out['txt_byt5'] = comfy.conds.CONDRegular(conditioning_byt5small)
+
+        guidance = kwargs.get("guidance", 6.0)
+        if guidance is not None:
+            out['guidance'] = comfy.conds.CONDRegular(torch.FloatTensor([guidance]))
+
+        clip_vision_output = kwargs.get("clip_vision_output", None)
+        if clip_vision_output is not None:
+            out['clip_fea'] = comfy.conds.CONDRegular(clip_vision_output.last_hidden_state)
+
+        return out
+
+class HunyuanVideo15_SR_Distilled(HunyuanVideo15):
+    def __init__(self, model_config, model_type=ModelType.FLOW, device=None):
+        super().__init__(model_config, model_type, device=device)
+
+    def concat_cond(self, **kwargs):
+        noise = kwargs.get("noise", None)
+        image = kwargs.get("concat_latent_image", None)
+        noise_augmentation = kwargs.get("noise_augmentation", 0.0)
+        device = kwargs["device"]
+
+        if image is None:
+            image = torch.zeros([noise.shape[0], noise.shape[1] * 2 + 2, noise.shape[-3], noise.shape[-2], noise.shape[-1]], device=comfy.model_management.intermediate_device())
+        else:
+            image = utils.common_upscale(image.to(device), noise.shape[-1], noise.shape[-2], "bilinear", "center")
+            #image = self.process_latent_in(image) # scaling wasn't applied in reference code
+            image = utils.resize_to_batch_size(image, noise.shape[0])
+            lq_image_slice = slice(noise.shape[1] + 1, 2 * noise.shape[1] + 1)
+            if noise_augmentation > 0:
+                generator = torch.Generator(device="cpu")
+                generator.manual_seed(kwargs.get("seed", 0) - 10)
+                noise = torch.randn(image[:, lq_image_slice].shape, generator=generator, dtype=image.dtype, device="cpu").to(image.device)
+                image[:, lq_image_slice] = noise_augmentation * noise + min(1.0 - noise_augmentation, 0.75) * image[:, lq_image_slice]
+            else:
+                image[:, lq_image_slice] = 0.75 * image[:, lq_image_slice]
+        return image
+
+    def extra_conds(self, **kwargs):
+        out = super().extra_conds(**kwargs)
+        out['disable_time_r'] = comfy.conds.CONDConstant(False)
+        return out
+
+class Kandinsky5(BaseModel):
+    def __init__(self, model_config, model_type=ModelType.FLOW, device=None):
+        super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.kandinsky5.model.Kandinsky5)
+
+    def encode_adm(self, **kwargs):
+        return kwargs["pooled_output"]
+
+    def concat_cond(self, **kwargs):
+        noise = kwargs.get("noise", None)
+        device = kwargs["device"]
+        image = torch.zeros_like(noise)
+
+        mask = kwargs.get("concat_mask", kwargs.get("denoise_mask", None))
+        if mask is None:
+            mask = torch.zeros_like(noise)[:, :1]
+        else:
+            mask = 1.0 - mask
+            mask = utils.common_upscale(mask.to(device), noise.shape[-1], noise.shape[-2], "bilinear", "center")
+            if mask.shape[-3] < noise.shape[-3]:
+                mask = torch.nn.functional.pad(mask, (0, 0, 0, 0, 0, noise.shape[-3] - mask.shape[-3]), mode='constant', value=0)
+            mask = utils.resize_to_batch_size(mask, noise.shape[0])
+
+        return torch.cat((image, mask), dim=1)
+
+    def extra_conds(self, **kwargs):
+        out = super().extra_conds(**kwargs)
+        attention_mask = kwargs.get("attention_mask", None)
+        if attention_mask is not None:
+            out['attention_mask'] = comfy.conds.CONDRegular(attention_mask)
+        cross_attn = kwargs.get("cross_attn", None)
+        if cross_attn is not None:
+            out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn)
+
+        time_dim_replace = kwargs.get("time_dim_replace", None)
+        if time_dim_replace is not None:
+            out['time_dim_replace'] = comfy.conds.CONDRegular(self.process_latent_in(time_dim_replace))
+
+        return out
+
+class Kandinsky5Image(Kandinsky5):
+    def __init__(self, model_config, model_type=ModelType.FLOW, device=None):
+        super().__init__(model_config, model_type, device=device)
+
+    def concat_cond(self, **kwargs):
+        return None

comfy/model_detection.py

@@ -6,20 +6,6 @@ import math
 import logging
 import torch
 
-
-def detect_layer_quantization(metadata):
-    quant_key = "_quantization_metadata"
-    if metadata is not None and quant_key in metadata:
-        quant_metadata = metadata.pop(quant_key)
-        quant_metadata = json.loads(quant_metadata)
-        if isinstance(quant_metadata, dict) and "layers" in quant_metadata:
-            logging.info(f"Found quantization metadata (version {quant_metadata.get('format_version', 'unknown')})")
-            return quant_metadata["layers"]
-        else:
-            raise ValueError("Invalid quantization metadata format")
-    return None
-
-
 def count_blocks(state_dict_keys, prefix_string):
     count = 0
     while True:
@@ -186,30 +172,75 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
 
         guidance_keys = list(filter(lambda a: a.startswith("{}guidance_in.".format(key_prefix)), state_dict_keys))
         dit_config["guidance_embed"] = len(guidance_keys) > 0
+
+        # HunyuanVideo 1.5
+        if '{}cond_type_embedding.weight'.format(key_prefix) in state_dict_keys:
+            dit_config["use_cond_type_embedding"] = True
+        else:
+            dit_config["use_cond_type_embedding"] = False
+        if '{}vision_in.proj.0.weight'.format(key_prefix) in state_dict_keys:
+            dit_config["vision_in_dim"] = state_dict['{}vision_in.proj.0.weight'.format(key_prefix)].shape[0]
+            dit_config["meanflow_sum"] = True
+        else:
+            dit_config["vision_in_dim"] = None
+            dit_config["meanflow_sum"] = False
         return dit_config
 
     if '{}double_blocks.0.img_attn.norm.key_norm.scale'.format(key_prefix) in state_dict_keys and ('{}img_in.weight'.format(key_prefix) in state_dict_keys or f"{key_prefix}distilled_guidance_layer.norms.0.scale" in state_dict_keys): #Flux, Chroma or Chroma Radiance (has no img_in.weight)
         dit_config = {}
-        dit_config["image_model"] = "flux"
+        if '{}double_stream_modulation_img.lin.weight'.format(key_prefix) in state_dict_keys:
+            dit_config["image_model"] = "flux2"
+            dit_config["axes_dim"] = [32, 32, 32, 32]
+            dit_config["num_heads"] = 48
+            dit_config["mlp_ratio"] = 3.0
+            dit_config["theta"] = 2000
+            dit_config["out_channels"] = 128
+            dit_config["global_modulation"] = True
+            dit_config["mlp_silu_act"] = True
+            dit_config["qkv_bias"] = False
+            dit_config["ops_bias"] = False
+            dit_config["default_ref_method"] = "index"
+            dit_config["ref_index_scale"] = 10.0
+            dit_config["txt_ids_dims"] = [3]
+            patch_size = 1
+        else:
+            dit_config["image_model"] = "flux"
+            dit_config["axes_dim"] = [16, 56, 56]
+            dit_config["num_heads"] = 24
+            dit_config["mlp_ratio"] = 4.0
+            dit_config["theta"] = 10000
+            dit_config["out_channels"] = 16
+            dit_config["qkv_bias"] = True
+            dit_config["txt_ids_dims"] = []
+            patch_size = 2
+
         dit_config["in_channels"] = 16
-        patch_size = 2
+        dit_config["hidden_size"] = 3072
+        dit_config["context_in_dim"] = 4096
+
         dit_config["patch_size"] = patch_size
         in_key = "{}img_in.weight".format(key_prefix)
         if in_key in state_dict_keys:
-            dit_config["in_channels"] = state_dict[in_key].shape[1] // (patch_size * patch_size)
-        dit_config["out_channels"] = 16
+            w = state_dict[in_key]
+            dit_config["in_channels"] = w.shape[1] // (patch_size * patch_size)
+            dit_config["hidden_size"] = w.shape[0]
+
+        txt_in_key = "{}txt_in.weight".format(key_prefix)
+        if txt_in_key in state_dict_keys:
+            w = state_dict[txt_in_key]
+            dit_config["context_in_dim"] = w.shape[1]
+            dit_config["hidden_size"] = w.shape[0]
+
         vec_in_key = '{}vector_in.in_layer.weight'.format(key_prefix)
         if vec_in_key in state_dict_keys:
             dit_config["vec_in_dim"] = state_dict[vec_in_key].shape[1]
-        dit_config["context_in_dim"] = 4096
-        dit_config["hidden_size"] = 3072
-        dit_config["mlp_ratio"] = 4.0
-        dit_config["num_heads"] = 24
+        else:
+            dit_config["vec_in_dim"] = None
+
+        dit_config["num_heads"] = dit_config["hidden_size"] // sum(dit_config["axes_dim"])
+
         dit_config["depth"] = count_blocks(state_dict_keys, '{}double_blocks.'.format(key_prefix) + '{}.')
         dit_config["depth_single_blocks"] = count_blocks(state_dict_keys, '{}single_blocks.'.format(key_prefix) + '{}.')
-        dit_config["axes_dim"] = [16, 56, 56]
-        dit_config["theta"] = 10000
-        dit_config["qkv_bias"] = True
         if '{}distilled_guidance_layer.0.norms.0.scale'.format(key_prefix) in state_dict_keys or '{}distilled_guidance_layer.norms.0.scale'.format(key_prefix) in state_dict_keys: #Chroma
             dit_config["image_model"] = "chroma"
             dit_config["in_channels"] = 64
@@ -222,7 +253,7 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
                 dit_config["image_model"] = "chroma_radiance"
                 dit_config["in_channels"] = 3
                 dit_config["out_channels"] = 3
-                dit_config["patch_size"] = 16
+                dit_config["patch_size"] = state_dict.get('{}img_in_patch.weight'.format(key_prefix)).size(dim=-1)
                 dit_config["nerf_hidden_size"] = 64
                 dit_config["nerf_mlp_ratio"] = 4
                 dit_config["nerf_depth"] = 4
@@ -230,8 +261,17 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
                 dit_config["nerf_tile_size"] = 512
                 dit_config["nerf_final_head_type"] = "conv" if f"{key_prefix}nerf_final_layer_conv.norm.scale" in state_dict_keys else "linear"
                 dit_config["nerf_embedder_dtype"] = torch.float32
+                if "{}__x0__".format(key_prefix) in state_dict_keys: # x0 pred
+                    dit_config["use_x0"] = True
+                else:
+                    dit_config["use_x0"] = False
         else:
             dit_config["guidance_embed"] = "{}guidance_in.in_layer.weight".format(key_prefix) in state_dict_keys
+            dit_config["yak_mlp"] = '{}double_blocks.0.img_mlp.gate_proj.weight'.format(key_prefix) in state_dict_keys
+            dit_config["txt_norm"] = "{}txt_norm.scale".format(key_prefix) in state_dict_keys
+            if dit_config["yak_mlp"] and dit_config["txt_norm"]:  # Ovis model
+                dit_config["txt_ids_dims"] = [1, 2]
+
         return dit_config
 
     if '{}t5_yproj.weight'.format(key_prefix) in state_dict_keys: #Genmo mochi preview
@@ -267,7 +307,7 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
 
     if '{}adaln_single.emb.timestep_embedder.linear_1.bias'.format(key_prefix) in state_dict_keys: #Lightricks ltxv
         dit_config = {}
-        dit_config["image_model"] = "ltxv"
+        dit_config["image_model"] = "ltxav" if f'{key_prefix}audio_adaln_single.linear.weight' in state_dict_keys else "ltxv"
         dit_config["num_layers"] = count_blocks(state_dict_keys, '{}transformer_blocks.'.format(key_prefix) + '{}.')
         shape = state_dict['{}transformer_blocks.0.attn2.to_k.weight'.format(key_prefix)].shape
         dit_config["attention_head_dim"] = shape[0] // 32
@@ -378,14 +418,42 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
         dit_config["image_model"] = "lumina2"
         dit_config["patch_size"] = 2
         dit_config["in_channels"] = 16
-        dit_config["dim"] = 2304
-        dit_config["cap_feat_dim"] = state_dict['{}cap_embedder.1.weight'.format(key_prefix)].shape[1]
+        w = state_dict['{}cap_embedder.1.weight'.format(key_prefix)]
+        dit_config["dim"] = w.shape[0]
+        dit_config["cap_feat_dim"] = w.shape[1]
         dit_config["n_layers"] = count_blocks(state_dict_keys, '{}layers.'.format(key_prefix) + '{}.')
-        dit_config["n_heads"] = 24
-        dit_config["n_kv_heads"] = 8
         dit_config["qk_norm"] = True
-        dit_config["axes_dims"] = [32, 32, 32]
-        dit_config["axes_lens"] = [300, 512, 512]
+
+        if dit_config["dim"] == 2304: # Original Lumina 2
+            dit_config["n_heads"] = 24
+            dit_config["n_kv_heads"] = 8
+            dit_config["axes_dims"] = [32, 32, 32]
+            dit_config["axes_lens"] = [300, 512, 512]
+            dit_config["rope_theta"] = 10000.0
+            dit_config["ffn_dim_multiplier"] = 4.0
+            ctd_weight = state_dict.get('{}clip_text_pooled_proj.0.weight'.format(key_prefix), None)
+            if ctd_weight is not None:  # NewBie
+                dit_config["clip_text_dim"] = ctd_weight.shape[0]
+                # NewBie also sets axes_lens = [1024, 512, 512] but it's not used in ComfyUI
+        elif dit_config["dim"] == 3840:  # Z image
+            dit_config["n_heads"] = 30
+            dit_config["n_kv_heads"] = 30
+            dit_config["axes_dims"] = [32, 48, 48]
+            dit_config["axes_lens"] = [1536, 512, 512]
+            dit_config["rope_theta"] = 256.0
+            dit_config["ffn_dim_multiplier"] = (8.0 / 3.0)
+            dit_config["z_image_modulation"] = True
+            dit_config["time_scale"] = 1000.0
+            try:
+                dit_config["allow_fp16"] = torch.std(state_dict['{}layers.{}.ffn_norm1.weight'.format(key_prefix, dit_config["n_layers"] - 2)], unbiased=False).item() < 0.42
+            except Exception:
+                pass
+            if '{}cap_pad_token'.format(key_prefix) in state_dict_keys:
+                dit_config["pad_tokens_multiple"] = 32
+            sig_weight = state_dict.get('{}siglip_embedder.0.weight'.format(key_prefix), None)
+            if sig_weight is not None:
+                dit_config["siglip_feat_dim"] = sig_weight.shape[0]
+
         return dit_config
 
     if '{}head.modulation'.format(key_prefix) in state_dict_keys:  # Wan 2.1
@@ -486,6 +554,8 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
     if '{}blocks.0.mlp.layer1.weight'.format(key_prefix) in state_dict_keys:  # Cosmos predict2
         dit_config = {}
         dit_config["image_model"] = "cosmos_predict2"
+        if "{}llm_adapter.blocks.0.cross_attn.q_proj.weight".format(key_prefix) in state_dict_keys:
+            dit_config["image_model"] = "anima"
         dit_config["max_img_h"] = 240
         dit_config["max_img_w"] = 240
         dit_config["max_frames"] = 128
@@ -560,6 +630,34 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
         dit_config["image_model"] = "qwen_image"
         dit_config["in_channels"] = state_dict['{}img_in.weight'.format(key_prefix)].shape[1]
         dit_config["num_layers"] = count_blocks(state_dict_keys, '{}transformer_blocks.'.format(key_prefix) + '{}.')
+        if "{}__index_timestep_zero__".format(key_prefix) in state_dict_keys:  # 2511
+            dit_config["default_ref_method"] = "index_timestep_zero"
+        if "{}time_text_embed.addition_t_embedding.weight".format(key_prefix) in state_dict_keys:  # Layered
+            dit_config["use_additional_t_cond"] = True
+            dit_config["default_ref_method"] = "negative_index"
+        return dit_config
+
+    if '{}visual_transformer_blocks.0.cross_attention.key_norm.weight'.format(key_prefix) in state_dict_keys: # Kandinsky 5
+        dit_config = {}
+        model_dim = state_dict['{}visual_embeddings.in_layer.bias'.format(key_prefix)].shape[0]
+        dit_config["model_dim"] = model_dim
+        if model_dim in [4096, 2560]: # pro video and lite image
+            dit_config["axes_dims"] = (32, 48, 48)
+            if model_dim == 2560: # lite image
+                dit_config["rope_scale_factor"] = (1.0, 1.0, 1.0)
+        elif model_dim == 1792: # lite video
+            dit_config["axes_dims"] = (16, 24, 24)
+        dit_config["time_dim"] = state_dict['{}time_embeddings.in_layer.bias'.format(key_prefix)].shape[0]
+        dit_config["image_model"] = "kandinsky5"
+        dit_config["ff_dim"] = state_dict['{}visual_transformer_blocks.0.feed_forward.in_layer.weight'.format(key_prefix)].shape[0]
+        dit_config["visual_embed_dim"] = state_dict['{}visual_embeddings.in_layer.weight'.format(key_prefix)].shape[1]
+        dit_config["num_text_blocks"] = count_blocks(state_dict_keys, '{}text_transformer_blocks.'.format(key_prefix) + '{}.')
+        dit_config["num_visual_blocks"] = count_blocks(state_dict_keys, '{}visual_transformer_blocks.'.format(key_prefix) + '{}.')
+        return dit_config
+
+    if '{}encoder.lyric_encoder.layers.0.input_layernorm.weight'.format(key_prefix) in state_dict_keys:
+        dit_config = {}
+        dit_config["audio_model"] = "ace1.5"
         return dit_config
 
     if '{}input_blocks.0.0.weight'.format(key_prefix) not in state_dict_keys:
@@ -704,22 +802,11 @@ def model_config_from_unet(state_dict, unet_key_prefix, use_base_if_no_match=Fal
     if model_config is None and use_base_if_no_match:
         model_config = comfy.supported_models_base.BASE(unet_config)
 
