t# This is a combination of 2 commits.

d

.ci/windows_amd_base_files/README_VERY_IMPORTANT.txt

@@ -1,27 +0,0 @@
-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
-
-HOW TO RUN:
-
-If you have a AMD gpu:
-
-run_amd_gpu.bat
-
-If you have memory issues you can try disabling the smart memory management by running comfyui with:
-
-run_amd_gpu_disable_smart_memory.bat
-
-IF YOU GET A RED ERROR IN THE UI MAKE SURE YOU HAVE A MODEL/CHECKPOINT IN: ComfyUI\models\checkpoints
-
-You can download the stable diffusion XL one from: https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/blob/main/sd_xl_base_1.0_0.9vae.safetensors
-
-
-RECOMMENDED WAY TO UPDATE:
-To update the ComfyUI code: update\update_comfyui.bat
-
-
-TO SHARE MODELS BETWEEN COMFYUI AND ANOTHER UI:
-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_base_files/run_nvidia_gpu_fast_fp16_accumulation.bat

@@ -1,2 +1,2 @@
-.\python_embeded\python.exe -s ComfyUI\main.py --windows-standalone-build --disable-smart-memory
+.\python_embeded\python.exe -s ComfyUI\main.py --windows-standalone-build --fast fp16_accumulation
 pause

.ci/windows_nvidia_base_files/advanced/run_nvidia_gpu_disable_api_nodes.bat

@@ -1,3 +0,0 @@
-..\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.
-pause

.ci/windows_nvidia_base_files/run_nvidia_gpu.bat

@@ -1,3 +0,0 @@
-.\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.
-pause

.ci/windows_nvidia_base_files/run_nvidia_gpu_fast_fp16_accumulation.bat

@@ -1,3 +0,0 @@
-.\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.
-pause

.github/ISSUE_TEMPLATE/bug-report.yml

@@ -8,15 +8,13 @@ body:
         Before submitting a **Bug Report**, please ensure the following:
 
         - **1:** You are running the latest version of ComfyUI.
-        - **2:** You have your ComfyUI logs and relevant workflow on hand and will post them in this bug report.
+        - **2:** You have looked at the existing bug reports and made sure this isn't already reported.
         - **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. If you have custom node try updating them to the latest version.
+        `--disable-all-custom-nodes` command line argument.
         - **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.
 
-        ## 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.
+        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.
   - type: checkboxes
     id: custom-nodes-test
     attributes:

.github/PULL_REQUEST_TEMPLATE/api-node.md

@@ -1,21 +0,0 @@
-<!-- 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

@@ -1,58 +0,0 @@
-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

@@ -1,78 +0,0 @@
-name: "Release Stable All Portable Versions"
-
-on:
-  workflow_dispatch:
-    inputs:
-      git_tag:
-        description: 'Git tag'
-        required: true
-        type: string
-
-jobs:
-  release_nvidia_default:
-    permissions:
-      contents: "write"
-      packages: "write"
-      pull-requests: "read"
-    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"
-      rel_name: "nvidia"
-      rel_extra_name: ""
-      test_release: true
-    secrets: inherit
-
-  release_nvidia_cu128:
-    permissions:
-      contents: "write"
-      packages: "write"
-      pull-requests: "read"
-    name: "Release NVIDIA cu128"
-    uses: ./.github/workflows/stable-release.yml
-    with:
-      git_tag: ${{ inputs.git_tag }}
-      cache_tag: "cu128"
-      python_minor: "12"
-      python_patch: "10"
-      rel_name: "nvidia"
-      rel_extra_name: "_cu128"
-      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"
-    uses: ./.github/workflows/stable-release.yml
-    with:
-      git_tag: ${{ inputs.git_tag }}
-      cache_tag: "rocm644"
-      python_minor: "12"
-      python_patch: "10"
-      rel_name: "amd"
-      rel_extra_name: ""
-      test_release: false
-    secrets: inherit

.github/workflows/ruff.yml

@@ -21,28 +21,3 @@ jobs:
 
     - name: Run Ruff
       run: ruff check .
-
-  pylint:
-    name: Run Pylint
-    runs-on: ubuntu-latest
-
-    steps:
-    - name: Checkout repository
-      uses: actions/checkout@v4
-
-    - name: Set up Python
-      uses: actions/setup-python@v4
-      with:
-        python-version: '3.12'
-
-    - name: Install requirements
-      run: |
-        python -m pip install --upgrade pip
-        pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cpu
-        pip install -r requirements.txt
-
-    - name: Install Pylint
-      run: pip install pylint
-
-    - name: Run Pylint
-      run: pylint comfy_api_nodes

.github/workflows/stable-release.yml

@@ -2,53 +2,17 @@
 name: "Release Stable Version"
 
 on:
-  workflow_call:
-    inputs:
-      git_tag:
-        description: 'Git tag'
-        required: true
-        type: string
-      cache_tag:
-        description: 'Cached dependencies tag'
-        required: true
-        type: string
-        default: "cu129"
-      python_minor:
-        description: 'Python minor version'
-        required: true
-        type: string
-        default: "13"
-      python_patch:
-        description: 'Python patch version'
-        required: true
-        type: string
-        default: "6"
-      rel_name:
-        description: 'Release name'
-        required: true
-        type: string
-        default: "nvidia"
-      rel_extra_name:
-        description: 'Release extra name'
-        required: false
-        type: string
-        default: ""
-      test_release:
-        description: 'Test Release'
-        required: true
-        type: boolean
-        default: true
   workflow_dispatch:
     inputs:
       git_tag:
         description: 'Git tag'
         required: true
         type: string
-      cache_tag:
-        description: 'Cached dependencies tag'
+      cu:
+        description: 'CUDA version'
         required: true
         type: string
-        default: "cu129"
+        default: "129"
       python_minor:
         description: 'Python minor version'
         required: true
@@ -59,21 +23,7 @@ on:
         required: true
         type: string
         default: "6"
-      rel_name:
-        description: 'Release name'
-        required: true
-        type: string
-        default: "nvidia"
-      rel_extra_name:
-        description: 'Release extra name'
-        required: false
-        type: string
-        default: ""
-      test_release:
-        description: 'Test Release'
-        required: true
-        type: boolean
-        default: true
+
 
 jobs:
   package_comfy_windows:
@@ -92,15 +42,15 @@ jobs:
         id: cache
         with:
           path: |
-            ${{ inputs.cache_tag }}_python_deps.tar
+            cu${{ inputs.cu }}_python_deps.tar
             update_comfyui_and_python_dependencies.bat
-          key: ${{ runner.os }}-build-${{ inputs.cache_tag }}-${{ inputs.python_minor }}
+          key: ${{ runner.os }}-build-cu${{ inputs.cu }}-${{ inputs.python_minor }}
       - shell: bash
         run: |
-          mv ${{ inputs.cache_tag }}_python_deps.tar ../
+          mv cu${{ inputs.cu }}_python_deps.tar ../
           mv update_comfyui_and_python_dependencies.bat ../
           cd ..
-          tar xf ${{ inputs.cache_tag }}_python_deps.tar
+          tar xf cu${{ inputs.cu }}_python_deps.tar
           pwd
           ls
 
@@ -115,19 +65,12 @@ jobs:
           echo 'import site' >> ./python3${{ inputs.python_minor }}._pth
           curl https://bootstrap.pypa.io/get-pip.py -o get-pip.py
           ./python.exe get-pip.py
-          ./python.exe -s -m pip install ../${{ inputs.cache_tag }}_python_deps/*
-
-          grep comfyui ../ComfyUI/requirements.txt > ./requirements_comfyui.txt
-          ./python.exe -s -m pip install -r requirements_comfyui.txt
-          rm requirements_comfyui.txt
-
+          ./python.exe -s -m pip install ../cu${{ inputs.cu }}_python_deps/*
           sed -i '1i../ComfyUI' ./python3${{ inputs.python_minor }}._pth
 
-          if test -f ./Lib/site-packages/torch/lib/dnnl.lib; then
-            rm ./Lib/site-packages/torch/lib/dnnl.lib #I don't think this is actually used and I need the space
-            rm ./Lib/site-packages/torch/lib/libprotoc.lib
-            rm ./Lib/site-packages/torch/lib/libprotobuf.lib
-          fi
+          rm ./Lib/site-packages/torch/lib/dnnl.lib #I don't think this is actually used and I need the space
+          rm ./Lib/site-packages/torch/lib/libprotoc.lib
+          rm ./Lib/site-packages/torch/lib/libprotobuf.lib
 
           cd ..
 
@@ -142,18 +85,14 @@ jobs:
 
           mkdir update
           cp -r ComfyUI/.ci/update_windows/* ./update/
-          cp -r ComfyUI/.ci/windows_${{ inputs.rel_name }}_base_files/* ./
+          cp -r ComfyUI/.ci/windows_base_files/* ./
           cp ../update_comfyui_and_python_dependencies.bat ./update/
 
           cd ..
 
           "C:\Program Files\7-Zip\7z.exe" a -t7z -m0=lzma2 -mx=9 -mfb=128 -md=768m -ms=on -mf=BCJ2 ComfyUI_windows_portable.7z ComfyUI_windows_portable
-          mv ComfyUI_windows_portable.7z ComfyUI/ComfyUI_windows_portable_${{ inputs.rel_name }}${{ inputs.rel_extra_name }}.7z
+          mv ComfyUI_windows_portable.7z ComfyUI/ComfyUI_windows_portable_nvidia.7z
 
-      - shell: bash
-        if: ${{ inputs.test_release }}
-        run: |
-          cd ..
           cd ComfyUI_windows_portable
           python_embeded/python.exe -s ComfyUI/main.py --quick-test-for-ci --cpu
 
@@ -162,9 +101,10 @@ jobs:
           ls
 
       - name: Upload binaries to release
-        uses: softprops/action-gh-release@v2
+        uses: svenstaro/upload-release-action@v2
         with:
-          files: ComfyUI_windows_portable_${{ inputs.rel_name }}${{ inputs.rel_extra_name }}.7z
-          tag_name: ${{ inputs.git_tag }}
+          repo_token: ${{ secrets.GITHUB_TOKEN }}
+          file: ComfyUI_windows_portable_nvidia.7z
+          tag: ${{ inputs.git_tag }}
+          overwrite: true
           draft: true
-          overwrite_files: true

.github/workflows/test-ci.yml

@@ -21,15 +21,14 @@ jobs:
       fail-fast: false
       matrix:
         # os: [macos, linux, windows]
-        # os: [macos, linux]
-        os: [linux]
-        python_version: ["3.10", "3.11", "3.12"]
+        os: [macos, linux]
+        python_version: ["3.9", "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: ""
@@ -74,15 +73,14 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        # os: [macos, linux]
-        os: [linux]
+        os: [macos, 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-unit.yml

@@ -10,7 +10,7 @@ jobs:
   test:
     strategy:
       matrix:
-        os: [ubuntu-latest, windows-2022, macos-latest]
+        os: [ubuntu-latest, windows-latest, macos-latest]
     runs-on: ${{ matrix.os }}
     continue-on-error: true
     steps:

.github/workflows/windows_release_dependencies.yml

@@ -17,7 +17,7 @@ on:
         description: 'cuda version'
         required: true
         type: string
-        default: "130"
+        default: "129"
 
       python_minor:
         description: 'python minor version'
@@ -29,7 +29,7 @@ on:
         description: 'python patch version'
         required: true
         type: string
-        default: "9"
+        default: "6"
 #  push:
 #    branches:
 #      - master
@@ -56,8 +56,7 @@ jobs:
             ..\python_embeded\python.exe -s -m pip install --upgrade torch torchvision torchaudio ${{ inputs.xformers }} --extra-index-url https://download.pytorch.org/whl/cu${{ inputs.cu }} -r ../ComfyUI/requirements.txt pygit2
             pause" > update_comfyui_and_python_dependencies.bat
 
-            grep -v comfyui requirements.txt > requirements_nocomfyui.txt
-            python -m pip wheel --no-cache-dir torch torchvision torchaudio ${{ inputs.xformers }} ${{ inputs.extra_dependencies }} --extra-index-url https://download.pytorch.org/whl/cu${{ inputs.cu }} -r requirements_nocomfyui.txt pygit2 -w ./temp_wheel_dir
+            python -m pip wheel --no-cache-dir torch torchvision torchaudio ${{ inputs.xformers }} ${{ inputs.extra_dependencies }} --extra-index-url https://download.pytorch.org/whl/cu${{ inputs.cu }} -r requirements.txt pygit2 -w ./temp_wheel_dir
             python -m pip install --no-cache-dir ./temp_wheel_dir/*
             echo installed basic
             ls -lah temp_wheel_dir

.github/workflows/windows_release_dependencies_manual.yml

@@ -1,64 +0,0 @@
-name: "Windows Release dependencies Manual"
-
-on:
-  workflow_dispatch:
-    inputs:
-      torch_dependencies:
-        description: 'torch dependencies'
-        required: false
-        type: string
-        default: "torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu128"
-      cache_tag:
-        description: 'Cached dependencies tag'
-        required: true
-        type: string
-        default: "cu128"
-
-      python_minor:
-        description: 'python minor version'
-        required: true
-        type: string
-        default: "12"
-
-      python_patch:
-        description: 'python patch version'
-        required: true
-        type: string
-        default: "10"
-
-jobs:
-  build_dependencies:
-    runs-on: windows-latest
-    steps:
-        - uses: actions/checkout@v4
-        - uses: actions/setup-python@v5
-          with:
-            python-version: 3.${{ inputs.python_minor }}.${{ inputs.python_patch }}
-
-        - shell: bash
-          run: |
-            echo "@echo off
-            call update_comfyui.bat nopause
-            echo -
-            echo This will try to update pytorch and all python dependencies.
-            echo -
-            echo If you just want to update normally, close this and run update_comfyui.bat instead.
-            echo -
-            pause
-            ..\python_embeded\python.exe -s -m pip install --upgrade ${{ inputs.torch_dependencies }} -r ../ComfyUI/requirements.txt pygit2
-            pause" > update_comfyui_and_python_dependencies.bat
-
-            grep -v comfyui requirements.txt > requirements_nocomfyui.txt
-            python -m pip wheel --no-cache-dir ${{ inputs.torch_dependencies }} -r requirements_nocomfyui.txt pygit2 -w ./temp_wheel_dir
-            python -m pip install --no-cache-dir ./temp_wheel_dir/*
-            echo installed basic
-            ls -lah temp_wheel_dir
-            mv temp_wheel_dir ${{ inputs.cache_tag }}_python_deps
-            tar cf ${{ inputs.cache_tag }}_python_deps.tar ${{ inputs.cache_tag }}_python_deps
-
-        - uses: actions/cache/save@v4
-          with:
-            path: |
-              ${{ inputs.cache_tag }}_python_deps.tar
-              update_comfyui_and_python_dependencies.bat
-            key: ${{ runner.os }}-build-${{ inputs.cache_tag }}-${{ inputs.python_minor }}

.github/workflows/windows_release_nightly_pytorch.yml

@@ -68,7 +68,7 @@ jobs:
 
             mkdir update
             cp -r ComfyUI/.ci/update_windows/* ./update/
-            cp -r ComfyUI/.ci/windows_nvidia_base_files/* ./
+            cp -r ComfyUI/.ci/windows_base_files/* ./
             cp -r ComfyUI/.ci/windows_nightly_base_files/* ./
 
             echo "call update_comfyui.bat nopause

.github/workflows/windows_release_package.yml

@@ -81,7 +81,7 @@ jobs:
 
             mkdir update
             cp -r ComfyUI/.ci/update_windows/* ./update/
-            cp -r ComfyUI/.ci/windows_nvidia_base_files/* ./
+            cp -r ComfyUI/.ci/windows_base_files/* ./
             cp ../update_comfyui_and_python_dependencies.bat ./update/
 
             cd ..

CODEOWNERS

@@ -1,3 +1,25 @@
 # Admins
 * @comfyanonymous
-* @kosinkadink
+
+# Note: Github teams syntax cannot be used here as the repo is not owned by Comfy-Org.
+# Inlined the team members for now.
+
+# Maintainers
+*.md @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne @guill
+/tests/ @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne @guill
+/tests-unit/ @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne @guill
+/notebooks/ @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne @guill
+/script_examples/ @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne @guill
+/.github/ @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne @guill
+/requirements.txt @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne @guill
+/pyproject.toml @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne @guill
+
+# Python web server
+/api_server/ @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @christian-byrne @guill
+/app/ @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @christian-byrne @guill
+/utils/ @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @christian-byrne @guill
+
+# Node developers
+/comfy_extras/ @yoland68 @robinjhuang @pythongosssss @ltdrdata @Kosinkadink @webfiltered @christian-byrne @guill
+/comfy/comfy_types/ @yoland68 @robinjhuang @pythongosssss @ltdrdata @Kosinkadink @webfiltered @christian-byrne @guill
+/comfy_api_nodes/ @yoland68 @robinjhuang @pythongosssss @ltdrdata @Kosinkadink @webfiltered @christian-byrne @guill

QUANTIZATION.md

@@ -1,168 +0,0 @@
-# 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

@@ -112,11 +112,10 @@ Workflow examples can be found on the [Examples page](https://comfyanonymous.git
 
 ## Release Process
 
-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:
+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:
 
 1. **[ComfyUI Core](https://github.com/comfyanonymous/ComfyUI)**
-   - Releases a new stable version (e.g., v0.7.0) roughly every week.
-   - Commits outside of the stable release tags may be very unstable and break many custom nodes.
+   - Releases a new stable version (e.g., v0.7.0)
    - Serves as the foundation for the desktop release
 
 2. **[ComfyUI Desktop](https://github.com/Comfy-Org/desktop)**
@@ -173,20 +172,10 @@ 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) 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\
+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
 
 If you have trouble extracting it, right click the file -> properties -> unblock
 
-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).
-
-[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?
 
 See the [Config file](extra_model_paths.yaml.example) to set the search paths for models. In the standalone windows build you can find this file in the ComfyUI directory. Rename this file to extra_model_paths.yaml and edit it with your favorite text editor.
@@ -202,11 +191,7 @@ comfy install
 
 ## Manual Install (Windows, Linux)
 
-Python 3.14 works but you may encounter issues with the torch compile node. The free threaded variant is still missing some dependencies.
-
-Python 3.13 is very well supported. If you have trouble with some custom node dependencies on 3.13 you can try 3.12
-
-### Instructions:
+Python 3.13 is very well supported. If you have trouble with some custom node dependencies you can try 3.12
 
 Git clone this repo.
 
@@ -215,36 +200,18 @@ Put your SD checkpoints (the huge ckpt/safetensors files) in: models/checkpoints
 Put your VAE in: models/vae
 