-    scaled_fp8_key = "{}scaled_fp8".format(unet_key_prefix)
-    if scaled_fp8_key in state_dict:
-        scaled_fp8_weight = state_dict.pop(scaled_fp8_key)
-        model_config.scaled_fp8 = scaled_fp8_weight.dtype
-        if model_config.scaled_fp8 == torch.float32:
-            model_config.scaled_fp8 = torch.float8_e4m3fn
-        if scaled_fp8_weight.nelement() == 2:
-            model_config.optimizations["fp8"] = False
-        else:
-            model_config.optimizations["fp8"] = True
-
     # Detect per-layer quantization (mixed precision)
-    layer_quant_config = detect_layer_quantization(metadata)
-    if layer_quant_config:
-        model_config.layer_quant_config = layer_quant_config
-        logging.info(f"Detected mixed precision quantization: {len(layer_quant_config)} layers quantized")
+    quant_config = comfy.utils.detect_layer_quantization(state_dict, unet_key_prefix)
+    if quant_config:
+        model_config.quant_config = quant_config
+        logging.info("Detected mixed precision quantization")
 
     return model_config
 

comfy/model_management.py

@@ -20,12 +20,20 @@ import psutil
 import logging
 from enum import Enum
 from comfy.cli_args import args, PerformanceFeature
+import threading
 import torch
 import sys
-import importlib
 import platform
 import weakref
 import gc
+import os
+from contextlib import nullcontext
+import comfy.memory_management
+import comfy.utils
+import comfy.quant_ops
+
+import comfy_aimdo.torch
+import comfy_aimdo.model_vbar
 
 class VRAMState(Enum):
     DISABLED = 0    #No vram present: no need to move models to vram
@@ -47,6 +55,11 @@ cpu_state = CPUState.GPU
 
 total_vram = 0
 
+
+# Training Related State
+in_training = False
+
+
 def get_supported_float8_types():
     float8_types = []
     try:
@@ -333,28 +346,42 @@ except:
 SUPPORT_FP8_OPS = args.supports_fp8_compute
 
 AMD_RDNA2_AND_OLDER_ARCH = ["gfx1030", "gfx1031", "gfx1010", "gfx1011", "gfx1012", "gfx906", "gfx900", "gfx803"]
+AMD_ENABLE_MIOPEN_ENV = 'COMFYUI_ENABLE_MIOPEN'
 
 try:
     if is_amd():
         arch = torch.cuda.get_device_properties(get_torch_device()).gcnArchName
         if not (any((a in arch) for a in AMD_RDNA2_AND_OLDER_ARCH)):
-            torch.backends.cudnn.enabled = False  # Seems to improve things a lot on AMD
-            logging.info("Set: torch.backends.cudnn.enabled = False for better AMD performance.")
+            if os.getenv(AMD_ENABLE_MIOPEN_ENV) != '1':
+                torch.backends.cudnn.enabled = False  # Seems to improve things a lot on AMD
+                logging.info("Set: torch.backends.cudnn.enabled = False for better AMD performance.")
 
         try:
             rocm_version = tuple(map(int, str(torch.version.hip).split(".")[:2]))
         except:
             rocm_version = (6, -1)
 
+        def aotriton_supported(gpu_arch):
+            path = torch.__path__[0]
+            path = os.path.join(os.path.join(path, "lib"), "aotriton.images")
+            gfx = set(map(lambda a: a[4:], filter(lambda a: a.startswith("amd-gfx"), os.listdir(path))))
+            if gpu_arch in gfx:
+                return True
+            if "{}x".format(gpu_arch[:-1]) in gfx:
+                return True
+            if "{}xx".format(gpu_arch[:-2]) in gfx:
+                return True
+            return False
+
         logging.info("AMD arch: {}".format(arch))
         logging.info("ROCm version: {}".format(rocm_version))
         if args.use_split_cross_attention == False and args.use_quad_cross_attention == False:
-            if importlib.util.find_spec('triton') is not None:  # AMD efficient attention implementation depends on triton. TODO: better way of detecting if it's compiled in or not.
+            if aotriton_supported(arch):  # AMD efficient attention implementation depends on aotriton.
                 if torch_version_numeric >= (2, 7):  # works on 2.6 but doesn't actually seem to improve much
                     if any((a in arch) for a in ["gfx90a", "gfx942", "gfx1100", "gfx1101", "gfx1151"]):  # TODO: more arches, TODO: gfx950
                         ENABLE_PYTORCH_ATTENTION = True
                 if rocm_version >= (7, 0):
-                   if any((a in arch) for a in ["gfx1201"]):
+                   if any((a in arch) for a in ["gfx1200", "gfx1201"]):
                        ENABLE_PYTORCH_ATTENTION = True
         if torch_version_numeric >= (2, 7) and rocm_version >= (6, 4):
             if any((a in arch) for a in ["gfx1200", "gfx1201", "gfx950"]):  # TODO: more arches, "gfx942" gives error on pytorch nightly 2.10 1013 rocm7.0
@@ -453,7 +480,7 @@ def module_size(module):
     sd = module.state_dict()
     for k in sd:
         t = sd[k]
-        module_mem += t.nelement() * t.element_size()
+        module_mem += t.nbytes
     return module_mem
 
 class LoadedModel:
@@ -504,6 +531,7 @@ class LoadedModel:
         if use_more_vram == 0:
             use_more_vram = 1e32
         self.model_use_more_vram(use_more_vram, force_patch_weights=force_patch_weights)
+
         real_model = self.model.model
 
         if is_intel_xpu() and not args.disable_ipex_optimize and 'ipex' in globals() and real_model is not None:
@@ -563,9 +591,15 @@ WINDOWS = any(platform.win32_ver())
 
 EXTRA_RESERVED_VRAM = 400 * 1024 * 1024
 if WINDOWS:
+    import comfy.windows
     EXTRA_RESERVED_VRAM = 600 * 1024 * 1024 #Windows is higher because of the shared vram issue
     if total_vram > (15 * 1024):  # more extra reserved vram on 16GB+ cards
         EXTRA_RESERVED_VRAM += 100 * 1024 * 1024
+    def get_free_ram():
+        return comfy.windows.get_free_ram()
+else:
+    def get_free_ram():
+        return psutil.virtual_memory().available
 
 if args.reserve_vram is not None:
     EXTRA_RESERVED_VRAM = args.reserve_vram * 1024 * 1024 * 1024
@@ -577,7 +611,7 @@ def extra_reserved_memory():
 def minimum_inference_memory():
     return (1024 * 1024 * 1024) * 0.8 + extra_reserved_memory()
 
-def free_memory(memory_required, device, keep_loaded=[]):
+def free_memory(memory_required, device, keep_loaded=[], for_dynamic=False, ram_required=0):
     cleanup_models_gc()
     unloaded_model = []
     can_unload = []
@@ -592,15 +626,23 @@ def free_memory(memory_required, device, keep_loaded=[]):
 
     for x in sorted(can_unload):
         i = x[-1]
-        memory_to_free = None
+        memory_to_free = 1e32
+        ram_to_free = 1e32
         if not DISABLE_SMART_MEMORY:
-            free_mem = get_free_memory(device)
-            if free_mem > memory_required:
-                break
-            memory_to_free = memory_required - free_mem
-        logging.debug(f"Unloading {current_loaded_models[i].model.model.__class__.__name__}")
-        if current_loaded_models[i].model_unload(memory_to_free):
+            memory_to_free = memory_required - get_free_memory(device)
+            ram_to_free = ram_required - get_free_ram()
+
+        if current_loaded_models[i].model.is_dynamic() and for_dynamic:
+            #don't actually unload dynamic models for the sake of other dynamic models
+            #as that works on-demand.
+            memory_required -= current_loaded_models[i].model.loaded_size()
+            memory_to_free = 0
+        if memory_to_free > 0 and current_loaded_models[i].model_unload(memory_to_free):
+            logging.debug(f"Unloading {current_loaded_models[i].model.model.__class__.__name__}")
             unloaded_model.append(i)
+        if ram_to_free > 0:
+            logging.debug(f"RAM Unloading {current_loaded_models[i].model.model.__class__.__name__}")
+            current_loaded_models[i].model.partially_unload_ram(ram_to_free)
 
     for i in sorted(unloaded_model, reverse=True):
         unloaded_models.append(current_loaded_models.pop(i))
@@ -635,7 +677,10 @@ def load_models_gpu(models, memory_required=0, force_patch_weights=False, minimu
 
     models_to_load = []
 
+    free_for_dynamic=True
     for x in models:
+        if not x.is_dynamic():
+            free_for_dynamic = False
         loaded_model = LoadedModel(x)
         try:
             loaded_model_index = current_loaded_models.index(loaded_model)
@@ -661,19 +706,25 @@ def load_models_gpu(models, memory_required=0, force_patch_weights=False, minimu
             model_to_unload.model.detach(unpatch_all=False)
             model_to_unload.model_finalizer.detach()
 
+
     total_memory_required = {}
+    total_ram_required = {}
     for loaded_model in models_to_load:
         total_memory_required[loaded_model.device] = total_memory_required.get(loaded_model.device, 0) + loaded_model.model_memory_required(loaded_model.device)
+        #x2, one to make sure the OS can fit the model for loading in disk cache, and for us to do any pinning we
+        #want to do.
+        #FIXME: This should subtract off the to_load current pin consumption.
+        total_ram_required[loaded_model.device] = total_ram_required.get(loaded_model.device, 0) + loaded_model.model_memory() * 2
 
     for device in total_memory_required:
         if device != torch.device("cpu"):
-            free_memory(total_memory_required[device] * 1.1 + extra_mem, device)
+            free_memory(total_memory_required[device] * 1.1 + extra_mem, device, for_dynamic=free_for_dynamic, ram_required=total_ram_required[device])
 
     for device in total_memory_required:
         if device != torch.device("cpu"):
             free_mem = get_free_memory(device)
             if free_mem < minimum_memory_required:
-                models_l = free_memory(minimum_memory_required, device)
+                models_l = free_memory(minimum_memory_required, device, for_dynamic=free_for_dynamic)
                 logging.info("{} models unloaded.".format(len(models_l)))
 
     for loaded_model in models_to_load:
@@ -688,8 +739,11 @@ def load_models_gpu(models, memory_required=0, force_patch_weights=False, minimu
             loaded_memory = loaded_model.model_loaded_memory()
             current_free_mem = get_free_memory(torch_dev) + loaded_memory
 
-            lowvram_model_memory = max(128 * 1024 * 1024, (current_free_mem - minimum_memory_required), min(current_free_mem * MIN_WEIGHT_MEMORY_RATIO, current_free_mem - minimum_inference_memory()))
-            lowvram_model_memory = max(0.1, lowvram_model_memory - loaded_memory)
+            lowvram_model_memory = max(0, (current_free_mem - minimum_memory_required), min(current_free_mem * MIN_WEIGHT_MEMORY_RATIO, current_free_mem - minimum_inference_memory()))
+            lowvram_model_memory = lowvram_model_memory - loaded_memory
+
+            if lowvram_model_memory == 0:
+                lowvram_model_memory = 0.1
 
         if vram_set_state == VRAMState.NO_VRAM:
             lowvram_model_memory = 0.1
@@ -714,6 +768,9 @@ def loaded_models(only_currently_used=False):
 
 def cleanup_models_gc():
     do_gc = False
+
+    reset_cast_buffers()
+
     for i in range(len(current_loaded_models)):
         cur = current_loaded_models[i]
         if cur.is_dead():
@@ -731,6 +788,11 @@ def cleanup_models_gc():
                 logging.warning("WARNING, memory leak with model {}. Please make sure it is not being referenced from somewhere.".format(cur.real_model().__class__.__name__))
 
 
+def archive_model_dtypes(model):
+    for name, module in model.named_modules():
+        for param_name, param in module.named_parameters(recurse=False):
+            setattr(module, f"{param_name}_comfy_model_dtype", param.dtype)
+
 
 def cleanup_models():
     to_delete = []
@@ -774,7 +836,7 @@ def unet_inital_load_device(parameters, dtype):
 
     mem_dev = get_free_memory(torch_dev)
     mem_cpu = get_free_memory(cpu_dev)
-    if mem_dev > mem_cpu and model_size < mem_dev:
+    if mem_dev > mem_cpu and model_size < mem_dev and comfy.memory_management.aimdo_allocator is None:
         return torch_dev
     else:
         return cpu_dev
@@ -1008,9 +1070,18 @@ def force_channels_last():
 
 
 STREAMS = {}
-NUM_STREAMS = 1
-if args.async_offload:
-    NUM_STREAMS = 2
+NUM_STREAMS = 0
+if args.async_offload is not None:
+    NUM_STREAMS = args.async_offload
+else:
+    #  Enable by default on Nvidia and AMD
+    if is_nvidia() or is_amd():
+        NUM_STREAMS = 2
+
+if args.disable_async_offload:
+    NUM_STREAMS = 0
+
+if NUM_STREAMS > 0:
     logging.info("Using async weight offloading with {} streams".format(NUM_STREAMS))
 
 def current_stream(device):
@@ -1024,9 +1095,57 @@ def current_stream(device):
         return None
 
 stream_counters = {}
+
+STREAM_CAST_BUFFERS = {}
+LARGEST_CASTED_WEIGHT = (None, 0)
+
+def get_cast_buffer(offload_stream, device, size, ref):
+    global LARGEST_CASTED_WEIGHT
+
+    if offload_stream is not None:
+        wf_context = offload_stream
+        if hasattr(wf_context, "as_context"):
+            wf_context = wf_context.as_context(offload_stream)
+    else:
+        wf_context = nullcontext()
+
+    cast_buffer = STREAM_CAST_BUFFERS.get(offload_stream, None)
+    if cast_buffer is None or cast_buffer.numel() < size:
+        if ref is LARGEST_CASTED_WEIGHT[0]:
+            #If there is one giant weight we do not want both streams to
+            #allocate a buffer for it. It's up to the caster to get the other
+            #offload stream in this corner case
+            return None
+        if cast_buffer is not None and cast_buffer.numel() > 50 * (1024 ** 2):
+            #I want my wrongly sized 50MB+ of VRAM back from the caching allocator right now
+            synchronize()
+            del STREAM_CAST_BUFFERS[offload_stream]
+            del cast_buffer
+            #FIXME: This doesn't work in Aimdo because mempool cant clear cache
+            soft_empty_cache()
+        with wf_context:
+            cast_buffer = torch.empty((size), dtype=torch.int8, device=device)
+            STREAM_CAST_BUFFERS[offload_stream] = cast_buffer
+
+        if  size > LARGEST_CASTED_WEIGHT[1]:
+            LARGEST_CASTED_WEIGHT = (ref, size)
+
+    return cast_buffer
+
+def reset_cast_buffers():
+    global LARGEST_CASTED_WEIGHT
+    LARGEST_CASTED_WEIGHT = (None, 0)
+    for offload_stream in STREAM_CAST_BUFFERS:
+        offload_stream.synchronize()
+    STREAM_CAST_BUFFERS.clear()
+    soft_empty_cache()
+
 def get_offload_stream(device):
     stream_counter = stream_counters.get(device, 0)
-    if NUM_STREAMS <= 1:
+    if NUM_STREAMS == 0:
+        return None
+
+    if torch.compiler.is_compiling():
         return None
 
     if device in STREAMS:
@@ -1039,7 +1158,9 @@ def get_offload_stream(device):
     elif is_device_cuda(device):
         ss = []
         for k in range(NUM_STREAMS):
-            ss.append(torch.cuda.Stream(device=device, priority=0))
+            s1 = torch.cuda.Stream(device=device, priority=0)
+            s1.as_context = torch.cuda.stream
+            ss.append(s1)
         STREAMS[device] = ss
         s = ss[stream_counter]
         stream_counters[device] = stream_counter
@@ -1047,7 +1168,9 @@ def get_offload_stream(device):
     elif is_device_xpu(device):
         ss = []
         for k in range(NUM_STREAMS):
-            ss.append(torch.xpu.Stream(device=device, priority=0))
+            s1 = torch.xpu.Stream(device=device, priority=0)
+            s1.as_context = torch.xpu.stream
+            ss.append(s1)
         STREAMS[device] = ss
         s = ss[stream_counter]
         stream_counters[device] = stream_counter
@@ -1059,22 +1182,81 @@ def sync_stream(device, stream):
         return
     current_stream(device).wait_stream(stream)
 
-def cast_to(weight, dtype=None, device=None, non_blocking=False, copy=False, stream=None):
+
+def cast_to_gathered(tensors, r, non_blocking=False, stream=None):
+    wf_context = nullcontext()
+    if stream is not None:
+       wf_context = stream
+       if hasattr(wf_context, "as_context"):
+           wf_context = wf_context.as_context(stream)
+
+    dest_views = comfy.memory_management.interpret_gathered_like(tensors, r)
+    with wf_context:
+        for tensor in tensors:
+            dest_view = dest_views.pop(0)
+            if tensor is None:
+                continue
+            dest_view.copy_(tensor, non_blocking=non_blocking)
+
+
+def cast_to(weight, dtype=None, device=None, non_blocking=False, copy=False, stream=None, r=None):
+    if hasattr(weight, "_v"):
+        #Unexpected usage patterns. There is no reason these don't work but they
+        #have no testing and no callers do this.
+        assert r is None
+        assert stream is None
+
+        cast_geometry = comfy.memory_management.tensors_to_geometries([ weight ])
+
+        if dtype is None:
+            dtype = weight._model_dtype
+
+        signature = comfy_aimdo.model_vbar.vbar_fault(weight._v)
+        if signature is not None:
+            v_tensor = comfy.memory_management.interpret_gathered_like(cast_geometry, weight._v_tensor)[0]
+            if not comfy_aimdo.model_vbar.vbar_signature_compare(signature, weight._v_signature):
+                weight._v_signature = signature
+                #Send it over
+                v_tensor.copy_(weight, non_blocking=non_blocking)
+            return v_tensor.to(dtype=dtype)
+
+        r = torch.empty_like(weight, dtype=dtype, device=device)
+
+        if weight.dtype != r.dtype and weight.dtype != weight._model_dtype:
+            #Offloaded casting could skip this, however it would make the quantizations
+            #inconsistent between loaded and offloaded weights. So force the double casting
+            #that would happen in regular flow to make offload deterministic.
+            cast_buffer = torch.empty_like(weight, dtype=weight._model_dtype, device=device)
+            cast_buffer.copy_(weight, non_blocking=non_blocking)
+            weight = cast_buffer
+        r.copy_(weight, non_blocking=non_blocking)
+
+        return r
+
     if device is None or weight.device == device:
         if not copy:
             if dtype is None or weight.dtype == dtype:
                 return weight
         if stream is not None:
-            with stream:
+            wf_context = stream
+            if hasattr(wf_context, "as_context"):
+                wf_context = wf_context.as_context(stream)
+            with wf_context:
                 return weight.to(dtype=dtype, copy=copy)
         return weight.to(dtype=dtype, copy=copy)
 
+
     if stream is not None:
-        with stream:
-            r = torch.empty_like(weight, dtype=dtype, device=device)
+        wf_context = stream
+        if hasattr(wf_context, "as_context"):
+            wf_context = wf_context.as_context(stream)
+        with wf_context:
+            if r is None:
+                r = torch.empty_like(weight, dtype=dtype, device=device)
             r.copy_(weight, non_blocking=non_blocking)
     else:
-        r = torch.empty_like(weight, dtype=dtype, device=device)
+        if r is None:
+            r = torch.empty_like(weight, dtype=dtype, device=device)
         r.copy_(weight, non_blocking=non_blocking)
     return r
 
@@ -1082,33 +1264,96 @@ def cast_to_device(tensor, device, dtype, copy=False):
     non_blocking = device_supports_non_blocking(device)
     return cast_to(tensor, dtype=dtype, device=device, non_blocking=non_blocking, copy=copy)
 
+
+PINNED_MEMORY = {}
+TOTAL_PINNED_MEMORY = 0
+MAX_PINNED_MEMORY = -1
+if not args.disable_pinned_memory:
+    if is_nvidia() or is_amd():
+        if WINDOWS:
+            MAX_PINNED_MEMORY = get_total_memory(torch.device("cpu")) * 0.45  # Windows limit is apparently 50%
+        else:
+            MAX_PINNED_MEMORY = get_total_memory(torch.device("cpu")) * 0.95
+        logging.info("Enabled pinned memory {}".format(MAX_PINNED_MEMORY // (1024 * 1024)))
+
+PINNING_ALLOWED_TYPES = set(["Tensor", "Parameter", "QuantizedTensor"])
+
+def discard_cuda_async_error():
+    try:
+        a = torch.tensor([1], dtype=torch.uint8, device=get_torch_device())
+        b = torch.tensor([1], dtype=torch.uint8, device=get_torch_device())
+        _ = a + b
+        synchronize()
+    except torch.AcceleratorError:
+        #Dump it! We already know about it from the synchronous return
+        pass
+
 def pin_memory(tensor):
-    if PerformanceFeature.PinnedMem not in args.fast:
+    global TOTAL_PINNED_MEMORY
+    if MAX_PINNED_MEMORY <= 0:
         return False
 
-    if not is_nvidia():
+    if type(tensor).__name__ not in PINNING_ALLOWED_TYPES:
         return False
 
     if not is_device_cpu(tensor.device):
         return False
 
-    if torch.cuda.cudart().cudaHostRegister(tensor.data_ptr(), tensor.numel() * tensor.element_size(), 1) == 0:
+    if tensor.is_pinned():
+        #NOTE: Cuda does detect when a tensor is already pinned and would
+        #error below, but there are proven cases where this also queues an error
+        #on the GPU async. So dont trust the CUDA API and guard here
+        return False
+
+    if not tensor.is_contiguous():
+        return False
+
+    size = tensor.nbytes
+    if (TOTAL_PINNED_MEMORY + size) > MAX_PINNED_MEMORY:
+        return False
+
+    ptr = tensor.data_ptr()
+    if ptr == 0:
+        return False
+
+    if torch.cuda.cudart().cudaHostRegister(ptr, size, 1) == 0:
+        PINNED_MEMORY[ptr] = size
+        TOTAL_PINNED_MEMORY += size
         return True
+    else:
+        logging.warning("Pin error.")
+        discard_cuda_async_error()
 
     return False
 
 def unpin_memory(tensor):
-    if PerformanceFeature.PinnedMem not in args.fast:
-        return False
-
-    if not is_nvidia():
+    global TOTAL_PINNED_MEMORY
+    if MAX_PINNED_MEMORY <= 0:
         return False
 
     if not is_device_cpu(tensor.device):
         return False
 
-    if torch.cuda.cudart().cudaHostUnregister(tensor.data_ptr()) == 0:
+    ptr = tensor.data_ptr()
+    size = tensor.nbytes
+
+    size_stored = PINNED_MEMORY.get(ptr, None)
+    if size_stored is None:
+        logging.warning("Tried to unpin tensor not pinned by ComfyUI")
+        return False
+
+    if size != size_stored:
+        logging.warning("Size of pinned tensor changed")
+        return False
+
+    if torch.cuda.cudart().cudaHostUnregister(ptr) == 0:
+        TOTAL_PINNED_MEMORY -= PINNED_MEMORY.pop(ptr)
+        if len(PINNED_MEMORY) == 0:
+            TOTAL_PINNED_MEMORY = 0
         return True
+    else:
+        logging.warning("Unpin error.")
+        discard_cuda_async_error()
 
     return False
 
@@ -1411,6 +1656,16 @@ def supports_fp8_compute(device=None):
 
     return True
 
+def supports_nvfp4_compute(device=None):
+    if not is_nvidia():
+        return False
+
+    props = torch.cuda.get_device_properties(device)
+    if props.major < 10:
+        return False
+
+    return True
+
 def extended_fp16_support():
     # TODO: check why some models work with fp16 on newer torch versions but not on older
     if torch_version_numeric < (2, 7):
@@ -1418,6 +1673,26 @@ def extended_fp16_support():
 
     return True
 
+LORA_COMPUTE_DTYPES = {}
+def lora_compute_dtype(device):
+    dtype = LORA_COMPUTE_DTYPES.get(device, None)
+    if dtype is not None:
+        return dtype
+
+    if should_use_fp16(device):
+        dtype = torch.float16
+    else:
+        dtype = torch.float32
+
+    LORA_COMPUTE_DTYPES[device] = dtype
+    return dtype
+
+def synchronize():
+    if is_intel_xpu():
+        torch.xpu.synchronize()
+    elif torch.cuda.is_available():
+        torch.cuda.synchronize()
+
 def soft_empty_cache(force=False):
     global cpu_state
     if cpu_state == CPUState.MPS:
@@ -1429,15 +1704,17 @@ def soft_empty_cache(force=False):
     elif is_mlu():
         torch.mlu.empty_cache()
     elif torch.cuda.is_available():
+        torch.cuda.synchronize()
         torch.cuda.empty_cache()
         torch.cuda.ipc_collect()
 
 def unload_all_models():
     free_memory(1e30, get_torch_device())
 
-
-#TODO: might be cleaner to put this somewhere else
-import threading
+def debug_memory_summary():
+    if is_amd() or is_nvidia():
+        return torch.cuda.memory.memory_summary()
+    return ""
 
 class InterruptProcessingException(Exception):
     pass

comfy/model_patcher.py

@@ -19,7 +19,6 @@
 from __future__ import annotations
 
 import collections
-import copy
 import inspect
 import logging
 import math
@@ -35,21 +34,10 @@ import comfy.model_management
 import comfy.patcher_extension
 import comfy.utils
 from comfy.comfy_types import UnetWrapperFunction
+from comfy.quant_ops import QuantizedTensor
 from comfy.patcher_extension import CallbacksMP, PatcherInjection, WrappersMP
 
-
-def string_to_seed(data):
-    crc = 0xFFFFFFFF
-    for byte in data:
-        if isinstance(byte, str):
-            byte = ord(byte)
-        crc ^= byte
-        for _ in range(8):
-            if crc & 1:
-                crc = (crc >> 1) ^ 0xEDB88320
-            else:
-                crc >>= 1
-    return crc ^ 0xFFFFFFFF
+import comfy_aimdo.model_vbar
 
 def set_model_options_patch_replace(model_options, patch, name, block_name, number, transformer_index=None):
     to = model_options["transformer_options"].copy()
@@ -122,31 +110,31 @@ def move_weight_functions(m, device):
                 memory += f.move_to(device=device)
     return memory
 
+def string_to_seed(data):
+    logging.warning("WARNING: string_to_seed has moved from comfy.model_patcher to comfy.utils")
+    return comfy.utils.string_to_seed(data)
+
 class LowVramPatch:
     def __init__(self, key, patches, convert_func=None, set_func=None):
         self.key = key
         self.patches = patches
-        self.convert_func = convert_func
+        self.convert_func = convert_func # TODO: remove
         self.set_func = set_func
 
     def __call__(self, weight):
-        intermediate_dtype = weight.dtype
-        if self.convert_func is not None:
-            weight = self.convert_func(weight.to(dtype=torch.float32, copy=True), inplace=True)
+        return comfy.lora.calculate_weight(self.patches[self.key], weight, self.key, intermediate_dtype=weight.dtype)
 
-        if intermediate_dtype not in [torch.float32, torch.float16, torch.bfloat16]: #intermediate_dtype has to be one that is supported in math ops
-            intermediate_dtype = torch.float32
-            out = comfy.lora.calculate_weight(self.patches[self.key], weight.to(intermediate_dtype), self.key, intermediate_dtype=intermediate_dtype)
-            if self.set_func is None:
-                return comfy.float.stochastic_rounding(out, weight.dtype, seed=string_to_seed(self.key))
-            else:
-                return self.set_func(out, seed=string_to_seed(self.key), return_weight=True)
+LOWVRAM_PATCH_ESTIMATE_MATH_FACTOR = 2
 
-        out = comfy.lora.calculate_weight(self.patches[self.key], weight, self.key, intermediate_dtype=intermediate_dtype)
-        if self.set_func is not None:
-            return self.set_func(out, seed=string_to_seed(self.key), return_weight=True).to(dtype=intermediate_dtype)
-        else:
-            return out
+def low_vram_patch_estimate_vram(model, key):
+    weight, set_func, convert_func = get_key_weight(model, key)
+    if weight is None:
+        return 0
+    model_dtype = getattr(model, "manual_cast_dtype", torch.float32)
+    if model_dtype is None:
+        model_dtype = weight.dtype
+
+    return weight.numel() * model_dtype.itemsize * LOWVRAM_PATCH_ESTIMATE_MATH_FACTOR
 
 def get_key_weight(model, key):
     set_func = None
@@ -172,6 +160,11 @@ def get_key_weight(model, key):
 
     return weight, set_func, convert_func
 
+def key_param_name_to_key(key, param):
+    if len(key) == 0:
+        return param
+    return "{}.{}".format(key, param)
+
 class AutoPatcherEjector:
     def __init__(self, model: 'ModelPatcher', skip_and_inject_on_exit_only=False):
         self.model = model
@@ -215,6 +208,27 @@ class MemoryCounter:
     def decrement(self, used: int):
         self.value -= used
 
+CustomTorchDevice = collections.namedtuple("FakeDevice", ["type", "index"])("comfy-lazy-caster", 0)
+
+class LazyCastingParam(torch.nn.Parameter):
+    def __new__(cls, model, key, tensor):
+        return super().__new__(cls, tensor)
+
+    def __init__(self, model, key, tensor):
+        self.model = model
+        self.key = key
+
+    @property
+    def device(self):
+        return CustomTorchDevice
+
+    #safetensors will .to() us to the cpu which we catch here to cast on demand. The returned tensor is
+    #then just a short lived thing in the safetensors serialization logic inside its big for loop over
+    #all weights getting garbage collected per-weight
+    def to(self, *args, **kwargs):
+        return self.model.patch_weight_to_device(self.key, device_to=self.model.load_device, return_weight=True).to("cpu")
+
+
 class ModelPatcher:
     def __init__(self, model, load_device, offload_device, size=0, weight_inplace_update=False):
         self.size = size
@@ -231,7 +245,6 @@ class ModelPatcher:
         self.object_patches_backup = {}
         self.weight_wrapper_patches = {}
         self.model_options = {"transformer_options":{}}
-        self.model_size()
         self.load_device = load_device
         self.offload_device = offload_device
         self.weight_inplace_update = weight_inplace_update
@@ -270,6 +283,12 @@ class ModelPatcher:
         if not hasattr(self.model, 'current_weight_patches_uuid'):
             self.model.current_weight_patches_uuid = None
 
+        if not hasattr(self.model, 'model_offload_buffer_memory'):
+            self.model.model_offload_buffer_memory = 0
+
+    def is_dynamic(self):
+        return False
+
     def model_size(self):
         if self.size > 0:
             return self.size
@@ -285,8 +304,11 @@ class ModelPatcher:
     def lowvram_patch_counter(self):
         return self.model.lowvram_patch_counter
 
+    def get_free_memory(self, device):
+        return comfy.model_management.get_free_memory(device)
+
     def clone(self):
-        n = self.__class__(self.model, self.load_device, self.offload_device, self.size, weight_inplace_update=self.weight_inplace_update)
+        n = self.__class__(self.model, self.load_device, self.offload_device, self.model_size(), weight_inplace_update=self.weight_inplace_update)
         n.patches = {}
         for k in self.patches:
             n.patches[k] = self.patches[k][:]
@@ -294,7 +316,7 @@ class ModelPatcher:
 
         n.object_patches = self.object_patches.copy()
         n.weight_wrapper_patches = self.weight_wrapper_patches.copy()
-        n.model_options = copy.deepcopy(self.model_options)
+        n.model_options = comfy.utils.deepcopy_list_dict(self.model_options)
         n.backup = self.backup
         n.object_patches_backup = self.object_patches_backup
         n.parent = self
@@ -455,6 +477,9 @@ class ModelPatcher:
     def set_model_post_input_patch(self, patch):
         self.set_model_patch(patch, "post_input")
 
+    def set_model_noise_refiner_patch(self, patch):
+        self.set_model_patch(patch, "noise_refiner")
+
     def set_model_rope_options(self, scale_x, shift_x, scale_y, shift_y, scale_t, shift_t, **kwargs):
         rope_options = self.model_options["transformer_options"].get("rope_options", {})
         rope_options["scale_x"] = scale_x
@@ -609,32 +634,35 @@ class ModelPatcher:
                         sd.pop(k)
             return sd
 
-    def patch_weight_to_device(self, key, device_to=None, inplace_update=False):
-        if key not in self.patches:
-            return
-
+    def patch_weight_to_device(self, key, device_to=None, inplace_update=False, return_weight=False):
         weight, set_func, convert_func = get_key_weight(self.model, key)
+        if key not in self.patches:
+            return weight
+
         inplace_update = self.weight_inplace_update or inplace_update
 
-        if key not in self.backup:
+        if key not in self.backup and not return_weight:
             self.backup[key] = collections.namedtuple('Dimension', ['weight', 'inplace_update'])(weight.to(device=self.offload_device, copy=inplace_update), inplace_update)
 
+        temp_dtype = comfy.model_management.lora_compute_dtype(device_to)
         if device_to is not None:
-            temp_weight = comfy.model_management.cast_to_device(weight, device_to, torch.float32, copy=True)
+            temp_weight = comfy.model_management.cast_to_device(weight, device_to, temp_dtype, copy=True)
         else:
-            temp_weight = weight.to(torch.float32, copy=True)
+            temp_weight = weight.to(temp_dtype, copy=True)
         if convert_func is not None:
             temp_weight = convert_func(temp_weight, inplace=True)
 
         out_weight = comfy.lora.calculate_weight(self.patches[key], temp_weight, key)
         if set_func is None:
-            out_weight = comfy.float.stochastic_rounding(out_weight, weight.dtype, seed=string_to_seed(key))
-            if inplace_update:
+            out_weight = comfy.float.stochastic_rounding(out_weight, weight.dtype, seed=comfy.utils.string_to_seed(key))
+            if return_weight:
+                return out_weight
+            elif inplace_update:
                 comfy.utils.copy_to_param(self.model, key, out_weight)
             else:
                 comfy.utils.set_attr_param(self.model, key, out_weight)
         else:
-            set_func(out_weight, inplace_update=inplace_update, seed=string_to_seed(key))
+            return set_func(out_weight, inplace_update=inplace_update, seed=comfy.utils.string_to_seed(key), return_weight=return_weight)
 
     def pin_weight_to_device(self, key):
         weight, set_func, convert_func = get_key_weight(self.model, key)
@@ -651,7 +679,7 @@ class ModelPatcher:
         for key in list(self.pinned):
             self.unpin_weight(key)
 