 
-### AMD GPUs (Linux)
-
+### AMD GPUs (Linux only)
 AMD users can install rocm and pytorch with pip if you don't have it already installed, this is the command to install the stable version:
 
 ```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 6.4 which might have some performance improvements:
 
-```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.
-
-These have less hardware support than the builds above but they work on windows. You also need to install the pytorch version specific to your hardware.
-
-RDNA 3 (RX 7000 series):
-
-```pip install --pre torch torchvision torchaudio --index-url https://rocm.nightlies.amd.com/v2/gfx110X-dgpu/```
-
-RDNA 3.5 (Strix halo/Ryzen AI Max+ 365):
-
-```pip install --pre torch torchvision torchaudio --index-url https://rocm.nightlies.amd.com/v2/gfx1151/```
-
-RDNA 4 (RX 9000 series):
-
-```pip install --pre torch torchvision torchaudio --index-url https://rocm.nightlies.amd.com/v2/gfx120X-all/```
+```pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/rocm6.4```
 
 ### Intel GPUs (Windows and Linux)
 
-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)
+(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)
 
 1. To install PyTorch xpu, use the following command:
 
@@ -254,15 +221,19 @@ 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:
 
-```pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu130```
+```pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu129```
 
 This is the command to install pytorch nightly instead which might have performance improvements.
 
-```pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu130```
+```pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu129```
 
 #### Troubleshooting
 
@@ -293,6 +264,12 @@ You can install ComfyUI in Apple Mac silicon (M1 or M2) with any recent macOS ve
 
 > **Note**: Remember to add your models, VAE, LoRAs etc. to the corresponding Comfy folders, as discussed in [ComfyUI manual installation](#manual-install-windows-linux).
 
+#### DirectML (AMD Cards on Windows)
+
+This is very badly supported and is not recommended. There are some unofficial builds of pytorch ROCm on windows that exist that will give you a much better experience than this. This readme will be updated once official pytorch ROCm builds for windows come out.
+
+```pip install torch-directml``` Then you can launch ComfyUI with: ```python main.py --directml```
+
 #### Ascend NPUs
 
 For models compatible with Ascend Extension for PyTorch (torch_npu). To get started, ensure your environment meets the prerequisites outlined on the [installation](https://ascend.github.io/docs/sources/ascend/quick_install.html) page. Here's a step-by-step guide tailored to your platform and installation method:

app/frontend_management.py

@@ -10,8 +10,7 @@ import importlib
 from dataclasses import dataclass
 from functools import cached_property
 from pathlib import Path
-from typing import Dict, TypedDict, Optional
-from aiohttp import web
+from typing import TypedDict, Optional
 from importlib.metadata import version
 
 import requests
@@ -43,7 +42,6 @@ def get_installed_frontend_version():
     frontend_version_str = version("comfyui-frontend-package")
     return frontend_version_str
 
-
 def get_required_frontend_version():
     """Get the required frontend version from requirements.txt."""
     try:
@@ -65,7 +63,6 @@ def get_required_frontend_version():
         logging.error(f"Error reading requirements.txt: {e}")
         return None
 
-
 def check_frontend_version():
     """Check if the frontend version is up to date."""
 
@@ -206,37 +203,6 @@ class FrontendManager:
         """Get the required frontend package version."""
         return get_required_frontend_version()
 
-    @classmethod
-    def get_installed_templates_version(cls) -> str:
-        """Get the currently installed workflow templates package version."""
-        try:
-            templates_version_str = version("comfyui-workflow-templates")
-            return templates_version_str
-        except Exception:
-            return None
-
-    @classmethod
-    def get_required_templates_version(cls) -> str:
-        """Get the required workflow templates version from requirements.txt."""
-        try:
-            with open(requirements_path, "r", encoding="utf-8") as f:
-                for line in f:
-                    line = line.strip()
-                    if line.startswith("comfyui-workflow-templates=="):
-                        version_str = line.split("==")[-1]
-                        if not is_valid_version(version_str):
-                            logging.error(f"Invalid templates version format in requirements.txt: {version_str}")
-                            return None
-                        return version_str
-                logging.error("comfyui-workflow-templates not found in requirements.txt")
-                return None
-        except FileNotFoundError:
-            logging.error("requirements.txt not found. Cannot determine required templates version.")
-            return None
-        except Exception as e:
-            logging.error(f"Error reading requirements.txt: {e}")
-            return None
-
     @classmethod
     def default_frontend_path(cls) -> str:
         try:
@@ -258,54 +224,7 @@ comfyui-frontend-package is not installed.
             sys.exit(-1)
 
     @classmethod
-    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."""
+    def templates_path(cls) -> str:
         try:
             import comfyui_workflow_templates
 
@@ -324,7 +243,6 @@ comfyui-workflow-templates is not installed.
 ********** ERROR ***********
 """.strip()
             )
-            return None
 
     @classmethod
     def embedded_docs_path(cls) -> str:
@@ -441,17 +359,3 @@ 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/subgraph_manager.py

@@ -1,112 +0,0 @@
-from __future__ import annotations
-
-from typing import TypedDict
-import os
-import folder_paths
-import glob
-from aiohttp import web
-import hashlib
-
-
-class Source:
-    custom_node = "custom_node"
-
-class SubgraphEntry(TypedDict):
-    source: str
-    """
-    Source of subgraph - custom_nodes vs templates.
-    """
-    path: str
-    """
-    Relative path of the subgraph file.
-    For custom nodes, will be the relative directory like <custom_node_dir>/subgraphs/<name>.json
-    """
-    name: str
-    """
-    Name of subgraph file.
-    """
-    info: CustomNodeSubgraphEntryInfo
-    """
-    Additional info about subgraph; in the case of custom_nodes, will contain nodepack name
-    """
-    data: str
-
-class CustomNodeSubgraphEntryInfo(TypedDict):
-    node_pack: str
-    """Node pack name."""
-
-class SubgraphManager:
-    def __init__(self):
-        self.cached_custom_node_subgraphs: dict[SubgraphEntry] | None = None
-
-    async def load_entry_data(self, entry: SubgraphEntry):
-        with open(entry['path'], 'r') as f:
-            entry['data'] = f.read()
-        return entry
-
-    async def sanitize_entry(self, entry: SubgraphEntry | None, remove_data=False) -> SubgraphEntry | None:
-        if entry is None:
-            return None
-        entry = entry.copy()
-        entry.pop('path', None)
-        if remove_data:
-            entry.pop('data', None)
-        return entry
-
-    async def sanitize_entries(self, entries: dict[str, SubgraphEntry], remove_data=False) -> dict[str, SubgraphEntry]:
-        entries = entries.copy()
-        for key in list(entries.keys()):
-            entries[key] = await self.sanitize_entry(entries[key], remove_data)
-        return entries
-
-    async def get_custom_node_subgraphs(self, loadedModules, force_reload=False):
-        # if not forced to reload and cached, return cache
-        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] = {}
-
-        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
-                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
-        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:
-            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
-            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)
-            return web.json_response(await self.sanitize_entry(subgraph))

comfy/cli_args.py

@@ -105,7 +105,6 @@ cache_group = parser.add_mutually_exclusive_group()
 cache_group.add_argument("--cache-classic", action="store_true", help="Use the old style (aggressive) caching.")
 cache_group.add_argument("--cache-lru", type=int, default=0, help="Use LRU caching with a maximum of N node results cached. May use more RAM/VRAM.")
 cache_group.add_argument("--cache-none", action="store_true", help="Reduced RAM/VRAM usage at the expense of executing every node for each run.")
-cache_group.add_argument("--cache-ram", nargs='?', const=4.0, type=float, default=0, help="Use RAM pressure caching with the specified headroom threshold. If available RAM drops below the threhold the cache remove large items to free RAM. Default 4GB")
 
 attn_group = parser.add_mutually_exclusive_group()
 attn_group.add_argument("--use-split-cross-attention", action="store_true", help="Use the split cross attention optimization. Ignored when xformers is used.")
@@ -146,9 +145,7 @@ class PerformanceFeature(enum.Enum):
     CublasOps = "cublas_ops"
     AutoTune = "autotune"
 
-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("--fast", nargs="*", type=PerformanceFeature, help="Enable some untested and potentially quality deteriorating optimizations. --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("--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.")
@@ -160,7 +157,7 @@ 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. Also prevents the frontend from communicating with the internet.")
+parser.add_argument("--disable-api-nodes", action="store_true", help="Disable loading all api nodes.")
 
 parser.add_argument("--multi-user", action="store_true", help="Enables per-user storage.")
 

comfy/controlnet.py

@@ -310,13 +310,11 @@ class ControlLoraOps:
             self.bias = None
 
         def forward(self, input):
-            weight, bias, offload_stream = comfy.ops.cast_bias_weight(self, input, offloadable=True)
+            weight, bias = comfy.ops.cast_bias_weight(self, input)
             if self.up is not None:
-                x = torch.nn.functional.linear(input, weight + (torch.mm(self.up.flatten(start_dim=1), self.down.flatten(start_dim=1))).reshape(self.weight.shape).type(input.dtype), bias)
+                return torch.nn.functional.linear(input, weight + (torch.mm(self.up.flatten(start_dim=1), self.down.flatten(start_dim=1))).reshape(self.weight.shape).type(input.dtype), bias)
             else:
-                x = torch.nn.functional.linear(input, weight, bias)
-            comfy.ops.uncast_bias_weight(self, weight, bias, offload_stream)
-            return x
+                return torch.nn.functional.linear(input, weight, bias)
 
     class Conv2d(torch.nn.Module, comfy.ops.CastWeightBiasOp):
         def __init__(
@@ -352,13 +350,12 @@ class ControlLoraOps:
 
 
         def forward(self, input):
-            weight, bias, offload_stream = comfy.ops.cast_bias_weight(self, input, offloadable=True)
+            weight, bias = comfy.ops.cast_bias_weight(self, input)
             if self.up is not None:
-                x = torch.nn.functional.conv2d(input, weight + (torch.mm(self.up.flatten(start_dim=1), self.down.flatten(start_dim=1))).reshape(self.weight.shape).type(input.dtype), bias, self.stride, self.padding, self.dilation, self.groups)
+                return torch.nn.functional.conv2d(input, weight + (torch.mm(self.up.flatten(start_dim=1), self.down.flatten(start_dim=1))).reshape(self.weight.shape).type(input.dtype), bias, self.stride, self.padding, self.dilation, self.groups)
             else:
-                x = torch.nn.functional.conv2d(input, weight, bias, self.stride, self.padding, self.dilation, self.groups)
-            comfy.ops.uncast_bias_weight(self, weight, bias, offload_stream)
-            return x
+                return torch.nn.functional.conv2d(input, weight, bias, self.stride, self.padding, self.dilation, self.groups)
+
 
 class ControlLora(ControlNet):
     def __init__(self, control_weights, global_average_pooling=False, model_options={}): #TODO? model_options

comfy/latent_formats.py

@@ -611,66 +611,6 @@ 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
-    scale_factor = 1.03682
-
 class Hunyuan3Dv2(LatentFormat):
     latent_channels = 64
     latent_dimensions = 1

comfy/ldm/ace/vae/music_dcae_pipeline.py

@@ -23,6 +23,8 @@ class MusicDCAE(torch.nn.Module):
         else:
             self.source_sample_rate = source_sample_rate
 
+        # self.resampler = torchaudio.transforms.Resample(source_sample_rate, 44100)
+
         self.transform = transforms.Compose([
             transforms.Normalize(0.5, 0.5),
         ])
@@ -35,6 +37,10 @@ class MusicDCAE(torch.nn.Module):
         self.scale_factor = 0.1786
         self.shift_factor = -1.9091
 
+    def load_audio(self, audio_path):
+        audio, sr = torchaudio.load(audio_path)
+        return audio, sr
+
     def forward_mel(self, audios):
         mels = []
         for i in range(len(audios)):
@@ -67,8 +73,10 @@ class MusicDCAE(torch.nn.Module):
             latent = self.dcae.encoder(mel.unsqueeze(0))
             latents.append(latent)
         latents = torch.cat(latents, dim=0)
+        # latent_lengths = (audio_lengths / sr * 44100 / 512 / self.time_dimention_multiple).long()
         latents = (latents - self.shift_factor) * self.scale_factor
         return latents
+        # return latents, latent_lengths
 
     @torch.no_grad()
     def decode(self, latents, audio_lengths=None, sr=None):
@@ -83,7 +91,9 @@ class MusicDCAE(torch.nn.Module):
             wav = self.vocoder.decode(mels[0]).squeeze(1)
 
             if sr is not None:
+                # resampler = torchaudio.transforms.Resample(44100, sr).to(latents.device).to(latents.dtype)
                 wav = torchaudio.functional.resample(wav, 44100, sr)
+                # wav = resampler(wav)
             else:
                 sr = 44100
             pred_wavs.append(wav)
@@ -91,6 +101,7 @@ class MusicDCAE(torch.nn.Module):
         if audio_lengths is not None:
             pred_wavs = [wav[:, :length].cpu() for wav, length in zip(pred_wavs, audio_lengths)]
         return torch.stack(pred_wavs)
+        # return sr, pred_wavs
 
     def forward(self, audios, audio_lengths=None, sr=None):
         latents, latent_lengths = self.encode(audios=audios, audio_lengths=audio_lengths, sr=sr)

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):
@@ -48,6 +48,124 @@ 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,
 )
@@ -90,7 +90,6 @@ 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)
@@ -99,7 +98,7 @@ class Chroma(nn.Module):
 
         self.single_blocks = nn.ModuleList(
             [
-                SingleStreamBlock(self.hidden_size, self.num_heads, mlp_ratio=params.mlp_ratio, modulation=False, dtype=dtype, device=device, operations=operations)
+                SingleStreamBlock(self.hidden_size, self.num_heads, mlp_ratio=params.mlp_ratio, dtype=dtype, device=device, operations=operations)
                 for _ in range(params.depth_single_blocks)
             ]
         )

comfy/ldm/chroma_radiance/model.py

@@ -10,10 +10,12 @@ from torch import Tensor, nn
 from einops import repeat
 import comfy.ldm.common_dit
 
-from comfy.ldm.flux.layers import EmbedND, DoubleStreamBlock, SingleStreamBlock
+from comfy.ldm.flux.layers import EmbedND
 
 from comfy.ldm.chroma.model import Chroma, ChromaParams
 from comfy.ldm.chroma.layers import (
+    DoubleStreamBlock,
+    SingleStreamBlock,
     Approximator,
 )
 from .layers import (
@@ -87,6 +89,7 @@ class ChromaRadiance(Chroma):
                     dtype=dtype, device=device, operations=operations
                 )
 
+
         self.double_blocks = nn.ModuleList(
             [
                 DoubleStreamBlock(
@@ -94,7 +97,6 @@ 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)
@@ -107,7 +109,6 @@ 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)
@@ -188,15 +189,15 @@ class ChromaRadiance(Chroma):
         nerf_pixels = nn.functional.unfold(img_orig, kernel_size=patch_size, stride=patch_size)
         nerf_pixels = nerf_pixels.transpose(1, 2) # -> [B, NumPatches, C * P * P]
 
-        # Reshape for per-patch processing
-        nerf_hidden = img_out.reshape(B * num_patches, params.hidden_size)
-        nerf_pixels = nerf_pixels.reshape(B * num_patches, C, patch_size**2).transpose(1, 2)
-
         if params.nerf_tile_size > 0 and num_patches > params.nerf_tile_size:
             # Enable tiling if nerf_tile_size isn't 0 and we actually have more patches than
             # the tile size.
-            img_dct = self.forward_tiled_nerf(nerf_hidden, nerf_pixels, B, C, num_patches, patch_size, params)
+            img_dct = self.forward_tiled_nerf(img_out, nerf_pixels, B, C, num_patches, patch_size, params)
         else:
+            # Reshape for per-patch processing
+            nerf_hidden = img_out.reshape(B * num_patches, params.hidden_size)
+            nerf_pixels = nerf_pixels.reshape(B * num_patches, C, patch_size**2).transpose(1, 2)
+
             # Get DCT-encoded pixel embeddings [pixel-dct]
             img_dct = self.nerf_image_embedder(nerf_pixels)
 
@@ -239,8 +240,17 @@ class ChromaRadiance(Chroma):
             end = min(i + tile_size, num_patches)
 
             # Slice the current tile from the input tensors
-            nerf_hidden_tile = nerf_hidden[i * batch:end * batch]
-            nerf_pixels_tile = nerf_pixels[i * batch:end * batch]
+            nerf_hidden_tile = nerf_hidden[:, i:end, :]
+            nerf_pixels_tile = nerf_pixels[:, i:end, :]
+
+            # Get the actual number of patches in this tile (can be smaller for the last tile)
+            num_patches_tile = nerf_hidden_tile.shape[1]
+
+            # Reshape the tile for per-patch processing
+            # [B, NumPatches_tile, D] -> [B * NumPatches_tile, D]
+            nerf_hidden_tile = nerf_hidden_tile.reshape(batch * num_patches_tile, params.hidden_size)
+            # [B, NumPatches_tile, C*P*P] -> [B*NumPatches_tile, C, P*P] -> [B*NumPatches_tile, P*P, C]
+            nerf_pixels_tile = nerf_pixels_tile.reshape(batch * num_patches_tile, channels, patch_size**2).transpose(1, 2)
 
             # get DCT-encoded pixel embeddings [pixel-dct]
             img_dct_tile = self.nerf_image_embedder(nerf_pixels_tile)

comfy/ldm/flux/layers.py

@@ -130,17 +130,13 @@ def apply_mod(tensor, m_mult, m_add=None, modulation_dims=None):
 
 
 class DoubleStreamBlock(nn.Module):
-    def __init__(self, hidden_size: int, num_heads: int, mlp_ratio: float, qkv_bias: bool = False, flipped_img_txt=False, modulation=True, 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, 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.modulation = modulation
-
-        if self.modulation:
-            self.img_mod = Modulation(hidden_size, double=True, dtype=dtype, device=device, operations=operations)
-
+        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)
 
@@ -151,9 +147,7 @@ class DoubleStreamBlock(nn.Module):
             operations.Linear(mlp_hidden_dim, hidden_size, bias=True, dtype=dtype, device=device),
         )
 
-        if self.modulation:
-            self.txt_mod = Modulation(hidden_size, double=True, dtype=dtype, device=device, operations=operations)
-
+        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)
 
@@ -166,65 +160,46 @@ class DoubleStreamBlock(nn.Module):
         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={}):
-        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
+        img_mod1, img_mod2 = self.img_mod(vec)
+        txt_mod1, txt_mod2 = self.txt_mod(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(q, k, v,
+            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),
                              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(q, k, v,
+            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)
-            del q, k, v
 
             txt_attn, img_attn = attn[:, : txt.shape[1]], attn[:, txt.shape[1]:]
 
         # calculate the img bloks
-        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)
+        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)
 
         # 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:
@@ -245,7 +220,6 @@ class SingleStreamBlock(nn.Module):
         num_heads: int,
         mlp_ratio: float = 4.0,
         qk_scale: float = None,
-        modulation=True,
         dtype=None,
         device=None,
         operations=None
@@ -268,29 +242,19 @@ class SingleStreamBlock(nn.Module):
         self.pre_norm = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
 
         self.mlp_act = nn.GELU(approximate="tanh")
-        if modulation:
-            self.modulation = Modulation(hidden_size, double=False, dtype=dtype, device=device, operations=operations)
-        else:
-            self.modulation = None
+        self.modulation = Modulation(hidden_size, double=False, dtype=dtype, device=device, operations=operations)
 
     def forward(self, x: Tensor, vec: Tensor, pe: Tensor, attn_mask=None, modulation_dims=None, transformer_options={}) -> Tensor:
-        if self.modulation:
-            mod, _ = self.modulation(vec)
-        else:
-            mod = vec
-
+        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)
 
         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
-        mlp = self.mlp_act(mlp)
-        output = self.linear2(torch.cat((attn, mlp), 2))
+        output = self.linear2(torch.cat((attn, self.mlp_act(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)

comfy/ldm/flux/math.py

@@ -7,8 +7,15 @@ import comfy.model_management
 
 
 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, k = apply_rope(q, k, pe)
+        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)
+
     heads = q.shape[1]
     x = optimized_attention(q, k, v, heads, skip_reshape=True, mask=mask, transformer_options=transformer_options)
     return x
@@ -30,10 +37,7 @@ def rope(pos: Tensor, dim: int, theta: int) -> 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]
-    x_out.addcmul_(freqs_cis[..., 1], x_[..., 1])
-
+    x_out = freqs_cis[..., 0] * x_[..., 0] + freqs_cis[..., 1] * x_[..., 1]
     return x_out.reshape(*x.shape).type_as(x)
 
 def apply_rope(xq: Tensor, xk: Tensor, freqs_cis: Tensor):

comfy/ldm/flux/model.py

@@ -210,7 +210,7 @@ class Flux(nn.Module):
         img = self.final_layer(img, vec)  # (N, T, patch_size ** 2 * out_channels)
         return img
 
-    def process_img(self, x, index=0, h_offset=0, w_offset=0, transformer_options={}):
+    def process_img(self, x, index=0, h_offset=0, w_offset=0):
         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))
@@ -222,22 +222,10 @@ class Flux(nn.Module):
         h_offset = ((h_offset + (patch_size // 2)) // patch_size)
         w_offset = ((w_offset + (patch_size // 2)) // patch_size)
 
-        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, 3), device=x.device, dtype=x.dtype)
+        img_ids = torch.zeros((h_len, w_len, 3), device=x.device, dtype=x.dtype)
         img_ids[:, :, 0] = img_ids[:, :, 1] + index
-        img_ids[:, :, 1] = img_ids[:, :, 1] + torch.linspace(h_offset, h_len - 1 + h_offset, steps=steps_h, device=x.device, dtype=x.dtype).unsqueeze(1)
-        img_ids[:, :, 2] = img_ids[:, :, 2] + torch.linspace(w_offset, w_len - 1 + w_offset, steps=steps_w, device=x.device, dtype=x.dtype).unsqueeze(0)
+        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)
         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):
@@ -253,7 +241,7 @@ 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, transformer_options=transformer_options)
+        img, img_ids = self.process_img(x)
         img_tokens = img.shape[1]
         if ref_latents is not None:
             h = 0

comfy/ldm/hunyuan_video/model.py

@@ -6,6 +6,7 @@ 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
 
@@ -41,8 +42,6 @@ class HunyuanVideoParams:
     guidance_embed: bool
     byt5: bool
     meanflow: bool
-    use_cond_type_embedding: bool
-    vision_in_dim: int
 
 
 class SelfAttentionRef(nn.Module):
@@ -158,10 +157,7 @@ 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
-        if x.dtype == torch.float16:
-            c = x.float().sum(dim=1) / x.shape[1]
-        else:
-            c = x.sum(dim=1) / x.shape[1]
+        c = x.sum(dim=1) / x.shape[1]
 
         c = t + self.c_embedder(c.to(x.dtype))
         x = self.input_embedder(x)
@@ -200,15 +196,11 @@ 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}"
@@ -274,18 +266,6 @@ 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,
@@ -296,7 +276,6 @@ 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,
@@ -352,31 +331,12 @@ 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)
 
@@ -470,14 +430,14 @@ 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, clip_fea=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, 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, clip_fea, guidance, attention_mask, guiding_frame_index, ref_latent, disable_time_r, control, transformer_options, **kwargs)
+        ).execute(x, timestep, context, y, txt_byt5, 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, clip_fea=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, 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)
@@ -485,5 +445,5 @@ 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, clip_fea, 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, 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

@@ -1,120 +0,0 @@
-import torch
-import torch.nn as nn
-import torch.nn.functional as F
-from comfy.ldm.hunyuan_video.vae_refiner import RMS_norm, ResnetBlock, VideoConv3d
-import model_management, 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 = model_management.vae_device()
-        offload_device = model_management.vae_offload_device()
-        self.dtype = model_management.vae_dtype(self.load_device)
-        self.model_class = UPSAMPLERS.get(model_type)
-        self.model = self.model_class(**config).eval()
-
-        self.patcher = model_patcher.ModelPatcher(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)
-
-    def get_sd(self):
-        return self.model.state_dict()
-
-    def resample_latent(self, latent):
-        model_management.load_model_gpu(self.patcher)
-        return self.model(latent.to(self.load_device))

comfy/ldm/hunyuan_video/vae_refiner.py

@@ -1,43 +1,11 @@
 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, VideoConv3d
 import comfy.ops
 import comfy.ldm.models.autoencoder
-import comfy.model_management
 ops = comfy.ops.disable_weight_init
 
-class NoPadConv3d(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 conv_carry_out is not None:
-        to_push = x[:, :, -2:, :, :].clone()
-        conv_carry_out.append(to_push)
-
-    if isinstance(op, NoPadConv3d):
-        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.cat([conv_carry_in.pop(0), x], dim=2)
-            x = torch.nn.functional.pad(x, (1, 1, 1, 1, 2 - carry_len, 0), mode = 'replicate')
-
-    out = op(x)
-
-    return out
-
-
 class RMS_norm(nn.Module):
     def __init__(self, dim):
         super().__init__()
@@ -46,25 +14,23 @@ class RMS_norm(nn.Module):
         self.gamma = nn.Parameter(torch.empty(shape))
 
     def forward(self, x):
-        return F.normalize(x, dim=1) * self.scale * comfy.model_management.cast_to(self.gamma, dtype=x.dtype, device=x.device)
+        return F.normalize(x, dim=1) * self.scale * self.gamma
 
 class DnSmpl(nn.Module):
-    def __init__(self, ic, oc, tds=True, refiner_vae=True, op=VideoConv3d):
+    def __init__(self, ic, oc, tds=True):
         super().__init__()
         fct = 2 * 2 * 2 if tds else 1 * 2 * 2
         assert oc % fct == 0
-        self.conv = op(ic, oc // fct, kernel_size=3, stride=1, padding=1)
-        self.refiner_vae = refiner_vae
+        self.conv = VideoConv3d(ic, oc // fct, kernel_size=3)
 
         self.tds = tds
         self.gs = fct * ic // oc
 
-    def forward(self, x, conv_carry_in=None, conv_carry_out=None):
+    def forward(self, x):
         r1 = 2 if self.tds else 1
-        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:
+        h = self.conv(x)
 
+        if self.tds:
             hf = h[:, :, :1, :, :]
             b, c, f, ht, wd = hf.shape
             hf = hf.reshape(b, c, f, ht // 2, 2, wd // 2, 2)
@@ -72,7 +38,14 @@ class DnSmpl(nn.Module):
             hf = hf.reshape(b, 2 * 2 * c, f, ht // 2, wd // 2)
             hf = torch.cat([hf, hf], dim=1)
 
-            h = h[:, :, 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)
 
             xf = x[:, :, :1, :, :]
             b, ci, f, ht, wd = xf.shape
@@ -80,49 +53,49 @@ 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, hf.shape[1], self.gs // 2, T, H, W).mean(dim=2)
+            xf = xf.view(B, h.shape[1], self.gs // 2, T, H, W).mean(dim=2)
 
-            x = x[:, :, 1:, :, :]
+            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
+            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)
-
-        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
+        return h + sc
 
 
 class UpSmpl(nn.Module):
-    def __init__(self, ic, oc, tus=True, refiner_vae=True, op=VideoConv3d):
+    def __init__(self, ic, oc, tus=True):
         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)
-        self.refiner_vae = refiner_vae
+        self.conv = VideoConv3d(ic, oc * fct, kernel_size=3)
 
         self.tus = tus
         self.rp = fct * oc // ic
 
-    def forward(self, x, conv_carry_in=None, conv_carry_out=None):
+    def forward(self, x):
         r1 = 2 if self.tus else 1
-        h = conv_carry_causal_3d([x], self.conv, conv_carry_in, conv_carry_out)
+        h = self.conv(x)
 
-        if self.tus and self.refiner_vae and conv_carry_in is None:
+        if self.tus:
             hf = h[:, :, :1, :, :]
             b, c, f, ht, wd = hf.shape
             nc = c // (2 * 2)
@@ -131,7 +104,14 @@ class UpSmpl(nn.Module):
             hf = hf.reshape(b, nc, f, ht * 2, wd * 2)
             hf = hf[:, : hf.shape[1] // 2]
 
-            h = h[:, :, 1:, :, :]
+            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)
 
             xf = x[:, :, :1, :, :]
             b, ci, f, ht, wd = xf.shape
@@ -142,165 +122,109 @@ class UpSmpl(nn.Module):
             xf = xf.permute(0, 3, 4, 5, 1, 6, 2)
             xf = xf.reshape(b, nc, f, ht * 2, wd * 2)
 
-            x = x[:, :, 1:, :, :]
+            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)
 
-        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)
+            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)
 
-        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 HunyuanRefinerResnetBlock(ResnetBlock):
-    def __init__(self, in_channels, out_channels, conv_op=NoPadConv3d, norm_op=RMS_norm):
-        super().__init__(in_channels=in_channels, out_channels=out_channels, temb_channels=0, conv_op=conv_op, norm_op=norm_op)
-
-    def forward(self, x, conv_carry_in=None, conv_carry_out=None):
-        h = x
-        h = [ self.swish(self.norm1(x)) ]
-        h = conv_carry_causal_3d(h, self.conv1, conv_carry_in=conv_carry_in, conv_carry_out=conv_carry_out)
-
-        h = [ self.dropout(self.swish(self.norm2(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:
-            x = self.nin_shortcut(x)
-
-        return x+h
+        return h + sc
 
 class Encoder(nn.Module):
     def __init__(self, in_channels, z_channels, block_out_channels, num_res_blocks,
-                 ffactor_spatial, ffactor_temporal, downsample_match_channel=True, refiner_vae=True, **_):
+                 ffactor_spatial, ffactor_temporal, downsample_match_channel=True, **_):
         super().__init__()
         self.z_channels = z_channels
         self.block_out_channels = block_out_channels
         self.num_res_blocks = num_res_blocks
-        self.ffactor_temporal = ffactor_temporal
-
-        self.refiner_vae = refiner_vae
-        if self.refiner_vae:
-            conv_op = NoPadConv3d
-            norm_op = RMS_norm
-        else:
-            conv_op = ops.Conv3d
-            norm_op = Normalize
-
-        self.conv_in = conv_op(in_channels, block_out_channels[0], 3, 1, 1)
+        self.conv_in = VideoConv3d(in_channels, block_out_channels[0], 3, 1, 1)
 
         self.down = nn.ModuleList()
         ch = block_out_channels[0]
         depth = (ffactor_spatial >> 1).bit_length()
-        depth_temporal = ((ffactor_spatial // self.ffactor_temporal) >> 1).bit_length()
+        depth_temporal = ((ffactor_spatial // ffactor_temporal) >> 1).bit_length()
 
         for i, tgt in enumerate(block_out_channels):
             stage = nn.Module()
-            stage.block = nn.ModuleList([HunyuanRefinerResnetBlock(in_channels=ch if j == 0 else tgt,
-                                                                   out_channels=tgt,
-                                                                   conv_op=conv_op, norm_op=norm_op)
+            stage.block = nn.ModuleList([ResnetBlock(in_channels=ch if j == 0 else tgt,
+                                                     out_channels=tgt,
+                                                     temb_channels=0,
+                                                     conv_op=VideoConv3d, norm_op=RMS_norm)
                                         for j in range(num_res_blocks)])