-    def _load_list(self):
+    def _load_list(self, prio_comfy_cast_weights=False):
         loading = []
         for n, m in self.model.named_modules():
             params = []
@@ -663,7 +691,23 @@ class ModelPatcher:
                     skip = True # skip random weights in non leaf modules
                     break
             if not skip and (hasattr(m, "comfy_cast_weights") or len(params) > 0):
-                loading.append((comfy.model_management.module_size(m), n, m, params))
+                module_mem = comfy.model_management.module_size(m)
+                module_offload_mem = module_mem
+                if hasattr(m, "comfy_cast_weights"):
+                    def check_module_offload_mem(key):
+                        if key in self.patches:
+                            return low_vram_patch_estimate_vram(self.model, key)
+                        model_dtype = getattr(self.model, "manual_cast_dtype", None)
+                        weight, _, _ = get_key_weight(self.model, key)
+                        if model_dtype is None or weight is None:
+                            return 0
+                        if (weight.dtype != model_dtype or isinstance(weight, QuantizedTensor)):
+                            return weight.numel() * model_dtype.itemsize
+                        return 0
+                    module_offload_mem += check_module_offload_mem("{}.weight".format(n))
+                    module_offload_mem += check_module_offload_mem("{}.bias".format(n))
+                prepend = (not hasattr(m, "comfy_cast_weights"),) if prio_comfy_cast_weights else ()
+                loading.append(prepend + (module_offload_mem, module_mem, n, m, params))
         return loading
 
     def load(self, device_to=None, lowvram_model_memory=0, force_patch_weights=False, full_load=False):
@@ -677,20 +721,22 @@ class ModelPatcher:
 
             load_completely = []
             offloaded = []
+            offload_buffer = 0
             loading.sort(reverse=True)
-            for x in loading:
-                n = x[1]
-                m = x[2]
-                params = x[3]
-                module_mem = x[0]
+            for i, x in enumerate(loading):
+                module_offload_mem, module_mem, n, m, params = x
 
                 lowvram_weight = False
 
+                potential_offload = max(offload_buffer, module_offload_mem + sum([ x1[1] for x1 in loading[i+1:i+1+comfy.model_management.NUM_STREAMS]]))
+                lowvram_fits = mem_counter + module_mem + potential_offload < lowvram_model_memory
+
                 weight_key = "{}.weight".format(n)
                 bias_key = "{}.bias".format(n)
 
                 if not full_load and hasattr(m, "comfy_cast_weights"):
-                    if mem_counter + module_mem >= lowvram_model_memory:
+                    if not lowvram_fits:
+                        offload_buffer = potential_offload
                         lowvram_weight = True
                         lowvram_counter += 1
                         lowvram_mem_counter += module_mem
@@ -698,6 +744,7 @@ class ModelPatcher:
                             continue
 
                 cast_weight = self.force_cast_weights
+                m.comfy_force_cast_weights = self.force_cast_weights
                 if lowvram_weight:
                     if hasattr(m, "comfy_cast_weights"):
                         m.weight_function = []
@@ -724,9 +771,11 @@ class ModelPatcher:
                     if hasattr(m, "comfy_cast_weights"):
                         wipe_lowvram_weight(m)
 
-                    if full_load or mem_counter + module_mem < lowvram_model_memory:
+                    if full_load or lowvram_fits:
                         mem_counter += module_mem
                         load_completely.append((module_mem, n, m, params))
+                    else:
+                        offload_buffer = potential_offload
 
                 if cast_weight and hasattr(m, "comfy_cast_weights"):
                     m.prev_comfy_cast_weights = m.comfy_cast_weights
@@ -750,9 +799,11 @@ class ModelPatcher:
                         continue
 
                 for param in params:
-                    key = "{}.{}".format(n, param)
+                    key = key_param_name_to_key(n, param)
                     self.unpin_weight(key)
                     self.patch_weight_to_device(key, device_to=device_to)
+                if comfy.model_management.is_device_cuda(device_to):
+                    torch.cuda.synchronize()
 
                 logging.debug("lowvram: loaded module regularly {} {}".format(n, m))
                 m.comfy_patched_weights = True
@@ -764,13 +815,14 @@ class ModelPatcher:
                 n = x[1]
                 params = x[3]
                 for param in params:
-                    self.pin_weight_to_device("{}.{}".format(n, param))
+                    self.pin_weight_to_device(key_param_name_to_key(n, param))
 
+            usable_stat = "{:.2f} MB usable,".format(lowvram_model_memory / (1024 * 1024)) if lowvram_model_memory < 1e32 else ""
             if lowvram_counter > 0:
-                logging.info("loaded partially; {:.2f} MB usable, {:.2f} MB loaded, {:.2f} MB offloaded, lowvram patches: {}".format(lowvram_model_memory / (1024 * 1024), mem_counter / (1024 * 1024), lowvram_mem_counter / (1024 * 1024), patch_counter))
+                logging.info("loaded partially; {} {:.2f} MB loaded, {:.2f} MB offloaded, {:.2f} MB buffer reserved, lowvram patches: {}".format(usable_stat, mem_counter / (1024 * 1024), lowvram_mem_counter / (1024 * 1024), offload_buffer / (1024 * 1024), patch_counter))
                 self.model.model_lowvram = True
             else:
-                logging.info("loaded completely; {:.2f} MB usable, {:.2f} MB loaded, full load: {}".format(lowvram_model_memory / (1024 * 1024), mem_counter / (1024 * 1024), full_load))
+                logging.info("loaded completely; {} {:.2f} MB loaded, full load: {}".format(usable_stat, mem_counter / (1024 * 1024), full_load))
                 self.model.model_lowvram = False
                 if full_load:
                     self.model.to(device_to)
@@ -779,6 +831,7 @@ class ModelPatcher:
             self.model.lowvram_patch_counter += patch_counter
             self.model.device = device_to
             self.model.model_loaded_weight_memory = mem_counter
+            self.model.model_offload_buffer_memory = offload_buffer
             self.model.current_weight_patches_uuid = self.patches_uuid
 
             for callback in self.get_all_callbacks(CallbacksMP.ON_LOAD):
@@ -832,6 +885,7 @@ class ModelPatcher:
                 self.model.to(device_to)
                 self.model.device = device_to
             self.model.model_loaded_weight_memory = 0
+            self.model.model_offload_buffer_memory = 0
 
             for m in self.model.modules():
                 if hasattr(m, "comfy_patched_weights"):
@@ -843,26 +897,31 @@ class ModelPatcher:
 
         self.object_patches_backup.clear()
 
-    def partially_unload(self, device_to, memory_to_free=0):
+    def partially_unload(self, device_to, memory_to_free=0, force_patch_weights=False):
         with self.use_ejected():
             hooks_unpatched = False
             memory_freed = 0
             patch_counter = 0
             unload_list = self._load_list()
             unload_list.sort()
+
+            offload_buffer = self.model.model_offload_buffer_memory
+            if len(unload_list) > 0:
+                NS = comfy.model_management.NUM_STREAMS
+                offload_weight_factor = [ min(offload_buffer / (NS + 1), unload_list[0][1]) ] * NS
+
             for unload in unload_list:
-                if memory_to_free < memory_freed:
+                if memory_to_free + offload_buffer - self.model.model_offload_buffer_memory < memory_freed:
                     break
-                module_mem = unload[0]
-                n = unload[1]
-                m = unload[2]
-                params = unload[3]
+                module_offload_mem, module_mem, n, m, params = unload
+
+                potential_offload = module_offload_mem + sum(offload_weight_factor)
 
                 lowvram_possible = hasattr(m, "comfy_cast_weights")
                 if hasattr(m, "comfy_patched_weights") and m.comfy_patched_weights == True:
                     move_weight = True
                     for param in params:
-                        key = "{}.{}".format(n, param)
+                        key = key_param_name_to_key(n, param)
                         bk = self.backup.get(key, None)
                         if bk is not None:
                             if not lowvram_possible:
@@ -887,28 +946,40 @@ class ModelPatcher:
                         module_mem += move_weight_functions(m, device_to)
                         if lowvram_possible:
                             if weight_key in self.patches:
-                                _, set_func, convert_func = get_key_weight(self.model, weight_key)
-                                m.weight_function.append(LowVramPatch(weight_key, self.patches, convert_func, set_func))
-                                patch_counter += 1
+                                if force_patch_weights:
+                                    self.patch_weight_to_device(weight_key)
+                                else:
+                                    _, set_func, convert_func = get_key_weight(self.model, weight_key)
+                                    m.weight_function.append(LowVramPatch(weight_key, self.patches, convert_func, set_func))
+                                    patch_counter += 1
                             if bias_key in self.patches:
-                                _, set_func, convert_func = get_key_weight(self.model, bias_key)
-                                m.bias_function.append(LowVramPatch(bias_key, self.patches, convert_func, set_func))
-                                patch_counter += 1
+                                if force_patch_weights:
+                                    self.patch_weight_to_device(bias_key)
+                                else:
+                                    _, set_func, convert_func = get_key_weight(self.model, bias_key)
+                                    m.bias_function.append(LowVramPatch(bias_key, self.patches, convert_func, set_func))
+                                    patch_counter += 1
                             cast_weight = True
 
-                        if cast_weight:
+                        if cast_weight and hasattr(m, "comfy_cast_weights"):
                             m.prev_comfy_cast_weights = m.comfy_cast_weights
                             m.comfy_cast_weights = True
                         m.comfy_patched_weights = False
                         memory_freed += module_mem
+                        offload_buffer = max(offload_buffer, potential_offload)
+                        offload_weight_factor.append(module_mem)
+                        offload_weight_factor.pop(0)
                         logging.debug("freed {}".format(n))
 
                         for param in params:
-                            self.pin_weight_to_device("{}.{}".format(n, param))
+                            self.pin_weight_to_device(key_param_name_to_key(n, param))
+
 
             self.model.model_lowvram = True
             self.model.lowvram_patch_counter += patch_counter
             self.model.model_loaded_weight_memory -= memory_freed
+            self.model.model_offload_buffer_memory = offload_buffer
+            logging.info("Unloaded partially: {:.2f} MB freed, {:.2f} MB remains loaded, {:.2f} MB buffer reserved, lowvram patches: {}".format(memory_freed / (1024 * 1024), self.model.model_loaded_weight_memory / (1024 * 1024), offload_buffer / (1024 * 1024), self.model.lowvram_patch_counter))
             return memory_freed
 
     def partially_load(self, device_to, extra_memory=0, force_patch_weights=False):
@@ -921,6 +992,9 @@ class ModelPatcher:
                 extra_memory += (used - self.model.model_loaded_weight_memory)
 
             self.patch_model(load_weights=False)
+            if extra_memory < 0 and not unpatch_weights:
+                self.partially_unload(self.offload_device, -extra_memory, force_patch_weights=force_patch_weights)
+                return 0
             full_load = False
             if self.model.model_lowvram == False and self.model.model_loaded_weight_memory > 0:
                 self.apply_hooks(self.forced_hooks, force_apply=True)
@@ -936,6 +1010,9 @@ class ModelPatcher:
 
             return self.model.model_loaded_weight_memory - current_used
 
+    def partially_unload_ram(self, ram_to_unload):
+        pass
+
     def detach(self, unpatch_all=True):
         self.eject_model()
         self.model_patches_to(self.offload_device)
@@ -1269,10 +1346,10 @@ class ModelPatcher:
                                                  key, original_weights=original_weights)
         del original_weights[key]
         if set_func is None:
-            out_weight = comfy.float.stochastic_rounding(out_weight, weight.dtype, seed=string_to_seed(key))
+            out_weight = comfy.float.stochastic_rounding(out_weight, weight.dtype, seed=comfy.utils.string_to_seed(key))
             comfy.utils.copy_to_param(self.model, key, out_weight)
         else:
-            set_func(out_weight, inplace_update=True, seed=string_to_seed(key))
+            set_func(out_weight, inplace_update=True, seed=comfy.utils.string_to_seed(key))
         if self.hook_mode == comfy.hooks.EnumHookMode.MaxSpeed:
             # TODO: disable caching if not enough system RAM to do so
             target_device = self.offload_device
@@ -1307,7 +1384,254 @@ class ModelPatcher:
         self.unpatch_hooks()
         self.clear_cached_hook_weights()
 
+    def state_dict_for_saving(self, clip_state_dict=None, vae_state_dict=None, clip_vision_state_dict=None):
+        unet_state_dict = self.model.diffusion_model.state_dict()
+        for k, v in unet_state_dict.items():
+            op_keys = k.rsplit('.', 1)
+            if (len(op_keys) < 2) or op_keys[1] not in ["weight", "bias"]:
+                continue
+            try:
+                op = comfy.utils.get_attr(self.model.diffusion_model, op_keys[0])
+            except:
+                continue
+            if not op or not hasattr(op, "comfy_cast_weights") or \
+                (hasattr(op, "comfy_patched_weights") and op.comfy_patched_weights == True):
+                continue
+            key = "diffusion_model." + k
+            unet_state_dict[k] = LazyCastingParam(self, key, comfy.utils.get_attr(self.model, key))
+        return self.model.state_dict_for_saving(unet_state_dict, clip_state_dict=clip_state_dict, vae_state_dict=vae_state_dict, clip_vision_state_dict=clip_vision_state_dict)
+
     def __del__(self):
         self.unpin_all_weights()
         self.detach(unpatch_all=False)
 
+class ModelPatcherDynamic(ModelPatcher):
+
+    def __new__(cls, model=None, load_device=None, offload_device=None, size=0, weight_inplace_update=False):
+        if load_device is not None and comfy.model_management.is_device_cpu(load_device):
+            #reroute to default MP for CPUs
+            return ModelPatcher(model, load_device, offload_device, size, weight_inplace_update)
+        return super().__new__(cls)
+
+    def __init__(self, model, load_device, offload_device, size=0, weight_inplace_update=False):
+        super().__init__(model, load_device, offload_device, size, weight_inplace_update)
+        #this is now way more dynamic and we dont support the same base model for both Dynamic
+        #and non-dynamic patchers.
+        if hasattr(self.model, "model_loaded_weight_memory"):
+            del self.model.model_loaded_weight_memory
+        if not hasattr(self.model, "dynamic_vbars"):
+            self.model.dynamic_vbars = {}
+        assert load_device is not None
+
+    def is_dynamic(self):
+        return True
+
+    def _vbar_get(self, create=False):
+        if self.load_device == torch.device("cpu"):
+            return None
+        vbar = self.model.dynamic_vbars.get(self.load_device, None)
+        if create and vbar is None:
+            # x10. We dont know what model defined type casts we have in the vbar, but virtual address
+            # space is pretty free. This will cover someone casting an entire model from FP4 to FP32
+            # with some left over.
+            vbar = comfy_aimdo.model_vbar.ModelVBAR(self.model_size() * 10, self.load_device.index)
+            self.model.dynamic_vbars[self.load_device] = vbar
+        return vbar
+
+    def loaded_size(self):
+        vbar = self._vbar_get()
+        if vbar is None:
+            return 0
+        return vbar.loaded_size()
+
+    def get_free_memory(self, device):
+        #NOTE: on high condition / batch counts, estimate should have already vacated
+        #all non-dynamic models so this is safe even if its not 100% true that this
+        #would all be avaiable for inference use.
+        return comfy.model_management.get_total_memory(device) - self.model_size()
+
+    #Pinning is deferred to ops time. Assert against this API to avoid pin leaks.
+
+    def pin_weight_to_device(self, key):
+        raise RuntimeError("pin_weight_to_device invalid for dymamic weight loading")
+
+    def unpin_weight(self, key):
+        raise RuntimeError("unpin_weight invalid for dymamic weight loading")
+
+    def unpin_all_weights(self):
+        self.partially_unload_ram(1e32)
+
+    def memory_required(self, input_shape):
+        #Pad this significantly. We are trying to get away from precise estimates. This
+        #estimate is only used when using the ModelPatcherDynamic after ModelPatcher. If you
+        #use all ModelPatcherDynamic this is ignored and its all done dynamically.
+        return super().memory_required(input_shape=input_shape) * 1.3 + (1024 ** 3)
+
+
+    def load(self, device_to=None, lowvram_model_memory=0, force_patch_weights=False, full_load=False, dirty=False):
+
+        #Force patching doesn't make sense in Dynamic loading, as you dont know what does and
+        #doesn't need to be forced at this stage. The only thing you could do would be patch
+        #it all on CPU which consumes huge RAM.
+        assert not force_patch_weights
+
+        #Full load doesn't make sense as we dont actually have any loader capability here and
+        #now.
+        assert not full_load
+
+        assert device_to == self.load_device
+
+        num_patches = 0
+        allocated_size = 0
+
+        with self.use_ejected():
+            self.unpatch_hooks()
+
+            vbar = self._vbar_get(create=True)
+            if vbar is not None:
+                vbar.prioritize()
+
+            #We force reserve VRAM for the non comfy-weight so we dont have to deal
+            #with pin and unpin syncrhonization which can be expensive for small weights
+            #with a high layer rate (e.g. autoregressive LLMs).
+            #prioritize the non-comfy weights (note the order reverse).
+            loading = self._load_list(prio_comfy_cast_weights=True)
+            loading.sort(reverse=True)
+
+            for x in loading:
+                _, _, _, n, m, params = x
+
+                def set_dirty(item, dirty):
+                    if dirty or not hasattr(item, "_v_signature"):
+                        item._v_signature = None
+
+                def setup_param(self, m, n, param_key):
+                    nonlocal num_patches
+                    key = key_param_name_to_key(n, param_key)
+
+                    weight_function = []
+
+                    weight, _, _ = get_key_weight(self.model, key)
+                    if weight is None:
+                        return 0
+                    if key in self.patches:
+                        setattr(m, param_key + "_lowvram_function", LowVramPatch(key, self.patches))
+                        num_patches += 1
+                    else:
+                        setattr(m, param_key + "_lowvram_function", None)
+
+                    if key in self.weight_wrapper_patches:
+                        weight_function.extend(self.weight_wrapper_patches[key])
+                    setattr(m, param_key + "_function", weight_function)
+                    geometry = weight
+                    if not isinstance(weight, QuantizedTensor):
+                        model_dtype = getattr(m, param_key + "_comfy_model_dtype", weight.dtype)
+                        weight._model_dtype = model_dtype
+                        geometry = comfy.memory_management.TensorGeometry(shape=weight.shape, dtype=model_dtype)
+                    return comfy.memory_management.vram_aligned_size(geometry)
+
+                if hasattr(m, "comfy_cast_weights"):
+                    m.comfy_cast_weights = True
+                    m.pin_failed = False
+                    m.seed_key = n
+                    set_dirty(m, dirty)
+
+                    v_weight_size = 0
+                    v_weight_size += setup_param(self, m, n, "weight")
+                    v_weight_size += setup_param(self, m, n, "bias")
+
+                    if vbar is not None and not hasattr(m, "_v"):
+                        m._v = vbar.alloc(v_weight_size)
+                        m._v_tensor = comfy_aimdo.torch.aimdo_to_tensor(m._v, device_to)
+                    allocated_size += v_weight_size
+
+                else:
+                    for param in params:
+                        key = key_param_name_to_key(n, param)
+                        weight, _, _ = get_key_weight(self.model, key)
+                        weight.seed_key = key
+                        set_dirty(weight, dirty)
+                        geometry = weight
+                        model_dtype = getattr(m, param + "_comfy_model_dtype", weight.dtype)
+                        geometry = comfy.memory_management.TensorGeometry(shape=weight.shape, dtype=model_dtype)
+                        weight_size = geometry.numel() * geometry.element_size()
+                        if vbar is not None and not hasattr(weight, "_v"):
+                            weight._v = vbar.alloc(weight_size)
+                            weight._v_tensor = comfy_aimdo.torch.aimdo_to_tensor(weight._v, device_to)
+                            weight._model_dtype = model_dtype
+                        allocated_size += weight_size
+                    vbar.set_watermark_limit(allocated_size)
+
+            logging.info(f"Model {self.model.__class__.__name__} prepared for dynamic VRAM loading. {allocated_size // (1024 ** 2)}MB Staged. {num_patches} patches attached.")
+
+            self.model.device = device_to
+            self.model.current_weight_patches_uuid = self.patches_uuid
+
+            for callback in self.get_all_callbacks(CallbacksMP.ON_LOAD):
+                #These are all super dangerous. Who knows what the custom nodes actually do here...
+                callback(self, device_to, lowvram_model_memory, force_patch_weights, full_load)
+
+            self.apply_hooks(self.forced_hooks, force_apply=True)
+
+    def partially_unload(self, device_to, memory_to_free=0, force_patch_weights=False):
+        assert not force_patch_weights #See above
+        assert self.load_device != torch.device("cpu")
+
+        vbar = self._vbar_get()
+        return 0 if vbar is None else vbar.free_memory(memory_to_free)
+
+    def partially_unload_ram(self, ram_to_unload):
+        loading = self._load_list(prio_comfy_cast_weights=True)
+        for x in loading:
+            _, _, _, _, m, _ = x
+            ram_to_unload -= comfy.pinned_memory.unpin_memory(m)
+            if ram_to_unload <= 0:
+                return
+
+    def patch_model(self, device_to=None, lowvram_model_memory=0, load_weights=True, force_patch_weights=False):
+        #This isn't used by the core at all and can only be to load a model out of
+        #the control of proper model_managment. If you are a custom node author reading
+        #this, the correct pattern is to call load_models_gpu() to get a proper
+        #managed load of your model.
+        assert not load_weights
+        return super().patch_model(load_weights=load_weights, force_patch_weights=force_patch_weights)
+
+    def unpatch_model(self, device_to=None, unpatch_weights=True):
+        super().unpatch_model(device_to=None, unpatch_weights=False)
+
+        if unpatch_weights:
+            self.partially_unload_ram(1e32)
+            self.partially_unload(None, 1e32)
+
+    def partially_load(self, device_to, extra_memory=0, force_patch_weights=False):
+        assert not force_patch_weights #See above
+        with self.use_ejected(skip_and_inject_on_exit_only=True):
+            dirty = self.model.current_weight_patches_uuid is not None and (self.model.current_weight_patches_uuid != self.patches_uuid)
+
+            self.unpatch_model(self.offload_device, unpatch_weights=False)
+            self.patch_model(load_weights=False)
+
+            try:
+                self.load(device_to, dirty=dirty)
+            except Exception as e:
+                self.detach()
+                raise e
+            #ModelPatcher::partially_load returns a number on what got loaded but
+            #nothing in core uses this and we have no data in the Dynamic world. Hit
+            #the custom node devs with a None rather than a 0 that would mislead any
+            #logic they might have.
+            return None
+
+    def patch_cached_hook_weights(self, cached_weights: dict, key: str, memory_counter: MemoryCounter):
+        assert False #Should be unreachable - we dont ever cache in the new implementation
+
+    def patch_hook_weight_to_device(self, hooks: comfy.hooks.HookGroup, combined_patches: dict, key: str, original_weights: dict, memory_counter: MemoryCounter):
+        if key not in combined_patches:
+            return
+
+        raise RuntimeError("Hooks not implemented in ModelPatcherDynamic. Please remove --fast arguments form ComfyUI startup")
+
+    def unpatch_hooks(self, whitelist_keys_set: set[str]=None) -> None:
+        pass
+
+CoreModelPatcher = ModelPatcher

comfy/ops.py

@@ -19,10 +19,16 @@
 import torch
 import logging
 import comfy.model_management
-from comfy.cli_args import args, PerformanceFeature
+from comfy.cli_args import args, PerformanceFeature, enables_dynamic_vram
 import comfy.float
 import comfy.rmsnorm
-import contextlib
+import json
+import comfy.memory_management
+import comfy.pinned_memory
+import comfy.utils
+
+import comfy_aimdo.model_vbar
+import comfy_aimdo.torch
 
 def run_every_op():
     if torch.compiler.is_compiling():
@@ -48,6 +54,8 @@ try:
             SDPA_BACKEND_PRIORITY.insert(0, SDPBackend.CUDNN_ATTENTION)
 
             def scaled_dot_product_attention(q, k, v, *args, **kwargs):
+                if q.nelement() < 1024 * 128:  # arbitrary number, for small inputs cudnn attention seems slower
+                    return torch.nn.functional.scaled_dot_product_attention(q, k, v, *args, **kwargs)
                 with sdpa_kernel(SDPA_BACKEND_PRIORITY, set_priority=True):
                     return torch.nn.functional.scaled_dot_product_attention(q, k, v, *args, **kwargs)
         else:
@@ -58,7 +66,8 @@ except (ModuleNotFoundError, TypeError):
 NVIDIA_MEMORY_CONV_BUG_WORKAROUND = False
 try:
     if comfy.model_management.is_nvidia():
-        if torch.backends.cudnn.version() >= 91002 and comfy.model_management.torch_version_numeric >= (2, 9) and comfy.model_management.torch_version_numeric <= (2, 10):
+        cudnn_version = torch.backends.cudnn.version()
+        if (cudnn_version >= 91002 and cudnn_version < 91500) and comfy.model_management.torch_version_numeric >= (2, 9) and comfy.model_management.torch_version_numeric <= (2, 10):
             #TODO: change upper bound version once it's fixed'
             NVIDIA_MEMORY_CONV_BUG_WORKAROUND = True
             logging.info("working around nvidia conv3d memory bug.")
@@ -71,54 +80,181 @@ def cast_to_input(weight, input, non_blocking=False, copy=True):
     return comfy.model_management.cast_to(weight, input.dtype, input.device, non_blocking=non_blocking, copy=copy)
 
 
-def cast_bias_weight(s, input=None, dtype=None, device=None, bias_dtype=None, offloadable=False):
+def cast_bias_weight_with_vbar(s, dtype, device, bias_dtype, non_blocking, compute_dtype):
+    offload_stream = None
+    xfer_dest = None
+    cast_geometry = comfy.memory_management.tensors_to_geometries([ s.weight, s.bias ])
+
+    signature = comfy_aimdo.model_vbar.vbar_fault(s._v)
+    if signature is not None:
+        xfer_dest = s._v_tensor
+    resident = comfy_aimdo.model_vbar.vbar_signature_compare(signature, s._v_signature)
+
+    if not resident:
+        cast_dest = None
+
+        xfer_source = [ s.weight, s.bias ]
+
+        pin = comfy.pinned_memory.get_pin(s)
+        if pin is not None:
+            xfer_source = [ pin ]
+
+        for data, geometry in zip([ s.weight, s.bias ], cast_geometry):
+            if data is None:
+                continue
+            if data.dtype != geometry.dtype:
+                cast_dest = xfer_dest
+                if cast_dest is None:
+                    cast_dest = torch.empty((comfy.memory_management.vram_aligned_size(cast_geometry),), dtype=torch.uint8, device=device)
+                xfer_dest = None
+                break
+
+        dest_size = comfy.memory_management.vram_aligned_size(xfer_source)
+        offload_stream = comfy.model_management.get_offload_stream(device)
+        if xfer_dest is None and offload_stream is not None:
+                xfer_dest = comfy.model_management.get_cast_buffer(offload_stream, device, dest_size, s)
+                if xfer_dest is None:
+                    offload_stream = comfy.model_management.get_offload_stream(device)
+                    xfer_dest = comfy.model_management.get_cast_buffer(offload_stream, device, dest_size, s)
+        if xfer_dest is None:
+            xfer_dest = torch.empty((dest_size,), dtype=torch.uint8, device=device)
+            offload_stream = None
+
+        if signature is None and pin is None:
+            comfy.pinned_memory.pin_memory(s)
+            pin = comfy.pinned_memory.get_pin(s)
+        else:
+            pin = None
+
+        if pin is not None:
+            comfy.model_management.cast_to_gathered(xfer_source, pin)
+            xfer_source = [ pin ]
+        #send it over
+        comfy.model_management.cast_to_gathered(xfer_source, xfer_dest, non_blocking=non_blocking, stream=offload_stream)
+        comfy.model_management.sync_stream(device, offload_stream)
+
+        if cast_dest is not None:
+            for pre_cast, post_cast in zip(comfy.memory_management.interpret_gathered_like([s.weight, s.bias ], xfer_dest),
+                                           comfy.memory_management.interpret_gathered_like(cast_geometry, cast_dest)):
+                if post_cast is not None:
+                    post_cast.copy_(pre_cast)
+            xfer_dest = cast_dest
+
+    params = comfy.memory_management.interpret_gathered_like(cast_geometry, xfer_dest)
+    weight = params[0]
+    bias = params[1]
+
+    def post_cast(s, param_key, x, dtype, resident, update_weight):
+        lowvram_fn = getattr(s, param_key + "_lowvram_function", None)
+        fns = getattr(s, param_key + "_function", [])
+
+        orig = x
+
+        def to_dequant(tensor, dtype):
+            tensor = tensor.to(dtype=dtype)
+            if isinstance(tensor, QuantizedTensor):
+                tensor = tensor.dequantize()
+            return tensor
+
+        if orig.dtype != dtype or len(fns) > 0:
+            x = to_dequant(x, dtype)
+        if not resident and lowvram_fn is not None:
+            x = to_dequant(x, dtype if compute_dtype is None else compute_dtype)
+            #FIXME: this is not accurate, we need to be sensitive to the compute dtype
+            x = lowvram_fn(x)
+            if (isinstance(orig, QuantizedTensor) and
+                (orig.dtype == dtype and len(fns) == 0 or update_weight)):
+                seed = comfy.utils.string_to_seed(s.seed_key)
+                y = QuantizedTensor.from_float(x, s.layout_type, scale="recalculate", stochastic_rounding=seed)
+                if orig.dtype == dtype and len(fns) == 0:
+                    #The layer actually wants our freshly saved QT
+                    x = y
+            elif update_weight:
+                y = comfy.float.stochastic_rounding(x, orig.dtype, seed = comfy.utils.string_to_seed(s.seed_key))
+            if update_weight:
+                orig.copy_(y)
+        for f in fns:
+            x = f(x)
+        return x
+
+    update_weight = signature is not None
+
+    weight = post_cast(s, "weight", weight, dtype, resident, update_weight)
+    if s.bias is not None:
+        bias = post_cast(s, "bias", bias, bias_dtype, resident, update_weight)
+    s._v_signature=signature
+
+    #FIXME: weird offload return protocol
+    return weight, bias, (offload_stream, device if signature is not None else None, None)
+
+
+def cast_bias_weight(s, input=None, dtype=None, device=None, bias_dtype=None, offloadable=False, compute_dtype=None):
     # NOTE: offloadable=False is a a legacy and if you are a custom node author reading this please pass
     # offloadable=True and call uncast_bias_weight() after your last usage of the weight/bias. This
     # will add async-offload support to your cast and improve performance.
     if input is not None:
         if dtype is None:
-            dtype = input.dtype
+            if isinstance(input, QuantizedTensor):
+                dtype = input.params.orig_dtype
+            else:
+                dtype = input.dtype
         if bias_dtype is None:
             bias_dtype = dtype
         if device is None:
             device = input.device
 
+    non_blocking = comfy.model_management.device_supports_non_blocking(device)
+
+    if hasattr(s, "_v"):
+        return cast_bias_weight_with_vbar(s, dtype, device, bias_dtype, non_blocking, compute_dtype)
+
     if offloadable and (device != s.weight.device or
                         (s.bias is not None and device != s.bias.device)):
         offload_stream = comfy.model_management.get_offload_stream(device)
     else:
         offload_stream = None
 
-    if offload_stream is not None:
-        wf_context = offload_stream
-    else:
-        wf_context = contextlib.nullcontext()
+    bias = None
+    weight = None
 
-    non_blocking = comfy.model_management.device_supports_non_blocking(device)
+    if offload_stream is not None and not args.cuda_malloc:
+        cast_buffer_size = comfy.memory_management.vram_aligned_size([ s.weight, s.bias ])
+        cast_buffer = comfy.model_management.get_cast_buffer(offload_stream, device, cast_buffer_size, s)
+        #The streams can be uneven in buffer capability and reject us. Retry to get the other stream
+        if cast_buffer is None:
+            offload_stream = comfy.model_management.get_offload_stream(device)
+            cast_buffer = comfy.model_management.get_cast_buffer(offload_stream, device, cast_buffer_size, s)
+        params = comfy.memory_management.interpret_gathered_like([ s.weight, s.bias ], cast_buffer)
+        weight = params[0]
+        bias = params[1]
 
     weight_has_function = len(s.weight_function) > 0
     bias_has_function = len(s.bias_function) > 0
 
-    weight = comfy.model_management.cast_to(s.weight, None, device, non_blocking=non_blocking, copy=weight_has_function, stream=offload_stream)
+    weight = comfy.model_management.cast_to(s.weight, None, device, non_blocking=non_blocking, copy=weight_has_function, stream=offload_stream, r=weight)
 
-    bias = None
     if s.bias is not None:
-        bias = comfy.model_management.cast_to(s.bias, bias_dtype, device, non_blocking=non_blocking, copy=bias_has_function, stream=offload_stream)
-
-        if bias_has_function:
-            with wf_context:
-                for f in s.bias_function:
-                    bias = f(bias)
-
-    weight = weight.to(dtype=dtype)
-    if weight_has_function:
-        with wf_context:
-            for f in s.weight_function:
-                weight = f(weight)
+        bias = comfy.model_management.cast_to(s.bias, None, device, non_blocking=non_blocking, copy=bias_has_function, stream=offload_stream, r=bias)
 
     comfy.model_management.sync_stream(device, offload_stream)
+
+    bias_a = bias
+    weight_a = weight
+
+    if s.bias is not None:
+        bias = bias.to(dtype=bias_dtype)
+        for f in s.bias_function:
+            bias = f(bias)
+
+    if weight_has_function or weight.dtype != dtype:
+        weight = weight.to(dtype=dtype)
+        if isinstance(weight, QuantizedTensor):
+            weight = weight.dequantize()
+        for f in s.weight_function:
+            weight = f(weight)
+
     if offloadable:
-        return weight, bias, offload_stream
+        return weight, bias, (offload_stream, weight_a, bias_a)
     else:
         #Legacy function signature
         return weight, bias
@@ -127,13 +263,22 @@ def cast_bias_weight(s, input=None, dtype=None, device=None, bias_dtype=None, of
 def uncast_bias_weight(s, weight, bias, offload_stream):
     if offload_stream is None:
         return
-    if weight is not None:
-        device = weight.device
-    else:
-        if bias is None:
-            return
-        device = bias.device
-    offload_stream.wait_stream(comfy.model_management.current_stream(device))
+    os, weight_a, bias_a = offload_stream
+    device=None
+    #FIXME: This is not good RTTI
+    if not isinstance(weight_a, torch.Tensor):
+        comfy_aimdo.model_vbar.vbar_unpin(s._v)
+        device = weight_a
+    if os is None:
+        return
+    if device is None:
+        if weight_a is not None:
+            device = weight_a.device
+        else:
+            if bias_a is None:
+                return
+            device = bias_a.device
+    os.wait_stream(comfy.model_management.current_stream(device))
 
 
 class CastWeightBiasOp:
@@ -143,6 +288,57 @@ class CastWeightBiasOp:
 
 class disable_weight_init:
     class Linear(torch.nn.Linear, CastWeightBiasOp):
+
+        def __init__(self, in_features, out_features, bias=True, device=None, dtype=None):
+            if not comfy.model_management.WINDOWS or not enables_dynamic_vram():
+                super().__init__(in_features, out_features, bias, device, dtype)
+                return
+
+            # Issue is with `torch.empty` still reserving the full memory for the layer.
+            # Windows doesn't over-commit memory so without this, We are momentarily commit
+            # charged for the weight even though we might zero-copy it when we load the
+            # state dict. If the commit charge exceeds the ceiling we can destabilize the
+            # system.
+            torch.nn.Module.__init__(self)
+            self.in_features = in_features
+            self.out_features = out_features
+            self.weight = None
+            self.bias = None
+            self.comfy_need_lazy_init_bias=bias
+            self.weight_comfy_model_dtype = dtype
+            self.bias_comfy_model_dtype = dtype
+
+        def _load_from_state_dict(self, state_dict, prefix, local_metadata,
+                                strict, missing_keys, unexpected_keys, error_msgs):
+
+            if not comfy.model_management.WINDOWS or not enables_dynamic_vram():
+                return super()._load_from_state_dict(state_dict, prefix, local_metadata, strict,
+                                                     missing_keys, unexpected_keys, error_msgs)
+            assign_to_params_buffers = local_metadata.get("assign_to_params_buffers", False)
+            prefix_len = len(prefix)
+            for k,v in state_dict.items():
+                if k[prefix_len:] == "weight":
+                    if not assign_to_params_buffers:
+                        v = v.clone()
+                    self.weight = torch.nn.Parameter(v, requires_grad=False)
+                elif k[prefix_len:] == "bias" and v is not None:
+                    if not assign_to_params_buffers:
+                        v = v.clone()
+                    self.bias = torch.nn.Parameter(v, requires_grad=False)
+                else:
+                    unexpected_keys.append(k)
+
+            #Reconcile default construction of the weight if its missing.
+            if self.weight is None:
+                v = torch.zeros(self.in_features, self.out_features)
+                self.weight = torch.nn.Parameter(v, requires_grad=False)
+                missing_keys.append(prefix+"weight")
+            if self.bias is None and self.comfy_need_lazy_init_bias:
+                v = torch.zeros(self.out_features,)
+                self.bias = torch.nn.Parameter(v, requires_grad=False)
+                missing_keys.append(prefix+"bias")
+
+
         def reset_parameters(self):
             return None
 