             ch = tgt
             if i < depth:
                 nxt = block_out_channels[i + 1] if i + 1 < len(block_out_channels) and downsample_match_channel else ch
-                stage.downsample = DnSmpl(ch, nxt, tds=i >= depth_temporal, refiner_vae=self.refiner_vae, op=conv_op)
+                stage.downsample = DnSmpl(ch, nxt, tds=i >= depth_temporal)
                 ch = nxt
             self.down.append(stage)
 
         self.mid = nn.Module()
-        self.mid.block_1 = HunyuanRefinerResnetBlock(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 = HunyuanRefinerResnetBlock(in_channels=ch, out_channels=ch, conv_op=conv_op, norm_op=norm_op)
+        self.mid.block_1 = ResnetBlock(in_channels=ch, out_channels=ch, temb_channels=0, conv_op=VideoConv3d, norm_op=RMS_norm)
+        self.mid.attn_1 = AttnBlock(ch, conv_op=ops.Conv3d, norm_op=RMS_norm)
+        self.mid.block_2 = ResnetBlock(in_channels=ch, out_channels=ch, temb_channels=0, conv_op=VideoConv3d, norm_op=RMS_norm)
 
-        self.norm_out = norm_op(ch)
-        self.conv_out = conv_op(ch, z_channels << 1, 3, 1, 1)
+        self.norm_out = RMS_norm(ch)
+        self.conv_out = VideoConv3d(ch, z_channels << 1, 3, 1, 1)
 
         self.regul = comfy.ldm.models.autoencoder.DiagonalGaussianRegularizer()
 
     def forward(self, x):
-        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
-
-        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, 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
-
-        if len(out) > 1:
-            out = torch.cat(out, dim=2)
-        else:
-            out = out[0]
-
-        x = self.mid.block_2(self.mid.attn_1(self.mid.block_1(out)))
-        del out
+        x = self.mid.block_2(self.mid.attn_1(self.mid.block_1(x)))
 
         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 = conv_carry_causal_3d([F.silu(self.norm_out(x))], self.conv_out) + skip
-
-        if self.refiner_vae:
-            out = self.regul(out)[0]
+        out = self.conv_out(F.silu(self.norm_out(x))) + skip
+        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):
     def __init__(self, z_channels, out_channels, block_out_channels, num_res_blocks,
-                 ffactor_spatial, ffactor_temporal, upsample_match_channel=True, refiner_vae=True, **_):
+                 ffactor_spatial, ffactor_temporal, upsample_match_channel=True, **_):
         super().__init__()
         block_out_channels = block_out_channels[::-1]
         self.z_channels = z_channels
         self.block_out_channels = block_out_channels
         self.num_res_blocks = num_res_blocks
 
-        self.refiner_vae = refiner_vae
-        if self.refiner_vae:
-            conv_op = NoPadConv3d
-            norm_op = RMS_norm
-        else:
-            conv_op = ops.Conv3d
-            norm_op = Normalize
-
         ch = block_out_channels[0]
-        self.conv_in = conv_op(z_channels, ch, kernel_size=3, stride=1, padding=1)
+        self.conv_in = VideoConv3d(z_channels, ch, 3)
 
         self.mid = nn.Module()
-        self.mid.block_1 = HunyuanRefinerResnetBlock(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 = HunyuanRefinerResnetBlock(in_channels=ch, out_channels=ch,  conv_op=conv_op, norm_op=norm_op)
+        self.mid.block_1 = ResnetBlock(in_channels=ch, out_channels=ch, temb_channels=0, conv_op=VideoConv3d, norm_op=RMS_norm)
+        self.mid.attn_1 = AttnBlock(ch, conv_op=ops.Conv3d, norm_op=RMS_norm)
+        self.mid.block_2 = ResnetBlock(in_channels=ch, out_channels=ch, temb_channels=0, conv_op=VideoConv3d, norm_op=RMS_norm)
 
         self.up = nn.ModuleList()
         depth = (ffactor_spatial >> 1).bit_length()
@@ -308,56 +232,36 @@ class Decoder(nn.Module):
 
         for i, tgt in enumerate(block_out_channels):
             stage = nn.Module()
-            stage.block = nn.ModuleList([HunyuanRefinerResnetBlock(in_channels=ch if j == 0 else tgt,
-                                                                   out_channels=tgt,
-                                                                   conv_op=conv_op, norm_op=norm_op)
+            stage.block = nn.ModuleList([ResnetBlock(in_channels=ch if j == 0 else tgt,
+                                                     out_channels=tgt,
+                                                     temb_channels=0,
+                                                     conv_op=VideoConv3d, norm_op=RMS_norm)
                                         for j in range(num_res_blocks + 1)])
             ch = tgt
             if i < depth:
                 nxt = block_out_channels[i + 1] if i + 1 < len(block_out_channels) and upsample_match_channel else ch
-                stage.upsample = UpSmpl(ch, nxt, tus=i < depth_temporal, refiner_vae=self.refiner_vae, op=conv_op)
+                stage.upsample = UpSmpl(ch, nxt, tus=i < depth_temporal)
                 ch = nxt
             self.up.append(stage)
 
-        self.norm_out = norm_op(ch)
-        self.conv_out = conv_op(ch, out_channels, 3, stride=1, padding=1)
+        self.norm_out = RMS_norm(ch)
+        self.conv_out = VideoConv3d(ch, out_channels, 3)
 
     def forward(self, z):
-        x = conv_carry_causal_3d([z], self.conv_in) + z.repeat_interleave(self.block_out_channels[0] // self.z_channels, 1)
+        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 = self.mid.block_2(self.mid.attn_1(self.mid.block_1(x)))
 
-        if self.refiner_vae:
-            x = torch.split(x, 2, dim=2)
-        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
-            for stage in self.up:
-                for blk in stage.block:
-                    x1 = blk(x1, 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
-
-        if len(out) > 1:
-            out = torch.cat(out, dim=2)
-        else:
-            out = out[0]
-
-        if not self.refiner_vae:
-            if z.shape[-3] == 1:
-                out = out[:, :, -1:]
-
-        return out
+        for stage in self.up:
+            for blk in stage.block:
+                x = blk(x)
+            if hasattr(stage, 'upsample'):
+                x = stage.upsample(x)
 
+        return self.conv_out(F.silu(self.norm_out(x)))

comfy/ldm/lightricks/model.py

@@ -3,11 +3,12 @@ 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
-from comfy.ldm.flux.math import apply_rope1
+
 
 def get_timestep_embedding(
     timesteps: torch.Tensor,
@@ -237,6 +238,20 @@ class FeedForward(nn.Module):
         return self.net(x)
 
 
+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]
+
+    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)
+    input_tensor_rot = rearrange(t_dup, "... d r -> ... (d r)")
+
+    out = input_tensor * cos_freqs + input_tensor_rot * sin_freqs
+
+    return out
+
+
 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):
         super().__init__()
@@ -266,8 +281,8 @@ class CrossAttention(nn.Module):
         k = self.k_norm(k)
 
         if pe is not None:
-            q = apply_rope1(q.unsqueeze(1), pe).squeeze(1)
-            k = apply_rope1(k.unsqueeze(1), pe).squeeze(1)
+            q = apply_rotary_emb(q, pe)
+            k = apply_rotary_emb(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)
@@ -291,17 +306,12 @@ class BasicTransformerBlock(nn.Module):
     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)
 
-        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.attn1(comfy.ldm.common_dit.rms_norm(x) * (1 + scale_msa) + shift_msa, pe=pe, transformer_options=transformer_options) * gate_msa
 
         x += self.attn2(x, context=context, mask=attention_mask, transformer_options=transformer_options)
 
-        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)
+        y = comfy.ldm.common_dit.rms_norm(x) * (1 + scale_mlp) + shift_mlp
+        x += self.ff(y) * gate_mlp
 
         return x
 
@@ -317,35 +327,41 @@ def get_fractional_positions(indices_grid, max_pos):
 
 
 def precompute_freqs_cis(indices_grid, dim, out_dtype, theta=10000.0, max_pos=[20, 2048, 2048]):
-    dtype = torch.float32
-    device = indices_grid.device
+    dtype = torch.float32 #self.dtype
 
-    # Get fractional positions and compute frequency indices
     fractional_positions = get_fractional_positions(indices_grid, max_pos)
-    indices = theta ** torch.linspace(0, 1, dim // 6, device=device, dtype=dtype) * math.pi / 2
 
-    # Compute frequencies and apply cos/sin
-    freqs = (indices * (fractional_positions.unsqueeze(-1) * 2 - 1)).transpose(-1, -2).flatten(2)
-    cos_vals = freqs.cos().repeat_interleave(2, dim=-1)
-    sin_vals = freqs.sin().repeat_interleave(2, dim=-1)
+    start = 1
+    end = theta
+    device = fractional_positions.device
 
-    # Pad if dim is not divisible by 6
+    indices = theta ** (
+        torch.linspace(
+            math.log(start, theta),
+            math.log(end, theta),
+            dim // 6,
+            device=device,
+            dtype=dtype,
+        )
+    )
+    indices = indices.to(dtype=dtype)
+
+    indices = indices * math.pi / 2
+
+    freqs = (
+        (indices * (fractional_positions.unsqueeze(-1) * 2 - 1))
+        .transpose(-1, -2)
+        .flatten(2)
+    )
+
+    cos_freq = freqs.cos().repeat_interleave(2, dim=-1)
+    sin_freq = freqs.sin().repeat_interleave(2, dim=-1)
     if dim % 6 != 0:
-        padding_size = dim % 6
-        cos_vals = torch.cat([torch.ones_like(cos_vals[:, :, :padding_size]), cos_vals], dim=-1)
-        sin_vals = torch.cat([torch.zeros_like(sin_vals[:, :, :padding_size]), sin_vals], dim=-1)
-
-    # Reshape and extract one value per pair (since repeat_interleave duplicates each value)
-    cos_vals = cos_vals.reshape(*cos_vals.shape[:2], -1, 2)[..., 0].to(out_dtype)  # [B, N, dim//2]
-    sin_vals = sin_vals.reshape(*sin_vals.shape[:2], -1, 2)[..., 0].to(out_dtype)  # [B, N, dim//2]
-
-    # Build rotation matrix [[cos, -sin], [sin, cos]] and add heads dimension
-    freqs_cis = torch.stack([
-        torch.stack([cos_vals, -sin_vals], dim=-1),
-        torch.stack([sin_vals, cos_vals], dim=-1)
-    ], dim=-2).unsqueeze(1)  # [B, 1, N, dim//2, 2, 2]
-
-    return freqs_cis
+        cos_padding = torch.ones_like(cos_freq[:, :, : dim % 6])
+        sin_padding = torch.zeros_like(cos_freq[:, :, : dim % 6])
+        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)
 
 
 class LTXVModel(torch.nn.Module):
@@ -485,7 +501,7 @@ class LTXVModel(torch.nn.Module):
         shift, scale = scale_shift_values[:, :, 0], scale_shift_values[:, :, 1]
         x = self.norm_out(x)
         # Modulation
-        x = torch.addcmul(x, x, scale).add_(shift)
+        x = x * (1 + scale) + shift
         x = self.proj_out(x)
 
         x = self.patchifier.unpatchify(

comfy/ldm/lumina/model.py

@@ -522,7 +522,7 @@ class NextDiT(nn.Module):
         max_cap_len = max(l_effective_cap_len)
         max_img_len = max(l_effective_img_len)
 
-        position_ids = torch.zeros(bsz, max_seq_len, 3, dtype=torch.float32, device=device)
+        position_ids = torch.zeros(bsz, max_seq_len, 3, dtype=torch.int32, device=device)
 
         for i in range(bsz):
             cap_len = l_effective_cap_len[i]
@@ -531,22 +531,10 @@ class NextDiT(nn.Module):
             H_tokens, W_tokens = H // pH, W // pW
             assert H_tokens * W_tokens == img_len
 
-            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)
-
-            position_ids[i, :cap_len, 0] = torch.arange(cap_len, dtype=torch.float32, device=device)
+            position_ids[i, :cap_len, 0] = torch.arange(cap_len, dtype=torch.int32, 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()
+            row_ids = torch.arange(H_tokens, dtype=torch.int32, device=device).view(-1, 1).repeat(1, W_tokens).flatten()
+            col_ids = torch.arange(W_tokens, dtype=torch.int32, device=device).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
 

comfy/ldm/mmaudio/vae/activations.py

@@ -1,120 +0,0 @@
-# Implementation adapted from https://github.com/EdwardDixon/snake under the MIT license.
-#   LICENSE is in incl_licenses directory.
-
-import torch
-from torch import nn, sin, pow
-from torch.nn import Parameter
-import comfy.model_management
-
-class Snake(nn.Module):
-    '''
-    Implementation of a sine-based periodic activation function
-    Shape:
-        - Input: (B, C, T)
-        - Output: (B, C, T), same shape as the input
-    Parameters:
-        - alpha - trainable parameter
-    References:
-        - This activation function is from this paper by Liu Ziyin, Tilman Hartwig, Masahito Ueda:
-        https://arxiv.org/abs/2006.08195
-    Examples:
-        >>> a1 = snake(256)
-        >>> x = torch.randn(256)
-        >>> x = a1(x)
-    '''
-    def __init__(self, in_features, alpha=1.0, alpha_trainable=True, alpha_logscale=False):
-        '''
-        Initialization.
-        INPUT:
-            - in_features: shape of the input
-            - alpha: trainable parameter
-            alpha is initialized to 1 by default, higher values = higher-frequency.
-            alpha will be trained along with the rest of your model.
-        '''
-        super(Snake, self).__init__()
-        self.in_features = in_features
-
-        # initialize alpha
-        self.alpha_logscale = alpha_logscale
-        if self.alpha_logscale:
-            self.alpha = Parameter(torch.empty(in_features))
-        else:
-            self.alpha = Parameter(torch.empty(in_features))
-
-        self.alpha.requires_grad = alpha_trainable
-
-        self.no_div_by_zero = 0.000000001
-
-    def forward(self, x):
-        '''
-        Forward pass of the function.
-        Applies the function to the input elementwise.
-        Snake ∶= x + 1/a * sin^2 (xa)
-        '''
-        alpha = comfy.model_management.cast_to(self.alpha, dtype=x.dtype, device=x.device).unsqueeze(0).unsqueeze(-1) # line up with x to [B, C, T]
-        if self.alpha_logscale:
-            alpha = torch.exp(alpha)
-        x = x + (1.0 / (alpha + self.no_div_by_zero)) * pow(sin(x * alpha), 2)
-
-        return x
-
-
-class SnakeBeta(nn.Module):
-    '''
-    A modified Snake function which uses separate parameters for the magnitude of the periodic components
-    Shape:
-        - Input: (B, C, T)
-        - Output: (B, C, T), same shape as the input
-    Parameters:
-        - alpha - trainable parameter that controls frequency
-        - beta - trainable parameter that controls magnitude
-    References:
-        - This activation function is a modified version based on this paper by Liu Ziyin, Tilman Hartwig, Masahito Ueda:
-        https://arxiv.org/abs/2006.08195
-    Examples:
-        >>> a1 = snakebeta(256)
-        >>> x = torch.randn(256)
-        >>> x = a1(x)
-    '''
-    def __init__(self, in_features, alpha=1.0, alpha_trainable=True, alpha_logscale=False):
-        '''
-        Initialization.
-        INPUT:
-            - in_features: shape of the input
-            - alpha - trainable parameter that controls frequency
-            - beta - trainable parameter that controls magnitude
-            alpha is initialized to 1 by default, higher values = higher-frequency.
-            beta is initialized to 1 by default, higher values = higher-magnitude.
-            alpha will be trained along with the rest of your model.
-        '''
-        super(SnakeBeta, self).__init__()
-        self.in_features = in_features
-
-        # initialize alpha
-        self.alpha_logscale = alpha_logscale
-        if self.alpha_logscale:
-            self.alpha = Parameter(torch.empty(in_features))
-            self.beta = Parameter(torch.empty(in_features))
-        else:
-            self.alpha = Parameter(torch.empty(in_features))
-            self.beta = Parameter(torch.empty(in_features))
-
-        self.alpha.requires_grad = alpha_trainable
-        self.beta.requires_grad = alpha_trainable
-
-        self.no_div_by_zero = 0.000000001
-
-    def forward(self, x):
-        '''
-        Forward pass of the function.
-        Applies the function to the input elementwise.
-        SnakeBeta ∶= x + 1/b * sin^2 (xa)
-        '''
-        alpha = comfy.model_management.cast_to(self.alpha, dtype=x.dtype, device=x.device).unsqueeze(0).unsqueeze(-1) # line up with x to [B, C, T]
-        beta = comfy.model_management.cast_to(self.beta, dtype=x.dtype, device=x.device).unsqueeze(0).unsqueeze(-1)
-        if self.alpha_logscale:
-            alpha = torch.exp(alpha)
-            beta = torch.exp(beta)
-        x = x + (1.0 / (beta + self.no_div_by_zero)) * pow(sin(x * alpha), 2)
-
-        return x

comfy/ldm/mmaudio/vae/alias_free_torch.py

@@ -1,157 +0,0 @@
-import torch
-import torch.nn as nn
-import torch.nn.functional as F
-import math
-import comfy.model_management
-
-if 'sinc' in dir(torch):
-    sinc = torch.sinc
-else:
-    # This code is adopted from adefossez's julius.core.sinc under the MIT License
-    # https://adefossez.github.io/julius/julius/core.html
-    #   LICENSE is in incl_licenses directory.
-    def sinc(x: torch.Tensor):
-        """
-        Implementation of sinc, i.e. sin(pi * x) / (pi * x)
-        __Warning__: Different to julius.sinc, the input is multiplied by `pi`!
-        """
-        return torch.where(x == 0,
-                           torch.tensor(1., device=x.device, dtype=x.dtype),
-                           torch.sin(math.pi * x) / math.pi / x)
-
-
-# This code is adopted from adefossez's julius.lowpass.LowPassFilters under the MIT License
-# https://adefossez.github.io/julius/julius/lowpass.html
-#   LICENSE is in incl_licenses directory.
-def kaiser_sinc_filter1d(cutoff, half_width, kernel_size): # return filter [1,1,kernel_size]
-    even = (kernel_size % 2 == 0)
-    half_size = kernel_size // 2
-
-    #For kaiser window
-    delta_f = 4 * half_width
-    A = 2.285 * (half_size - 1) * math.pi * delta_f + 7.95
-    if A > 50.:
-        beta = 0.1102 * (A - 8.7)
-    elif A >= 21.:
-        beta = 0.5842 * (A - 21)**0.4 + 0.07886 * (A - 21.)
-    else:
-        beta = 0.
-    window = torch.kaiser_window(kernel_size, beta=beta, periodic=False)
-
-    # ratio = 0.5/cutoff -> 2 * cutoff = 1 / ratio
-    if even:
-        time = (torch.arange(-half_size, half_size) + 0.5)
-    else:
-        time = torch.arange(kernel_size) - half_size
-    if cutoff == 0:
-        filter_ = torch.zeros_like(time)
-    else:
-        filter_ = 2 * cutoff * window * sinc(2 * cutoff * time)
-        # Normalize filter to have sum = 1, otherwise we will have a small leakage
-        # of the constant component in the input signal.
-        filter_ /= filter_.sum()
-        filter = filter_.view(1, 1, kernel_size)
-
-    return filter
-
-
-class LowPassFilter1d(nn.Module):
-    def __init__(self,
-                 cutoff=0.5,
-                 half_width=0.6,
-                 stride: int = 1,
-                 padding: bool = True,
-                 padding_mode: str = 'replicate',
-                 kernel_size: int = 12):
-        # kernel_size should be even number for stylegan3 setup,
-        # in this implementation, odd number is also possible.
-        super().__init__()
-        if cutoff < -0.:
-            raise ValueError("Minimum cutoff must be larger than zero.")
-        if cutoff > 0.5:
-            raise ValueError("A cutoff above 0.5 does not make sense.")
-        self.kernel_size = kernel_size
-        self.even = (kernel_size % 2 == 0)
-        self.pad_left = kernel_size // 2 - int(self.even)
-        self.pad_right = kernel_size // 2
-        self.stride = stride
-        self.padding = padding
-        self.padding_mode = padding_mode
-        filter = kaiser_sinc_filter1d(cutoff, half_width, kernel_size)
-        self.register_buffer("filter", filter)
-
-    #input [B, C, T]
-    def forward(self, x):
-        _, C, _ = x.shape
-
-        if self.padding:
-            x = F.pad(x, (self.pad_left, self.pad_right),
-                      mode=self.padding_mode)
-        out = F.conv1d(x, comfy.model_management.cast_to(self.filter.expand(C, -1, -1), dtype=x.dtype, device=x.device),
-                       stride=self.stride, groups=C)
-
-        return out
-
-
-class UpSample1d(nn.Module):
-    def __init__(self, ratio=2, kernel_size=None):
-        super().__init__()
-        self.ratio = ratio
-        self.kernel_size = int(6 * ratio // 2) * 2 if kernel_size is None else kernel_size
-        self.stride = ratio
-        self.pad = self.kernel_size // ratio - 1
-        self.pad_left = self.pad * self.stride + (self.kernel_size - self.stride) // 2
-        self.pad_right = self.pad * self.stride + (self.kernel_size - self.stride + 1) // 2
-        filter = kaiser_sinc_filter1d(cutoff=0.5 / ratio,
-                                      half_width=0.6 / ratio,
-                                      kernel_size=self.kernel_size)
-        self.register_buffer("filter", filter)
-
-    # x: [B, C, T]
-    def forward(self, x):
-        _, C, _ = x.shape
-
-        x = F.pad(x, (self.pad, self.pad), mode='replicate')
-        x = self.ratio * F.conv_transpose1d(
-            x, comfy.model_management.cast_to(self.filter.expand(C, -1, -1), dtype=x.dtype, device=x.device), stride=self.stride, groups=C)
-        x = x[..., self.pad_left:-self.pad_right]
-
-        return x
-
-
-class DownSample1d(nn.Module):
-    def __init__(self, ratio=2, kernel_size=None):
-        super().__init__()
-        self.ratio = ratio
-        self.kernel_size = int(6 * ratio // 2) * 2 if kernel_size is None else kernel_size
-        self.lowpass = LowPassFilter1d(cutoff=0.5 / ratio,
-                                       half_width=0.6 / ratio,
-                                       stride=ratio,
-                                       kernel_size=self.kernel_size)
-
-    def forward(self, x):
-        xx = self.lowpass(x)
-
-        return xx
-
-class Activation1d(nn.Module):
-    def __init__(self,
-                 activation,
-                 up_ratio: int = 2,
-                 down_ratio: int = 2,
-                 up_kernel_size: int = 12,
-                 down_kernel_size: int = 12):
-        super().__init__()
-        self.up_ratio = up_ratio
-        self.down_ratio = down_ratio
-        self.act = activation
-        self.upsample = UpSample1d(up_ratio, up_kernel_size)
-        self.downsample = DownSample1d(down_ratio, down_kernel_size)
-
-    # x: [B,C,T]
-    def forward(self, x):
-        x = self.upsample(x)
-        x = self.act(x)
-        x = self.downsample(x)
-
-        return x

comfy/ldm/mmaudio/vae/autoencoder.py

@@ -1,156 +0,0 @@
-from typing import Literal
-
-import torch
-import torch.nn as nn
-
-from .distributions import DiagonalGaussianDistribution
-from .vae import VAE_16k
-from .bigvgan import BigVGANVocoder
-import logging
-
-try:
-    import torchaudio
-except:
-    logging.warning("torchaudio missing, MMAudio VAE model will be broken")
-
-def dynamic_range_compression_torch(x, C=1, clip_val=1e-5, *, norm_fn):
-    return norm_fn(torch.clamp(x, min=clip_val) * C)
-
-
-def spectral_normalize_torch(magnitudes, norm_fn):
-    output = dynamic_range_compression_torch(magnitudes, norm_fn=norm_fn)
-    return output
-
-class MelConverter(nn.Module):
-
-    def __init__(
-        self,
-        *,
-        sampling_rate: float,
-        n_fft: int,
-        num_mels: int,
-        hop_size: int,
-        win_size: int,
-        fmin: float,
-        fmax: float,
-        norm_fn,
-    ):
-        super().__init__()
-        self.sampling_rate = sampling_rate
-        self.n_fft = n_fft
-        self.num_mels = num_mels
-        self.hop_size = hop_size
-        self.win_size = win_size
-        self.fmin = fmin
-        self.fmax = fmax
-        self.norm_fn = norm_fn
-
-        # mel = librosa_mel_fn(sr=self.sampling_rate,
-        #                      n_fft=self.n_fft,
-        #                      n_mels=self.num_mels,
-        #                      fmin=self.fmin,
-        #                      fmax=self.fmax)
-        # mel_basis = torch.from_numpy(mel).float()
-        mel_basis = torch.empty((num_mels, 1 + n_fft // 2))
-        hann_window = torch.hann_window(self.win_size)
-
-        self.register_buffer('mel_basis', mel_basis)
-        self.register_buffer('hann_window', hann_window)
-
-    @property
-    def device(self):
-        return self.mel_basis.device
-
-    def forward(self, waveform: torch.Tensor, center: bool = False) -> torch.Tensor:
-        waveform = waveform.clamp(min=-1., max=1.).to(self.device)
-
-        waveform = torch.nn.functional.pad(
-            waveform.unsqueeze(1),
-            [int((self.n_fft - self.hop_size) / 2),
-             int((self.n_fft - self.hop_size) / 2)],
-            mode='reflect')
-        waveform = waveform.squeeze(1)
-
-        spec = torch.stft(waveform,
-                          self.n_fft,
-                          hop_length=self.hop_size,
-                          win_length=self.win_size,
-                          window=self.hann_window,
-                          center=center,
-                          pad_mode='reflect',
-                          normalized=False,
-                          onesided=True,
-                          return_complex=True)
-
-        spec = torch.view_as_real(spec)
-        spec = torch.sqrt(spec.pow(2).sum(-1) + (1e-9))
-        spec = torch.matmul(self.mel_basis, spec)
-        spec = spectral_normalize_torch(spec, self.norm_fn)
-
-        return spec
-
-class AudioAutoencoder(nn.Module):
-
-    def __init__(
-        self,
-        *,
-        # ckpt_path: str,
-        mode=Literal['16k', '44k'],
-        need_vae_encoder: bool = True,
-    ):
-        super().__init__()
-
-        assert mode == "16k", "Only 16k mode is supported currently."
-        self.mel_converter = MelConverter(sampling_rate=16_000,
-                            n_fft=1024,
-                            num_mels=80,
-                            hop_size=256,
-                            win_size=1024,
-                            fmin=0,
-                            fmax=8_000,
-                            norm_fn=torch.log10)
-
-        self.vae = VAE_16k().eval()
-
-        bigvgan_config = {
-            "resblock": "1",
-            "num_mels": 80,
-            "upsample_rates": [4, 4, 2, 2, 2, 2],
-            "upsample_kernel_sizes": [8, 8, 4, 4, 4, 4],
-            "upsample_initial_channel": 1536,
-            "resblock_kernel_sizes": [3, 7, 11],
-            "resblock_dilation_sizes": [
-                [1, 3, 5],
-                [1, 3, 5],
-                [1, 3, 5],
-            ],
-            "activation": "snakebeta",
-            "snake_logscale": True,
-        }
-
-        self.vocoder = BigVGANVocoder(
-            bigvgan_config
-        ).eval()
-
-    @torch.inference_mode()
-    def encode_audio(self, x) -> DiagonalGaussianDistribution:
-        # x: (B * L)
-        mel = self.mel_converter(x)
-        dist = self.vae.encode(mel)
-
-        return dist
-
-    @torch.no_grad()
-    def decode(self, z):
-        mel_decoded = self.vae.decode(z)
-        audio = self.vocoder(mel_decoded)
-
-        audio = torchaudio.functional.resample(audio, 16000, 44100)
-        return audio
-
-    @torch.no_grad()
-    def encode(self, audio):
-        audio = audio.mean(dim=1)
-        audio = torchaudio.functional.resample(audio, 44100, 16000)
-        dist = self.encode_audio(audio)
-        return dist.mean

comfy/ldm/mmaudio/vae/bigvgan.py

@@ -1,219 +0,0 @@
-# Copyright (c) 2022 NVIDIA CORPORATION.
-#   Licensed under the MIT license.
-
-# Adapted from https://github.com/jik876/hifi-gan under the MIT license.
-#   LICENSE is in incl_licenses directory.
-
-import torch
-import torch.nn as nn
-from types import SimpleNamespace
-from . import activations
-from .alias_free_torch import Activation1d
-import comfy.ops
-ops = comfy.ops.disable_weight_init
-
-def get_padding(kernel_size, dilation=1):
-    return int((kernel_size * dilation - dilation) / 2)
-
-class AMPBlock1(torch.nn.Module):
-
-    def __init__(self, h, channels, kernel_size=3, dilation=(1, 3, 5), activation=None):
-        super(AMPBlock1, self).__init__()
-        self.h = h
-
-        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))
-        ])
-
-        self.num_layers = len(self.convs1) + len(self.convs2)  # total number of conv layers
-
-        if activation == 'snake':  # periodic nonlinearity with snake function and anti-aliasing
-            self.activations = nn.ModuleList([
-                Activation1d(
-                    activation=activations.Snake(channels, alpha_logscale=h.snake_logscale))
-                for _ in range(self.num_layers)
-            ])
-        elif activation == 'snakebeta':  # periodic nonlinearity with snakebeta function and anti-aliasing
-            self.activations = nn.ModuleList([
-                Activation1d(
-                    activation=activations.SnakeBeta(channels, alpha_logscale=h.snake_logscale))
-                for _ in range(self.num_layers)
-            ])
-        else:
-            raise NotImplementedError(
-                "activation incorrectly specified. check the config file and look for 'activation'."
-            )
-
-    def forward(self, x):
-        acts1, acts2 = self.activations[::2], self.activations[1::2]
-        for c1, c2, a1, a2 in zip(self.convs1, self.convs2, acts1, acts2):
-            xt = a1(x)
-            xt = c1(xt)
-            xt = a2(xt)
-            xt = c2(xt)
-            x = xt + x
-
-        return x
-
-
-class AMPBlock2(torch.nn.Module):
-
-    def __init__(self, h, channels, kernel_size=3, dilation=(1, 3), activation=None):
-        super(AMPBlock2, self).__init__()
-        self.h = h
-
-        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]))
-        ])
-
-        self.num_layers = len(self.convs)  # total number of conv layers
-
-        if activation == 'snake':  # periodic nonlinearity with snake function and anti-aliasing
-            self.activations = nn.ModuleList([
-                Activation1d(
-                    activation=activations.Snake(channels, alpha_logscale=h.snake_logscale))
-                for _ in range(self.num_layers)
-            ])
-        elif activation == 'snakebeta':  # periodic nonlinearity with snakebeta function and anti-aliasing
-            self.activations = nn.ModuleList([
-                Activation1d(
-                    activation=activations.SnakeBeta(channels, alpha_logscale=h.snake_logscale))
-                for _ in range(self.num_layers)
-            ])
-        else:
-            raise NotImplementedError(
-                "activation incorrectly specified. check the config file and look for 'activation'."
-            )
-
-    def forward(self, x):
-        for c, a in zip(self.convs, self.activations):
-            xt = a(x)
-            xt = c(xt)
-            x = xt + x
-
-        return x
-
-
-class BigVGANVocoder(torch.nn.Module):
-    # this is our main BigVGAN model. Applies anti-aliased periodic activation for resblocks.
-    def __init__(self, h):
-        super().__init__()
-        if isinstance(h, dict):
-            h = SimpleNamespace(**h)
-        self.h = h
-
-        self.num_kernels = len(h.resblock_kernel_sizes)
-        self.num_upsamples = len(h.upsample_rates)
-
-        # pre conv
-        self.conv_pre = ops.Conv1d(h.num_mels, h.upsample_initial_channel, 7, 1, padding=3)
-
-        # define which AMPBlock to use. BigVGAN uses AMPBlock1 as default
-        resblock = AMPBlock1 if h.resblock == '1' else AMPBlock2
-
-        # transposed conv-based upsamplers. does not apply anti-aliasing
-        self.ups = nn.ModuleList()
-        for i, (u, k) in enumerate(zip(h.upsample_rates, h.upsample_kernel_sizes)):
-            self.ups.append(
-                nn.ModuleList([
-                        ops.ConvTranspose1d(h.upsample_initial_channel // (2**i),
-                                        h.upsample_initial_channel // (2**(i + 1)),
-                                        k,
-                                        u,
-                                        padding=(k - u) // 2)
-                ]))
-
-        # residual blocks using anti-aliased multi-periodicity composition modules (AMP)
-        self.resblocks = nn.ModuleList()
-        for i in range(len(self.ups)):
-            ch = h.upsample_initial_channel // (2**(i + 1))
-            for j, (k, d) in enumerate(zip(h.resblock_kernel_sizes, h.resblock_dilation_sizes)):
-                self.resblocks.append(resblock(h, ch, k, d, activation=h.activation))
-
-        # post conv
-        if h.activation == "snake":  # periodic nonlinearity with snake function and anti-aliasing
-            activation_post = activations.Snake(ch, alpha_logscale=h.snake_logscale)
-            self.activation_post = Activation1d(activation=activation_post)
-        elif h.activation == "snakebeta":  # periodic nonlinearity with snakebeta function and anti-aliasing
-            activation_post = activations.SnakeBeta(ch, alpha_logscale=h.snake_logscale)
-            self.activation_post = Activation1d(activation=activation_post)
-        else:
-            raise NotImplementedError(
-                "activation incorrectly specified. check the config file and look for 'activation'."
-            )
-
-        self.conv_post = ops.Conv1d(ch, 1, 7, 1, padding=3)
-
-
-    def forward(self, x):
-        # pre conv
-        x = self.conv_pre(x)
-
-        for i in range(self.num_upsamples):
-            # upsampling
-            for i_up in range(len(self.ups[i])):
-                x = self.ups[i][i_up](x)
-            # AMP blocks
-            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
-
-        # post conv
-        x = self.activation_post(x)
-        x = self.conv_post(x)
-        x = torch.tanh(x)
-
-        return x

comfy/ldm/mmaudio/vae/distributions.py

@@ -1,92 +0,0 @@
-import torch
-import numpy as np
-
-
-class AbstractDistribution:
-    def sample(self):
-        raise NotImplementedError()
-
-    def mode(self):
-        raise NotImplementedError()
-
-
-class DiracDistribution(AbstractDistribution):
-    def __init__(self, value):
-        self.value = value
-
-    def sample(self):
-        return self.value
-
-    def mode(self):
-        return self.value
-
-
-class DiagonalGaussianDistribution(object):
-    def __init__(self, parameters, deterministic=False):
-        self.parameters = parameters
-        self.mean, self.logvar = torch.chunk(parameters, 2, dim=1)
-        self.logvar = torch.clamp(self.logvar, -30.0, 20.0)
-        self.deterministic = deterministic
-        self.std = torch.exp(0.5 * self.logvar)
-        self.var = torch.exp(self.logvar)
-        if self.deterministic:
-            self.var = self.std = torch.zeros_like(self.mean, device=self.parameters.device)
-
-    def sample(self):
-        x = self.mean + self.std * torch.randn(self.mean.shape, device=self.parameters.device)
-        return x
-
-    def kl(self, other=None):
-        if self.deterministic:
-            return torch.Tensor([0.])
-        else:
-            if other is None:
-                return 0.5 * torch.sum(torch.pow(self.mean, 2)
-                                       + self.var - 1.0 - self.logvar,
-                                       dim=[1, 2, 3])
-            else:
-                return 0.5 * torch.sum(
-                    torch.pow(self.mean - other.mean, 2) / other.var
-                    + self.var / other.var - 1.0 - self.logvar + other.logvar,
-                    dim=[1, 2, 3])
-
-    def nll(self, sample, dims=[1,2,3]):
-        if self.deterministic:
-            return torch.Tensor([0.])
-        logtwopi = np.log(2.0 * np.pi)
-        return 0.5 * torch.sum(
-            logtwopi + self.logvar + torch.pow(sample - self.mean, 2) / self.var,
-            dim=dims)
-
-    def mode(self):
-        return self.mean
-
-
-def normal_kl(mean1, logvar1, mean2, logvar2):
-    """
-    source: https://github.com/openai/guided-diffusion/blob/27c20a8fab9cb472df5d6bdd6c8d11c8f430b924/guided_diffusion/losses.py#L12
-    Compute the KL divergence between two gaussians.
-    Shapes are automatically broadcasted, so batches can be compared to
-    scalars, among other use cases.
-    """
-    tensor = None
-    for obj in (mean1, logvar1, mean2, logvar2):
-        if isinstance(obj, torch.Tensor):
-            tensor = obj
-            break
-    assert tensor is not None, "at least one argument must be a Tensor"
-
-    # Force variances to be Tensors. Broadcasting helps convert scalars to
-    # Tensors, but it does not work for torch.exp().
-    logvar1, logvar2 = [
-        x if isinstance(x, torch.Tensor) else torch.tensor(x).to(tensor)
-        for x in (logvar1, logvar2)
-    ]
-
-    return 0.5 * (
-        -1.0
-        + logvar2
-        - logvar1
-        + torch.exp(logvar1 - logvar2)
-        + ((mean1 - mean2) ** 2) * torch.exp(-logvar2)
-    )

comfy/ldm/mmaudio/vae/vae.py

@@ -1,358 +0,0 @@
-import logging
-from typing import Optional
-
-import torch
-import torch.nn as nn
-
-from .vae_modules import (AttnBlock1D, Downsample1D, ResnetBlock1D,
-                                                 Upsample1D, nonlinearity)
-from .distributions import DiagonalGaussianDistribution
-
-import comfy.ops
-ops = comfy.ops.disable_weight_init
-
-log = logging.getLogger()
-
-DATA_MEAN_80D = [
-    -1.6058, -1.3676, -1.2520, -1.2453, -1.2078, -1.2224, -1.2419, -1.2439, -1.2922, -1.2927,
-    -1.3170, -1.3543, -1.3401, -1.3836, -1.3907, -1.3912, -1.4313, -1.4152, -1.4527, -1.4728,
-    -1.4568, -1.5101, -1.5051, -1.5172, -1.5623, -1.5373, -1.5746, -1.5687, -1.6032, -1.6131,
-    -1.6081, -1.6331, -1.6489, -1.6489, -1.6700, -1.6738, -1.6953, -1.6969, -1.7048, -1.7280,
-    -1.7361, -1.7495, -1.7658, -1.7814, -1.7889, -1.8064, -1.8221, -1.8377, -1.8417, -1.8643,
-    -1.8857, -1.8929, -1.9173, -1.9379, -1.9531, -1.9673, -1.9824, -2.0042, -2.0215, -2.0436,
-    -2.0766, -2.1064, -2.1418, -2.1855, -2.2319, -2.2767, -2.3161, -2.3572, -2.3954, -2.4282,
-    -2.4659, -2.5072, -2.5552, -2.6074, -2.6584, -2.7107, -2.7634, -2.8266, -2.8981, -2.9673
-]
-
-DATA_STD_80D = [
-    1.0291, 1.0411, 1.0043, 0.9820, 0.9677, 0.9543, 0.9450, 0.9392, 0.9343, 0.9297, 0.9276, 0.9263,
-    0.9242, 0.9254, 0.9232, 0.9281, 0.9263, 0.9315, 0.9274, 0.9247, 0.9277, 0.9199, 0.9188, 0.9194,
-    0.9160, 0.9161, 0.9146, 0.9161, 0.9100, 0.9095, 0.9145, 0.9076, 0.9066, 0.9095, 0.9032, 0.9043,
-    0.9038, 0.9011, 0.9019, 0.9010, 0.8984, 0.8983, 0.8986, 0.8961, 0.8962, 0.8978, 0.8962, 0.8973,
-    0.8993, 0.8976, 0.8995, 0.9016, 0.8982, 0.8972, 0.8974, 0.8949, 0.8940, 0.8947, 0.8936, 0.8939,
-    0.8951, 0.8956, 0.9017, 0.9167, 0.9436, 0.9690, 1.0003, 1.0225, 1.0381, 1.0491, 1.0545, 1.0604,
-    1.0761, 1.0929, 1.1089, 1.1196, 1.1176, 1.1156, 1.1117, 1.1070
-]
-
-DATA_MEAN_128D = [
-    -3.3462, -2.6723, -2.4893, -2.3143, -2.2664, -2.3317, -2.1802, -2.4006, -2.2357, -2.4597,
-    -2.3717, -2.4690, -2.5142, -2.4919, -2.6610, -2.5047, -2.7483, -2.5926, -2.7462, -2.7033,
-    -2.7386, -2.8112, -2.7502, -2.9594, -2.7473, -3.0035, -2.8891, -2.9922, -2.9856, -3.0157,
-    -3.1191, -2.9893, -3.1718, -3.0745, -3.1879, -3.2310, -3.1424, -3.2296, -3.2791, -3.2782,
-    -3.2756, -3.3134, -3.3509, -3.3750, -3.3951, -3.3698, -3.4505, -3.4509, -3.5089, -3.4647,
-    -3.5536, -3.5788, -3.5867, -3.6036, -3.6400, -3.6747, -3.7072, -3.7279, -3.7283, -3.7795,
-    -3.8259, -3.8447, -3.8663, -3.9182, -3.9605, -3.9861, -4.0105, -4.0373, -4.0762, -4.1121,
-    -4.1488, -4.1874, -4.2461, -4.3170, -4.3639, -4.4452, -4.5282, -4.6297, -4.7019, -4.7960,
-    -4.8700, -4.9507, -5.0303, -5.0866, -5.1634, -5.2342, -5.3242, -5.4053, -5.4927, -5.5712,
-    -5.6464, -5.7052, -5.7619, -5.8410, -5.9188, -6.0103, -6.0955, -6.1673, -6.2362, -6.3120,
-    -6.3926, -6.4797, -6.5565, -6.6511, -6.8130, -6.9961, -7.1275, -7.2457, -7.3576, -7.4663,
-    -7.6136, -7.7469, -7.8815, -8.0132, -8.1515, -8.3071, -8.4722, -8.7418, -9.3975, -9.6628,
-    -9.7671, -9.8863, -9.9992, -10.0860, -10.1709, -10.5418, -11.2795, -11.3861
-]
-
-DATA_STD_128D = [
-    2.3804, 2.4368, 2.3772, 2.3145, 2.2803, 2.2510, 2.2316, 2.2083, 2.1996, 2.1835, 2.1769, 2.1659,
-    2.1631, 2.1618, 2.1540, 2.1606, 2.1571, 2.1567, 2.1612, 2.1579, 2.1679, 2.1683, 2.1634, 2.1557,
-    2.1668, 2.1518, 2.1415, 2.1449, 2.1406, 2.1350, 2.1313, 2.1415, 2.1281, 2.1352, 2.1219, 2.1182,
-    2.1327, 2.1195, 2.1137, 2.1080, 2.1179, 2.1036, 2.1087, 2.1036, 2.1015, 2.1068, 2.0975, 2.0991,
-    2.0902, 2.1015, 2.0857, 2.0920, 2.0893, 2.0897, 2.0910, 2.0881, 2.0925, 2.0873, 2.0960, 2.0900,
-    2.0957, 2.0958, 2.0978, 2.0936, 2.0886, 2.0905, 2.0845, 2.0855, 2.0796, 2.0840, 2.0813, 2.0817,
-    2.0838, 2.0840, 2.0917, 2.1061, 2.1431, 2.1976, 2.2482, 2.3055, 2.3700, 2.4088, 2.4372, 2.4609,
-    2.4731, 2.4847, 2.5072, 2.5451, 2.5772, 2.6147, 2.6529, 2.6596, 2.6645, 2.6726, 2.6803, 2.6812,
-    2.6899, 2.6916, 2.6931, 2.6998, 2.7062, 2.7262, 2.7222, 2.7158, 2.7041, 2.7485, 2.7491, 2.7451,
-    2.7485, 2.7233, 2.7297, 2.7233, 2.7145, 2.6958, 2.6788, 2.6439, 2.6007, 2.4786, 2.2469, 2.1877,
-    2.1392, 2.0717, 2.0107, 1.9676, 1.9140, 1.7102, 0.9101, 0.7164
-]
-
-
-class VAE(nn.Module):
-
-    def __init__(
-        self,
-        *,
-        data_dim: int,
-        embed_dim: int,
-        hidden_dim: int,
-    ):
-        super().__init__()
-
-        if data_dim == 80:
-            self.data_mean = nn.Buffer(torch.tensor(DATA_MEAN_80D, dtype=torch.float32))
-            self.data_std = nn.Buffer(torch.tensor(DATA_STD_80D, dtype=torch.float32))
-        elif data_dim == 128:
-            self.data_mean = nn.Buffer(torch.tensor(DATA_MEAN_128D, dtype=torch.float32))
-            self.data_std = nn.Buffer(torch.tensor(DATA_STD_128D, dtype=torch.float32))
-
-        self.data_mean = self.data_mean.view(1, -1, 1)
-        self.data_std = self.data_std.view(1, -1, 1)
-
-        self.encoder = Encoder1D(
-            dim=hidden_dim,
-            ch_mult=(1, 2, 4),
-            num_res_blocks=2,
-            attn_layers=[3],
-            down_layers=[0],
-            in_dim=data_dim,
-            embed_dim=embed_dim,
-        )
-        self.decoder = Decoder1D(
-            dim=hidden_dim,
-            ch_mult=(1, 2, 4),
-            num_res_blocks=2,
-            attn_layers=[3],
-            down_layers=[0],
-            in_dim=data_dim,
-            out_dim=data_dim,
-            embed_dim=embed_dim,
-        )
-
-        self.embed_dim = embed_dim
-        # self.quant_conv = nn.Conv1d(2 * embed_dim, 2 * embed_dim, 1)
-        # self.post_quant_conv = nn.Conv1d(embed_dim, embed_dim, 1)
-
-        self.initialize_weights()
-
-    def initialize_weights(self):
-        pass
-
-    def encode(self, x: torch.Tensor, normalize: bool = True) -> DiagonalGaussianDistribution:
-        if normalize:
-            x = self.normalize(x)
-        moments = self.encoder(x)
-        posterior = DiagonalGaussianDistribution(moments)
-        return posterior
-
-    def decode(self, z: torch.Tensor, unnormalize: bool = True) -> torch.Tensor:
-        dec = self.decoder(z)
-        if unnormalize:
-            dec = self.unnormalize(dec)
-        return dec
-
-    def normalize(self, x: torch.Tensor) -> torch.Tensor:
-        return (x - comfy.model_management.cast_to(self.data_mean, dtype=x.dtype, device=x.device)) / comfy.model_management.cast_to(self.data_std, dtype=x.dtype, device=x.device)
-
-    def unnormalize(self, x: torch.Tensor) -> torch.Tensor:
-        return x * comfy.model_management.cast_to(self.data_std, dtype=x.dtype, device=x.device) + comfy.model_management.cast_to(self.data_mean, dtype=x.dtype, device=x.device)
-
-    def forward(
-        self,
-        x: torch.Tensor,
-        sample_posterior: bool = True,
-        rng: Optional[torch.Generator] = None,
-        normalize: bool = True,
-        unnormalize: bool = True,
-    ) -> tuple[torch.Tensor, DiagonalGaussianDistribution]:
-
-        posterior = self.encode(x, normalize=normalize)
-        if sample_posterior:
-            z = posterior.sample(rng)
-        else:
-            z = posterior.mode()
-        dec = self.decode(z, unnormalize=unnormalize)
-        return dec, posterior
-
-    def load_weights(self, src_dict) -> None:
-        self.load_state_dict(src_dict, strict=True)
-
-    @property
-    def device(self) -> torch.device:
-        return next(self.parameters()).device
-
-    def get_last_layer(self):
-        return self.decoder.conv_out.weight
-
-    def remove_weight_norm(self):
-        return self
-
-
-class Encoder1D(nn.Module):
-
-    def __init__(self,
-                 *,
-                 dim: int,
-                 ch_mult: tuple[int] = (1, 2, 4, 8),
-                 num_res_blocks: int,
-                 attn_layers: list[int] = [],
-                 down_layers: list[int] = [],
-                 resamp_with_conv: bool = True,
-                 in_dim: int,
-                 embed_dim: int,
-                 double_z: bool = True,
-                 kernel_size: int = 3,
-                 clip_act: float = 256.0):
-        super().__init__()
-        self.dim = dim
-        self.num_layers = len(ch_mult)
-        self.num_res_blocks = num_res_blocks
-        self.in_channels = in_dim
-        self.clip_act = clip_act
-        self.down_layers = down_layers
-        self.attn_layers = attn_layers
-        self.conv_in = ops.Conv1d(in_dim, self.dim, kernel_size=kernel_size, padding=kernel_size // 2, bias=False)
-
-        in_ch_mult = (1, ) + tuple(ch_mult)
-        self.in_ch_mult = in_ch_mult
-        # downsampling
-        self.down = nn.ModuleList()
-        for i_level in range(self.num_layers):
-            block = nn.ModuleList()
-            attn = nn.ModuleList()
-            block_in = dim * in_ch_mult[i_level]
-            block_out = dim * ch_mult[i_level]
-            for i_block in range(self.num_res_blocks):
-                block.append(
-                    ResnetBlock1D(in_dim=block_in,
-                                  out_dim=block_out,
-                                  kernel_size=kernel_size,
-                                  use_norm=True))
-                block_in = block_out
-                if i_level in attn_layers:
-                    attn.append(AttnBlock1D(block_in))
-            down = nn.Module()
-            down.block = block
-            down.attn = attn
-            if i_level in down_layers:
-                down.downsample = Downsample1D(block_in, resamp_with_conv)
-            self.down.append(down)
-
-        # middle
-        self.mid = nn.Module()
-        self.mid.block_1 = ResnetBlock1D(in_dim=block_in,
-                                         out_dim=block_in,
-                                         kernel_size=kernel_size,
-                                         use_norm=True)
-        self.mid.attn_1 = AttnBlock1D(block_in)
-        self.mid.block_2 = ResnetBlock1D(in_dim=block_in,
-                                         out_dim=block_in,
-                                         kernel_size=kernel_size,
-                                         use_norm=True)
-
-        # end
-        self.conv_out = ops.Conv1d(block_in,
-                                 2 * embed_dim if double_z else embed_dim,
-                                 kernel_size=kernel_size, padding=kernel_size // 2, bias=False)
-
-        self.learnable_gain = nn.Parameter(torch.zeros([]))
-
-    def forward(self, x):
-
-        # downsampling
-        h = self.conv_in(x)
-        for i_level in range(self.num_layers):
-            for i_block in range(self.num_res_blocks):
-                h = self.down[i_level].block[i_block](h)
-                if len(self.down[i_level].attn) > 0:
-                    h = self.down[i_level].attn[i_block](h)
-                h = h.clamp(-self.clip_act, self.clip_act)
-            if i_level in self.down_layers:
-                h = self.down[i_level].downsample(h)
-
-        # middle
-        h = self.mid.block_1(h)
-        h = self.mid.attn_1(h)
-        h = self.mid.block_2(h)
-        h = h.clamp(-self.clip_act, self.clip_act)
-
-        # end
-        h = nonlinearity(h)
-        h = self.conv_out(h) * (self.learnable_gain + 1)
-        return h
-
-
-class Decoder1D(nn.Module):
-
-    def __init__(self,
-                 *,
-                 dim: int,
-                 out_dim: int,
-                 ch_mult: tuple[int] = (1, 2, 4, 8),
-                 num_res_blocks: int,
-                 attn_layers: list[int] = [],
-                 down_layers: list[int] = [],
-                 kernel_size: int = 3,
-                 resamp_with_conv: bool = True,
-                 in_dim: int,
-                 embed_dim: int,
-                 clip_act: float = 256.0):
-        super().__init__()
-        self.ch = dim
-        self.num_layers = len(ch_mult)
-        self.num_res_blocks = num_res_blocks
-        self.in_channels = in_dim
-        self.clip_act = clip_act
-        self.down_layers = [i + 1 for i in down_layers]  # each downlayer add one
-
-        # compute in_ch_mult, block_in and curr_res at lowest res
-        block_in = dim * ch_mult[self.num_layers - 1]
-
-        # z to block_in
-        self.conv_in = ops.Conv1d(embed_dim, block_in, kernel_size=kernel_size, padding=kernel_size // 2, bias=False)
-
-        # middle
-        self.mid = nn.Module()
-        self.mid.block_1 = ResnetBlock1D(in_dim=block_in, out_dim=block_in, use_norm=True)
-        self.mid.attn_1 = AttnBlock1D(block_in)
-        self.mid.block_2 = ResnetBlock1D(in_dim=block_in, out_dim=block_in, use_norm=True)
-
-        # upsampling
-        self.up = nn.ModuleList()
-        for i_level in reversed(range(self.num_layers)):
-            block = nn.ModuleList()
-            attn = nn.ModuleList()
-            block_out = dim * ch_mult[i_level]
-            for i_block in range(self.num_res_blocks + 1):
-                block.append(ResnetBlock1D(in_dim=block_in, out_dim=block_out, use_norm=True))
-                block_in = block_out
-                if i_level in attn_layers:
-                    attn.append(AttnBlock1D(block_in))
-            up = nn.Module()
-            up.block = block
-            up.attn = attn
-            if i_level in self.down_layers:
-                up.upsample = Upsample1D(block_in, resamp_with_conv)
-            self.up.insert(0, up)  # prepend to get consistent order
-
-        # end
-        self.conv_out = ops.Conv1d(block_in, out_dim, kernel_size=kernel_size, padding=kernel_size // 2, bias=False)
-        self.learnable_gain = nn.Parameter(torch.zeros([]))
-
-    def forward(self, z):
-        # z to block_in
-        h = self.conv_in(z)
-
-        # middle
-        h = self.mid.block_1(h)
-        h = self.mid.attn_1(h)
-        h = self.mid.block_2(h)
-        h = h.clamp(-self.clip_act, self.clip_act)
-
-        # upsampling
-        for i_level in reversed(range(self.num_layers)):
-            for i_block in range(self.num_res_blocks + 1):
-                h = self.up[i_level].block[i_block](h)
-                if len(self.up[i_level].attn) > 0:
-                    h = self.up[i_level].attn[i_block](h)
-                h = h.clamp(-self.clip_act, self.clip_act)
-            if i_level in self.down_layers:
-                h = self.up[i_level].upsample(h)
-
-        h = nonlinearity(h)
-        h = self.conv_out(h) * (self.learnable_gain + 1)
-        return h
-
-
-def VAE_16k(**kwargs) -> VAE:
-    return VAE(data_dim=80, embed_dim=20, hidden_dim=384, **kwargs)
-
-
-def VAE_44k(**kwargs) -> VAE:
-    return VAE(data_dim=128, embed_dim=40, hidden_dim=512, **kwargs)
-
-
-def get_my_vae(name: str, **kwargs) -> VAE:
-    if name == '16k':
-        return VAE_16k(**kwargs)
-    if name == '44k':
-        return VAE_44k(**kwargs)
-    raise ValueError(f'Unknown model: {name}')
-

comfy/ldm/mmaudio/vae/vae_modules.py

@@ -1,121 +0,0 @@
-import torch
-import torch.nn as nn
-import torch.nn.functional as F
-from comfy.ldm.modules.diffusionmodules.model import vae_attention
-import math
-import comfy.ops
-ops = comfy.ops.disable_weight_init
-
-def nonlinearity(x):
-    # swish
-    return torch.nn.functional.silu(x) / 0.596
-
-def mp_sum(a, b, t=0.5):
-    return a.lerp(b, t) / math.sqrt((1 - t)**2 + t**2)
-
-def normalize(x, dim=None, eps=1e-4):
-    if dim is None:
-        dim = list(range(1, x.ndim))
-    norm = torch.linalg.vector_norm(x, dim=dim, keepdim=True, dtype=torch.float32)
-    norm = torch.add(eps, norm, alpha=math.sqrt(norm.numel() / x.numel()))
-    return x / norm.to(x.dtype)
-
-class ResnetBlock1D(nn.Module):
-
-    def __init__(self, *, in_dim, out_dim=None, conv_shortcut=False, kernel_size=3, use_norm=True):
-        super().__init__()
-        self.in_dim = in_dim
-        out_dim = in_dim if out_dim is None else out_dim
-        self.out_dim = out_dim
-        self.use_conv_shortcut = conv_shortcut
-        self.use_norm = use_norm
-
-        self.conv1 = ops.Conv1d(in_dim, out_dim, kernel_size=kernel_size, padding=kernel_size // 2, bias=False)
-        self.conv2 = ops.Conv1d(out_dim, out_dim, kernel_size=kernel_size, padding=kernel_size // 2, bias=False)
-        if self.in_dim != self.out_dim:
-            if self.use_conv_shortcut:
-                self.conv_shortcut = ops.Conv1d(in_dim, out_dim, kernel_size=kernel_size, padding=kernel_size // 2, bias=False)
-            else:
-                self.nin_shortcut = ops.Conv1d(in_dim, out_dim, kernel_size=1, padding=0, bias=False)
-
-    def forward(self, x: torch.Tensor) -> torch.Tensor:
-
-        # pixel norm
-        if self.use_norm:
-            x = normalize(x, dim=1)
-
-        h = x
-        h = nonlinearity(h)
-        h = self.conv1(h)
-
-        h = nonlinearity(h)
-        h = self.conv2(h)
-
-        if self.in_dim != self.out_dim:
-            if self.use_conv_shortcut:
-                x = self.conv_shortcut(x)
-            else:
-                x = self.nin_shortcut(x)
-
-        return mp_sum(x, h, t=0.3)
-
-
-class AttnBlock1D(nn.Module):
-
-    def __init__(self, in_channels, num_heads=1):
-        super().__init__()
-        self.in_channels = in_channels
-
-        self.num_heads = num_heads
-        self.qkv = ops.Conv1d(in_channels, in_channels * 3, kernel_size=1, padding=0, bias=False)
-        self.proj_out = ops.Conv1d(in_channels, in_channels, kernel_size=1, padding=0, bias=False)
-        self.optimized_attention = vae_attention()
-
-    def forward(self, x):
-        h = x
-        y = self.qkv(h)
-        y = y.reshape(y.shape[0], -1, 3, y.shape[-1])
-        q, k, v = normalize(y, dim=1).unbind(2)
-
-        h = self.optimized_attention(q, k, v)
-        h = self.proj_out(h)
-
-        return mp_sum(x, h, t=0.3)
-
-
-class Upsample1D(nn.Module):
-
-    def __init__(self, in_channels, with_conv):
-        super().__init__()
-        self.with_conv = with_conv
-        if self.with_conv:
-            self.conv = ops.Conv1d(in_channels, in_channels, kernel_size=3, padding=1, bias=False)
-
-    def forward(self, x):
-        x = F.interpolate(x, scale_factor=2.0, mode='nearest-exact')  # support 3D tensor(B,C,T)
-        if self.with_conv:
-            x = self.conv(x)
-        return x
-
-
-class Downsample1D(nn.Module):
-
-    def __init__(self, in_channels, with_conv):
-        super().__init__()
-        self.with_conv = with_conv
-        if self.with_conv:
-            # no asymmetric padding in torch conv, must do it ourselves
-            self.conv1 = ops.Conv1d(in_channels, in_channels, kernel_size=1, padding=0, bias=False)
-            self.conv2 = ops.Conv1d(in_channels, in_channels, kernel_size=1, padding=0, bias=False)
-
-    def forward(self, x):
-
-        if self.with_conv:
-            x = self.conv1(x)
-
-        x = F.avg_pool1d(x, kernel_size=2, stride=2)
-
-        if self.with_conv:
-            x = self.conv2(x)
-
-        return x

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).to(x.dtype).contiguous()
+        image_rotary_emb = self.pe_embedder(ids).squeeze(1).unsqueeze(2).to(x.dtype)
         del ids, txt_ids, img_ids
 
         hidden_states = self.img_in(hidden_states) + self.controlnet_x_embedder(hint)

comfy/ldm/qwen_image/model.py

@@ -10,7 +10,6 @@ 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):
@@ -135,34 +134,33 @@ 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]
 
-        # 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)
+        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))
 
-        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)
+        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))
 
         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=2)
-        joint_key = torch.cat([txt_key, img_key], dim=2)
-        joint_value = torch.cat([txt_value, img_value], dim=2)
+        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 = apply_rope1(joint_query, image_rotary_emb)
-        joint_key = apply_rope1(joint_key, image_rotary_emb)
+        joint_query = apply_rotary_emb(joint_query, image_rotary_emb)
+        joint_key = apply_rotary_emb(joint_key, image_rotary_emb)
 
-        joint_hidden_states = optimized_attention_masked(joint_query, joint_key, joint_value, self.heads,
-                                                         attention_mask, transformer_options=transformer_options,
-                                                         skip_reshape=True)
+        joint_query = joint_query.flatten(start_dim=2)
+        joint_key = joint_key.flatten(start_dim=2)
+        joint_value = joint_value.flatten(start_dim=2)
+
+        joint_hidden_states = optimized_attention_masked(joint_query, joint_key, joint_value, self.heads, attention_mask, transformer_options=transformer_options)
 
         txt_attn_output = joint_hidden_states[:, :seq_txt, :]
         img_attn_output = joint_hidden_states[:, seq_txt:, :]
@@ -236,10 +234,10 @@ class QwenImageTransformerBlock(nn.Module):
         img_mod1, img_mod2 = img_mod_params.chunk(2, dim=-1)
         txt_mod1, txt_mod2 = txt_mod_params.chunk(2, dim=-1)
 
-        img_modulated, img_gate1 = self._modulate(self.img_norm1(hidden_states), img_mod1)
-        del img_mod1
-        txt_modulated, txt_gate1 = self._modulate(self.txt_norm1(encoder_hidden_states), txt_mod1)
-        del txt_mod1
+        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_attn_output, txt_attn_output = self.attn(
             hidden_states=img_modulated,
@@ -248,20 +246,16 @@ 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
         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_modulated2, img_gate2 = self._modulate(self.img_norm2(hidden_states), img_mod2)
+        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))
 
-        txt_modulated2, txt_gate2 = self._modulate(self.txt_norm2(encoder_hidden_states), txt_mod2)
+        txt_normed2 = self.txt_norm2(encoder_hidden_states)
+        txt_modulated2, txt_gate2 = self._modulate(txt_normed2, txt_mod2)
         encoder_hidden_states = torch.addcmul(encoder_hidden_states, txt_gate2, self.txt_mlp(txt_modulated2))
 
         return encoder_hidden_states, hidden_states
@@ -419,7 +413,7 @@ class QwenImageTransformer2DModel(nn.Module):
         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).to(x.dtype).contiguous()
+        image_rotary_emb = self.pe_embedder(ids).squeeze(1).unsqueeze(2).to(x.dtype)
         del ids, txt_ids, img_ids
 
         hidden_states = self.img_in(hidden_states)

comfy/ldm/wan/model.py

@@ -232,13 +232,11 @@ 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)
 
         x = torch.addcmul(x, y, repeat_e(e[2], x))
-        del y
 
         # cross-attention & ffn
         x = x + self.cross_attn(self.norm3(x), context, context_img_len=context_img_len, transformer_options=transformer_options)
@@ -589,7 +587,7 @@ class WanModel(torch.nn.Module):
         x = self.unpatchify(x, grid_sizes)
         return x
 
-    def rope_encode(self, t, h, w, t_start=0, steps_t=None, steps_h=None, steps_w=None, device=None, dtype=None, transformer_options={}):
+    def rope_encode(self, t, h, w, t_start=0, steps_t=None, steps_h=None, steps_w=None, device=None, dtype=None):
         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])
@@ -602,22 +600,10 @@ class WanModel(torch.nn.Module):
         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)
-
         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[:, :, :, 1] = img_ids[:, :, :, 1] + torch.linspace(0, h_len - 1, steps=steps_h, device=device, dtype=dtype).reshape(1, -1, 1)
+        img_ids[:, :, :, 2] = img_ids[:, :, :, 2] + torch.linspace(0, 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(img_ids).movedim(1, 2)
@@ -643,7 +629,7 @@ class WanModel(torch.nn.Module):
         if self.ref_conv is not None and "reference_latent" in kwargs:
             t_len += 1
 
-        freqs = self.rope_encode(t_len, h, w, device=x.device, dtype=x.dtype, transformer_options=transformer_options)
+        freqs = self.rope_encode(t_len, h, w, device=x.device, dtype=x.dtype)
         return self.forward_orig(x, timestep, context, clip_fea=clip_fea, freqs=freqs, transformer_options=transformer_options, **kwargs)[:, :, :t, :h, :w]
 
     def unpatchify(self, x, grid_sizes):
@@ -916,7 +902,7 @@ class MotionEncoder_tc(nn.Module):
     def __init__(self,
                  in_dim: int,
                  hidden_dim: int,
-                 num_heads: int,
+                 num_heads=int,
                  need_global=True,
                  dtype=None,
                  device=None,
@@ -1369,7 +1355,7 @@ class WanT2VCrossAttentionGather(WanSelfAttention):
 
         x = optimized_attention(q, k, v, heads=self.num_heads, skip_reshape=True, skip_output_reshape=True, transformer_options=transformer_options)
 
-        x = x.transpose(1, 2).reshape(b, -1, n * d)
+        x = x.transpose(1, 2).view(b, -1, n, d).flatten(2)
         x = self.o(x)
         return x
 

comfy/ldm/wan/vae.py

@@ -468,46 +468,55 @@ class WanVAE(nn.Module):
                                  attn_scales, self.temperal_upsample, dropout)
 
     def encode(self, x):
-        conv_idx = [0]
-        feat_map = [None] * count_conv3d(self.decoder)
+        self.clear_cache()
         ## cache
         t = x.shape[2]
         iter_ = 1 + (t - 1) // 4
         ## 对encode输入的x，按时间拆分为1、4、4、4....
         for i in range(iter_):
-            conv_idx = [0]
+            self._enc_conv_idx = [0]
             if i == 0:
                 out = self.encoder(
                     x[:, :, :1, :, :],
-                    feat_cache=feat_map,
-                    feat_idx=conv_idx)
+                    feat_cache=self._enc_feat_map,
+                    feat_idx=self._enc_conv_idx)
             else:
                 out_ = self.encoder(
                     x[:, :, 1 + 4 * (i - 1):1 + 4 * i, :, :],
-                    feat_cache=feat_map,
-                    feat_idx=conv_idx)
+                    feat_cache=self._enc_feat_map,
+                    feat_idx=self._enc_conv_idx)
                 out = torch.cat([out, out_], 2)
         mu, log_var = self.conv1(out).chunk(2, dim=1)
+        self.clear_cache()
         return mu
 
     def decode(self, z):
-        conv_idx = [0]
-        feat_map = [None] * count_conv3d(self.decoder)
+        self.clear_cache()
         # z: [b,c,t,h,w]
 
         iter_ = z.shape[2]
         x = self.conv2(z)
         for i in range(iter_):
-            conv_idx = [0]
+            self._conv_idx = [0]
             if i == 0:
                 out = self.decoder(
                     x[:, :, i:i + 1, :, :],
-                    feat_cache=feat_map,
-                    feat_idx=conv_idx)
+                    feat_cache=self._feat_map,
+                    feat_idx=self._conv_idx)
             else:
                 out_ = self.decoder(
                     x[:, :, i:i + 1, :, :],
-                    feat_cache=feat_map,
-                    feat_idx=conv_idx)
+                    feat_cache=self._feat_map,
+                    feat_idx=self._conv_idx)
                 out = torch.cat([out, out_], 2)
+        self.clear_cache()
         return out
+
+    def clear_cache(self):
+        self._conv_num = count_conv3d(self.decoder)
+        self._conv_idx = [0]
+        self._feat_map = [None] * self._conv_num
+        #cache encode
+        self._enc_conv_num = count_conv3d(self.encoder)
+        self._enc_conv_idx = [0]
+        self._enc_feat_map = [None] * self._enc_conv_num

comfy/ldm/wan/vae2_2.py

@@ -657,51 +657,51 @@ class WanVAE(nn.Module):
         )
 
     def encode(self, x):
-        conv_idx = [0]
-        feat_map = [None] * count_conv3d(self.encoder)
+        self.clear_cache()
         x = patchify(x, patch_size=2)
         t = x.shape[2]
         iter_ = 1 + (t - 1) // 4
         for i in range(iter_):
-            conv_idx = [0]
+            self._enc_conv_idx = [0]
             if i == 0:
                 out = self.encoder(
                     x[:, :, :1, :, :],
-                    feat_cache=feat_map,
-                    feat_idx=conv_idx,
+                    feat_cache=self._enc_feat_map,
+                    feat_idx=self._enc_conv_idx,
                 )
             else:
                 out_ = self.encoder(
                     x[:, :, 1 + 4 * (i - 1):1 + 4 * i, :, :],
-                    feat_cache=feat_map,
-                    feat_idx=conv_idx,
+                    feat_cache=self._enc_feat_map,
+                    feat_idx=self._enc_conv_idx,
                 )
                 out = torch.cat([out, out_], 2)
         mu, log_var = self.conv1(out).chunk(2, dim=1)
+        self.clear_cache()
         return mu
 
     def decode(self, z):
-        conv_idx = [0]
-        feat_map = [None] * count_conv3d(self.decoder)
+        self.clear_cache()
         iter_ = z.shape[2]
         x = self.conv2(z)
         for i in range(iter_):
-            conv_idx = [0]
+            self._conv_idx = [0]
             if i == 0:
                 out = self.decoder(
                     x[:, :, i:i + 1, :, :],
-                    feat_cache=feat_map,
-                    feat_idx=conv_idx,
+                    feat_cache=self._feat_map,
+                    feat_idx=self._conv_idx,
                     first_chunk=True,
                 )
             else:
                 out_ = self.decoder(
                     x[:, :, i:i + 1, :, :],
-                    feat_cache=feat_map,
-                    feat_idx=conv_idx,
+                    feat_cache=self._feat_map,
+                    feat_idx=self._conv_idx,
                 )
                 out = torch.cat([out, out_], 2)
         out = unpatchify(out, patch_size=2)
+        self.clear_cache()
         return out
 
     def reparameterize(self, mu, log_var):
@@ -715,3 +715,12 @@ class WanVAE(nn.Module):
             return mu
         std = torch.exp(0.5 * log_var.clamp(-30.0, 20.0))
         return mu + std * torch.randn_like(std)
+
+    def clear_cache(self):
+        self._conv_num = count_conv3d(self.decoder)
+        self._conv_idx = [0]
+        self._feat_map = [None] * self._conv_num
+        # cache encode
+        self._enc_conv_num = count_conv3d(self.encoder)
+        self._enc_conv_idx = [0]
+        self._enc_feat_map = [None] * self._enc_conv_num