@@ -197,7 +393,9 @@ class disable_weight_init:
         def reset_parameters(self):
             return None
 
-        def _conv_forward(self, input, weight, bias, *args, **kwargs):
+        def _conv_forward(self, input, weight, bias, autopad=None, *args, **kwargs):
+            if autopad == "causal_zero":
+                weight = weight[:, :, -input.shape[2]:, :, :]
             if NVIDIA_MEMORY_CONV_BUG_WORKAROUND and weight.dtype in (torch.float16, torch.bfloat16):
                 out = torch.cudnn_convolution(input, weight, self.padding, self.stride, self.dilation, self.groups, benchmark=False, deterministic=False, allow_tf32=True)
                 if bias is not None:
@@ -206,15 +404,15 @@ class disable_weight_init:
             else:
                 return super()._conv_forward(input, weight, bias, *args, **kwargs)
 
-        def forward_comfy_cast_weights(self, input):
+        def forward_comfy_cast_weights(self, input, autopad=None):
             weight, bias, offload_stream = cast_bias_weight(self, input, offloadable=True)
-            x = self._conv_forward(input, weight, bias)
+            x = self._conv_forward(input, weight, bias, autopad=autopad)
             uncast_bias_weight(self, weight, bias, offload_stream)
             return x
 
         def forward(self, *args, **kwargs):
             run_every_op()
-            if self.comfy_cast_weights or len(self.weight_function) > 0 or len(self.bias_function) > 0:
+            if self.comfy_cast_weights or len(self.weight_function) > 0 or len(self.bias_function) > 0 or "autopad" in kwargs:
                 return self.forward_comfy_cast_weights(*args, **kwargs)
             else:
                 return super().forward(*args, **kwargs)
@@ -406,36 +604,34 @@ def fp8_linear(self, input):
         return None
 
     input_dtype = input.dtype
+    input_shape = input.shape
+    tensor_3d = input.ndim == 3
 
-    if input.ndim == 3 or input.ndim == 2:
-        w, bias, offload_stream = cast_bias_weight(self, input, dtype=dtype, bias_dtype=input_dtype, offloadable=True)
+    if tensor_3d:
+        input = input.reshape(-1, input_shape[2])
 
-        scale_weight = self.scale_weight
-        scale_input = self.scale_input
-        if scale_weight is None:
-            scale_weight = torch.ones((), device=input.device, dtype=torch.float32)
-        else:
-            scale_weight = scale_weight.to(input.device)
+    if input.ndim != 2:
+        return None
+    w, bias, offload_stream = cast_bias_weight(self, input, dtype=dtype, bias_dtype=input_dtype, offloadable=True)
+    scale_weight = torch.ones((), device=input.device, dtype=torch.float32)
 
-        if scale_input is None:
-            scale_input = torch.ones((), device=input.device, dtype=torch.float32)
-            input = torch.clamp(input, min=-448, max=448, out=input)
-            layout_params_weight = {'scale': scale_input, 'orig_dtype': input_dtype}
-            quantized_input = QuantizedTensor(input.to(dtype).contiguous(), "TensorCoreFP8Layout", layout_params_weight)
-        else:
-            scale_input = scale_input.to(input.device)
-            quantized_input = QuantizedTensor.from_float(input, "TensorCoreFP8Layout", scale=scale_input, dtype=dtype)
+    scale_input = torch.ones((), device=input.device, dtype=torch.float32)
+    input = torch.clamp(input, min=-448, max=448, out=input)
+    input_fp8 = input.to(dtype).contiguous()
+    layout_params_input = TensorCoreFP8Layout.Params(scale=scale_input, orig_dtype=input_dtype, orig_shape=tuple(input_fp8.shape))
+    quantized_input = QuantizedTensor(input_fp8, "TensorCoreFP8Layout", layout_params_input)
 
-        # Wrap weight in QuantizedTensor - this enables unified dispatch
-        # Call F.linear - __torch_dispatch__ routes to fp8_linear handler in quant_ops.py!
-        layout_params_weight = {'scale': scale_weight, 'orig_dtype': input_dtype}
-        quantized_weight = QuantizedTensor(w, "TensorCoreFP8Layout", layout_params_weight)
-        o = torch.nn.functional.linear(quantized_input, quantized_weight, bias)
+    # Wrap weight in QuantizedTensor - this enables unified dispatch
+    # Call F.linear - __torch_dispatch__ routes to fp8_linear handler in quant_ops.py!
+    layout_params_weight = TensorCoreFP8Layout.Params(scale=scale_weight, orig_dtype=input_dtype, orig_shape=tuple(w.shape))
+    quantized_weight = QuantizedTensor(w, "TensorCoreFP8Layout", layout_params_weight)
+    o = torch.nn.functional.linear(quantized_input, quantized_weight, bias)
 
-        uncast_bias_weight(self, w, bias, offload_stream)
-        return o
+    uncast_bias_weight(self, w, bias, offload_stream)
+    if tensor_3d:
+        o = o.reshape((input_shape[0], input_shape[1], w.shape[0]))
 
-    return None
+    return o
 
 class fp8_ops(manual_cast):
     class Linear(manual_cast.Linear):
@@ -445,7 +641,7 @@ class fp8_ops(manual_cast):
             return None
 
         def forward_comfy_cast_weights(self, input):
-            if not self.training:
+            if len(self.weight_function) == 0 and len(self.bias_function) == 0:
                 try:
                     out = fp8_linear(self, input)
                     if out is not None:
@@ -458,59 +654,6 @@ class fp8_ops(manual_cast):
             uncast_bias_weight(self, weight, bias, offload_stream)
             return x
 
-def scaled_fp8_ops(fp8_matrix_mult=False, scale_input=False, override_dtype=None):
-    logging.info("Using scaled fp8: fp8 matrix mult: {}, scale input: {}".format(fp8_matrix_mult, scale_input))
-    class scaled_fp8_op(manual_cast):
-        class Linear(manual_cast.Linear):
-            def __init__(self, *args, **kwargs):
-                if override_dtype is not None:
-                    kwargs['dtype'] = override_dtype
-                super().__init__(*args, **kwargs)
-
-            def reset_parameters(self):
-                if not hasattr(self, 'scale_weight'):
-                    self.scale_weight = torch.nn.parameter.Parameter(data=torch.ones((), device=self.weight.device, dtype=torch.float32), requires_grad=False)
-
-                if not scale_input:
-                    self.scale_input = None
-
-                if not hasattr(self, 'scale_input'):
-                    self.scale_input = torch.nn.parameter.Parameter(data=torch.ones((), device=self.weight.device, dtype=torch.float32), requires_grad=False)
-                return None
-
-            def forward_comfy_cast_weights(self, input):
-                if fp8_matrix_mult:
-                    out = fp8_linear(self, input)
-                    if out is not None:
-                        return out
-
-                weight, bias, offload_stream = cast_bias_weight(self, input, offloadable=True)
-
-                if weight.numel() < input.numel(): #TODO: optimize
-                    x = torch.nn.functional.linear(input, weight * self.scale_weight.to(device=weight.device, dtype=weight.dtype), bias)
-                else:
-                    x = torch.nn.functional.linear(input * self.scale_weight.to(device=weight.device, dtype=weight.dtype), weight, bias)
-                uncast_bias_weight(self, weight, bias, offload_stream)
-                return x
-
-            def convert_weight(self, weight, inplace=False, **kwargs):
-                if inplace:
-                    weight *= self.scale_weight.to(device=weight.device, dtype=weight.dtype)
-                    return weight
-                else:
-                    return weight * self.scale_weight.to(device=weight.device, dtype=weight.dtype)
-
-            def set_weight(self, weight, inplace_update=False, seed=None, return_weight=False, **kwargs):
-                weight = comfy.float.stochastic_rounding(weight / self.scale_weight.to(device=weight.device, dtype=weight.dtype), self.weight.dtype, seed=seed)
-                if return_weight:
-                    return weight
-                if inplace_update:
-                    self.weight.data.copy_(weight)
-                else:
-                    self.weight = torch.nn.Parameter(weight, requires_grad=False)
-
-    return scaled_fp8_op
-
 CUBLAS_IS_AVAILABLE = False
 try:
     from cublas_ops import CublasLinear
@@ -534,129 +677,261 @@ if CUBLAS_IS_AVAILABLE:
 # ==============================================================================
 # Mixed Precision Operations
 # ==============================================================================
-from .quant_ops import QuantizedTensor
+from .quant_ops import (
+    QuantizedTensor,
+    QUANT_ALGOS,
+    TensorCoreFP8Layout,
+    get_layout_class,
+)
 
-QUANT_FORMAT_MIXINS = {
-    "float8_e4m3fn": {
-        "dtype": torch.float8_e4m3fn,
-        "layout_type": "TensorCoreFP8Layout",
-        "parameters": {
-            "weight_scale": torch.nn.Parameter(torch.zeros((), dtype=torch.float32), requires_grad=False),
-            "input_scale": torch.nn.Parameter(torch.zeros((), dtype=torch.float32), requires_grad=False),
-        }
-    }
-}
 
-class MixedPrecisionOps(disable_weight_init):
-    _layer_quant_config = {}
-    _compute_dtype = torch.bfloat16
+def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_precision_mm=False, disabled=[]):
+    class MixedPrecisionOps(manual_cast):
+        _quant_config = quant_config
+        _compute_dtype = compute_dtype
+        _full_precision_mm = full_precision_mm
+        _disabled = disabled
 
-    class Linear(torch.nn.Module, CastWeightBiasOp):
-        def __init__(
-            self,
-            in_features: int,
-            out_features: int,
-            bias: bool = True,
-            device=None,
-            dtype=None,
-        ) -> None:
-            super().__init__()
+        class Linear(torch.nn.Module, CastWeightBiasOp):
+            def __init__(
+                self,
+                in_features: int,
+                out_features: int,
+                bias: bool = True,
+                device=None,
+                dtype=None,
+            ) -> None:
+                super().__init__()
 
-            self.factory_kwargs = {"device": device, "dtype": MixedPrecisionOps._compute_dtype}
-            # self.factory_kwargs = {"device": device, "dtype": dtype}
+                self.factory_kwargs = {"device": device, "dtype": MixedPrecisionOps._compute_dtype}
+                # self.factory_kwargs = {"device": device, "dtype": dtype}
 
-            self.in_features = in_features
-            self.out_features = out_features
-            if bias:
-                self.bias = torch.nn.Parameter(torch.empty(out_features, **self.factory_kwargs))
-            else:
-                self.register_parameter("bias", None)
+                self.in_features = in_features
+                self.out_features = out_features
+                if bias:
+                    self.bias = torch.nn.Parameter(torch.empty(out_features, **self.factory_kwargs))
+                else:
+                    self.register_parameter("bias", None)
 
-            self.tensor_class = None
+                self.tensor_class = None
+                self._full_precision_mm = MixedPrecisionOps._full_precision_mm
+                self._full_precision_mm_config = False
 
-        def reset_parameters(self):
-            return None
+            def reset_parameters(self):
+                return None
 
-        def _load_from_state_dict(self, state_dict, prefix, local_metadata,
-                                  strict, missing_keys, unexpected_keys, error_msgs):
+            def _load_scale_param(self, state_dict, prefix, param_name, device, manually_loaded_keys, dtype=None):
+                key = f"{prefix}{param_name}"
+                value = state_dict.pop(key, None)
+                if value is not None:
+                    value = value.to(device=device)
+                    if dtype is not None:
+                        value = value.view(dtype=dtype)
+                    manually_loaded_keys.append(key)
+                return value
 
-            device = self.factory_kwargs["device"]
-            layer_name = prefix.rstrip('.')
-            weight_key = f"{prefix}weight"
-            weight = state_dict.pop(weight_key, None)
-            if weight is None:
-                raise ValueError(f"Missing weight for layer {layer_name}")
+            def _load_from_state_dict(self, state_dict, prefix, local_metadata,
+                                    strict, missing_keys, unexpected_keys, error_msgs):
 
-            manually_loaded_keys = [weight_key]
+                device = self.factory_kwargs["device"]
+                layer_name = prefix.rstrip('.')
+                weight_key = f"{prefix}weight"
+                weight = state_dict.pop(weight_key, None)
+                if weight is None:
+                    logging.warning(f"Missing weight for layer {layer_name}")
+                    return
 
-            if layer_name not in MixedPrecisionOps._layer_quant_config:
-                self.weight = torch.nn.Parameter(weight.to(device=device, dtype=MixedPrecisionOps._compute_dtype), requires_grad=False)
-            else:
-                quant_format = MixedPrecisionOps._layer_quant_config[layer_name].get("format", None)
-                if quant_format is None:
-                    raise ValueError(f"Unknown quantization format for layer {layer_name}")
+                manually_loaded_keys = [weight_key]
 
-                mixin = QUANT_FORMAT_MIXINS[quant_format]
-                self.layout_type = mixin["layout_type"]
+                layer_conf = state_dict.pop(f"{prefix}comfy_quant", None)
+                if layer_conf is not None:
+                    layer_conf = json.loads(layer_conf.numpy().tobytes())
 
-                scale_key = f"{prefix}weight_scale"
-                layout_params = {
-                    'scale': state_dict.pop(scale_key, None),
-                    'orig_dtype': MixedPrecisionOps._compute_dtype
-                }
-                if layout_params['scale'] is not None:
-                    manually_loaded_keys.append(scale_key)
+                if layer_conf is None:
+                    self.weight = torch.nn.Parameter(weight.to(device=device, dtype=MixedPrecisionOps._compute_dtype), requires_grad=False)
+                else:
+                    self.quant_format = layer_conf.get("format", None)
+                    self._full_precision_mm_config = layer_conf.get("full_precision_matrix_mult", False)
+                    if not self._full_precision_mm:
+                        self._full_precision_mm = self._full_precision_mm_config
 
-                self.weight = torch.nn.Parameter(
-                    QuantizedTensor(weight.to(device=device, dtype=mixin["dtype"]), self.layout_type, layout_params),
-                    requires_grad=False
-                )
+                    if self.quant_format in MixedPrecisionOps._disabled:
+                        self._full_precision_mm = True
 