comfy/model_base.py

@@ -134,11 +134,10 @@ 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, scaled_fp8=model_config.scaled_fp8)
             else:
                 operations = model_config.custom_operations
             self.diffusion_model = unet_model(**unet_config, device=device, operations=operations)
-            self.diffusion_model.eval()
             if comfy.model_management.force_channels_last():
                 self.diffusion_model.to(memory_format=torch.channels_last)
                 logging.debug("using channels last mode for diffusion model")
@@ -197,14 +196,8 @@ class BaseModel(torch.nn.Module):
             extra_conds[o] = extra
 
         t = self.process_timestep(t, x=x, **extra_conds)
-        if "latent_shapes" in extra_conds:
-            xc = utils.unpack_latents(xc, extra_conds.pop("latent_shapes"))
-
-        model_output = self.diffusion_model(xc, t, context=context, control=control, transformer_options=transformer_options, **extra_conds)
-        if len(model_output) > 1 and not torch.is_tensor(model_output):
-            model_output, _ = utils.pack_latents(model_output)
-
-        return self.model_sampling.calculate_denoised(sigma, model_output.float(), x)
+        model_output = self.diffusion_model(xc, t, context=context, control=control, transformer_options=transformer_options, **extra_conds).float()
+        return self.model_sampling.calculate_denoised(sigma, model_output, x)
 
     def process_timestep(self, timestep, **kwargs):
         return timestep
@@ -333,14 +326,6 @@ class BaseModel(torch.nn.Module):
         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:
@@ -684,6 +669,7 @@ class Lotus(BaseModel):
 class StableCascade_C(BaseModel):
     def __init__(self, model_config, model_type=ModelType.STABLE_CASCADE, device=None):
         super().__init__(model_config, model_type, device=device, unet_model=StageC)
+        self.diffusion_model.eval().requires_grad_(False)
 
     def extra_conds(self, **kwargs):
         out = {}
@@ -712,6 +698,7 @@ class StableCascade_C(BaseModel):
 class StableCascade_B(BaseModel):
     def __init__(self, model_config, model_type=ModelType.STABLE_CASCADE, device=None):
         super().__init__(model_config, model_type, device=device, unet_model=StageB)
+        self.diffusion_model.eval().requires_grad_(False)
 
     def extra_conds(self, **kwargs):
         out = {}
@@ -1536,94 +1523,3 @@ 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

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:
@@ -186,16 +172,6 @@ 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]
-        else:
-            dit_config["vision_in_dim"] = None
         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)
@@ -237,7 +213,7 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
                 dit_config["nerf_mlp_ratio"] = 4
                 dit_config["nerf_depth"] = 4
                 dit_config["nerf_max_freqs"] = 8
-                dit_config["nerf_tile_size"] = 512
+                dit_config["nerf_tile_size"] = 32
                 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
         else:
@@ -389,8 +365,8 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
         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]
-        dit_config["n_layers"] = count_blocks(state_dict_keys, '{}layers.'.format(key_prefix) + '{}.')
+        dit_config["cap_feat_dim"] = 2304
+        dit_config["n_layers"] = 26
         dit_config["n_heads"] = 24
         dit_config["n_kv_heads"] = 8
         dit_config["qk_norm"] = True
@@ -725,12 +701,6 @@ def model_config_from_unet(state_dict, unet_key_prefix, use_base_if_no_match=Fal
         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")
-
     return model_config
 
 def unet_prefix_from_state_dict(state_dict):

comfy/model_management.py

@@ -89,7 +89,6 @@ if args.deterministic:
 
 directml_enabled = False
 if args.directml is not None:
-    logging.warning("WARNING: torch-directml barely works, is very slow, has not been updated in over 1 year and might be removed soon, please don't use it, there are better options.")
     import torch_directml
     directml_enabled = True
     device_index = args.directml
@@ -331,21 +330,13 @@ except:
 
 
 SUPPORT_FP8_OPS = args.supports_fp8_compute
-
-AMD_RDNA2_AND_OLDER_ARCH = ["gfx1030", "gfx1031", "gfx1010", "gfx1011", "gfx1012", "gfx906", "gfx900", "gfx803"]
-
 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.")
-
         try:
             rocm_version = tuple(map(int, str(torch.version.hip).split(".")[:2]))
         except:
             rocm_version = (6, -1)
-
+        arch = torch.cuda.get_device_properties(get_torch_device()).gcnArchName
         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:
@@ -353,11 +344,11 @@ try:
                 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"]):
-                       ENABLE_PYTORCH_ATTENTION = True
+#                if torch_version_numeric >= (2, 8):
+#                    if any((a in arch) for a in ["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
+            if any((a in arch) for a in ["gfx1200", "gfx1201", "gfx942", "gfx950"]):  # TODO: more arches
                 SUPPORT_FP8_OPS = True
 
 except:
@@ -379,9 +370,6 @@ try:
 except:
     pass
 
-if torch.cuda.is_available() and torch.backends.cudnn.is_available() and PerformanceFeature.AutoTune in args.fast:
-    torch.backends.cudnn.benchmark = True
-
 try:
     if torch_version_numeric >= (2, 5):
         torch.backends.cuda.allow_fp16_bf16_reduction_math_sdp(True)
@@ -504,7 +492,6 @@ 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:
@@ -658,9 +645,7 @@ def load_models_gpu(models, memory_required=0, force_patch_weights=False, minimu
             if loaded_model.model.is_clone(current_loaded_models[i].model):
                 to_unload = [i] + to_unload
         for i in to_unload:
-            model_to_unload = current_loaded_models.pop(i)
-            model_to_unload.model.detach(unpatch_all=False)
-            model_to_unload.model_finalizer.detach()
+            current_loaded_models.pop(i).model.detach(unpatch_all=False)
 
     total_memory_required = {}
     for loaded_model in models_to_load:
@@ -690,10 +675,7 @@ def load_models_gpu(models, memory_required=0, force_patch_weights=False, minimu
             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 = lowvram_model_memory - loaded_memory
-
-            if lowvram_model_memory == 0:
-                lowvram_model_memory = 0.1
+            lowvram_model_memory = max(0.1, lowvram_model_memory - loaded_memory)
 
         if vram_set_state == VRAMState.NO_VRAM:
             lowvram_model_memory = 0.1
@@ -941,7 +923,11 @@ def vae_dtype(device=None, allowed_dtypes=[]):
         if d == torch.float16 and should_use_fp16(device):
             return d
 
-        if d == torch.bfloat16 and should_use_bf16(device):
+        # NOTE: bfloat16 seems to work on AMD for the VAE but is extremely slow in some cases compared to fp32
+        # slowness still a problem on pytorch nightly 2.9.0.dev20250720+rocm6.4 tested on RDNA3
+        # also a problem on RDNA4 except fp32 is also slow there.
+        # This is due to large bf16 convolutions being extremely slow.
+        if d == torch.bfloat16 and ((not is_amd()) or amd_min_version(device, min_rdna_version=4)) and should_use_bf16(device):
             return d
 
     return torch.float32
@@ -1003,6 +989,12 @@ def device_supports_non_blocking(device):
         return False
     return True
 
+def device_should_use_non_blocking(device):
+    if not device_supports_non_blocking(device):
+        return False
+    return False
+    # return True #TODO: figure out why this causes memory issues on Nvidia and possibly others
+
 def force_channels_last():
     if args.force_channels_last:
         return True
@@ -1017,16 +1009,6 @@ if args.async_offload:
     NUM_STREAMS = 2
     logging.info("Using async weight offloading with {} streams".format(NUM_STREAMS))
 
-def current_stream(device):
-    if device is None:
-        return None
-    if is_device_cuda(device):
-        return torch.cuda.current_stream()
-    elif is_device_xpu(device):
-        return torch.xpu.current_stream()
-    else:
-        return None
-
 stream_counters = {}
 def get_offload_stream(device):
     stream_counter = stream_counters.get(device, 0)
@@ -1035,17 +1017,21 @@ def get_offload_stream(device):
 
     if device in STREAMS:
         ss = STREAMS[device]
-        #Sync the oldest stream in the queue with the current
-        ss[stream_counter].wait_stream(current_stream(device))
+        s = ss[stream_counter]
         stream_counter = (stream_counter + 1) % len(ss)
+        if is_device_cuda(device):
+            ss[stream_counter].wait_stream(torch.cuda.current_stream())
+        elif is_device_xpu(device):
+            ss[stream_counter].wait_stream(torch.xpu.current_stream())
         stream_counters[device] = stream_counter
-        return ss[stream_counter]
+        return s
     elif is_device_cuda(device):
         ss = []
         for k in range(NUM_STREAMS):
             ss.append(torch.cuda.Stream(device=device, priority=0))
         STREAMS[device] = ss
         s = ss[stream_counter]
+        stream_counter = (stream_counter + 1) % len(ss)
         stream_counters[device] = stream_counter
         return s
     elif is_device_xpu(device):
@@ -1054,14 +1040,18 @@ def get_offload_stream(device):
             ss.append(torch.xpu.Stream(device=device, priority=0))
         STREAMS[device] = ss
         s = ss[stream_counter]
+        stream_counter = (stream_counter + 1) % len(ss)
         stream_counters[device] = stream_counter
         return s
     return None
 
 def sync_stream(device, stream):
-    if stream is None or current_stream(device) is None:
+    if stream is None:
         return
-    current_stream(device).wait_stream(stream)
+    if is_device_cuda(device):
+        torch.cuda.current_stream().wait_stream(stream)
+    elif is_device_xpu(device):
+        torch.xpu.current_stream().wait_stream(stream)
 
 def cast_to(weight, dtype=None, device=None, non_blocking=False, copy=False, stream=None):
     if device is None or weight.device == device:
@@ -1086,79 +1076,6 @@ 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)))
-
-
-def pin_memory(tensor):
-    global TOTAL_PINNED_MEMORY
-    if MAX_PINNED_MEMORY <= 0:
-        return False
-
-    if type(tensor) is not torch.nn.parameter.Parameter:
-        return False
-
-    if not is_device_cpu(tensor.device):
-        return False
-
-    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.numel() * tensor.element_size()
-    if (TOTAL_PINNED_MEMORY + size) > MAX_PINNED_MEMORY:
-        return False
-
-    ptr = tensor.data_ptr()
-    if torch.cuda.cudart().cudaHostRegister(ptr, size, 1) == 0:
-        PINNED_MEMORY[ptr] = size
-        TOTAL_PINNED_MEMORY += size
-        return True
-
-    return False
-
-def unpin_memory(tensor):
-    global TOTAL_PINNED_MEMORY
-    if MAX_PINNED_MEMORY <= 0:
-        return False
-
-    if not is_device_cpu(tensor.device):
-        return False
-
-    ptr = tensor.data_ptr()
-    size = tensor.numel() * tensor.element_size()
-
-    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
-
-    return False
-
 def sage_attention_enabled():
     return args.use_sage_attention
 
@@ -1411,7 +1328,7 @@ def should_use_bf16(device=None, model_params=0, prioritize_performance=True, ma
 
     if is_amd():
         arch = torch.cuda.get_device_properties(device).gcnArchName
-        if any((a in arch) for a in AMD_RDNA2_AND_OLDER_ARCH):  # RDNA2 and older don't support bf16
+        if any((a in arch) for a in ["gfx1030", "gfx1031", "gfx1010", "gfx1011", "gfx1012", "gfx906", "gfx900", "gfx803"]):  # RDNA2 and older don't support bf16
             if manual_cast:
                 return True
             return False

comfy/model_patcher.py

@@ -123,30 +123,16 @@ def move_weight_functions(m, device):
     return memory
 
 class LowVramPatch:
-    def __init__(self, key, patches, convert_func=None, set_func=None):
+    def __init__(self, key, patches):
         self.key = key
         self.patches = patches
-        self.convert_func = convert_func
-        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)
-
         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)
+            return comfy.float.stochastic_rounding(comfy.lora.calculate_weight(self.patches[self.key], weight.to(intermediate_dtype), self.key, intermediate_dtype=intermediate_dtype), weight.dtype, seed=string_to_seed(self.key))
 
-        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
+        return comfy.lora.calculate_weight(self.patches[self.key], weight, self.key, intermediate_dtype=intermediate_dtype)
 
 def get_key_weight(model, key):
     set_func = None
@@ -238,7 +224,6 @@ class ModelPatcher:
         self.force_cast_weights = False
         self.patches_uuid = uuid.uuid4()
         self.parent = None
-        self.pinned = set()
 
         self.attachments: dict[str] = {}
         self.additional_models: dict[str, list[ModelPatcher]] = {}
@@ -276,9 +261,6 @@ class ModelPatcher:
         self.size = comfy.model_management.module_size(self.model)
         return self.size
 
-    def get_ram_usage(self):
-        return self.model_size()
-
     def loaded_size(self):
         return self.model.model_loaded_weight_memory
 
@@ -298,7 +280,6 @@ class ModelPatcher:
         n.backup = self.backup
         n.object_patches_backup = self.object_patches_backup
         n.parent = self
-        n.pinned = self.pinned
 
         n.force_cast_weights = self.force_cast_weights
 
@@ -455,19 +436,6 @@ class ModelPatcher:
     def set_model_post_input_patch(self, patch):
         self.set_model_patch(patch, "post_input")
 
-    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
-        rope_options["scale_y"] = scale_y
-        rope_options["scale_t"] = scale_t
-
-        rope_options["shift_x"] = shift_x
-        rope_options["shift_y"] = shift_y
-        rope_options["shift_t"] = shift_t
-
-        self.model_options["transformer_options"]["rope_options"] = rope_options
-
-
     def add_object_patch(self, name, obj):
         self.object_patches[name] = obj
 
@@ -636,21 +604,6 @@ class ModelPatcher:
         else:
             set_func(out_weight, inplace_update=inplace_update, seed=string_to_seed(key))
 
-    def pin_weight_to_device(self, key):
-        weight, set_func, convert_func = get_key_weight(self.model, key)
-        if comfy.model_management.pin_memory(weight):
-            self.pinned.add(key)
-
-    def unpin_weight(self, key):
-        if key in self.pinned:
-            weight, set_func, convert_func = get_key_weight(self.model, key)
-            comfy.model_management.unpin_memory(weight)
-            self.pinned.remove(key)
-
-    def unpin_all_weights(self):
-        for key in list(self.pinned):
-            self.unpin_weight(key)
-
     def _load_list(self):
         loading = []
         for n, m in self.model.named_modules():
@@ -672,11 +625,9 @@ class ModelPatcher:
             mem_counter = 0
             patch_counter = 0
             lowvram_counter = 0
-            lowvram_mem_counter = 0
             loading = self._load_list()
 
             load_completely = []
-            offloaded = []
             loading.sort(reverse=True)
             for x in loading:
                 n = x[1]
@@ -693,7 +644,6 @@ class ModelPatcher:
                     if mem_counter + module_mem >= lowvram_model_memory:
                         lowvram_weight = True
                         lowvram_counter += 1
-                        lowvram_mem_counter += module_mem
                         if hasattr(m, "prev_comfy_cast_weights"): #Already lowvramed
                             continue
 
@@ -707,19 +657,16 @@ class ModelPatcher:
                         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 = [LowVramPatch(weight_key, self.patches, convert_func, set_func)]
+                            m.weight_function = [LowVramPatch(weight_key, self.patches)]
                             patch_counter += 1
                     if bias_key in self.patches:
                         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 = [LowVramPatch(bias_key, self.patches, convert_func, set_func)]
+                            m.bias_function = [LowVramPatch(bias_key, self.patches)]
                             patch_counter += 1
 
                     cast_weight = True
-                    offloaded.append((module_mem, n, m, params))
                 else:
                     if hasattr(m, "comfy_cast_weights"):
                         wipe_lowvram_weight(m)
@@ -750,9 +697,7 @@ class ModelPatcher:
                         continue
 
                 for param in params:
-                    key = "{}.{}".format(n, param)
-                    self.unpin_weight(key)
-                    self.patch_weight_to_device(key, device_to=device_to)
+                    self.patch_weight_to_device("{}.{}".format(n, param), device_to=device_to)
 
                 logging.debug("lowvram: loaded module regularly {} {}".format(n, m))
                 m.comfy_patched_weights = True
@@ -760,17 +705,11 @@ class ModelPatcher:
             for x in load_completely:
                 x[2].to(device_to)
 
-            for x in offloaded:
-                n = x[1]
-                params = x[3]
-                for param in params:
-                    self.pin_weight_to_device("{}.{}".format(n, param))
-
             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 {} {} {}".format(lowvram_model_memory / (1024 * 1024), mem_counter / (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 {} {} {}".format(lowvram_model_memory / (1024 * 1024), mem_counter / (1024 * 1024), full_load))
                 self.model.model_lowvram = False
                 if full_load:
                     self.model.to(device_to)
@@ -807,7 +746,6 @@ class ModelPatcher:
         self.eject_model()
         if unpatch_weights:
             self.unpatch_hooks()
-            self.unpin_all_weights()
             if self.model.model_lowvram:
                 for m in self.model.modules():
                     move_weight_functions(m, device_to)
@@ -843,7 +781,7 @@ class ModelPatcher:
 
         self.object_patches_backup.clear()
 
-    def partially_unload(self, device_to, memory_to_free=0, force_patch_weights=False):
+    def partially_unload(self, device_to, memory_to_free=0):
         with self.use_ejected():
             hooks_unpatched = False
             memory_freed = 0
@@ -887,19 +825,11 @@ class ModelPatcher:
                         module_mem += move_weight_functions(m, device_to)
                         if lowvram_possible:
                             if weight_key in self.patches:
-                                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
+                                m.weight_function.append(LowVramPatch(weight_key, self.patches))
+                                patch_counter += 1
                             if bias_key in self.patches:
-                                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
+                                m.bias_function.append(LowVramPatch(bias_key, self.patches))
+                                patch_counter += 1
                             cast_weight = True
 
                         if cast_weight:
@@ -909,13 +839,9 @@ class ModelPatcher:
                         memory_freed += module_mem
                         logging.debug("freed {}".format(n))
 
-                        for param in params:
-                            self.pin_weight_to_device("{}.{}".format(n, param))
-
             self.model.model_lowvram = True
             self.model.lowvram_patch_counter += patch_counter
             self.model.model_loaded_weight_memory -= memory_freed
-            logging.info("loaded partially: {:.2f} MB loaded, lowvram patches: {}".format(self.model.model_loaded_weight_memory / (1024 * 1024), self.model.lowvram_patch_counter))
             return memory_freed
 
     def partially_load(self, device_to, extra_memory=0, force_patch_weights=False):
@@ -928,9 +854,6 @@ 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)
@@ -1318,6 +1241,5 @@ class ModelPatcher:
         self.clear_cached_hook_weights()
 
     def __del__(self):
-        self.unpin_all_weights()
         self.detach(unpatch_all=False)
 

comfy/model_sampling.py

@@ -21,23 +21,17 @@ def rescale_zero_terminal_snr_sigmas(sigmas):
     alphas_bar[-1] = 4.8973451890853435e-08
     return ((1 - alphas_bar) / alphas_bar) ** 0.5
 
-def reshape_sigma(sigma, noise_dim):
-    if sigma.nelement() == 1:
-        return sigma.view(())
-    else:
-        return sigma.view(sigma.shape[:1] + (1,) * (noise_dim - 1))
-
 class EPS:
     def calculate_input(self, sigma, noise):
-        sigma = reshape_sigma(sigma, noise.ndim)
+        sigma = sigma.view(sigma.shape[:1] + (1,) * (noise.ndim - 1))
         return noise / (sigma ** 2 + self.sigma_data ** 2) ** 0.5
 
     def calculate_denoised(self, sigma, model_output, model_input):
-        sigma = reshape_sigma(sigma, model_output.ndim)
+        sigma = sigma.view(sigma.shape[:1] + (1,) * (model_output.ndim - 1))
         return model_input - model_output * sigma
 
     def noise_scaling(self, sigma, noise, latent_image, max_denoise=False):
-        sigma = reshape_sigma(sigma, noise.ndim)
+        sigma = sigma.view(sigma.shape[:1] + (1,) * (noise.ndim - 1))
         if max_denoise:
             noise = noise * torch.sqrt(1.0 + sigma ** 2.0)
         else:
@@ -51,12 +45,12 @@ class EPS:
 
 class V_PREDICTION(EPS):
     def calculate_denoised(self, sigma, model_output, model_input):
-        sigma = reshape_sigma(sigma, model_output.ndim)
+        sigma = sigma.view(sigma.shape[:1] + (1,) * (model_output.ndim - 1))
         return model_input * self.sigma_data ** 2 / (sigma ** 2 + self.sigma_data ** 2) - model_output * sigma * self.sigma_data / (sigma ** 2 + self.sigma_data ** 2) ** 0.5
 
 class EDM(V_PREDICTION):
     def calculate_denoised(self, sigma, model_output, model_input):
-        sigma = reshape_sigma(sigma, model_output.ndim)
+        sigma = sigma.view(sigma.shape[:1] + (1,) * (model_output.ndim - 1))
         return model_input * self.sigma_data ** 2 / (sigma ** 2 + self.sigma_data ** 2) + model_output * sigma * self.sigma_data / (sigma ** 2 + self.sigma_data ** 2) ** 0.5
 
 class CONST:
@@ -64,15 +58,15 @@ class CONST:
         return noise
 
     def calculate_denoised(self, sigma, model_output, model_input):
-        sigma = reshape_sigma(sigma, model_output.ndim)
+        sigma = sigma.view(sigma.shape[:1] + (1,) * (model_output.ndim - 1))
         return model_input - model_output * sigma
 
     def noise_scaling(self, sigma, noise, latent_image, max_denoise=False):
-        sigma = reshape_sigma(sigma, noise.ndim)
+        sigma = sigma.view(sigma.shape[:1] + (1,) * (noise.ndim - 1))
         return sigma * noise + (1.0 - sigma) * latent_image
 
     def inverse_noise_scaling(self, sigma, latent):
-        sigma = reshape_sigma(sigma, latent.ndim)
+        sigma = sigma.view(sigma.shape[:1] + (1,) * (latent.ndim - 1))
         return latent / (1.0 - sigma)
 
 class X0(EPS):
@@ -86,16 +80,16 @@ class IMG_TO_IMG(X0):
 class COSMOS_RFLOW:
     def calculate_input(self, sigma, noise):
         sigma = (sigma / (sigma + 1))
-        sigma = reshape_sigma(sigma, noise.ndim)
+        sigma = sigma.view(sigma.shape[:1] + (1,) * (noise.ndim - 1))
         return noise * (1.0 - sigma)
 
     def calculate_denoised(self, sigma, model_output, model_input):
         sigma = (sigma / (sigma + 1))
-        sigma = reshape_sigma(sigma, model_output.ndim)
+        sigma = sigma.view(sigma.shape[:1] + (1,) * (model_output.ndim - 1))
         return model_input * (1.0 - sigma) - model_output * sigma
 
     def noise_scaling(self, sigma, noise, latent_image, max_denoise=False):
-        sigma = reshape_sigma(sigma, noise.ndim)
+        sigma = sigma.view(sigma.shape[:1] + (1,) * (noise.ndim - 1))
         noise = noise * sigma
         noise += latent_image
         return noise

comfy/nested_tensor.py

@@ -1,91 +0,0 @@
-import torch
-
-class NestedTensor:
-    def __init__(self, tensors):
-        self.tensors = list(tensors)
-        self.is_nested = True
-
-    def _copy(self):
-        return NestedTensor(self.tensors)
-
-    def apply_operation(self, other, operation):
-        o = self._copy()
-        if isinstance(other, NestedTensor):
-            for i, t in enumerate(o.tensors):
-                o.tensors[i] = operation(t, other.tensors[i])
-        else:
-            for i, t in enumerate(o.tensors):
-                o.tensors[i] = operation(t, other)
-        return o
-
-    def __add__(self, b):
-        return self.apply_operation(b, lambda x, y: x + y)
-
-    def __sub__(self, b):
-        return self.apply_operation(b, lambda x, y: x - y)
-
-    def __mul__(self, b):
-        return self.apply_operation(b, lambda x, y: x * y)
-
-    # def __itruediv__(self, b):
-    #     return self.apply_operation(b, lambda x, y: x / y)
-
-    def __truediv__(self, b):
-        return self.apply_operation(b, lambda x, y: x / y)
-
-    def __getitem__(self, *args, **kwargs):
-        return self.apply_operation(None, lambda x, y: x.__getitem__(*args, **kwargs))
-
-    def unbind(self):
-        return self.tensors
-
-    def to(self, *args, **kwargs):
-        o = self._copy()
-        for i, t in enumerate(o.tensors):
-            o.tensors[i] = t.to(*args, **kwargs)
-        return o
-
-    def new_ones(self, *args, **kwargs):
-        return self.tensors[0].new_ones(*args, **kwargs)
-
-    def float(self):
-        return self.to(dtype=torch.float)
-
-    def chunk(self, *args, **kwargs):
-        return self.apply_operation(None, lambda x, y: x.chunk(*args, **kwargs))
-
-    def size(self):
-        return self.tensors[0].size()
-
-    @property
-    def shape(self):
-        return self.tensors[0].shape
-
-    @property
-    def ndim(self):
-        dims = 0
-        for t in self.tensors:
-            dims = max(t.ndim, dims)
-        return dims
-
-    @property
-    def device(self):
-        return self.tensors[0].device
-
-    @property
-    def dtype(self):
-        return self.tensors[0].dtype
-
-    @property
-    def layout(self):
-        return self.tensors[0].layout
-
-
-def cat_nested(tensors, *args, **kwargs):
-    cated_tensors = []
-    for i in range(len(tensors[0].tensors)):
-        tens = []
-        for j in range(len(tensors)):
-            tens.append(tensors[j].tensors[i])
-        cated_tensors.append(torch.cat(tens, *args, **kwargs))
-    return NestedTensor(cated_tensors)

comfy/ops.py

@@ -24,18 +24,13 @@ import comfy.float
 import comfy.rmsnorm
 import contextlib
 
-def run_every_op():
-    if torch.compiler.is_compiling():
-        return
-
-    comfy.model_management.throw_exception_if_processing_interrupted()
 
 def scaled_dot_product_attention(q, k, v, *args, **kwargs):
     return torch.nn.functional.scaled_dot_product_attention(q, k, v, *args, **kwargs)
 
 
 try:
-    if torch.cuda.is_available() and comfy.model_management.WINDOWS:
+    if torch.cuda.is_available():
         from torch.nn.attention import SDPBackend, sdpa_kernel
         import inspect
         if "set_priority" in inspect.signature(sdpa_kernel).parameters:
@@ -55,90 +50,49 @@ try:
 except (ModuleNotFoundError, TypeError):
     logging.warning("Could not set sdpa backend priority.")
 
-NVIDIA_MEMORY_CONV_BUG_WORKAROUND = False
-try:
-    if comfy.model_management.is_nvidia():
-        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.")
-except:
-    pass
-
 cast_to = comfy.model_management.cast_to #TODO: remove once no more references
 
+if torch.cuda.is_available() and torch.backends.cudnn.is_available() and PerformanceFeature.AutoTune in args.fast:
+    torch.backends.cudnn.benchmark = True
+
 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):
-    # 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.
+def cast_bias_weight(s, input=None, dtype=None, device=None, bias_dtype=None):
     if input is not None:
         if dtype is None:
-            if isinstance(input, QuantizedTensor):
-                dtype = input._layout_params["orig_dtype"]
-            else:
-                dtype = input.dtype
+            dtype = input.dtype
         if bias_dtype is None:
             bias_dtype = dtype
         if device is None:
             device = input.device
 
-    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
-
+    offload_stream = comfy.model_management.get_offload_stream(device)
     if offload_stream is not None:
         wf_context = offload_stream
     else:
         wf_context = contextlib.nullcontext()
 
-    non_blocking = comfy.model_management.device_supports_non_blocking(device)
-
-    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)
-
     bias = None
+    non_blocking = comfy.model_management.device_supports_non_blocking(device)
     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)
+        has_function = len(s.bias_function) > 0
+        bias = comfy.model_management.cast_to(s.bias, bias_dtype, device, non_blocking=non_blocking, copy=has_function, stream=offload_stream)
 
-        if bias_has_function:
+        if has_function:
             with wf_context:
                 for f in s.bias_function:
                     bias = f(bias)
 
-    if weight_has_function or weight.dtype != dtype:
+    has_function = len(s.weight_function) > 0
+    weight = comfy.model_management.cast_to(s.weight, dtype, device, non_blocking=non_blocking, copy=has_function, stream=offload_stream)
+    if has_function:
         with wf_context:
-            weight = weight.to(dtype=dtype)
             for f in s.weight_function:
                 weight = f(weight)
 
     comfy.model_management.sync_stream(device, offload_stream)
-    if offloadable:
-        return weight, bias, offload_stream
-    else:
-        #Legacy function signature
-        return weight, bias
-
-
-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))
-
+    return weight, bias
 
 class CastWeightBiasOp:
     comfy_cast_weights = False
@@ -151,13 +105,10 @@ class disable_weight_init:
             return None
 
         def forward_comfy_cast_weights(self, input):
-            weight, bias, offload_stream = cast_bias_weight(self, input, offloadable=True)
-            x = torch.nn.functional.linear(input, weight, bias)
-            uncast_bias_weight(self, weight, bias, offload_stream)
-            return x
+            weight, bias = cast_bias_weight(self, input)
+            return torch.nn.functional.linear(input, weight, bias)
 
         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:
                 return self.forward_comfy_cast_weights(*args, **kwargs)
             else:
@@ -168,13 +119,10 @@ class disable_weight_init:
             return None
 
         def forward_comfy_cast_weights(self, input):
-            weight, bias, offload_stream = cast_bias_weight(self, input, offloadable=True)
-            x = self._conv_forward(input, weight, bias)
-            uncast_bias_weight(self, weight, bias, offload_stream)
-            return x
+            weight, bias = cast_bias_weight(self, input)
+            return self._conv_forward(input, weight, bias)
 
         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:
                 return self.forward_comfy_cast_weights(*args, **kwargs)
             else:
@@ -185,13 +133,10 @@ class disable_weight_init:
             return None
 
         def forward_comfy_cast_weights(self, input):
-            weight, bias, offload_stream = cast_bias_weight(self, input, offloadable=True)
-            x = self._conv_forward(input, weight, bias)
-            uncast_bias_weight(self, weight, bias, offload_stream)
-            return x
+            weight, bias = cast_bias_weight(self, input)
+            return self._conv_forward(input, weight, bias)
 
         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:
                 return self.forward_comfy_cast_weights(*args, **kwargs)
             else:
@@ -201,23 +146,11 @@ class disable_weight_init:
         def reset_parameters(self):
             return None
 
-        def _conv_forward(self, input, weight, bias, *args, **kwargs):
-            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:
-                    out += bias.reshape((1, -1) + (1,) * (out.ndim - 2))
-                return out
-            else:
-                return super()._conv_forward(input, weight, bias, *args, **kwargs)
-
         def forward_comfy_cast_weights(self, input):
-            weight, bias, offload_stream = cast_bias_weight(self, input, offloadable=True)
-            x = self._conv_forward(input, weight, bias)
-            uncast_bias_weight(self, weight, bias, offload_stream)
-            return x
+            weight, bias = cast_bias_weight(self, input)
+            return self._conv_forward(input, weight, bias)
 
         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:
                 return self.forward_comfy_cast_weights(*args, **kwargs)
             else:
@@ -228,13 +161,10 @@ class disable_weight_init:
             return None
 
         def forward_comfy_cast_weights(self, input):
-            weight, bias, offload_stream = cast_bias_weight(self, input, offloadable=True)
-            x = torch.nn.functional.group_norm(input, self.num_groups, weight, bias, self.eps)
-            uncast_bias_weight(self, weight, bias, offload_stream)
-            return x
+            weight, bias = cast_bias_weight(self, input)
+            return torch.nn.functional.group_norm(input, self.num_groups, weight, bias, self.eps)
 
         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:
                 return self.forward_comfy_cast_weights(*args, **kwargs)
             else:
@@ -246,17 +176,13 @@ class disable_weight_init:
 
         def forward_comfy_cast_weights(self, input):
             if self.weight is not None:
-                weight, bias, offload_stream = cast_bias_weight(self, input, offloadable=True)
+                weight, bias = cast_bias_weight(self, input)
             else:
                 weight = None
                 bias = None
-                offload_stream = None
-            x = torch.nn.functional.layer_norm(input, self.normalized_shape, weight, bias, self.eps)
-            uncast_bias_weight(self, weight, bias, offload_stream)
-            return x
+            return torch.nn.functional.layer_norm(input, self.normalized_shape, weight, bias, self.eps)
 
         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:
                 return self.forward_comfy_cast_weights(*args, **kwargs)
             else:
@@ -269,18 +195,13 @@ class disable_weight_init:
 
         def forward_comfy_cast_weights(self, input):
             if self.weight is not None:
-                weight, bias, offload_stream = cast_bias_weight(self, input, offloadable=True)
+                weight, bias = cast_bias_weight(self, input)
             else:
                 weight = None
-                bias = None
-                offload_stream = None
-            x = comfy.rmsnorm.rms_norm(input, weight, self.eps)  # TODO: switch to commented out line when old torch is deprecated
-            # x = torch.nn.functional.rms_norm(input, self.normalized_shape, weight, self.eps)
-            uncast_bias_weight(self, weight, bias, offload_stream)
-            return x
+            return comfy.rmsnorm.rms_norm(input, weight, self.eps)  # TODO: switch to commented out line when old torch is deprecated
+            # return torch.nn.functional.rms_norm(input, self.normalized_shape, weight, self.eps)
 
         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:
                 return self.forward_comfy_cast_weights(*args, **kwargs)
             else:
@@ -296,15 +217,12 @@ class disable_weight_init:
                 input, output_size, self.stride, self.padding, self.kernel_size,
                 num_spatial_dims, self.dilation)
 
-            weight, bias, offload_stream = cast_bias_weight(self, input, offloadable=True)
-            x = torch.nn.functional.conv_transpose2d(
+            weight, bias = cast_bias_weight(self, input)
+            return torch.nn.functional.conv_transpose2d(
                 input, weight, bias, self.stride, self.padding,
                 output_padding, self.groups, self.dilation)
-            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:
                 return self.forward_comfy_cast_weights(*args, **kwargs)
             else:
@@ -320,15 +238,12 @@ class disable_weight_init:
                 input, output_size, self.stride, self.padding, self.kernel_size,
                 num_spatial_dims, self.dilation)
 
-            weight, bias, offload_stream = cast_bias_weight(self, input, offloadable=True)
-            x = torch.nn.functional.conv_transpose1d(
+            weight, bias = cast_bias_weight(self, input)
+            return torch.nn.functional.conv_transpose1d(
                 input, weight, bias, self.stride, self.padding,
                 output_padding, self.groups, self.dilation)
-            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:
                 return self.forward_comfy_cast_weights(*args, **kwargs)
             else:
@@ -343,14 +258,10 @@ class disable_weight_init:
             output_dtype = out_dtype
             if self.weight.dtype == torch.float16 or self.weight.dtype == torch.bfloat16:
                 out_dtype = None
-            weight, bias, offload_stream = cast_bias_weight(self, device=input.device, dtype=out_dtype, offloadable=True)
-            x = torch.nn.functional.embedding(input, weight, self.padding_idx, self.max_norm, self.norm_type, self.scale_grad_by_freq, self.sparse).to(dtype=output_dtype)
-            uncast_bias_weight(self, weight, bias, offload_stream)
-            return x
-
+            weight, bias = cast_bias_weight(self, device=input.device, dtype=out_dtype)
+            return torch.nn.functional.embedding(input, weight, self.padding_idx, self.max_norm, self.norm_type, self.scale_grad_by_freq, self.sparse).to(dtype=output_dtype)
 
         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:
                 return self.forward_comfy_cast_weights(*args, **kwargs)
             else:
@@ -401,18 +312,20 @@ class manual_cast(disable_weight_init):
 
 
 def fp8_linear(self, input):
-    """
-    Legacy FP8 linear function for backward compatibility.
-    Uses QuantizedTensor subclass for dispatch.
-    """
     dtype = self.weight.dtype
     if dtype not in [torch.float8_e4m3fn]:
         return None
 
-    input_dtype = input.dtype
+    tensor_2d = False
+    if len(input.shape) == 2:
+        tensor_2d = True
+        input = input.unsqueeze(1)
 
-    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)
+    input_shape = input.shape
+    input_dtype = input.dtype
+    if len(input.shape) == 3:
+        w, bias = cast_bias_weight(self, input, dtype=dtype, bias_dtype=input_dtype)
+        w = w.t()
 
         scale_weight = self.scale_weight
         scale_input = self.scale_input
@@ -424,20 +337,23 @@ def fp8_linear(self, input):
         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)
+            input = input.reshape(-1, input_shape[2]).to(dtype).contiguous()
         else:
             scale_input = scale_input.to(input.device)
-            quantized_input = QuantizedTensor.from_float(input, "TensorCoreFP8Layout", scale=scale_input, dtype=dtype)
+            input = (input * (1.0 / scale_input).to(input_dtype)).reshape(-1, input_shape[2]).to(dtype).contiguous()
 
-        # 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)
+        if bias is not None:
+            o = torch._scaled_mm(input, w, out_dtype=input_dtype, bias=bias, scale_a=scale_input, scale_b=scale_weight)
+        else:
+            o = torch._scaled_mm(input, w, out_dtype=input_dtype, scale_a=scale_input, scale_b=scale_weight)
 
-        uncast_bias_weight(self, w, bias, offload_stream)
-        return o
+        if isinstance(o, tuple):
+            o = o[0]
+
+        if tensor_2d:
+            return o.reshape(input_shape[0], -1)
+
+        return o.reshape((-1, input_shape[1], self.weight.shape[0]))
 
     return None
 
@@ -457,10 +373,8 @@ class fp8_ops(manual_cast):
                 except Exception as e:
                     logging.info("Exception during fp8 op: {}".format(e))
 
-            weight, bias, offload_stream = cast_bias_weight(self, input, offloadable=True)
-            x = torch.nn.functional.linear(input, weight, bias)
-            uncast_bias_weight(self, weight, bias, offload_stream)
-            return x
+            weight, bias = cast_bias_weight(self, input)
+            return torch.nn.functional.linear(input, weight, bias)
 
 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))
@@ -488,14 +402,12 @@ def scaled_fp8_ops(fp8_matrix_mult=False, scale_input=False, override_dtype=None
                     if out is not None:
                         return out
 
-                weight, bias, offload_stream = cast_bias_weight(self, input, offloadable=True)
+                weight, bias = cast_bias_weight(self, input)
 
                 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)
+                    return 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
+                    return torch.nn.functional.linear(input * self.scale_weight.to(device=weight.device, dtype=weight.dtype), weight, bias)
 
             def convert_weight(self, weight, inplace=False, **kwargs):
                 if inplace:
@@ -504,10 +416,8 @@ def scaled_fp8_ops(fp8_matrix_mult=False, scale_input=False, override_dtype=None
                 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):
+            def set_weight(self, weight, inplace_update=False, seed=None, **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:
@@ -534,120 +444,7 @@ if CUBLAS_IS_AVAILABLE:
             def forward(self, *args, **kwargs):
                 return super().forward(*args, **kwargs)
 
-
-# ==============================================================================
-# Mixed Precision Operations
-# ==============================================================================
-from .quant_ops import QuantizedTensor, QUANT_ALGOS
-
-class MixedPrecisionOps(disable_weight_init):
-    _layer_quant_config = {}
-    _compute_dtype = torch.bfloat16
-
-    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.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
-
-        def reset_parameters(self):
-            return None
-
-        def _load_from_state_dict(self, state_dict, prefix, local_metadata,
-                                  strict, missing_keys, unexpected_keys, error_msgs):
-
-            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}")
-
-            manually_loaded_keys = [weight_key]
-
-            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}")
-
-                qconfig = QUANT_ALGOS[quant_format]
-                self.layout_type = qconfig["comfy_tensor_layout"]
-
-                weight_scale_key = f"{prefix}weight_scale"
-                layout_params = {
-                    'scale': state_dict.pop(weight_scale_key, None),
-                    'orig_dtype': MixedPrecisionOps._compute_dtype,
-                    'block_size': qconfig.get("group_size", None),
-                }
-                if layout_params['scale'] is not None:
-                    manually_loaded_keys.append(weight_scale_key)
-
-                self.weight = torch.nn.Parameter(
-                    QuantizedTensor(weight.to(device=device), self.layout_type, layout_params),
-                    requires_grad=False
-                )
-
-                for param_name in qconfig["parameters"]:
-                    param_key = f"{prefix}{param_name}"
-                    _v = state_dict.pop(param_key, None)
-                    if _v is None:
-                        continue
-                    setattr(self, 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 _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, 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
-
+def pick_operations(weight_dtype, compute_dtype, load_device=None, disable_fast_fp8=False, fp8_optimizations=False, scaled_fp8=None):
     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)

comfy/patcher_extension.py

@@ -150,7 +150,7 @@ def merge_nested_dicts(dict1: dict, dict2: dict, copy_dict1=True):
     for key, value in dict2.items():
         if isinstance(value, dict):
             curr_value = merged_dict.setdefault(key, {})
-            merged_dict[key] = merge_nested_dicts(curr_value, value)
+            merged_dict[key] = merge_nested_dicts(value, curr_value)
         elif isinstance(value, list):
             merged_dict.setdefault(key, []).extend(value)
         else:

comfy/quant_ops.py

@@ -1,545 +0,0 @@
-import torch
-import logging
-from typing import Tuple, Dict
-
-_LAYOUT_REGISTRY = {}
-_GENERIC_UTILS = {}
-
-
-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
-
-
-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)
-
-    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 _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
-
-def _copy_layout_params_inplace(src, dst, non_blocking=False):
-    for k, v in src.items():
-        if isinstance(v, torch.Tensor):
-            dst[k].copy_(v, non_blocking=non_blocking)
-        else:
-            dst[k] = v
-
-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)
-
-
-# ==============================================================================
-# Generic Utilities (Layout-Agnostic Operations)
-# ==============================================================================
-
-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)
-
-
-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]
-    non_blocking = args[2] if len(args) > 2 else False
-    if isinstance(qt_dest, QuantizedTensor):
-        if isinstance(src, QuantizedTensor):
-            # Copy from another quantized tensor
-            qt_dest._qdata.copy_(src._qdata, non_blocking=non_blocking)
-            qt_dest._layout_type = src._layout_type
-            _copy_layout_params_inplace(src._layout_params, qt_dest._layout_params, non_blocking=non_blocking)
-        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
-
-
-@register_generic_util(torch.ops.aten.empty_like.default)
-def generic_empty_like(func, args, kwargs):
-    """Empty_like operation - creates an empty tensor with the same quantized structure."""
-    qt = args[0]
-    if isinstance(qt, QuantizedTensor):
-        # Create empty tensor with same shape and dtype as the quantized data
-        hp_dtype = kwargs.pop('dtype', qt._layout_params["orig_dtype"])
-        new_qdata = torch.empty_like(qt._qdata, **kwargs)
-
-        # Handle device transfer for layout params
-        target_device = kwargs.get('device', new_qdata.device)
-        new_params = _move_layout_params_to_device(qt._layout_params, target_device)
-
-        # Update orig_dtype if dtype is specified
-        new_params['orig_dtype'] = hp_dtype
-
-        return QuantizedTensor(new_qdata, qt._layout_type, new_params)
-    return func(*args, **kwargs)
-
-# ==============================================================================
-# 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):
-        orig_dtype = tensor.dtype
-
-        if scale is None:
-            scale = torch.amax(tensor.abs()) / torch.finfo(dtype).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)
-
-        layout_params = {
-            'scale': scale,
-            'orig_dtype': orig_dtype
-        }
-        return qdata, layout_params
-
-    @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']
-
-QUANT_ALGOS = {
-    "float8_e4m3fn": {
-        "storage_t": torch.float8_e4m3fn,
-        "parameters": {"weight_scale", "input_scale"},
-        "comfy_tensor_layout": "TensorCoreFP8Layout",
-    },
-}
-
-LAYOUTS = {
-    "TensorCoreFP8Layout": TensorCoreFP8Layout,
-}
-
-
-@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
-
-    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)

comfy/sample.py

@@ -4,9 +4,13 @@ import comfy.samplers
 import comfy.utils
 import numpy as np
 import logging
-import comfy.nested_tensor
 
-def prepare_noise_inner(latent_image, generator, noise_inds=None):
+def prepare_noise(latent_image, seed, noise_inds=None):
+    """
+    creates random noise given a latent image and a seed.
+    optional arg skip can be used to skip and discard x number of noise generations for a given seed
+    """
+    generator = torch.manual_seed(seed)
     if noise_inds is None:
         return torch.randn(latent_image.size(), dtype=latent_image.dtype, layout=latent_image.layout, generator=generator, device="cpu")
 
@@ -17,29 +21,10 @@ def prepare_noise_inner(latent_image, generator, noise_inds=None):
         if i in unique_inds:
             noises.append(noise)
     noises = [noises[i] for i in inverse]
-    return torch.cat(noises, axis=0)
-
-def prepare_noise(latent_image, seed, noise_inds=None):
-    """
-    creates random noise given a latent image and a seed.
-    optional arg skip can be used to skip and discard x number of noise generations for a given seed
-    """
-    generator = torch.manual_seed(seed)
-
-    if latent_image.is_nested:
-        tensors = latent_image.unbind()
-        noises = []
-        for t in tensors:
-            noises.append(prepare_noise_inner(t, generator, noise_inds))
-        noises = comfy.nested_tensor.NestedTensor(noises)
-    else:
-        noises = prepare_noise_inner(latent_image, generator, noise_inds)
-
+    noises = torch.cat(noises, axis=0)
     return noises
 
 def fix_empty_latent_channels(model, latent_image):
-    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)

comfy/samplers.py

@@ -306,10 +306,17 @@ def _calc_cond_batch(model: BaseModel, conds: list[list[dict]], x_in: torch.Tens
                                                                                  copy_dict1=False)
 
             if patches is not None:
-                transformer_options["patches"] = comfy.patcher_extension.merge_nested_dicts(
-                    transformer_options.get("patches", {}),
-                    patches
-                )
+                # TODO: replace with merge_nested_dicts function
+                if "patches" in transformer_options:
+                    cur_patches = transformer_options["patches"].copy()
+                    for p in patches:
+                        if p in cur_patches:
+                            cur_patches[p] = cur_patches[p] + patches[p]
+                        else:
+                            cur_patches[p] = patches[p]
+                    transformer_options["patches"] = cur_patches
+                else:
+                    transformer_options["patches"] = patches
 
             transformer_options["cond_or_uncond"] = cond_or_uncond[:]
             transformer_options["uuids"] = uuids[:]
@@ -353,7 +360,7 @@ def calc_cond_uncond_batch(model, cond, uncond, x_in, timestep, model_options):
 def cfg_function(model, cond_pred, uncond_pred, cond_scale, x, timestep, model_options={}, cond=None, uncond=None):
     if "sampler_cfg_function" in model_options:
         args = {"cond": x - cond_pred, "uncond": x - uncond_pred, "cond_scale": cond_scale, "timestep": timestep, "input": x, "sigma": timestep,
-                "cond_denoised": cond_pred, "uncond_denoised": uncond_pred, "model": model, "model_options": model_options, "input_cond": cond, "input_uncond": uncond}
+                "cond_denoised": cond_pred, "uncond_denoised": uncond_pred, "model": model, "model_options": model_options}
         cfg_result = x - model_options["sampler_cfg_function"](args)
     else:
         cfg_result = uncond_pred + (cond_pred - uncond_pred) * cond_scale
@@ -383,7 +390,7 @@ def sampling_function(model, x, timestep, uncond, cond, cond_scale, model_option
     for fn in model_options.get("sampler_pre_cfg_function", []):
         args = {"conds":conds, "conds_out": out, "cond_scale": cond_scale, "timestep": timestep,
                 "input": x, "sigma": timestep, "model": model, "model_options": model_options}
-        out = fn(args)
+        out  = fn(args)
 
     return cfg_function(model, out[0], out[1], cond_scale, x, timestep, model_options=model_options, cond=cond, uncond=uncond_)
 
@@ -782,7 +789,7 @@ def ksampler(sampler_name, extra_options={}, inpaint_options={}):
     return KSAMPLER(sampler_function, extra_options, inpaint_options)
 
 
-def process_conds(model, noise, conds, device, latent_image=None, denoise_mask=None, seed=None, latent_shapes=None):
+def process_conds(model, noise, conds, device, latent_image=None, denoise_mask=None, seed=None):
     for k in conds:
         conds[k] = conds[k][:]
         resolve_areas_and_cond_masks_multidim(conds[k], noise.shape[2:], device)
@@ -792,7 +799,7 @@ def process_conds(model, noise, conds, device, latent_image=None, denoise_mask=N
 
     if hasattr(model, 'extra_conds'):
         for k in conds:
-            conds[k] = encode_model_conds(model.extra_conds, conds[k], noise, device, k, latent_image=latent_image, denoise_mask=denoise_mask, seed=seed, latent_shapes=latent_shapes)
+            conds[k] = encode_model_conds(model.extra_conds, conds[k], noise, device, k, latent_image=latent_image, denoise_mask=denoise_mask, seed=seed)
 
     #make sure each cond area has an opposite one with the same area
     for k in conds:
@@ -962,11 +969,11 @@ class CFGGuider:
     def predict_noise(self, x, timestep, model_options={}, seed=None):
         return sampling_function(self.inner_model, x, timestep, self.conds.get("negative", None), self.conds.get("positive", None), self.cfg, model_options=model_options, seed=seed)
 
-    def inner_sample(self, noise, latent_image, device, sampler, sigmas, denoise_mask, callback, disable_pbar, seed, latent_shapes=None):
+    def inner_sample(self, noise, latent_image, device, sampler, sigmas, denoise_mask, callback, disable_pbar, seed):
         if latent_image is not None and torch.count_nonzero(latent_image) > 0: #Don't shift the empty latent image.
             latent_image = self.inner_model.process_latent_in(latent_image)
 
-        self.conds = process_conds(self.inner_model, noise, self.conds, device, latent_image, denoise_mask, seed, latent_shapes=latent_shapes)
+        self.conds = process_conds(self.inner_model, noise, self.conds, device, latent_image, denoise_mask, seed)
 
         extra_model_options = comfy.model_patcher.create_model_options_clone(self.model_options)
         extra_model_options.setdefault("transformer_options", {})["sample_sigmas"] = sigmas
@@ -980,7 +987,7 @@ class CFGGuider:
         samples = executor.execute(self, sigmas, extra_args, callback, noise, latent_image, denoise_mask, disable_pbar)
         return self.inner_model.process_latent_out(samples.to(torch.float32))
 
-    def outer_sample(self, noise, latent_image, sampler, sigmas, denoise_mask=None, callback=None, disable_pbar=False, seed=None, latent_shapes=None):
+    def outer_sample(self, noise, latent_image, sampler, sigmas, denoise_mask=None, callback=None, disable_pbar=False, seed=None):
         self.inner_model, self.conds, self.loaded_models = comfy.sampler_helpers.prepare_sampling(self.model_patcher, noise.shape, self.conds, self.model_options)
         device = self.model_patcher.load_device
 
@@ -994,7 +1001,7 @@ class CFGGuider:
 
         try:
             self.model_patcher.pre_run()
-            output = self.inner_sample(noise, latent_image, device, sampler, sigmas, denoise_mask, callback, disable_pbar, seed, latent_shapes=latent_shapes)
+            output = self.inner_sample(noise, latent_image, device, sampler, sigmas, denoise_mask, callback, disable_pbar, seed)
         finally:
             self.model_patcher.cleanup()
 
@@ -1007,12 +1014,6 @@ class CFGGuider:
         if sigmas.shape[-1] == 0:
             return latent_image
 
-        if latent_image.is_nested:
-            latent_image, latent_shapes = comfy.utils.pack_latents(latent_image.unbind())
-            noise, _ = comfy.utils.pack_latents(noise.unbind())
-        else:
-            latent_shapes = [latent_image.shape]
-
         self.conds = {}
         for k in self.original_conds:
             self.conds[k] = list(map(lambda a: a.copy(), self.original_conds[k]))
@@ -1032,7 +1033,7 @@ class CFGGuider:
                 self,
                 comfy.patcher_extension.get_all_wrappers(comfy.patcher_extension.WrappersMP.OUTER_SAMPLE, self.model_options, is_model_options=True)
             )
-            output = executor.execute(noise, latent_image, sampler, sigmas, denoise_mask, callback, disable_pbar, seed, latent_shapes=latent_shapes)
+            output = executor.execute(noise, latent_image, sampler, sigmas, denoise_mask, callback, disable_pbar, seed)
         finally:
             cast_to_load_options(self.model_options, device=self.model_patcher.offload_device)
             self.model_options = orig_model_options
@@ -1040,9 +1041,6 @@ class CFGGuider:
             self.model_patcher.restore_hook_patches()
 
         del self.conds
-
-        if len(latent_shapes) > 1:
-            output = comfy.nested_tensor.NestedTensor(comfy.utils.unpack_latents(output, latent_shapes))
         return output
 
 

comfy/sd.py

@@ -18,7 +18,6 @@ import comfy.ldm.wan.vae2_2
 import comfy.ldm.hunyuan3d.vae
 import comfy.ldm.ace.vae.music_dcae_pipeline
 import comfy.ldm.hunyuan_video.vae
-import comfy.ldm.mmaudio.vae.autoencoder
 import comfy.pixel_space_convert
 import yaml
 import math
@@ -143,9 +142,6 @@ class CLIP:
         n.apply_hooks_to_conds = self.apply_hooks_to_conds
         return n
 
-    def get_ram_usage(self):
-        return self.patcher.get_ram_usage()
-
     def add_patches(self, patches, strength_patch=1.0, strength_model=1.0):
         return self.patcher.add_patches(patches, strength_patch, strength_model)
 
@@ -279,13 +275,8 @@ class VAE:
         if 'decoder.up_blocks.0.resnets.0.norm1.weight' in sd.keys(): #diffusers format
             sd = diffusers_convert.convert_vae_state_dict(sd)
 
-        if model_management.is_amd():
-            VAE_KL_MEM_RATIO = 2.73
-        else:
-            VAE_KL_MEM_RATIO = 1.0
-
-        self.memory_used_encode = lambda shape, dtype: (1767 * shape[2] * shape[3]) * model_management.dtype_size(dtype) * VAE_KL_MEM_RATIO #These are for AutoencoderKL and need tweaking (should be lower)
-        self.memory_used_decode = lambda shape, dtype: (2178 * shape[2] * shape[3] * 64) * model_management.dtype_size(dtype) * VAE_KL_MEM_RATIO
+        self.memory_used_encode = lambda shape, dtype: (1767 * shape[2] * shape[3]) * model_management.dtype_size(dtype) #These are for AutoencoderKL and need tweaking (should be lower)
+        self.memory_used_decode = lambda shape, dtype: (2178 * shape[2] * shape[3] * 64) * model_management.dtype_size(dtype)
         self.downscale_ratio = 8
         self.upscale_ratio = 8
         self.latent_channels = 4
@@ -296,12 +287,10 @@ class VAE:
         self.working_dtypes = [torch.bfloat16, torch.float32]
         self.disable_offload = False
         self.not_video = False
-        self.size = None
 
         self.downscale_index_formula = None
         self.upscale_index_formula = None
         self.extra_1d_channel = None
-        self.crop_input = True
 
         if config is None:
             if "decoder.mid.block_1.mix_factor" in sd:
@@ -343,51 +332,35 @@ class VAE:
                 self.first_stage_model = StageC_coder()
                 self.downscale_ratio = 32
                 self.latent_channels = 16
+            elif "decoder.conv_in.weight" in sd and sd['decoder.conv_in.weight'].shape[1] == 64:
+                ddconfig = {"block_out_channels": [128, 256, 512, 512, 1024, 1024], "in_channels": 3, "out_channels": 3, "num_res_blocks": 2, "ffactor_spatial": 32, "downsample_match_channel": True, "upsample_match_channel": True}
+                self.latent_channels = ddconfig['z_channels'] = sd["decoder.conv_in.weight"].shape[1]
+                self.downscale_ratio = 32
+                self.upscale_ratio = 32
+                self.working_dtypes = [torch.float16, torch.bfloat16, torch.float32]
+                self.first_stage_model = AutoencodingEngine(regularizer_config={'target': "comfy.ldm.models.autoencoder.DiagonalGaussianRegularizer"},
+                                                            encoder_config={'target': "comfy.ldm.hunyuan_video.vae.Encoder", 'params': ddconfig},
+                                                            decoder_config={'target': "comfy.ldm.hunyuan_video.vae.Decoder", 'params': ddconfig})
+
+                self.memory_used_encode = lambda shape, dtype: (700 * shape[2] * shape[3]) * model_management.dtype_size(dtype)
+                self.memory_used_decode = lambda shape, dtype: (700 * shape[2] * shape[3] * 32 * 32) * model_management.dtype_size(dtype)
+
             elif "decoder.conv_in.weight" in sd:
-                if sd['decoder.conv_in.weight'].shape[1] == 64:
-                    ddconfig = {"block_out_channels": [128, 256, 512, 512, 1024, 1024], "in_channels": 3, "out_channels": 3, "num_res_blocks": 2, "ffactor_spatial": 32, "downsample_match_channel": True, "upsample_match_channel": True}
-                    self.latent_channels = ddconfig['z_channels'] = sd["decoder.conv_in.weight"].shape[1]
-                    self.downscale_ratio = 32
-                    self.upscale_ratio = 32
-                    self.working_dtypes = [torch.float16, torch.bfloat16, torch.float32]
-                    self.first_stage_model = AutoencodingEngine(regularizer_config={'target': "comfy.ldm.models.autoencoder.DiagonalGaussianRegularizer"},
-                                                                encoder_config={'target': "comfy.ldm.hunyuan_video.vae.Encoder", 'params': ddconfig},
-                                                                decoder_config={'target': "comfy.ldm.hunyuan_video.vae.Decoder", 'params': ddconfig})
+                #default SD1.x/SD2.x VAE parameters
+                ddconfig = {'double_z': True, 'z_channels': 4, 'resolution': 256, 'in_channels': 3, 'out_ch': 3, 'ch': 128, 'ch_mult': [1, 2, 4, 4], 'num_res_blocks': 2, 'attn_resolutions': [], 'dropout': 0.0}
 
-                    self.memory_used_encode = lambda shape, dtype: (700 * shape[2] * shape[3]) * model_management.dtype_size(dtype)
-                    self.memory_used_decode = lambda shape, dtype: (700 * shape[2] * shape[3] * 32 * 32) * model_management.dtype_size(dtype)
-                elif sd['decoder.conv_in.weight'].shape[1] == 32:
-                    ddconfig = {"block_out_channels": [128, 256, 512, 1024, 1024], "in_channels": 3, "out_channels": 3, "num_res_blocks": 2, "ffactor_spatial": 16, "ffactor_temporal": 4, "downsample_match_channel": True, "upsample_match_channel": True, "refiner_vae": False}
-                    self.latent_channels = ddconfig['z_channels'] = sd["decoder.conv_in.weight"].shape[1]
-                    self.working_dtypes = [torch.float16, torch.bfloat16, torch.float32]
-                    self.upscale_ratio = (lambda a: max(0, a * 4 - 3), 16, 16)
-                    self.upscale_index_formula = (4, 16, 16)
-                    self.downscale_ratio = (lambda a: max(0, math.floor((a + 3) / 4)), 16, 16)
-                    self.downscale_index_formula = (4, 16, 16)
-                    self.latent_dim = 3
-                    self.not_video = True
-                    self.first_stage_model = AutoencodingEngine(regularizer_config={'target': "comfy.ldm.models.autoencoder.DiagonalGaussianRegularizer"},
-                                                                encoder_config={'target': "comfy.ldm.hunyuan_video.vae_refiner.Encoder", 'params': ddconfig},
-                                                                decoder_config={'target': "comfy.ldm.hunyuan_video.vae_refiner.Decoder", 'params': ddconfig})
+                if 'encoder.down.2.downsample.conv.weight' not in sd and 'decoder.up.3.upsample.conv.weight' not in sd: #Stable diffusion x4 upscaler VAE
+                    ddconfig['ch_mult'] = [1, 2, 4]
+                    self.downscale_ratio = 4
+                    self.upscale_ratio = 4
 
-                    self.memory_used_encode = lambda shape, dtype: (2800 * shape[-2] * shape[-1]) * model_management.dtype_size(dtype)
-                    self.memory_used_decode = lambda shape, dtype: (2800 * shape[-3] * shape[-2] * shape[-1] * 16 * 16) * model_management.dtype_size(dtype)
+                self.latent_channels = ddconfig['z_channels'] = sd["decoder.conv_in.weight"].shape[1]
+                if 'post_quant_conv.weight' in sd:
+                    self.first_stage_model = AutoencoderKL(ddconfig=ddconfig, embed_dim=sd['post_quant_conv.weight'].shape[1])
                 else:
-                    #default SD1.x/SD2.x VAE parameters
-                    ddconfig = {'double_z': True, 'z_channels': 4, 'resolution': 256, 'in_channels': 3, 'out_ch': 3, 'ch': 128, 'ch_mult': [1, 2, 4, 4], 'num_res_blocks': 2, 'attn_resolutions': [], 'dropout': 0.0}
-
-                    if 'encoder.down.2.downsample.conv.weight' not in sd and 'decoder.up.3.upsample.conv.weight' not in sd: #Stable diffusion x4 upscaler VAE
-                        ddconfig['ch_mult'] = [1, 2, 4]
-                        self.downscale_ratio = 4
-                        self.upscale_ratio = 4
-
-                    self.latent_channels = ddconfig['z_channels'] = sd["decoder.conv_in.weight"].shape[1]
-                    if 'post_quant_conv.weight' in sd:
-                        self.first_stage_model = AutoencoderKL(ddconfig=ddconfig, embed_dim=sd['post_quant_conv.weight'].shape[1])
-                    else:
-                        self.first_stage_model = AutoencodingEngine(regularizer_config={'target': "comfy.ldm.models.autoencoder.DiagonalGaussianRegularizer"},
-                                                                    encoder_config={'target': "comfy.ldm.modules.diffusionmodules.model.Encoder", 'params': ddconfig},
-                                                                    decoder_config={'target': "comfy.ldm.modules.diffusionmodules.model.Decoder", 'params': ddconfig})
+                    self.first_stage_model = AutoencodingEngine(regularizer_config={'target': "comfy.ldm.models.autoencoder.DiagonalGaussianRegularizer"},
+                                                                encoder_config={'target': "comfy.ldm.modules.diffusionmodules.model.Encoder", 'params': ddconfig},
+                                                                decoder_config={'target': "comfy.ldm.modules.diffusionmodules.model.Decoder", 'params': ddconfig})
             elif "decoder.layers.1.layers.0.beta" in sd:
                 self.first_stage_model = AudioOobleckVAE()
                 self.memory_used_encode = lambda shape, dtype: (1000 * shape[2]) * model_management.dtype_size(dtype)