-                for param_name, param_value in mixin["parameters"].items():
-                    param_key = f"{prefix}{param_name}"
-                    _v = state_dict.pop(param_key, None)
-                    if _v is None:
+                    if self.quant_format is None:
+                        raise ValueError(f"Unknown quantization format for layer {layer_name}")
+
+                    qconfig = QUANT_ALGOS[self.quant_format]
+                    self.layout_type = qconfig["comfy_tensor_layout"]
+                    layout_cls = get_layout_class(self.layout_type)
+
+                    # Load format-specific parameters
+                    if self.quant_format in ["float8_e4m3fn", "float8_e5m2"]:
+                        # FP8: single tensor scale
+                        scale = self._load_scale_param(state_dict, prefix, "weight_scale", device, manually_loaded_keys)
+
+                        params = layout_cls.Params(
+                            scale=scale,
+                            orig_dtype=MixedPrecisionOps._compute_dtype,
+                            orig_shape=(self.out_features, self.in_features),
+                        )
+
+                    elif self.quant_format == "nvfp4":
+                        # NVFP4: tensor_scale (weight_scale_2) + block_scale (weight_scale)
+                        tensor_scale = self._load_scale_param(state_dict, prefix, "weight_scale_2", device, manually_loaded_keys)
+                        block_scale = self._load_scale_param(state_dict, prefix, "weight_scale", device, manually_loaded_keys,
+                                                             dtype=torch.float8_e4m3fn)
+
+                        if tensor_scale is None or block_scale is None:
+                            raise ValueError(f"Missing NVFP4 scales for layer {layer_name}")
+
+                        params = layout_cls.Params(
+                            scale=tensor_scale,
+                            block_scale=block_scale,
+                            orig_dtype=MixedPrecisionOps._compute_dtype,
+                            orig_shape=(self.out_features, self.in_features),
+                        )
+                    else:
+                        raise ValueError(f"Unsupported quantization format: {self.quant_format}")
+
+                    self.weight = torch.nn.Parameter(
+                        QuantizedTensor(weight.to(device=device, dtype=qconfig["storage_t"]), self.layout_type, params),
+                        requires_grad=False
+                    )
+
+                    for param_name in qconfig["parameters"]:
+                        if param_name in {"weight_scale", "weight_scale_2"}:
+                            continue  # Already handled above
+
+                        param_key = f"{prefix}{param_name}"
+                        _v = state_dict.pop(param_key, None)
+                        if _v is None:
+                            continue
+                        self.register_parameter(param_name, torch.nn.Parameter(_v.to(device=device), requires_grad=False))
+                        manually_loaded_keys.append(param_key)
+
+                super()._load_from_state_dict(state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs)
+
+                for key in manually_loaded_keys:
+                    if key in missing_keys:
+                        missing_keys.remove(key)
+
+            def state_dict(self, *args, destination=None, prefix="", **kwargs):
+                if destination is not None:
+                    sd = destination
+                else:
+                    sd = {}
+
+                if self.bias is not None:
+                    sd["{}bias".format(prefix)] = self.bias
+
+                if isinstance(self.weight, QuantizedTensor):
+                    sd_out = self.weight.state_dict("{}weight".format(prefix))
+                    for k in sd_out:
+                        sd[k] = sd_out[k]
+
+                    quant_conf = {"format": self.quant_format}
+                    if self._full_precision_mm_config:
+                        quant_conf["full_precision_matrix_mult"] = True
+                    sd["{}comfy_quant".format(prefix)] = torch.tensor(list(json.dumps(quant_conf).encode('utf-8')), dtype=torch.uint8)
+
+                    input_scale = getattr(self, 'input_scale', None)
+                    if input_scale is not None:
+                        sd["{}input_scale".format(prefix)] = input_scale
+                else:
+                    sd["{}weight".format(prefix)] = self.weight
+                return sd
+
+            def _forward(self, input, weight, bias):
+                return torch.nn.functional.linear(input, weight, bias)
+
+            def forward_comfy_cast_weights(self, input, compute_dtype=None):
+                weight, bias, offload_stream = cast_bias_weight(self, input, offloadable=True, compute_dtype=compute_dtype)
+                x = self._forward(input, weight, bias)
+                uncast_bias_weight(self, weight, bias, offload_stream)
+                return x
+
+            def forward(self, input, *args, **kwargs):
+                run_every_op()
+
+                input_shape = input.shape
+                reshaped_3d = False
+                #If cast needs to apply lora, it should be done in the compute dtype
+                compute_dtype = input.dtype
+
+                if (getattr(self, 'layout_type', None) is not None and
+                    not isinstance(input, QuantizedTensor) and not self._full_precision_mm and
+                    not getattr(self, 'comfy_force_cast_weights', False) and
+                    len(self.weight_function) == 0 and len(self.bias_function) == 0):
+
+                    # Reshape 3D tensors to 2D for quantization (needed for NVFP4 and others)
+                    input_reshaped = input.reshape(-1, input_shape[2]) if input.ndim == 3 else input
+
+                    # Fall back to non-quantized for non-2D tensors
+                    if input_reshaped.ndim == 2:
+                        reshaped_3d = input.ndim == 3
+                        # dtype is now implicit in the layout class
+                        scale = getattr(self, 'input_scale', None)
+                        if scale is not None:
+                            scale = comfy.model_management.cast_to_device(scale, input.device, None)
+                        input = QuantizedTensor.from_float(input_reshaped, self.layout_type, scale=scale)
+
+
+                output = self.forward_comfy_cast_weights(input, compute_dtype)
+
+                # Reshape output back to 3D if input was 3D
+                if reshaped_3d:
+                    output = output.reshape((input_shape[0], input_shape[1], self.weight.shape[0]))
+
+                return output
+
+            def convert_weight(self, weight, inplace=False, **kwargs):
+                if isinstance(weight, QuantizedTensor):
+                    return weight.dequantize()
+                else:
+                    return weight
+
+            def set_weight(self, weight, inplace_update=False, seed=None, return_weight=False, **kwargs):
+                if getattr(self, 'layout_type', None) is not None:
+                    # dtype is now implicit in the layout class
+                    weight = QuantizedTensor.from_float(weight, self.layout_type, scale="recalculate", stochastic_rounding=seed, inplace_ops=True).to(self.weight.dtype)
+                else:
+                    weight = weight.to(self.weight.dtype)
+                if return_weight:
+                    return weight
+
+                assert inplace_update is False  # TODO: eventually remove the inplace_update stuff
+                self.weight = torch.nn.Parameter(weight, requires_grad=False)
+
+            def _apply(self, fn, recurse=True):  # This is to get torch.compile + moving weights to another device working
+                if recurse:
+                    for module in self.children():
+                        module._apply(fn)
+
+                for key, param in self._parameters.items():
+                    if param is None:
                         continue
-                    setattr(self, param_name, torch.nn.Parameter(_v.to(device=device), requires_grad=False))
-                    manually_loaded_keys.append(param_key)
+                    self.register_parameter(key, torch.nn.Parameter(fn(param), requires_grad=False))
+                for key, buf in self._buffers.items():
+                    if buf is not None:
+                        self._buffers[key] = fn(buf)
+                return self
 
-            super()._load_from_state_dict(state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs)
+    return MixedPrecisionOps
 
-            for key in manually_loaded_keys:
-                if key in missing_keys:
-                    missing_keys.remove(key)
+def pick_operations(weight_dtype, compute_dtype, load_device=None, disable_fast_fp8=False, fp8_optimizations=False, model_config=None):
+    fp8_compute = comfy.model_management.supports_fp8_compute(load_device) # TODO: if we support more ops this needs to be more granular
+    nvfp4_compute = comfy.model_management.supports_nvfp4_compute(load_device)
 
-        def _forward(self, input, weight, bias):
-            return torch.nn.functional.linear(input, weight, bias)
-
-        def forward_comfy_cast_weights(self, input):
-            weight, bias, offload_stream = cast_bias_weight(self, input, offloadable=True)
-            x = self._forward(input, weight, bias)
-            uncast_bias_weight(self, weight, bias, offload_stream)
-            return x
-
-        def forward(self, input, *args, **kwargs):
-            run_every_op()
-
-            if self.comfy_cast_weights or len(self.weight_function) > 0 or len(self.bias_function) > 0:
-                return self.forward_comfy_cast_weights(input, *args, **kwargs)
-            if (getattr(self, 'layout_type', None) is not None and
-                getattr(self, 'input_scale', None) is not None and
-                not isinstance(input, QuantizedTensor)):
-                input = QuantizedTensor.from_float(input, self.layout_type, scale=self.input_scale, fp8_dtype=self.weight.dtype)
-            return self._forward(input, self.weight, self.bias)
-
-
-def pick_operations(weight_dtype, compute_dtype, load_device=None, disable_fast_fp8=False, fp8_optimizations=False, scaled_fp8=None, model_config=None):
-    if model_config and hasattr(model_config, 'layer_quant_config') and model_config.layer_quant_config:
-        MixedPrecisionOps._layer_quant_config = model_config.layer_quant_config
-        MixedPrecisionOps._compute_dtype = compute_dtype
-        logging.info(f"Using mixed precision operations: {len(model_config.layer_quant_config)} quantized layers")
-        return MixedPrecisionOps
-
-    fp8_compute = comfy.model_management.supports_fp8_compute(load_device)
-    if scaled_fp8 is not None:
-        return scaled_fp8_ops(fp8_matrix_mult=fp8_compute and fp8_optimizations, scale_input=fp8_optimizations, override_dtype=scaled_fp8)
+    if model_config and hasattr(model_config, 'quant_config') and model_config.quant_config:
+        logging.info("Using mixed precision operations")
+        disabled = set()
+        if not nvfp4_compute:
+            disabled.add("nvfp4")
+        if not fp8_compute:
+            disabled.add("float8_e4m3fn")
+            disabled.add("float8_e5m2")
+        return mixed_precision_ops(model_config.quant_config, compute_dtype, disabled=disabled)
 
     if (
         fp8_compute and

comfy/pinned_memory.py

@@ -0,0 +1,29 @@
+import torch
+import comfy.model_management
+import comfy.memory_management
+
+from comfy.cli_args import args
+
+def get_pin(module):
+    return getattr(module, "_pin", None)
+
+def pin_memory(module):
+    if module.pin_failed or args.disable_pinned_memory or get_pin(module) is not None:
+        return
+    #FIXME: This is a RAM cache trigger event
+    size = comfy.memory_management.vram_aligned_size([ module.weight, module.bias ])
+    pin = torch.empty((size,), dtype=torch.uint8)
+    if comfy.model_management.pin_memory(pin):
+        module._pin = pin
+    else:
+        module.pin_failed = True
+        return False
+    return True
+
+def unpin_memory(module):
+    if get_pin(module) is None:
+        return 0
+    size = module._pin.numel() * module._pin.element_size()
+    comfy.model_management.unpin_memory(module._pin)
+    del module._pin
+    return size

comfy/quant_ops.py

@@ -1,512 +1,174 @@
 import torch
 import logging
-from typing import Tuple, Dict
 
-_LAYOUT_REGISTRY = {}
-_GENERIC_UTILS = {}
+try:
+    import comfy_kitchen as ck
+    from comfy_kitchen.tensor import (
+        QuantizedTensor,
+        QuantizedLayout,
+        TensorCoreFP8Layout as _CKFp8Layout,
+        TensorCoreNVFP4Layout as _CKNvfp4Layout,
+        register_layout_op,
+        register_layout_class,
+        get_layout_class,
+    )
+    _CK_AVAILABLE = True
+    if torch.version.cuda is None:
+        ck.registry.disable("cuda")
+    else:
+        cuda_version = tuple(map(int, str(torch.version.cuda).split('.')))
+        if cuda_version < (13,):
+            ck.registry.disable("cuda")
+            logging.warning("WARNING: You need pytorch with cu130 or higher to use optimized CUDA operations.")
 
+    ck.registry.disable("triton")
+    for k, v in ck.list_backends().items():
+        logging.info(f"Found comfy_kitchen backend {k}: {v}")
+except ImportError as e:
+    logging.error(f"Failed to import comfy_kitchen, Error: {e}, fp8 and fp4 support will not be available.")
+    _CK_AVAILABLE = False
 
-def register_layout_op(torch_op, layout_type):
-    """
-    Decorator to register a layout-specific operation handler.
-    Args:
-        torch_op: PyTorch operation (e.g., torch.ops.aten.linear.default)
-        layout_type: Layout class (e.g., TensorCoreFP8Layout)
-    Example:
-        @register_layout_op(torch.ops.aten.linear.default, TensorCoreFP8Layout)
-        def fp8_linear(func, args, kwargs):
-            # FP8-specific linear implementation
-            ...
-    """
-    def decorator(handler_func):
-        if torch_op not in _LAYOUT_REGISTRY:
-            _LAYOUT_REGISTRY[torch_op] = {}
-        _LAYOUT_REGISTRY[torch_op][layout_type] = handler_func
-        return handler_func
-    return decorator
+    class QuantizedTensor:
+        pass
 
+    class _CKFp8Layout:
+        pass
 
-def register_generic_util(torch_op):
-    """
-    Decorator to register a generic utility that works for all layouts.
-    Args:
-        torch_op: PyTorch operation (e.g., torch.ops.aten.detach.default)
+    class _CKNvfp4Layout:
+        pass
 
-    Example:
-        @register_generic_util(torch.ops.aten.detach.default)
-        def generic_detach(func, args, kwargs):
-            # Works for any layout
-            ...
-    """
-    def decorator(handler_func):
-        _GENERIC_UTILS[torch_op] = handler_func
-        return handler_func
-    return decorator
+    def register_layout_class(name, cls):
+        pass
 
+    def get_layout_class(name):
+        return None
 
-def _get_layout_from_args(args):
-    for arg in args:
-        if isinstance(arg, QuantizedTensor):
-            return arg._layout_type
-        elif isinstance(arg, (list, tuple)):
-            for item in arg:
-                if isinstance(item, QuantizedTensor):
-                    return item._layout_type
-    return None
-
-
-def _move_layout_params_to_device(params, device):
-    new_params = {}
-    for k, v in params.items():
-        if isinstance(v, torch.Tensor):
-            new_params[k] = v.to(device=device)
-        else:
-            new_params[k] = v
-    return new_params
-
-
-def _copy_layout_params(params):
-    new_params = {}
-    for k, v in params.items():
-        if isinstance(v, torch.Tensor):
-            new_params[k] = v.clone()
-        else:
-            new_params[k] = v
-    return new_params
-
-
-class QuantizedLayout:
-    """
-    Base class for quantization layouts.
-
-    A layout encapsulates the format-specific logic for quantization/dequantization
-    and provides a uniform interface for extracting raw tensors needed for computation.
-
-    New quantization formats should subclass this and implement the required methods.
-    """
-    @classmethod
-    def quantize(cls, tensor, **kwargs) -> Tuple[torch.Tensor, Dict]:
-        raise NotImplementedError(f"{cls.__name__} must implement quantize()")
-
-    @staticmethod
-    def dequantize(qdata, **layout_params) -> torch.Tensor:
-        raise NotImplementedError("TensorLayout must implement dequantize()")
-
-    @classmethod
-    def get_plain_tensors(cls, qtensor) -> torch.Tensor:
-        raise NotImplementedError(f"{cls.__name__} must implement get_plain_tensors()")
-
-
-class QuantizedTensor(torch.Tensor):
-    """
-    Universal quantized tensor that works with any layout.
-
-    This tensor subclass uses a pluggable layout system to support multiple
-    quantization formats (FP8, INT4, INT8, etc.) without code duplication.
-
-    The layout_type determines format-specific behavior, while common operations
-    (detach, clone, to) are handled generically.
-
-    Attributes:
-        _qdata: The quantized tensor data
-        _layout_type: Layout class (e.g., TensorCoreFP8Layout)
-        _layout_params: Dict with layout-specific params (scale, zero_point, etc.)
-    """
-
-    @staticmethod
-    def __new__(cls, qdata, layout_type, layout_params):
-        """
-        Create a quantized tensor.
-
-        Args:
-            qdata: The quantized data tensor
-            layout_type: Layout class (subclass of QuantizedLayout)
-            layout_params: Dict with layout-specific parameters
-        """
-        return torch.Tensor._make_wrapper_subclass(cls, qdata.shape, device=qdata.device, dtype=qdata.dtype, requires_grad=False)
-
-    def __init__(self, qdata, layout_type, layout_params):
-        self._qdata = qdata
-        self._layout_type = layout_type
-        self._layout_params = layout_params
-
-    def __repr__(self):
-        layout_name = self._layout_type
-        param_str = ", ".join(f"{k}={v}" for k, v in list(self._layout_params.items())[:2])
-        return f"QuantizedTensor(shape={self.shape}, layout={layout_name}, {param_str})"
-
-    @property
-    def layout_type(self):
-        return self._layout_type
-
-    def __tensor_flatten__(self):
-        """
-        Tensor flattening protocol for proper device movement.
-        """
-        inner_tensors = ["_qdata"]
-        ctx = {
-            "layout_type": self._layout_type,
-        }
-
-        tensor_params = {}
-        non_tensor_params = {}
-        for k, v in self._layout_params.items():
-            if isinstance(v, torch.Tensor):
-                tensor_params[k] = v
-            else:
-                non_tensor_params[k] = v
-
-        ctx["tensor_param_keys"] = list(tensor_params.keys())
-        ctx["non_tensor_params"] = non_tensor_params
-
-        for k, v in tensor_params.items():
-            attr_name = f"_layout_param_{k}"
-            object.__setattr__(self, attr_name, v)
-            inner_tensors.append(attr_name)
-
-        return inner_tensors, ctx
-
-    @staticmethod
-    def __tensor_unflatten__(inner_tensors, ctx, outer_size, outer_stride):
-        """
-        Tensor unflattening protocol for proper device movement.
-        Reconstructs the QuantizedTensor after device movement.
-        """
-        layout_type = ctx["layout_type"]
-        layout_params = dict(ctx["non_tensor_params"])
-
-        for key in ctx["tensor_param_keys"]:
-            attr_name = f"_layout_param_{key}"
-            layout_params[key] = inner_tensors[attr_name]
-
-        return QuantizedTensor(inner_tensors["_qdata"], layout_type, layout_params)
-
-    @classmethod
-    def from_float(cls, tensor, layout_type, **quantize_kwargs) -> 'QuantizedTensor':
-        qdata, layout_params = LAYOUTS[layout_type].quantize(tensor, **quantize_kwargs)
-        return cls(qdata, layout_type, layout_params)
-
-    def dequantize(self) -> torch.Tensor:
-        return LAYOUTS[self._layout_type].dequantize(self._qdata, **self._layout_params)
-
-    @classmethod
-    def __torch_dispatch__(cls, func, types, args=(), kwargs=None):
-        kwargs = kwargs or {}
-
-        # Step 1: Check generic utilities first (detach, clone, to, etc.)
-        if func in _GENERIC_UTILS:
-            return _GENERIC_UTILS[func](func, args, kwargs)
-
-        # Step 2: Check layout-specific handlers (linear, matmul, etc.)
-        layout_type = _get_layout_from_args(args)
-        if layout_type and func in _LAYOUT_REGISTRY:
-            handler = _LAYOUT_REGISTRY[func].get(layout_type)
-            if handler:
-                return handler(func, args, kwargs)
-
-        # Step 3: Fallback to dequantization
-        if isinstance(args[0] if args else None, QuantizedTensor):
-            logging.info(f"QuantizedTensor: Unhandled operation {func}, falling back to dequantization. kwargs={kwargs}")
-        return cls._dequant_and_fallback(func, args, kwargs)
-
-    @classmethod
-    def _dequant_and_fallback(cls, func, args, kwargs):
-        def dequant_arg(arg):
-            if isinstance(arg, QuantizedTensor):
-                return arg.dequantize()
-            elif isinstance(arg, (list, tuple)):
-                return type(arg)(dequant_arg(a) for a in arg)
-            return arg
-
-        new_args = dequant_arg(args)
-        new_kwargs = dequant_arg(kwargs)
-        return func(*new_args, **new_kwargs)
-
+import comfy.float
 
 # ==============================================================================
-# Generic Utilities (Layout-Agnostic Operations)
+# FP8 Layouts with Comfy-Specific Extensions
 # ==============================================================================
 