@@ -441,20 +414,20 @@ class VAE:
             elif "decoder.conv_in.conv.weight" in sd and sd['decoder.conv_in.conv.weight'].shape[1] == 32:
                 ddconfig = {"block_out_channels": [128, 256, 512, 1024, 1024], "in_channels": 3, "out_channels": 3, "num_res_blocks": 2, "ffactor_spatial": 16, "ffactor_temporal": 4, "downsample_match_channel": True, "upsample_match_channel": True}
                 ddconfig['z_channels'] = sd["decoder.conv_in.conv.weight"].shape[1]
-                self.latent_channels = 32
+                self.latent_channels = 64
                 self.upscale_ratio = (lambda a: max(0, a * 4 - 3), 16, 16)
                 self.upscale_index_formula = (4, 16, 16)
                 self.downscale_ratio = (lambda a: max(0, math.floor((a + 3) / 4)), 16, 16)
                 self.downscale_index_formula = (4, 16, 16)
                 self.latent_dim = 3
-                self.not_video = False
+                self.not_video = True
                 self.working_dtypes = [torch.float16, torch.bfloat16, torch.float32]
                 self.first_stage_model = AutoencodingEngine(regularizer_config={'target': "comfy.ldm.models.autoencoder.EmptyRegularizer"},
                                                             encoder_config={'target': "comfy.ldm.hunyuan_video.vae_refiner.Encoder", 'params': ddconfig},
                                                             decoder_config={'target': "comfy.ldm.hunyuan_video.vae_refiner.Decoder", 'params': ddconfig})
 
-                self.memory_used_encode = lambda shape, dtype: (1400 * 9 * shape[-2] * shape[-1]) * model_management.dtype_size(dtype)
-                self.memory_used_decode = lambda shape, dtype: (2800 * 4 * shape[-2] * shape[-1] * 16 * 16) * model_management.dtype_size(dtype)
+                self.memory_used_encode = lambda shape, dtype: (1400 * shape[-2] * shape[-1]) * model_management.dtype_size(dtype)
+                self.memory_used_decode = lambda shape, dtype: (1400 * shape[-3] * shape[-2] * shape[-1] * 16 * 16) * model_management.dtype_size(dtype)
             elif "decoder.conv_in.conv.weight" in sd:
                 ddconfig = {'double_z': True, 'z_channels': 4, 'resolution': 256, 'in_channels': 3, 'out_ch': 3, 'ch': 128, 'ch_mult': [1, 2, 4, 4], 'num_res_blocks': 2, 'attn_resolutions': [], 'dropout': 0.0}
                 ddconfig["conv3d"] = True
@@ -553,25 +526,6 @@ class VAE:
                 self.latent_channels = 3
                 self.latent_dim = 2
                 self.output_channels = 3
-            elif "vocoder.activation_post.downsample.lowpass.filter" in sd: #MMAudio VAE
-                sample_rate = 16000
-                if sample_rate == 16000:
-                    mode = '16k'
-                else:
-                    mode = '44k'
-
-                self.first_stage_model = comfy.ldm.mmaudio.vae.autoencoder.AudioAutoencoder(mode=mode)
-                self.memory_used_encode = lambda shape, dtype: (30 * shape[2]) * model_management.dtype_size(dtype)
-                self.memory_used_decode = lambda shape, dtype: (90 * shape[2] * 1411.2) * model_management.dtype_size(dtype)
-                self.latent_channels = 20
-                self.output_channels = 2
-                self.upscale_ratio = 512 * (44100 / sample_rate)
-                self.downscale_ratio = 512 * (44100 / sample_rate)
-                self.latent_dim = 1
-                self.process_output = lambda audio: audio
-                self.process_input = lambda audio: audio
-                self.working_dtypes = [torch.float32]
-                self.crop_input = False
             else:
                 logging.warning("WARNING: No VAE weights detected, VAE not initalized.")
                 self.first_stage_model = None
@@ -599,25 +553,12 @@ class VAE:
 
         self.patcher = comfy.model_patcher.ModelPatcher(self.first_stage_model, load_device=self.device, offload_device=offload_device)
         logging.info("VAE load device: {}, offload device: {}, dtype: {}".format(self.device, offload_device, self.vae_dtype))
-        self.model_size()
-
-    def model_size(self):
-        if self.size is not None:
-            return self.size
-        self.size = comfy.model_management.module_size(self.first_stage_model)
-        return self.size
-
-    def get_ram_usage(self):
-        return self.model_size()
 
     def throw_exception_if_invalid(self):
         if self.first_stage_model is None:
             raise RuntimeError("ERROR: VAE is invalid: None\n\nIf the VAE is from a checkpoint loader node your checkpoint does not contain a valid VAE.")
 
     def vae_encode_crop_pixels(self, pixels):
-        if not self.crop_input:
-            return pixels
-
         downscale_ratio = self.spacial_compression_encode()
 
         dims = pixels.shape[1:-1]
@@ -695,7 +636,6 @@ class VAE:
     def decode(self, samples_in, vae_options={}):
         self.throw_exception_if_invalid()
         pixel_samples = None
-        do_tile = False
         try:
             memory_used = self.memory_used_decode(samples_in.shape, self.vae_dtype)
             model_management.load_models_gpu([self.patcher], memory_required=memory_used, force_full_load=self.disable_offload)
@@ -711,13 +651,6 @@ class VAE:
                 pixel_samples[x:x+batch_number] = out
         except model_management.OOM_EXCEPTION:
             logging.warning("Warning: Ran out of memory when regular VAE decoding, retrying with tiled VAE decoding.")
-            #NOTE: We don't know what tensors were allocated to stack variables at the time of the
-            #exception and the exception itself refs them all until we get out of this except block.
-            #So we just set a flag for tiler fallback so that tensor gc can happen once the
-            #exception is fully off the books.
-            do_tile = True
-
-        if do_tile:
             dims = samples_in.ndim - 2
             if dims == 1 or self.extra_1d_channel is not None:
                 pixel_samples = self.decode_tiled_1d(samples_in)
@@ -764,7 +697,6 @@ class VAE:
         self.throw_exception_if_invalid()
         pixel_samples = self.vae_encode_crop_pixels(pixel_samples)
         pixel_samples = pixel_samples.movedim(-1, 1)
-        do_tile = False
         if self.latent_dim == 3 and pixel_samples.ndim < 5:
             if not self.not_video:
                 pixel_samples = pixel_samples.movedim(1, 0).unsqueeze(0)
@@ -786,13 +718,6 @@ class VAE:
 
         except model_management.OOM_EXCEPTION:
             logging.warning("Warning: Ran out of memory when regular VAE encoding, retrying with tiled VAE encoding.")
-            #NOTE: We don't know what tensors were allocated to stack variables at the time of the
-            #exception and the exception itself refs them all until we get out of this except block.
-            #So we just set a flag for tiler fallback so that tensor gc can happen once the
-            #exception is fully off the books.
-            do_tile = True
-
-        if do_tile:
             if self.latent_dim == 3:
                 tile = 256
                 overlap = tile // 4
@@ -911,7 +836,6 @@ class CLIPType(Enum):
     OMNIGEN2 = 17
     QWEN_IMAGE = 18
     HUNYUAN_IMAGE = 19
-    HUNYUAN_VIDEO_15 = 20
 
 
 def load_clip(ckpt_paths, embedding_directory=None, clip_type=CLIPType.STABLE_DIFFUSION, model_options={}):
@@ -934,7 +858,6 @@ class TEModel(Enum):
     QWEN25_3B = 10
     QWEN25_7B = 11
     BYT5_SMALL_GLYPH = 12
-    GEMMA_3_4B = 13
 
 def detect_te_model(sd):
     if "text_model.encoder.layers.30.mlp.fc1.weight" in sd:
@@ -957,8 +880,6 @@ def detect_te_model(sd):
             return TEModel.BYT5_SMALL_GLYPH
         return TEModel.T5_BASE
     if 'model.layers.0.post_feedforward_layernorm.weight' in sd:
-        if 'model.layers.0.self_attn.q_norm.weight' in sd:
-            return TEModel.GEMMA_3_4B
         return TEModel.GEMMA_2_2B
     if 'model.layers.0.self_attn.k_proj.bias' in sd:
         weight = sd['model.layers.0.self_attn.k_proj.bias']
@@ -1063,10 +984,6 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip
             clip_target.clip = comfy.text_encoders.lumina2.te(**llama_detect(clip_data))
             clip_target.tokenizer = comfy.text_encoders.lumina2.LuminaTokenizer
             tokenizer_data["spiece_model"] = clip_data[0].get("spiece_model", None)
-        elif te_model == TEModel.GEMMA_3_4B:
-            clip_target.clip = comfy.text_encoders.lumina2.te(**llama_detect(clip_data), model_type="gemma3_4b")
-            clip_target.tokenizer = comfy.text_encoders.lumina2.NTokenizer
-            tokenizer_data["spiece_model"] = clip_data[0].get("spiece_model", None)
         elif te_model == TEModel.LLAMA3_8:
             clip_target.clip = comfy.text_encoders.hidream.hidream_clip(**llama_detect(clip_data),
                                                                         clip_l=False, clip_g=False, t5=False, llama=True, dtype_t5=None, t5xxl_scaled_fp8=None)
@@ -1127,9 +1044,6 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip
         elif clip_type == CLIPType.HUNYUAN_IMAGE:
             clip_target.clip = comfy.text_encoders.hunyuan_image.te(**llama_detect(clip_data))
             clip_target.tokenizer = comfy.text_encoders.hunyuan_image.HunyuanImageTokenizer
-        elif clip_type == CLIPType.HUNYUAN_VIDEO_15:
-            clip_target.clip = comfy.text_encoders.hunyuan_image.te(**llama_detect(clip_data))
-            clip_target.tokenizer = comfy.text_encoders.hunyuan_video.HunyuanVideo15Tokenizer
         else:
             clip_target.clip = sdxl_clip.SDXLClipModel
             clip_target.tokenizer = sdxl_clip.SDXLTokenizer
@@ -1280,7 +1194,7 @@ def load_state_dict_guess_config(sd, output_vae=True, output_clip=True, output_c
     return (model_patcher, clip, vae, clipvision)
 
 
-def load_diffusion_model_state_dict(sd, model_options={}, metadata=None):
+def load_diffusion_model_state_dict(sd, model_options={}):
     """
     Loads a UNet diffusion model from a state dictionary, supporting both diffusers and regular formats.
 
@@ -1314,7 +1228,7 @@ def load_diffusion_model_state_dict(sd, model_options={}, metadata=None):
     weight_dtype = comfy.utils.weight_dtype(sd)
 
     load_device = model_management.get_torch_device()
-    model_config = model_detection.model_config_from_unet(sd, "", metadata=metadata)
+    model_config = model_detection.model_config_from_unet(sd, "")
 
     if model_config is not None:
         new_sd = sd
@@ -1348,10 +1262,7 @@ def load_diffusion_model_state_dict(sd, model_options={}, metadata=None):
     else:
         unet_dtype = dtype
 
-    if model_config.layer_quant_config is not None:
-        manual_cast_dtype = model_management.unet_manual_cast(None, load_device, model_config.supported_inference_dtypes)
-    else:
-        manual_cast_dtype = model_management.unet_manual_cast(unet_dtype, load_device, model_config.supported_inference_dtypes)
+    manual_cast_dtype = model_management.unet_manual_cast(unet_dtype, load_device, model_config.supported_inference_dtypes)
     model_config.set_inference_dtype(unet_dtype, manual_cast_dtype)
     model_config.custom_operations = model_options.get("custom_operations", model_config.custom_operations)
     if model_options.get("fp8_optimizations", False):
@@ -1367,8 +1278,8 @@ def load_diffusion_model_state_dict(sd, model_options={}, metadata=None):
 
 
 def load_diffusion_model(unet_path, model_options={}):
-    sd, metadata = comfy.utils.load_torch_file(unet_path, return_metadata=True)
-    model = load_diffusion_model_state_dict(sd, model_options=model_options, metadata=metadata)
+    sd = comfy.utils.load_torch_file(unet_path)
+    model = load_diffusion_model_state_dict(sd, model_options=model_options)
     if model is None:
         logging.error("ERROR UNSUPPORTED DIFFUSION MODEL {}".format(unet_path))
         raise RuntimeError("ERROR: Could not detect model type of: {}\n{}".format(unet_path, model_detection_error_hint(unet_path, sd)))

comfy/sd1_clip.py

@@ -460,7 +460,7 @@ def load_embed(embedding_name, embedding_directory, embedding_size, embed_key=No
     return embed_out
 
 class SDTokenizer:
-    def __init__(self, tokenizer_path=None, max_length=77, pad_with_end=True, embedding_directory=None, embedding_size=768, embedding_key='clip_l', tokenizer_class=CLIPTokenizer, has_start_token=True, has_end_token=True, pad_to_max_length=True, min_length=None, pad_token=None, end_token=None, min_padding=None, pad_left=False, tokenizer_data={}, tokenizer_args={}):
+    def __init__(self, tokenizer_path=None, max_length=77, pad_with_end=True, embedding_directory=None, embedding_size=768, embedding_key='clip_l', tokenizer_class=CLIPTokenizer, has_start_token=True, has_end_token=True, pad_to_max_length=True, min_length=None, pad_token=None, end_token=None, min_padding=None, tokenizer_data={}, tokenizer_args={}):
         if tokenizer_path is None:
             tokenizer_path = os.path.join(os.path.dirname(os.path.realpath(__file__)), "sd1_tokenizer")
         self.tokenizer = tokenizer_class.from_pretrained(tokenizer_path, **tokenizer_args)
@@ -468,7 +468,6 @@ class SDTokenizer:
         self.min_length = tokenizer_data.get("{}_min_length".format(embedding_key), min_length)
         self.end_token = None
         self.min_padding = min_padding
-        self.pad_left = pad_left
 
         empty = self.tokenizer('')["input_ids"]
         self.tokenizer_adds_end_token = has_end_token
@@ -523,12 +522,6 @@ class SDTokenizer:
                 return (embed, "{} {}".format(embedding_name[len(stripped):], leftover))
         return (embed, leftover)
 
-    def pad_tokens(self, tokens, amount):
-        if self.pad_left:
-            for i in range(amount):
-                tokens.insert(0, (self.pad_token, 1.0, 0))
-        else:
-            tokens.extend([(self.pad_token, 1.0, 0)] * amount)
 
     def tokenize_with_weights(self, text:str, return_word_ids=False, tokenizer_options={}, **kwargs):
         '''
@@ -607,7 +600,7 @@ class SDTokenizer:
                         if self.end_token is not None:
                             batch.append((self.end_token, 1.0, 0))
                         if self.pad_to_max_length:
-                            self.pad_tokens(batch, remaining_length)
+                            batch.extend([(self.pad_token, 1.0, 0)] * (remaining_length))
                     #start new batch
                     batch = []
                     if self.start_token is not None:
@@ -621,11 +614,11 @@ class SDTokenizer:
         if self.end_token is not None:
             batch.append((self.end_token, 1.0, 0))
         if min_padding is not None:
-            self.pad_tokens(batch, min_padding)
+            batch.extend([(self.pad_token, 1.0, 0)] * min_padding)
         if self.pad_to_max_length and len(batch) < self.max_length:
-            self.pad_tokens(batch, self.max_length - len(batch))
+            batch.extend([(self.pad_token, 1.0, 0)] * (self.max_length - len(batch)))
         if min_length is not None and len(batch) < min_length:
-            self.pad_tokens(batch, min_length - len(batch))
+            batch.extend([(self.pad_token, 1.0, 0)] * (min_length - len(batch)))
 
         if not return_word_ids:
             batched_tokens = [[(t, w) for t, w,_ in x] for x in batched_tokens]

comfy/supported_models.py

@@ -1374,54 +1374,6 @@ class HunyuanImage21Refiner(HunyuanVideo):
         out = model_base.HunyuanImage21Refiner(self, device=device)
         return out
 
-class HunyuanVideo15(HunyuanVideo):
-    unet_config = {
-        "image_model": "hunyuan_video",
-        "vision_in_dim": 1152,
-    }
-
-    sampling_settings = {
-        "shift": 7.0,
-    }
-    memory_usage_factor = 4.0 #TODO
-    supported_inference_dtypes = [torch.float16, torch.bfloat16, torch.float32]
-
-    latent_format = latent_formats.HunyuanVideo15
-
-    def get_model(self, state_dict, prefix="", device=None):
-        out = model_base.HunyuanVideo15(self, device=device)
-        return out
-
-    def clip_target(self, state_dict={}):
-        pref = self.text_encoder_key_prefix[0]
-        hunyuan_detect = comfy.text_encoders.hunyuan_video.llama_detect(state_dict, "{}qwen25_7b.transformer.".format(pref))
-        return supported_models_base.ClipTarget(comfy.text_encoders.hunyuan_video.HunyuanVideo15Tokenizer, comfy.text_encoders.hunyuan_image.te(**hunyuan_detect))
-
-
-class HunyuanVideo15_SR_Distilled(HunyuanVideo):
-    unet_config = {
-        "image_model": "hunyuan_video",
-        "vision_in_dim": 1152,
-        "in_channels": 98,
-    }
-
-    sampling_settings = {
-        "shift": 2.0,
-    }
-    memory_usage_factor = 4.0 #TODO
-    supported_inference_dtypes = [torch.float16, torch.bfloat16, torch.float32]
-
-    latent_format = latent_formats.HunyuanVideo15
-
-    def get_model(self, state_dict, prefix="", device=None):
-        out = model_base.HunyuanVideo15_SR_Distilled(self, device=device)
-        return out
-
-    def clip_target(self, state_dict={}):
-        pref = self.text_encoder_key_prefix[0]
-        hunyuan_detect = comfy.text_encoders.hunyuan_video.llama_detect(state_dict, "{}qwen25_7b.transformer.".format(pref))
-        return supported_models_base.ClipTarget(comfy.text_encoders.hunyuan_video.HunyuanVideo15Tokenizer, comfy.text_encoders.hunyuan_image.te(**hunyuan_detect))
-
-models = [LotusD, Stable_Zero123, SD15_instructpix2pix, SD15, SD20, SD21UnclipL, SD21UnclipH, SDXL_instructpix2pix, SDXLRefiner, SDXL, SSD1B, KOALA_700M, KOALA_1B, Segmind_Vega, SD_X4Upscaler, Stable_Cascade_C, Stable_Cascade_B, SV3D_u, SV3D_p, SD3, StableAudio, AuraFlow, PixArtAlpha, PixArtSigma, HunyuanDiT, HunyuanDiT1, FluxInpaint, Flux, FluxSchnell, GenmoMochi, LTXV, HunyuanVideo15_SR_Distilled, HunyuanVideo15, HunyuanImage21Refiner, HunyuanImage21, HunyuanVideoSkyreelsI2V, HunyuanVideoI2V, HunyuanVideo, CosmosT2V, CosmosI2V, CosmosT2IPredict2, CosmosI2VPredict2, Lumina2, WAN22_T2V, WAN21_T2V, WAN21_I2V, WAN21_FunControl2V, WAN21_Vace, WAN21_Camera, WAN22_Camera, WAN22_S2V, WAN21_HuMo, WAN22_Animate, Hunyuan3Dv2mini, Hunyuan3Dv2, Hunyuan3Dv2_1, HiDream, Chroma, ChromaRadiance, ACEStep, Omnigen2, QwenImage]
+models = [LotusD, Stable_Zero123, SD15_instructpix2pix, SD15, SD20, SD21UnclipL, SD21UnclipH, SDXL_instructpix2pix, SDXLRefiner, SDXL, SSD1B, KOALA_700M, KOALA_1B, Segmind_Vega, SD_X4Upscaler, Stable_Cascade_C, Stable_Cascade_B, SV3D_u, SV3D_p, SD3, StableAudio, AuraFlow, PixArtAlpha, PixArtSigma, HunyuanDiT, HunyuanDiT1, FluxInpaint, Flux, FluxSchnell, GenmoMochi, LTXV, HunyuanImage21Refiner, HunyuanImage21, HunyuanVideoSkyreelsI2V, HunyuanVideoI2V, HunyuanVideo, CosmosT2V, CosmosI2V, CosmosT2IPredict2, CosmosI2VPredict2, Lumina2, WAN22_T2V, WAN21_T2V, WAN21_I2V, WAN21_FunControl2V, WAN21_Vace, WAN21_Camera, WAN22_Camera, WAN22_S2V, WAN21_HuMo, WAN22_Animate, Hunyuan3Dv2mini, Hunyuan3Dv2, Hunyuan3Dv2_1, HiDream, Chroma, ChromaRadiance, ACEStep, Omnigen2, QwenImage]
 
 models += [SVD_img2vid]

comfy/supported_models_base.py

@@ -50,7 +50,6 @@ class BASE:
     manual_cast_dtype = None
     custom_operations = None
     scaled_fp8 = None
-    layer_quant_config = None  # Per-layer quantization configuration for mixed precision
     optimizations = {"fp8": False}
 
     @classmethod

comfy/text_encoders/hunyuan_image.py

@@ -63,13 +63,7 @@ class HunyuanImageTEModel(QwenImageTEModel):
             self.byt5_small = None
 
     def encode_token_weights(self, token_weight_pairs):
-        tok_pairs = token_weight_pairs["qwen25_7b"][0]
-        template_end = -1
-        if tok_pairs[0][0] == 27:
-            if len(tok_pairs) > 36:  # refiner prompt uses a fixed 36 template_end
-                template_end = 36
-
-        cond, p, extra = super().encode_token_weights(token_weight_pairs, template_end=template_end)
+        cond, p, extra = super().encode_token_weights(token_weight_pairs)
         if self.byt5_small is not None and "byt5" in token_weight_pairs:
             out = self.byt5_small.encode_token_weights(token_weight_pairs["byt5"])
             extra["conditioning_byt5small"] = out[0]

comfy/text_encoders/hunyuan_video.py

@@ -1,7 +1,6 @@
 from comfy import sd1_clip
 import comfy.model_management
 import comfy.text_encoders.llama
-from .hunyuan_image import HunyuanImageTokenizer
 from transformers import LlamaTokenizerFast
 import torch
 import os
@@ -74,14 +73,6 @@ class HunyuanVideoTokenizer:
         return {}
 
 
-class HunyuanVideo15Tokenizer(HunyuanImageTokenizer):
-    def __init__(self, embedding_directory=None, tokenizer_data={}):
-        super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data)
-        self.llama_template = "<|im_start|>system\nYou are a helpful assistant. Describe the video by detailing the following aspects:\n1. The main content and theme of the video.\n2. The color, shape, size, texture, quantity, text, and spatial relationships of the objects.\n3. Actions, events, behaviors temporal relationships, physical movement changes of the objects.\n4. background environment, light, style and atmosphere.\n5. camera angles, movements, and transitions used in the video.<|im_end|>\n<|im_start|>user\n{}<|im_end|>\n<|im_start|>assistant\n"
-
-    def tokenize_with_weights(self, text:str, return_word_ids=False, **kwargs):
-        return super().tokenize_with_weights(text, return_word_ids, prevent_empty_text=True, **kwargs)
-
 class HunyuanVideoClipModel(torch.nn.Module):
     def __init__(self, dtype_llama=None, device="cpu", dtype=None, model_options={}):
         super().__init__()

comfy/text_encoders/llama.py

@@ -3,7 +3,6 @@ import torch.nn as nn
 from dataclasses import dataclass
 from typing import Optional, Any
 import math
-import logging
 
 from comfy.ldm.modules.attention import optimized_attention_for_device
 import comfy.model_management
@@ -29,10 +28,6 @@ class Llama2Config:
     mlp_activation = "silu"
     qkv_bias = False
     rope_dims = None
-    q_norm = None
-    k_norm = None
-    rope_scale = None
-    final_norm: bool = True
 
 @dataclass
 class Qwen25_3BConfig:
@@ -51,10 +46,6 @@ class Qwen25_3BConfig:
     mlp_activation = "silu"
     qkv_bias = True
     rope_dims = None
-    q_norm = None
-    k_norm = None
-    rope_scale = None
-    final_norm: bool = True
 
 @dataclass
 class Qwen25_7BVLI_Config:
@@ -73,10 +64,6 @@ class Qwen25_7BVLI_Config:
     mlp_activation = "silu"
     qkv_bias = True
     rope_dims = [16, 24, 24]
-    q_norm = None
-    k_norm = None
-    rope_scale = None
-    final_norm: bool = True
 
 @dataclass
 class Gemma2_2B_Config:
@@ -95,34 +82,6 @@ class Gemma2_2B_Config:
     mlp_activation = "gelu_pytorch_tanh"
     qkv_bias = False
     rope_dims = None
-    q_norm = None
-    k_norm = None
-    sliding_attention = None
-    rope_scale = None
-    final_norm: bool = True
-
-@dataclass
-class Gemma3_4B_Config:
-    vocab_size: int = 262208
-    hidden_size: int = 2560
-    intermediate_size: int = 10240
-    num_hidden_layers: int = 34
-    num_attention_heads: int = 8
-    num_key_value_heads: int = 4
-    max_position_embeddings: int = 131072
-    rms_norm_eps: float = 1e-6
-    rope_theta = [10000.0, 1000000.0]
-    transformer_type: str = "gemma3"
-    head_dim = 256
-    rms_norm_add = True
-    mlp_activation = "gelu_pytorch_tanh"
-    qkv_bias = False
-    rope_dims = None
-    q_norm = "gemma3"
-    k_norm = "gemma3"
-    sliding_attention = [False, False, False, False, False, 1024]
-    rope_scale = [1.0, 8.0]
-    final_norm: bool = True
 
 class RMSNorm(nn.Module):
     def __init__(self, dim: int, eps: float = 1e-5, add=False, device=None, dtype=None):
@@ -147,40 +106,25 @@ def rotate_half(x):
     return torch.cat((-x2, x1), dim=-1)
 
 
-def precompute_freqs_cis(head_dim, position_ids, theta, rope_scale=None, rope_dims=None, device=None):
-    if not isinstance(theta, list):
-        theta = [theta]
+def precompute_freqs_cis(head_dim, position_ids, theta, rope_dims=None, device=None):
+    theta_numerator = torch.arange(0, head_dim, 2, device=device).float()
+    inv_freq = 1.0 / (theta ** (theta_numerator / head_dim))
 
-    out = []
-    for index, t in enumerate(theta):
-        theta_numerator = torch.arange(0, head_dim, 2, device=device).float()
-        inv_freq = 1.0 / (t ** (theta_numerator / head_dim))
+    inv_freq_expanded = inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1)
+    position_ids_expanded = position_ids[:, None, :].float()
+    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()
+    if rope_dims is not None and position_ids.shape[0] > 1:
+        mrope_section = rope_dims * 2
+        cos = torch.cat([m[i % 3] for i, m in enumerate(cos.split(mrope_section, dim=-1))], dim=-1).unsqueeze(0)
+        sin = torch.cat([m[i % 3] for i, m in enumerate(sin.split(mrope_section, dim=-1))], dim=-1).unsqueeze(0)
+    else:
+        cos = cos.unsqueeze(1)
+        sin = sin.unsqueeze(1)
 
-        if rope_scale is not None:
-            if isinstance(rope_scale, list):
-                inv_freq /= rope_scale[index]
-            else:
-                inv_freq /= rope_scale
-
-        inv_freq_expanded = inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1)
-        position_ids_expanded = position_ids[:, None, :].float()
-        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()
-        if rope_dims is not None and position_ids.shape[0] > 1:
-            mrope_section = rope_dims * 2
-            cos = torch.cat([m[i % 3] for i, m in enumerate(cos.split(mrope_section, dim=-1))], dim=-1).unsqueeze(0)
-            sin = torch.cat([m[i % 3] for i, m in enumerate(sin.split(mrope_section, dim=-1))], dim=-1).unsqueeze(0)
-        else:
-            cos = cos.unsqueeze(1)
-            sin = sin.unsqueeze(1)
-        out.append((cos, sin))
-
-    if len(out) == 1:
-        return out[0]
-
-    return out
+    return (cos, sin)
 
 
 def apply_rope(xq, xk, freqs_cis):
@@ -208,14 +152,6 @@ class Attention(nn.Module):
         self.v_proj = ops.Linear(config.hidden_size, self.num_kv_heads * self.head_dim, bias=config.qkv_bias, device=device, dtype=dtype)
         self.o_proj = ops.Linear(self.inner_size, config.hidden_size, bias=False, device=device, dtype=dtype)
 
-        self.q_norm = None
-        self.k_norm = None
-
-        if config.q_norm == "gemma3":
-            self.q_norm = RMSNorm(self.head_dim, eps=config.rms_norm_eps, add=config.rms_norm_add, device=device, dtype=dtype)
-        if config.k_norm == "gemma3":
-            self.k_norm = RMSNorm(self.head_dim, eps=config.rms_norm_eps, add=config.rms_norm_add, device=device, dtype=dtype)
-
     def forward(
         self,
         hidden_states: torch.Tensor,
@@ -232,11 +168,6 @@ class Attention(nn.Module):
         xk = xk.view(batch_size, seq_length, self.num_kv_heads, self.head_dim).transpose(1, 2)
         xv = xv.view(batch_size, seq_length, self.num_kv_heads, self.head_dim).transpose(1, 2)
 
-        if self.q_norm is not None:
-            xq = self.q_norm(xq)
-        if self.k_norm is not None:
-            xk = self.k_norm(xk)
-
         xq, xk = apply_rope(xq, xk, freqs_cis=freqs_cis)
 
         xk = xk.repeat_interleave(self.num_heads // self.num_kv_heads, dim=1)
@@ -261,7 +192,7 @@ class MLP(nn.Module):
         return self.down_proj(self.activation(self.gate_proj(x)) * self.up_proj(x))
 
 class TransformerBlock(nn.Module):
-    def __init__(self, config: Llama2Config, index, device=None, dtype=None, ops: Any = None):
+    def __init__(self, config: Llama2Config, device=None, dtype=None, ops: Any = None):
         super().__init__()
         self.self_attn = Attention(config, device=device, dtype=dtype, ops=ops)
         self.mlp = MLP(config, device=device, dtype=dtype, ops=ops)
@@ -295,7 +226,7 @@ class TransformerBlock(nn.Module):
         return x
 
 class TransformerBlockGemma2(nn.Module):
-    def __init__(self, config: Llama2Config, index, device=None, dtype=None, ops: Any = None):
+    def __init__(self, config: Llama2Config, device=None, dtype=None, ops: Any = None):
         super().__init__()
         self.self_attn = Attention(config, device=device, dtype=dtype, ops=ops)
         self.mlp = MLP(config, device=device, dtype=dtype, ops=ops)
@@ -304,13 +235,6 @@ class TransformerBlockGemma2(nn.Module):
         self.pre_feedforward_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps, add=config.rms_norm_add, device=device, dtype=dtype)
         self.post_feedforward_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps, add=config.rms_norm_add, device=device, dtype=dtype)
 
-        if config.sliding_attention is not None:  # TODO: implement. (Not that necessary since models are trained on less than 1024 tokens)
-            self.sliding_attention = config.sliding_attention[index % len(config.sliding_attention)]
-        else:
-            self.sliding_attention = False
-
-        self.transformer_type = config.transformer_type
-
     def forward(
         self,
         x: torch.Tensor,
@@ -318,14 +242,6 @@ class TransformerBlockGemma2(nn.Module):
         freqs_cis: Optional[torch.Tensor] = None,
         optimized_attention=None,
     ):
-        if self.transformer_type == 'gemma3':
-            if self.sliding_attention:
-                if x.shape[1] > self.sliding_attention:
-                    logging.warning("Warning: sliding attention not implemented, results may be incorrect")
-                freqs_cis = freqs_cis[1]
-            else:
-                freqs_cis = freqs_cis[0]
-
         # Self Attention
         residual = x
         x = self.input_layernorm(x)
@@ -360,7 +276,7 @@ class Llama2_(nn.Module):
             device=device,
             dtype=dtype
         )
-        if self.config.transformer_type == "gemma2" or self.config.transformer_type == "gemma3":
+        if self.config.transformer_type == "gemma2":
             transformer = TransformerBlockGemma2
             self.normalize_in = True
         else:
@@ -368,15 +284,10 @@ class Llama2_(nn.Module):
             self.normalize_in = False
 
         self.layers = nn.ModuleList([
-            transformer(config, index=i, device=device, dtype=dtype, ops=ops)
-            for i in range(config.num_hidden_layers)
+            transformer(config, device=device, dtype=dtype, ops=ops)
+            for _ in range(config.num_hidden_layers)
         ])
-
-        if config.final_norm:
-            self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps, add=config.rms_norm_add, device=device, dtype=dtype)
-        else:
-            self.norm = None
-
+        self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps, add=config.rms_norm_add, device=device, dtype=dtype)
         # self.lm_head = ops.Linear(config.hidden_size, config.vocab_size, bias=False, device=device, dtype=dtype)
 
     def forward(self, x, attention_mask=None, embeds=None, num_tokens=None, intermediate_output=None, final_layer_norm_intermediate=True, dtype=None, position_ids=None, embeds_info=[]):
@@ -394,7 +305,6 @@ class Llama2_(nn.Module):
         freqs_cis = precompute_freqs_cis(self.config.head_dim,
                                          position_ids,
                                          self.config.rope_theta,
-                                         self.config.rope_scale,
                                          self.config.rope_dims,
                                          device=x.device)
 
@@ -431,16 +341,14 @@ class Llama2_(nn.Module):
             if i == intermediate_output:
                 intermediate = x.clone()
 
-        if self.norm is not None:
-            x = self.norm(x)
-
+        x = self.norm(x)
         if all_intermediate is not None:
             all_intermediate.append(x.unsqueeze(1).clone())
 
         if all_intermediate is not None:
             intermediate = torch.cat(all_intermediate, dim=1)
 
-        if intermediate is not None and final_layer_norm_intermediate and self.norm is not None:
+        if intermediate is not None and final_layer_norm_intermediate:
             intermediate = self.norm(intermediate)
 
         return x, intermediate
@@ -525,12 +433,3 @@ class Gemma2_2B(BaseLlama, torch.nn.Module):
 
         self.model = Llama2_(config, device=device, dtype=dtype, ops=operations)
         self.dtype = dtype
-
-class Gemma3_4B(BaseLlama, torch.nn.Module):
-    def __init__(self, config_dict, dtype, device, operations):
-        super().__init__()
-        config = Gemma3_4B_Config(**config_dict)
-        self.num_layers = config.num_hidden_layers
-
-        self.model = Llama2_(config, device=device, dtype=dtype, ops=operations)
-        self.dtype = dtype

comfy/text_encoders/lumina2.py

@@ -11,41 +11,23 @@ class Gemma2BTokenizer(sd1_clip.SDTokenizer):
     def state_dict(self):
         return {"spiece_model": self.tokenizer.serialize_model()}
 
-class Gemma3_4BTokenizer(sd1_clip.SDTokenizer):
-    def __init__(self, embedding_directory=None, tokenizer_data={}):
-        tokenizer = tokenizer_data.get("spiece_model", None)
-        super().__init__(tokenizer, pad_with_end=False, embedding_size=2560, embedding_key='gemma3_4b', tokenizer_class=SPieceTokenizer, has_end_token=False, pad_to_max_length=False, max_length=99999999, min_length=1, tokenizer_args={"add_bos": True, "add_eos": False}, tokenizer_data=tokenizer_data)
-
-    def state_dict(self):
-        return {"spiece_model": self.tokenizer.serialize_model()}
 
 class LuminaTokenizer(sd1_clip.SD1Tokenizer):
     def __init__(self, embedding_directory=None, tokenizer_data={}):
         super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, name="gemma2_2b", tokenizer=Gemma2BTokenizer)
 