-def _create_transformed_qtensor(qt, transform_fn):
-    new_data = transform_fn(qt._qdata)
-    new_params = _copy_layout_params(qt._layout_params)
-    return QuantizedTensor(new_data, qt._layout_type, new_params)
+class _TensorCoreFP8LayoutBase(_CKFp8Layout):
+    FP8_DTYPE = None  # Must be overridden in subclass
 
-
-def _handle_device_transfer(qt, target_device, target_dtype=None, target_layout=None, op_name="to"):
-    if target_dtype is not None and target_dtype != qt.dtype:
-        logging.warning(
-            f"QuantizedTensor: dtype conversion requested to {target_dtype}, "
-            f"but not supported for quantized tensors. Ignoring dtype."
-        )
-
-    if target_layout is not None and target_layout != torch.strided:
-        logging.warning(
-            f"QuantizedTensor: layout change requested to {target_layout}, "
-            f"but not supported. Ignoring layout."
-        )
-
-    # Handle device transfer
-    current_device = qt._qdata.device
-    if target_device is not None:
-        # Normalize device for comparison
-        if isinstance(target_device, str):
-            target_device = torch.device(target_device)
-        if isinstance(current_device, str):
-            current_device = torch.device(current_device)
-
-        if target_device != current_device:
-            logging.debug(f"QuantizedTensor.{op_name}: Moving from {current_device} to {target_device}")
-            new_q_data = qt._qdata.to(device=target_device)
-            new_params = _move_layout_params_to_device(qt._layout_params, target_device)
-            new_qt = QuantizedTensor(new_q_data, qt._layout_type, new_params)
-            logging.debug(f"QuantizedTensor.{op_name}: Created new tensor on {target_device}")
-            return new_qt
-
-    logging.debug(f"QuantizedTensor.{op_name}: No device change needed, returning original")
-    return qt
-
-
-@register_generic_util(torch.ops.aten.detach.default)
-def generic_detach(func, args, kwargs):
-    """Detach operation - creates a detached copy of the quantized tensor."""
-    qt = args[0]
-    if isinstance(qt, QuantizedTensor):
-        return _create_transformed_qtensor(qt, lambda x: x.detach())
-    return func(*args, **kwargs)
-
-
-@register_generic_util(torch.ops.aten.clone.default)
-def generic_clone(func, args, kwargs):
-    """Clone operation - creates a deep copy of the quantized tensor."""
-    qt = args[0]
-    if isinstance(qt, QuantizedTensor):
-        return _create_transformed_qtensor(qt, lambda x: x.clone())
-    return func(*args, **kwargs)
-
-
-@register_generic_util(torch.ops.aten._to_copy.default)
-def generic_to_copy(func, args, kwargs):
-    """Device/dtype transfer operation - handles .to(device) calls."""
-    qt = args[0]
-    if isinstance(qt, QuantizedTensor):
-        return _handle_device_transfer(
-            qt,
-            target_device=kwargs.get('device', None),
-            target_dtype=kwargs.get('dtype', None),
-            op_name="_to_copy"
-        )
-    return func(*args, **kwargs)
-
-
-@register_generic_util(torch.ops.aten.to.dtype_layout)
-def generic_to_dtype_layout(func, args, kwargs):
-    """Handle .to(device) calls using the dtype_layout variant."""
-    qt = args[0]
-    if isinstance(qt, QuantizedTensor):
-        return _handle_device_transfer(
-            qt,
-            target_device=kwargs.get('device', None),
-            target_dtype=kwargs.get('dtype', None),
-            target_layout=kwargs.get('layout', None),
-            op_name="to"
-        )
-    return func(*args, **kwargs)
-
-
-@register_generic_util(torch.ops.aten.copy_.default)
-def generic_copy_(func, args, kwargs):
-    qt_dest = args[0]
-    src = args[1]
-
-    if isinstance(qt_dest, QuantizedTensor):
-        if isinstance(src, QuantizedTensor):
-            # Copy from another quantized tensor
-            qt_dest._qdata.copy_(src._qdata)
-            qt_dest._layout_type = src._layout_type
-            qt_dest._layout_params = _copy_layout_params(src._layout_params)
-        else:
-            # Copy from regular tensor - just copy raw data
-            qt_dest._qdata.copy_(src)
-        return qt_dest
-    return func(*args, **kwargs)
-
-
-@register_generic_util(torch.ops.aten._has_compatible_shallow_copy_type.default)
-def generic_has_compatible_shallow_copy_type(func, args, kwargs):
-    return True
-
-# ==============================================================================
-# FP8 Layout + Operation Handlers
-# ==============================================================================
-class TensorCoreFP8Layout(QuantizedLayout):
-    """
-    Storage format:
-    - qdata: FP8 tensor (torch.float8_e4m3fn or torch.float8_e5m2)
-    - scale: Scalar tensor (float32) for dequantization
-    - orig_dtype: Original dtype before quantization (for casting back)
-    """
     @classmethod
-    def quantize(cls, tensor, scale=None, dtype=torch.float8_e4m3fn):
+    def quantize(cls, tensor, scale=None, stochastic_rounding=0, inplace_ops=False):
+        if cls.FP8_DTYPE is None:
+            raise NotImplementedError(f"{cls.__name__} must define FP8_DTYPE")
+
         orig_dtype = tensor.dtype
+        orig_shape = tuple(tensor.shape)
+
+        if isinstance(scale, str) and scale == "recalculate":
+            scale = torch.amax(tensor.abs()).to(dtype=torch.float32) / torch.finfo(cls.FP8_DTYPE).max
+            if tensor.dtype not in [torch.float32, torch.bfloat16]:  # Prevent scale from being too small
+                tensor_info = torch.finfo(tensor.dtype)
+                scale = (1.0 / torch.clamp((1.0 / scale), min=tensor_info.min, max=tensor_info.max))
 
         if scale is None:
-            scale = torch.amax(tensor.abs()) / torch.finfo(dtype).max
+            scale = torch.ones((), device=tensor.device, dtype=torch.float32)
+        if not isinstance(scale, torch.Tensor):
+            scale = torch.tensor(scale, device=tensor.device, dtype=torch.float32)
+
+        if stochastic_rounding > 0:
+            if inplace_ops:
+                tensor *= (1.0 / scale).to(tensor.dtype)
+            else:
+                tensor = tensor * (1.0 / scale).to(tensor.dtype)
+            qdata = comfy.float.stochastic_rounding(tensor, dtype=cls.FP8_DTYPE, seed=stochastic_rounding)
+        else:
+            qdata = ck.quantize_per_tensor_fp8(tensor, scale, cls.FP8_DTYPE)
+
+        params = cls.Params(scale=scale.float(), orig_dtype=orig_dtype, orig_shape=orig_shape)
+        return qdata, params
+
+
+class TensorCoreNVFP4Layout(_CKNvfp4Layout):
+    @classmethod
+    def quantize(cls, tensor, scale=None, stochastic_rounding=0, inplace_ops=False):
+        if tensor.dim() != 2:
+            raise ValueError(f"NVFP4 requires 2D tensor, got {tensor.dim()}D")
+
+        orig_dtype = tensor.dtype
+        orig_shape = tuple(tensor.shape)
+
+        if scale is None or (isinstance(scale, str) and scale == "recalculate"):
+            scale = torch.amax(tensor.abs()) / (ck.float_utils.F8_E4M3_MAX * ck.float_utils.F4_E2M1_MAX)
 
         if not isinstance(scale, torch.Tensor):
             scale = torch.tensor(scale)
         scale = scale.to(device=tensor.device, dtype=torch.float32)
 
-        tensor_scaled = tensor * (1.0 / scale).to(tensor.dtype)
-        # TODO: uncomment this if it's actually needed because the clamp has a small performance penality'
-        # lp_amax = torch.finfo(dtype).max
-        # torch.clamp(tensor_scaled, min=-lp_amax, max=lp_amax, out=tensor_scaled)
-        qdata = tensor_scaled.to(dtype, memory_format=torch.contiguous_format)
+        padded_shape = cls.get_padded_shape(orig_shape)
+        needs_padding = padded_shape != orig_shape
 
-        layout_params = {
-            'scale': scale,
-            'orig_dtype': orig_dtype
-        }
-        return qdata, layout_params
+        if stochastic_rounding > 0:
+            qdata, block_scale = comfy.float.stochastic_round_quantize_nvfp4_by_block(tensor, scale, pad_16x=needs_padding, seed=stochastic_rounding)
+        else:
+            qdata, block_scale = ck.quantize_nvfp4(tensor, scale, pad_16x=needs_padding)
 
-    @staticmethod
-    def dequantize(qdata, scale, orig_dtype, **kwargs):
-        plain_tensor = torch.ops.aten._to_copy.default(qdata, dtype=orig_dtype)
-        return plain_tensor * scale
-
-    @classmethod
-    def get_plain_tensors(cls, qtensor):
-        return qtensor._qdata, qtensor._layout_params['scale']
+        params = cls.Params(
+            scale=scale,
+            orig_dtype=orig_dtype,
+            orig_shape=orig_shape,
+            block_scale=block_scale,
+        )
+        return qdata, params
 
 
-LAYOUTS = {
-    "TensorCoreFP8Layout": TensorCoreFP8Layout,
+class TensorCoreFP8E4M3Layout(_TensorCoreFP8LayoutBase):
+    FP8_DTYPE = torch.float8_e4m3fn
+
+
+class TensorCoreFP8E5M2Layout(_TensorCoreFP8LayoutBase):
+    FP8_DTYPE = torch.float8_e5m2
+
+
+# Backward compatibility alias - default to E4M3
+TensorCoreFP8Layout = TensorCoreFP8E4M3Layout
+
+
+# ==============================================================================
+# Registry
+# ==============================================================================
+
+register_layout_class("TensorCoreFP8Layout", TensorCoreFP8Layout)
+register_layout_class("TensorCoreFP8E4M3Layout", TensorCoreFP8E4M3Layout)
+register_layout_class("TensorCoreFP8E5M2Layout", TensorCoreFP8E5M2Layout)
+register_layout_class("TensorCoreNVFP4Layout", TensorCoreNVFP4Layout)
+
+QUANT_ALGOS = {
+    "float8_e4m3fn": {
+        "storage_t": torch.float8_e4m3fn,
+        "parameters": {"weight_scale", "input_scale"},
+        "comfy_tensor_layout": "TensorCoreFP8E4M3Layout",
+    },
+    "float8_e5m2": {
+        "storage_t": torch.float8_e5m2,
+        "parameters": {"weight_scale", "input_scale"},
+        "comfy_tensor_layout": "TensorCoreFP8E5M2Layout",
+    },
+    "nvfp4": {
+        "storage_t": torch.uint8,
+        "parameters": {"weight_scale", "weight_scale_2", "input_scale"},
+        "comfy_tensor_layout": "TensorCoreNVFP4Layout",
+        "group_size": 16,
+    },
 }
 
 
-@register_layout_op(torch.ops.aten.linear.default, "TensorCoreFP8Layout")
-def fp8_linear(func, args, kwargs):
-    input_tensor = args[0]
-    weight = args[1]
-    bias = args[2] if len(args) > 2 else None
+# ==============================================================================
+# Re-exports for backward compatibility
+# ==============================================================================
 
-    if isinstance(input_tensor, QuantizedTensor) and isinstance(weight, QuantizedTensor):
-        plain_input, scale_a = TensorCoreFP8Layout.get_plain_tensors(input_tensor)
-        plain_weight, scale_b = TensorCoreFP8Layout.get_plain_tensors(weight)
-
-        out_dtype = kwargs.get("out_dtype")
-        if out_dtype is None:
-            out_dtype = input_tensor._layout_params['orig_dtype']
-
-        weight_t = plain_weight.t()
-
-        tensor_2d = False
-        if len(plain_input.shape) == 2:
-            tensor_2d = True
-            plain_input = plain_input.unsqueeze(1)
-
-        input_shape = plain_input.shape
-        if len(input_shape) != 3:
-            return None
-
-        try:
-            output = torch._scaled_mm(
-                plain_input.reshape(-1, input_shape[2]).contiguous(),
-                weight_t,
-                bias=bias,
-                scale_a=scale_a,
-                scale_b=scale_b,
-                out_dtype=out_dtype,
-            )
-
-            if isinstance(output, tuple):  # TODO: remove when we drop support for torch 2.4
-                output = output[0]
-
-            if not tensor_2d:
-                output = output.reshape((-1, input_shape[1], weight.shape[0]))
-
-            if output.dtype in [torch.float8_e4m3fn, torch.float8_e5m2]:
-                output_scale = scale_a * scale_b
-                output_params = {
-                    'scale': output_scale,
-                    'orig_dtype': input_tensor._layout_params['orig_dtype']
-                }
-                return QuantizedTensor(output, "TensorCoreFP8Layout", output_params)
-            else:
-                return output
-
-        except Exception as e:
-            raise RuntimeError(f"FP8 _scaled_mm failed, falling back to dequantization: {e}")
-
-    # Case 2: DQ Fallback
-    if isinstance(weight, QuantizedTensor):
-        weight = weight.dequantize()
-    if isinstance(input_tensor, QuantizedTensor):
-        input_tensor = input_tensor.dequantize()
-
-    return torch.nn.functional.linear(input_tensor, weight, bias)
-
-def fp8_mm_(input_tensor, weight, bias=None, out_dtype=None):
-    if out_dtype is None:
-        out_dtype = input_tensor._layout_params['orig_dtype']
-
-    plain_input, scale_a = TensorCoreFP8Layout.get_plain_tensors(input_tensor)
-    plain_weight, scale_b = TensorCoreFP8Layout.get_plain_tensors(weight)
-
-    output = torch._scaled_mm(
-        plain_input.contiguous(),
-        plain_weight,
-        bias=bias,
-        scale_a=scale_a,
-        scale_b=scale_b,
-        out_dtype=out_dtype,
-    )
-
-    if isinstance(output, tuple):  # TODO: remove when we drop support for torch 2.4
-        output = output[0]
-    return output
-
-@register_layout_op(torch.ops.aten.addmm.default, "TensorCoreFP8Layout")
-def fp8_addmm(func, args, kwargs):
-    input_tensor = args[1]
-    weight = args[2]
-    bias = args[0]
-
-    if isinstance(input_tensor, QuantizedTensor) and isinstance(weight, QuantizedTensor):
-        return fp8_mm_(input_tensor, weight, bias=bias, out_dtype=kwargs.get("out_dtype", None))
-
-    a = list(args)
-    if isinstance(args[0], QuantizedTensor):
-        a[0] = args[0].dequantize()
-    if isinstance(args[1], QuantizedTensor):
-        a[1] = args[1].dequantize()
-    if isinstance(args[2], QuantizedTensor):
-        a[2] = args[2].dequantize()
-
-    return func(*a, **kwargs)
-
-@register_layout_op(torch.ops.aten.mm.default, "TensorCoreFP8Layout")
-def fp8_mm(func, args, kwargs):
-    input_tensor = args[0]
-    weight = args[1]
-
-    if isinstance(input_tensor, QuantizedTensor) and isinstance(weight, QuantizedTensor):
-        return fp8_mm_(input_tensor, weight, bias=None, out_dtype=kwargs.get("out_dtype", None))
-
-    a = list(args)
-    if isinstance(args[0], QuantizedTensor):
-        a[0] = args[0].dequantize()
-    if isinstance(args[1], QuantizedTensor):
-        a[1] = args[1].dequantize()
-    return func(*a, **kwargs)
-
-@register_layout_op(torch.ops.aten.view.default, "TensorCoreFP8Layout")
-@register_layout_op(torch.ops.aten.t.default, "TensorCoreFP8Layout")
-def fp8_func(func, args, kwargs):
-    input_tensor = args[0]
-    if isinstance(input_tensor, QuantizedTensor):
-        plain_input, scale_a = TensorCoreFP8Layout.get_plain_tensors(input_tensor)
-        ar = list(args)
-        ar[0] = plain_input
-        return QuantizedTensor(func(*ar, **kwargs), "TensorCoreFP8Layout", input_tensor._layout_params)
-    return func(*args, **kwargs)
+__all__ = [
+    "QuantizedTensor",
+    "QuantizedLayout",
+    "TensorCoreFP8Layout",
+    "TensorCoreFP8E4M3Layout",
+    "TensorCoreFP8E5M2Layout",
+    "TensorCoreNVFP4Layout",
+    "QUANT_ALGOS",
+    "register_layout_op",
+]

comfy/sample.py

@@ -37,12 +37,18 @@ def prepare_noise(latent_image, seed, noise_inds=None):
 
     return noises
 
-def fix_empty_latent_channels(model, latent_image):
+def fix_empty_latent_channels(model, latent_image, downscale_ratio_spacial=None):
     if latent_image.is_nested:
         return latent_image
     latent_format = model.get_model_object("latent_format") #Resize the empty latent image so it has the right number of channels
-    if latent_format.latent_channels != latent_image.shape[1] and torch.count_nonzero(latent_image) == 0:
-        latent_image = comfy.utils.repeat_to_batch_size(latent_image, latent_format.latent_channels, dim=1)
+    if torch.count_nonzero(latent_image) == 0:
+        if latent_format.latent_channels != latent_image.shape[1]:
+            latent_image = comfy.utils.repeat_to_batch_size(latent_image, latent_format.latent_channels, dim=1)
+        if downscale_ratio_spacial is not None:
+            if downscale_ratio_spacial != latent_format.spacial_downscale_ratio:
+                ratio = downscale_ratio_spacial / latent_format.spacial_downscale_ratio
+                latent_image = comfy.utils.common_upscale(latent_image, round(latent_image.shape[-1] * ratio), round(latent_image.shape[-2] * ratio), "nearest-exact", crop="disabled")
+
     if latent_format.latent_dimensions == 3 and latent_image.ndim == 4:
         latent_image = latent_image.unsqueeze(2)
     return latent_image