-class NTokenizer(sd1_clip.SD1Tokenizer):
-    def __init__(self, embedding_directory=None, tokenizer_data={}):
-        super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, name="gemma3_4b", tokenizer=Gemma3_4BTokenizer)
 
 class Gemma2_2BModel(sd1_clip.SDClipModel):
     def __init__(self, device="cpu", layer="hidden", layer_idx=-2, dtype=None, attention_mask=True, model_options={}):
         super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config={}, dtype=dtype, special_tokens={"start": 2, "pad": 0}, layer_norm_hidden_state=False, model_class=comfy.text_encoders.llama.Gemma2_2B, enable_attention_masks=attention_mask, return_attention_masks=attention_mask, model_options=model_options)
 
-class Gemma3_4BModel(sd1_clip.SDClipModel):
-    def __init__(self, device="cpu", layer="hidden", layer_idx=-2, dtype=None, attention_mask=True, model_options={}):
-        super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config={}, dtype=dtype, special_tokens={"start": 2, "pad": 0}, layer_norm_hidden_state=False, model_class=comfy.text_encoders.llama.Gemma3_4B, enable_attention_masks=attention_mask, return_attention_masks=attention_mask, model_options=model_options)
 
 class LuminaModel(sd1_clip.SD1ClipModel):
-    def __init__(self, device="cpu", dtype=None, model_options={}, name="gemma2_2b", clip_model=Gemma2_2BModel):
-        super().__init__(device=device, dtype=dtype, name=name, clip_model=clip_model, model_options=model_options)
+    def __init__(self, device="cpu", dtype=None, model_options={}):
+        super().__init__(device=device, dtype=dtype, name="gemma2_2b", clip_model=Gemma2_2BModel, model_options=model_options)
 
 
-def te(dtype_llama=None, llama_scaled_fp8=None, model_type="gemma2_2b"):
-    if model_type == "gemma2_2b":
-        model = Gemma2_2BModel
-    elif model_type == "gemma3_4b":
-        model = Gemma3_4BModel
-
+def te(dtype_llama=None, llama_scaled_fp8=None):
     class LuminaTEModel_(LuminaModel):
         def __init__(self, device="cpu", dtype=None, model_options={}):
             if llama_scaled_fp8 is not None and "scaled_fp8" not in model_options:
@@ -53,5 +35,5 @@ def te(dtype_llama=None, llama_scaled_fp8=None, model_type="gemma2_2b"):
                 model_options["scaled_fp8"] = llama_scaled_fp8
             if dtype_llama is not None:
                 dtype = dtype_llama
-            super().__init__(device=device, dtype=dtype, name=model_type, model_options=model_options, clip_model=model)
+            super().__init__(device=device, dtype=dtype, model_options=model_options)
     return LuminaTEModel_

comfy/text_encoders/qwen_image.py

@@ -17,25 +17,14 @@ class QwenImageTokenizer(sd1_clip.SD1Tokenizer):
         self.llama_template = "<|im_start|>system\nDescribe the image by detailing the color, shape, size, texture, quantity, text, spatial relationships of the objects and background:<|im_end|>\n<|im_start|>user\n{}<|im_end|>\n<|im_start|>assistant\n"
         self.llama_template_images = "<|im_start|>system\nDescribe the key features of the input image (color, shape, size, texture, objects, background), then explain how the user's text instruction should alter or modify the image. Generate a new image that meets the user's requirements while maintaining consistency with the original input where appropriate.<|im_end|>\n<|im_start|>user\n<|vision_start|><|image_pad|><|vision_end|>{}<|im_end|>\n<|im_start|>assistant\n"
 
-    def tokenize_with_weights(self, text, return_word_ids=False, llama_template=None, images=[], prevent_empty_text=False, **kwargs):
-        skip_template = False
-        if text.startswith('<|im_start|>'):
-            skip_template = True
-        if text.startswith('<|start_header_id|>'):
-            skip_template = True
-        if prevent_empty_text and text == '':
-            text = ' '
-
-        if skip_template:
-            llama_text = text
-        else:
-            if llama_template is None:
-                if len(images) > 0:
-                    llama_text = self.llama_template_images.format(text)
-                else:
-                    llama_text = self.llama_template.format(text)
+    def tokenize_with_weights(self, text, return_word_ids=False, llama_template=None, images=[], **kwargs):
+        if llama_template is None:
+            if len(images) > 0:
+                llama_text = self.llama_template_images.format(text)
             else:
-                llama_text = llama_template.format(text)
+                llama_text = self.llama_template.format(text)
+        else:
+            llama_text = llama_template.format(text)
         tokens = super().tokenize_with_weights(llama_text, return_word_ids=return_word_ids, disable_weights=True, **kwargs)
         key_name = next(iter(tokens))
         embed_count = 0
@@ -58,23 +47,22 @@ class QwenImageTEModel(sd1_clip.SD1ClipModel):
     def __init__(self, device="cpu", dtype=None, model_options={}):
         super().__init__(device=device, dtype=dtype, name="qwen25_7b", clip_model=Qwen25_7BVLIModel, model_options=model_options)
 
-    def encode_token_weights(self, token_weight_pairs, template_end=-1):
+    def encode_token_weights(self, token_weight_pairs):
         out, pooled, extra = super().encode_token_weights(token_weight_pairs)
         tok_pairs = token_weight_pairs["qwen25_7b"][0]
         count_im_start = 0
-        if template_end == -1:
-            for i, v in enumerate(tok_pairs):
-                elem = v[0]
-                if not torch.is_tensor(elem):
-                    if isinstance(elem, numbers.Integral):
-                        if elem == 151644 and count_im_start < 2:
-                            template_end = i
-                            count_im_start += 1
+        for i, v in enumerate(tok_pairs):
+            elem = v[0]
+            if not torch.is_tensor(elem):
+                if isinstance(elem, numbers.Integral):
+                    if elem == 151644 and count_im_start < 2:
+                        template_end = i
+                        count_im_start += 1
 
-            if out.shape[1] > (template_end + 3):
-                if tok_pairs[template_end + 1][0] == 872:
-                    if tok_pairs[template_end + 2][0] == 198:
-                        template_end += 3
+        if out.shape[1] > (template_end + 3):
+            if tok_pairs[template_end + 1][0] == 872:
+                if tok_pairs[template_end + 2][0] == 198:
+                    template_end += 3
 
         out = out[:, template_end:]
 

comfy/utils.py

@@ -39,11 +39,7 @@ if hasattr(torch.serialization, "add_safe_globals"):  # TODO: this was added in
         pass
     ModelCheckpoint.__module__ = "pytorch_lightning.callbacks.model_checkpoint"
 
-    def scalar(*args, **kwargs):
-        from numpy.core.multiarray import scalar as sc
-        return sc(*args, **kwargs)
-    scalar.__module__ = "numpy.core.multiarray"
-
+    from numpy.core.multiarray import scalar
     from numpy import dtype
     from numpy.dtypes import Float64DType
     from _codecs import encode
@@ -1106,25 +1102,3 @@ def upscale_dit_mask(mask: torch.Tensor, img_size_in, img_size_out):
             dim=1
         )
         return out
-
-def pack_latents(latents):
-    latent_shapes = []
-    tensors = []
-    for tensor in latents:
-        latent_shapes.append(tensor.shape)
-        tensors.append(tensor.reshape(tensor.shape[0], 1, -1))
-
-    latent = torch.cat(tensors, dim=-1)
-    return latent, latent_shapes
-
-def unpack_latents(combined_latent, latent_shapes):
-    if len(latent_shapes) > 1:
-        output_tensors = []
-        for shape in latent_shapes:
-            cut = math.prod(shape[1:])
-            tens = combined_latent[:, :, :cut]
-            combined_latent = combined_latent[:, :, cut:]
-            output_tensors.append(tens.reshape([tens.shape[0]] + list(shape)[1:]))
-    else:
-        output_tensors = combined_latent
-    return output_tensors

comfy_api/latest/__init__.py

@@ -7,9 +7,9 @@ from comfy_api.internal.singleton import ProxiedSingleton
 from comfy_api.internal.async_to_sync import create_sync_class
 from comfy_api.latest._input import ImageInput, AudioInput, MaskInput, LatentInput, VideoInput
 from comfy_api.latest._input_impl import VideoFromFile, VideoFromComponents
-from comfy_api.latest._util import VideoCodec, VideoContainer, VideoComponents, MESH, VOXEL
-from . import _io as io
-from . import _ui as ui
+from comfy_api.latest._util import VideoCodec, VideoContainer, VideoComponents
+from comfy_api.latest._io import _IO as io  #noqa: F401
+from comfy_api.latest._ui import _UI as ui  #noqa: F401
 # from comfy_api.latest._resources import _RESOURCES as resources  #noqa: F401
 from comfy_execution.utils import get_executing_context
 from comfy_execution.progress import get_progress_state, PreviewImageTuple
@@ -104,8 +104,6 @@ class Types:
     VideoCodec = VideoCodec
     VideoContainer = VideoContainer
     VideoComponents = VideoComponents
-    MESH = MESH
-    VOXEL = VOXEL
 
 ComfyAPI = ComfyAPI_latest
 
@@ -116,10 +114,6 @@ if TYPE_CHECKING:
     ComfyAPISync: Type[comfy_api.latest.generated.ComfyAPISyncStub.ComfyAPISyncStub]
 ComfyAPISync = create_sync_class(ComfyAPI_latest)
 
-# create new aliases for io and ui
-IO = io
-UI = ui
-
 __all__ = [
     "ComfyAPI",
     "ComfyAPISync",
@@ -127,8 +121,4 @@ __all__ = [
     "InputImpl",
     "Types",
     "ComfyExtension",
-    "io",
-    "IO",
-    "ui",
-    "UI",
 ]

comfy_api/latest/_input/video_types.py

@@ -1,6 +1,6 @@
 from __future__ import annotations
 from abc import ABC, abstractmethod
-from typing import Optional, Union, IO
+from typing import Optional, Union
 import io
 import av
 from comfy_api.util import VideoContainer, VideoCodec, VideoComponents
@@ -23,7 +23,7 @@ class VideoInput(ABC):
     @abstractmethod
     def save_to(
         self,
-        path: Union[str, IO[bytes]],
+        path: str,
         format: VideoContainer = VideoContainer.AUTO,
         codec: VideoCodec = VideoCodec.AUTO,
         metadata: Optional[dict] = None

comfy_api/latest/_io.py

@@ -27,7 +27,6 @@ from comfy_api.internal import (_ComfyNodeInternal, _NodeOutputInternal, classpr
     prune_dict, shallow_clone_class)
 from comfy_api.latest._resources import Resources, ResourcesLocal
 from comfy_execution.graph_utils import ExecutionBlocker
-from ._util import MESH, VOXEL
 
 # from comfy_extras.nodes_images import SVG as SVG_ # NOTE: needs to be moved before can be imported due to circular reference
 
@@ -337,25 +336,11 @@ class Combo(ComfyTypeIO):
     class Input(WidgetInput):
         """Combo input (dropdown)."""
         Type = str
-        def __init__(
-            self,
-            id: str,
-            options: list[str] | list[int] | type[Enum] = None,
-            display_name: str=None,
-            optional=False,
-            tooltip: str=None,
-            lazy: bool=None,
-            default: str | int | Enum = None,
-            control_after_generate: bool=None,
-            upload: UploadType=None,
-            image_folder: FolderType=None,
-            remote: RemoteOptions=None,
-            socketless: bool=None,
-        ):
-            if isinstance(options, type) and issubclass(options, Enum):
-                options = [v.value for v in options]
-            if isinstance(default, Enum):
-                default = default.value
+        def __init__(self, id: str, options: list[str]=None, display_name: str=None, optional=False, tooltip: str=None, lazy: bool=None,
+                    default: str=None, control_after_generate: bool=None,
+                    upload: UploadType=None, image_folder: FolderType=None,
+                    remote: RemoteOptions=None,
+                    socketless: bool=None):
             super().__init__(id, display_name, optional, tooltip, lazy, default, socketless)
             self.multiselect = False
             self.options = options
@@ -629,10 +614,6 @@ class UpscaleModel(ComfyTypeIO):
     if TYPE_CHECKING:
         Type = ImageModelDescriptor
 
-@comfytype(io_type="LATENT_UPSCALE_MODEL")
-class LatentUpscaleModel(ComfyTypeIO):
-    Type = Any
-
 @comfytype(io_type="AUDIO")
 class Audio(ComfyTypeIO):
     class AudioDict(TypedDict):
@@ -661,11 +642,11 @@ class LossMap(ComfyTypeIO):
 
 @comfytype(io_type="VOXEL")
 class Voxel(ComfyTypeIO):
-    Type = VOXEL
+    Type = Any # TODO: VOXEL class is defined in comfy_extras/nodes_hunyuan3d.py; should be moved to somewhere else before referenced directly in v3
 
 @comfytype(io_type="MESH")
 class Mesh(ComfyTypeIO):
-    Type = MESH
+    Type = Any # TODO: MESH class is defined in comfy_extras/nodes_hunyuan3d.py; should be moved to somewhere else before referenced directly in v3
 
 @comfytype(io_type="HOOKS")
 class Hooks(ComfyTypeIO):
@@ -1587,78 +1568,77 @@ class _UIOutput(ABC):
         ...
 
 
-__all__ = [
-    "FolderType",
-    "UploadType",
-    "RemoteOptions",
-    "NumberDisplay",
+class _IO:
+    FolderType = FolderType
+    UploadType = UploadType
+    RemoteOptions = RemoteOptions
+    NumberDisplay = NumberDisplay
 
-    "comfytype",
-    "Custom",
-    "Input",
-    "WidgetInput",
-    "Output",
-    "ComfyTypeI",
-    "ComfyTypeIO",
+    comfytype = staticmethod(comfytype)
+    Custom = staticmethod(Custom)
+    Input = Input
+    WidgetInput = WidgetInput
+    Output = Output
+    ComfyTypeI = ComfyTypeI
+    ComfyTypeIO = ComfyTypeIO
+    #---------------------------------
     # Supported Types
-    "Boolean",
-    "Int",
-    "Float",
-    "String",
-    "Combo",
-    "MultiCombo",
-    "Image",
-    "WanCameraEmbedding",
-    "Webcam",
-    "Mask",
-    "Latent",
-    "Conditioning",
-    "Sampler",
-    "Sigmas",
-    "Noise",
-    "Guider",
-    "Clip",
-    "ControlNet",
-    "Vae",
-    "Model",
-    "ClipVision",
-    "ClipVisionOutput",
-    "AudioEncoder",
-    "AudioEncoderOutput",
-    "StyleModel",
-    "Gligen",
-    "UpscaleModel",
-    "Audio",
-    "Video",
-    "SVG",
-    "LoraModel",
-    "LossMap",
-    "Voxel",
-    "Mesh",
-    "Hooks",
-    "HookKeyframes",
-    "TimestepsRange",
-    "LatentOperation",
-    "FlowControl",
-    "Accumulation",
-    "Load3DCamera",
-    "Load3D",
-    "Load3DAnimation",
-    "Photomaker",
-    "Point",
-    "FaceAnalysis",
-    "BBOX",
-    "SEGS",
-    "AnyType",
-    "MultiType",
-    # Other classes
-    "HiddenHolder",
-    "Hidden",
-    "NodeInfoV1",
-    "NodeInfoV3",
-    "Schema",
-    "ComfyNode",
-    "NodeOutput",
-    "add_to_dict_v1",
-    "add_to_dict_v3",
-]
+    Boolean = Boolean
+    Int = Int
+    Float = Float
+    String = String
+    Combo = Combo
+    MultiCombo = MultiCombo
+    Image = Image
+    WanCameraEmbedding = WanCameraEmbedding
+    Webcam = Webcam
+    Mask = Mask
+    Latent = Latent
+    Conditioning = Conditioning
+    Sampler = Sampler
+    Sigmas = Sigmas
+    Noise = Noise
+    Guider = Guider
+    Clip = Clip
+    ControlNet = ControlNet
+    Vae = Vae
+    Model = Model
+    ClipVision = ClipVision
+    ClipVisionOutput = ClipVisionOutput
+    AudioEncoderOutput = AudioEncoderOutput
+    StyleModel = StyleModel
+    Gligen = Gligen
+    UpscaleModel = UpscaleModel
+    Audio = Audio
+    Video = Video
+    SVG = SVG
+    LoraModel = LoraModel
+    LossMap = LossMap
+    Voxel = Voxel
+    Mesh = Mesh
+    Hooks = Hooks
+    HookKeyframes = HookKeyframes
+    TimestepsRange = TimestepsRange
+    LatentOperation = LatentOperation
+    FlowControl = FlowControl
+    Accumulation = Accumulation
+    Load3DCamera = Load3DCamera
+    Load3D = Load3D
+    Load3DAnimation = Load3DAnimation
+    Photomaker = Photomaker
+    Point = Point
+    FaceAnalysis = FaceAnalysis
+    BBOX = BBOX
+    SEGS = SEGS
+    AnyType = AnyType
+    MultiType = MultiType
+    #---------------------------------
+    HiddenHolder = HiddenHolder
+    Hidden = Hidden
+    NodeInfoV1 = NodeInfoV1
+    NodeInfoV3 = NodeInfoV3
+    Schema = Schema
+    ComfyNode = ComfyNode
+    NodeOutput = NodeOutput
+    add_to_dict_v1 = staticmethod(add_to_dict_v1)
+    add_to_dict_v3 = staticmethod(add_to_dict_v3)

comfy_api/latest/_ui.py

@@ -449,16 +449,15 @@ class PreviewText(_UIOutput):
         return {"text": (self.value,)}
 
 
-__all__ = [
-    "SavedResult",
-    "SavedImages",
-    "SavedAudios",
-    "ImageSaveHelper",
-    "AudioSaveHelper",
-    "PreviewImage",
-    "PreviewMask",
-    "PreviewAudio",
-    "PreviewVideo",
-    "PreviewUI3D",
-    "PreviewText",
-]
+class _UI:
+    SavedResult = SavedResult
+    SavedImages = SavedImages
+    SavedAudios = SavedAudios
+    ImageSaveHelper = ImageSaveHelper
+    AudioSaveHelper = AudioSaveHelper
+    PreviewImage = PreviewImage
+    PreviewMask = PreviewMask
+    PreviewAudio = PreviewAudio
+    PreviewVideo = PreviewVideo
+    PreviewUI3D = PreviewUI3D
+    PreviewText = PreviewText

comfy_api/latest/_util/__init__.py

@@ -1,11 +1,8 @@
 from .video_types import VideoContainer, VideoCodec, VideoComponents
-from .geometry_types import VOXEL, MESH
 
 __all__ = [
     # Utility Types
     "VideoContainer",
     "VideoCodec",
     "VideoComponents",
-    "VOXEL",
-    "MESH",
 ]

comfy_api/latest/_util/geometry_types.py

@@ -1,12 +0,0 @@
-import torch
-
-
-class VOXEL:
-    def __init__(self, data: torch.Tensor):
-        self.data = data
-
-
-class MESH:
-    def __init__(self, vertices: torch.Tensor, faces: torch.Tensor):
-        self.vertices = vertices
-        self.faces = faces

comfy_api_nodes/apinode_utils.py

@@ -0,0 +1,691 @@
+from __future__ import annotations
+import aiohttp
+import io
+import logging
+import mimetypes
+from typing import Optional, Union
+from comfy.utils import common_upscale
+from comfy_api.input_impl import VideoFromFile
+from comfy_api.util import VideoContainer, VideoCodec
+from comfy_api.input.video_types import VideoInput
+from comfy_api.input.basic_types import AudioInput
+from comfy_api_nodes.apis.client import (
+    ApiClient,
+    ApiEndpoint,
+    HttpMethod,
+    SynchronousOperation,
+    UploadRequest,
+    UploadResponse,
+)
+from server import PromptServer
+
+
+import numpy as np
+from PIL import Image
+import torch
+import math
+import base64
+import uuid
+from io import BytesIO
+import av
+
+
+async def download_url_to_video_output(video_url: str, timeout: int = None) -> VideoFromFile:
+    """Downloads a video from a URL and returns a `VIDEO` output.
+
+    Args:
+        video_url: The URL of the video to download.
+
+    Returns:
+        A Comfy node `VIDEO` output.
+    """
+    video_io = await download_url_to_bytesio(video_url, timeout)
+    if video_io is None:
+        error_msg = f"Failed to download video from {video_url}"
+        logging.error(error_msg)
+        raise ValueError(error_msg)
+    return VideoFromFile(video_io)
+
+
+def downscale_image_tensor(image, total_pixels=1536 * 1024) -> torch.Tensor:
+    """Downscale input image tensor to roughly the specified total pixels."""
+    samples = image.movedim(-1, 1)
+    total = int(total_pixels)
+    scale_by = math.sqrt(total / (samples.shape[3] * samples.shape[2]))
+    if scale_by >= 1:
+        return image
+    width = round(samples.shape[3] * scale_by)
+    height = round(samples.shape[2] * scale_by)
+
+    s = common_upscale(samples, width, height, "lanczos", "disabled")
+    s = s.movedim(1, -1)
+    return s
+
+
+async def validate_and_cast_response(
+    response, timeout: int = None, node_id: Union[str, None] = None
+) -> torch.Tensor:
+    """Validates and casts a response to a torch.Tensor.
+
+    Args:
+        response: The response to validate and cast.
+        timeout: Request timeout in seconds. Defaults to None (no timeout).
+
+    Returns:
+        A torch.Tensor representing the image (1, H, W, C).
+
+    Raises:
+        ValueError: If the response is not valid.
+    """
+    # validate raw JSON response
+    data = response.data
+    if not data or len(data) == 0:
+        raise ValueError("No images returned from API endpoint")
+
+    # Initialize list to store image tensors
+    image_tensors: list[torch.Tensor] = []
+
+    # Process each image in the data array
+    async with aiohttp.ClientSession(timeout=aiohttp.ClientTimeout(total=timeout)) as session:
+        for img_data in data:
+            img_bytes: bytes
+            if img_data.b64_json:
+                img_bytes = base64.b64decode(img_data.b64_json)
+            elif img_data.url:
+                if node_id:
+                    PromptServer.instance.send_progress_text(f"Result URL: {img_data.url}", node_id)
+                async with session.get(img_data.url) as resp:
+                    if resp.status != 200:
+                        raise ValueError("Failed to download generated image")
+                    img_bytes = await resp.read()
+            else:
+                raise ValueError("Invalid image payload – neither URL nor base64 data present.")
+
+            pil_img = Image.open(BytesIO(img_bytes)).convert("RGBA")
+            arr = np.asarray(pil_img).astype(np.float32) / 255.0
+            image_tensors.append(torch.from_numpy(arr))
+
+    return torch.stack(image_tensors, dim=0)
+
+
+def validate_aspect_ratio(
+    aspect_ratio: str,
+    minimum_ratio: float,
+    maximum_ratio: float,
+    minimum_ratio_str: str,
+    maximum_ratio_str: str,
+) -> float:
+    """Validates and casts an aspect ratio string to a float.
+
+    Args:
+        aspect_ratio: The aspect ratio string to validate.
+        minimum_ratio: The minimum aspect ratio.
+        maximum_ratio: The maximum aspect ratio.
+        minimum_ratio_str: The minimum aspect ratio string.
+        maximum_ratio_str: The maximum aspect ratio string.
+
+    Returns:
+        The validated and cast aspect ratio.
+
+    Raises:
+        Exception: If the aspect ratio is not valid.
+    """
+    # get ratio values
+    numbers = aspect_ratio.split(":")
+    if len(numbers) != 2:
+        raise TypeError(
+            f"Aspect ratio must be in the format X:Y, such as 16:9, but was {aspect_ratio}."
+        )
+    try:
+        numerator = int(numbers[0])
+        denominator = int(numbers[1])
+    except ValueError as exc:
+        raise TypeError(
+            f"Aspect ratio must contain numbers separated by ':', such as 16:9, but was {aspect_ratio}."
+        ) from exc
+    calculated_ratio = numerator / denominator
+    # if not close to minimum and maximum, check bounds
+    if not math.isclose(calculated_ratio, minimum_ratio) or not math.isclose(
+        calculated_ratio, maximum_ratio
+    ):
+        if calculated_ratio < minimum_ratio:
+            raise TypeError(
+                f"Aspect ratio cannot reduce to any less than {minimum_ratio_str} ({minimum_ratio}), but was {aspect_ratio} ({calculated_ratio})."
+            )
+        elif calculated_ratio > maximum_ratio:
+            raise TypeError(
+                f"Aspect ratio cannot reduce to any greater than {maximum_ratio_str} ({maximum_ratio}), but was {aspect_ratio} ({calculated_ratio})."
+            )
+    return aspect_ratio
+
+
+def mimetype_to_extension(mime_type: str) -> str:
+    """Converts a MIME type to a file extension."""
+    return mime_type.split("/")[-1].lower()
+
+
+async def download_url_to_bytesio(url: str, timeout: int = None) -> BytesIO:
+    """Downloads content from a URL using requests and returns it as BytesIO.
+
+    Args:
+        url: The URL to download.
+        timeout: Request timeout in seconds. Defaults to None (no timeout).
+
+    Returns:
+        BytesIO object containing the downloaded content.
+    """
+    timeout_cfg = aiohttp.ClientTimeout(total=timeout) if timeout else None
+    async with aiohttp.ClientSession(timeout=timeout_cfg) as session:
+        async with session.get(url) as resp:
+            resp.raise_for_status()  # Raises HTTPError for bad responses (4XX or 5XX)
+            return BytesIO(await resp.read())
+
+
+def bytesio_to_image_tensor(image_bytesio: BytesIO, mode: str = "RGBA") -> torch.Tensor:
+    """Converts image data from BytesIO to a torch.Tensor.
+
+    Args:
+        image_bytesio: BytesIO object containing the image data.
+        mode: The PIL mode to convert the image to (e.g., "RGB", "RGBA").
+
+    Returns:
+        A torch.Tensor representing the image (1, H, W, C).
+
+    Raises:
+        PIL.UnidentifiedImageError: If the image data cannot be identified.
+        ValueError: If the specified mode is invalid.
+    """
+    image = Image.open(image_bytesio)
+    image = image.convert(mode)
+    image_array = np.array(image).astype(np.float32) / 255.0
+    return torch.from_numpy(image_array).unsqueeze(0)
+
+
+async def download_url_to_image_tensor(url: str, timeout: int = None) -> torch.Tensor:
+    """Downloads an image from a URL and returns a [B, H, W, C] tensor."""
+    image_bytesio = await download_url_to_bytesio(url, timeout)
+    return bytesio_to_image_tensor(image_bytesio)
+
+
+def process_image_response(response_content: bytes | str) -> torch.Tensor:
+    """Uses content from a Response object and converts it to a torch.Tensor"""
+    return bytesio_to_image_tensor(BytesIO(response_content))
+
+
+def _tensor_to_pil(image: torch.Tensor, total_pixels: int = 2048 * 2048) -> Image.Image:
+    """Converts a single torch.Tensor image [H, W, C] to a PIL Image, optionally downscaling."""
+    if len(image.shape) > 3:
+        image = image[0]
+    # TODO: remove alpha if not allowed and present
+    input_tensor = image.cpu()
+    input_tensor = downscale_image_tensor(
+        input_tensor.unsqueeze(0), total_pixels=total_pixels
+    ).squeeze()
+    image_np = (input_tensor.numpy() * 255).astype(np.uint8)
+    img = Image.fromarray(image_np)
+    return img
+
+
+def _pil_to_bytesio(img: Image.Image, mime_type: str = "image/png") -> BytesIO:
+    """Converts a PIL Image to a BytesIO object."""
+    if not mime_type:
+        mime_type = "image/png"
+
+    img_byte_arr = io.BytesIO()
+    # Derive PIL format from MIME type (e.g., 'image/png' -> 'PNG')
+    pil_format = mime_type.split("/")[-1].upper()
+    if pil_format == "JPG":
+        pil_format = "JPEG"
+    img.save(img_byte_arr, format=pil_format)
+    img_byte_arr.seek(0)
+    return img_byte_arr
+
+
+def tensor_to_bytesio(
+    image: torch.Tensor,
+    name: Optional[str] = None,
+    total_pixels: int = 2048 * 2048,
+    mime_type: str = "image/png",
+) -> BytesIO:
+    """Converts a torch.Tensor image to a named BytesIO object.
+
+    Args:
+        image: Input torch.Tensor image.
+        name: Optional filename for the BytesIO object.
+        total_pixels: Maximum total pixels for potential downscaling.
+        mime_type: Target image MIME type (e.g., 'image/png', 'image/jpeg', 'image/webp', 'video/mp4').
+
+    Returns:
+        Named BytesIO object containing the image data.
+    """
+    if not mime_type:
+        mime_type = "image/png"
+
+    pil_image = _tensor_to_pil(image, total_pixels=total_pixels)
+    img_binary = _pil_to_bytesio(pil_image, mime_type=mime_type)
+    img_binary.name = (
+        f"{name if name else uuid.uuid4()}.{mimetype_to_extension(mime_type)}"
+    )
+    return img_binary
+
+
+def tensor_to_base64_string(
+    image_tensor: torch.Tensor,
+    total_pixels: int = 2048 * 2048,
+    mime_type: str = "image/png",
+) -> str:
+    """Convert [B, H, W, C] or [H, W, C] tensor to a base64 string.
+
+    Args:
+        image_tensor: Input torch.Tensor image.
+        total_pixels: Maximum total pixels for potential downscaling.
+        mime_type: Target image MIME type (e.g., 'image/png', 'image/jpeg', 'image/webp', 'video/mp4').
+
+    Returns:
+        Base64 encoded string of the image.
+    """
+    pil_image = _tensor_to_pil(image_tensor, total_pixels=total_pixels)
+    img_byte_arr = _pil_to_bytesio(pil_image, mime_type=mime_type)
+    img_bytes = img_byte_arr.getvalue()
+    # Encode bytes to base64 string
+    base64_encoded_string = base64.b64encode(img_bytes).decode("utf-8")
+    return base64_encoded_string
+
+
+def tensor_to_data_uri(
+    image_tensor: torch.Tensor,
+    total_pixels: int = 2048 * 2048,
+    mime_type: str = "image/png",
+) -> str:
+    """Converts a tensor image to a Data URI string.
+
+    Args:
+        image_tensor: Input torch.Tensor image.
+        total_pixels: Maximum total pixels for potential downscaling.
+        mime_type: Target image MIME type (e.g., 'image/png', 'image/jpeg', 'image/webp').
+
+    Returns:
+        Data URI string (e.g., 'data:image/png;base64,...').
+    """
+    base64_string = tensor_to_base64_string(image_tensor, total_pixels, mime_type)
+    return f"data:{mime_type};base64,{base64_string}"
+
+
+def text_filepath_to_base64_string(filepath: str) -> str:
+    """Converts a text file to a base64 string."""
+    with open(filepath, "rb") as f:
+        file_content = f.read()
+    return base64.b64encode(file_content).decode("utf-8")
+
+
+def text_filepath_to_data_uri(filepath: str) -> str:
+    """Converts a text file to a data URI."""
+    base64_string = text_filepath_to_base64_string(filepath)
+    mime_type, _ = mimetypes.guess_type(filepath)
+    if mime_type is None:
+        mime_type = "application/octet-stream"
+    return f"data:{mime_type};base64,{base64_string}"
+
+
+async def upload_file_to_comfyapi(
+    file_bytes_io: BytesIO,
+    filename: str,
+    upload_mime_type: Optional[str],
+    auth_kwargs: Optional[dict[str, str]] = None,
+) -> str:
+    """
+    Uploads a single file to ComfyUI API and returns its download URL.
+
+    Args:
+        file_bytes_io: BytesIO object containing the file data.
+        filename: The filename of the file.
+        upload_mime_type: MIME type of the file.
+        auth_kwargs: Optional authentication token(s).
+
+    Returns:
+        The download URL for the uploaded file.
+    """
+    if upload_mime_type is None:
+        request_object = UploadRequest(file_name=filename)
+    else:
+        request_object = UploadRequest(file_name=filename, content_type=upload_mime_type)
+    operation = SynchronousOperation(
+        endpoint=ApiEndpoint(
+            path="/customers/storage",
+            method=HttpMethod.POST,
+            request_model=UploadRequest,
+            response_model=UploadResponse,
+        ),
+        request=request_object,
+        auth_kwargs=auth_kwargs,
+    )
+
+    response: UploadResponse = await operation.execute()
+    await ApiClient.upload_file(response.upload_url, file_bytes_io, content_type=upload_mime_type)
+    return response.download_url
+
+
+def video_to_base64_string(
+    video: VideoInput,
+    container_format: VideoContainer = None,
+    codec: VideoCodec = None
+) -> str:
+    """
+    Converts a video input to a base64 string.
+
+    Args:
+        video: The video input to convert
+        container_format: Optional container format to use (defaults to video.container if available)
+        codec: Optional codec to use (defaults to video.codec if available)
+    """
+    video_bytes_io = io.BytesIO()
+
+    # Use provided format/codec if specified, otherwise use video's own if available
+    format_to_use = container_format if container_format is not None else getattr(video, 'container', VideoContainer.MP4)
+    codec_to_use = codec if codec is not None else getattr(video, 'codec', VideoCodec.H264)
+
+    video.save_to(video_bytes_io, format=format_to_use, codec=codec_to_use)
+    video_bytes_io.seek(0)
+    return base64.b64encode(video_bytes_io.getvalue()).decode("utf-8")
+
+
+async def upload_video_to_comfyapi(
+    video: VideoInput,
+    auth_kwargs: Optional[dict[str, str]] = None,
+    container: VideoContainer = VideoContainer.MP4,
+    codec: VideoCodec = VideoCodec.H264,
+    max_duration: Optional[int] = None,
+) -> str:
+    """
+    Uploads a single video to ComfyUI API and returns its download URL.
+    Uses the specified container and codec for saving the video before upload.
+
+    Args:
+        video: VideoInput object (Comfy VIDEO type).
+        auth_kwargs: Optional authentication token(s).
+        container: The video container format to use (default: MP4).
+        codec: The video codec to use (default: H264).
+        max_duration: Optional maximum duration of the video in seconds. If the video is longer than this, an error will be raised.
+
+    Returns:
+        The download URL for the uploaded video file.
+    """
+    if max_duration is not None:
+        try:
+            actual_duration = video.duration_seconds
+            if actual_duration is not None and actual_duration > max_duration:
+                raise ValueError(
+                    f"Video duration ({actual_duration:.2f}s) exceeds the maximum allowed ({max_duration}s)."
+                )
+        except Exception as e:
+            logging.error(f"Error getting video duration: {e}")
+            raise ValueError(f"Could not verify video duration from source: {e}") from e
+
+    upload_mime_type = f"video/{container.value.lower()}"
+    filename = f"uploaded_video.{container.value.lower()}"
+
+    # Convert VideoInput to BytesIO using specified container/codec
+    video_bytes_io = io.BytesIO()
+    video.save_to(video_bytes_io, format=container, codec=codec)
+    video_bytes_io.seek(0)
+
+    return await upload_file_to_comfyapi(video_bytes_io, filename, upload_mime_type, auth_kwargs)
+
+
+def audio_tensor_to_contiguous_ndarray(waveform: torch.Tensor) -> np.ndarray:
+    """
+    Prepares audio waveform for av library by converting to a contiguous numpy array.
+
+    Args:
+        waveform: a tensor of shape (1, channels, samples) derived from a Comfy `AUDIO` type.
+
+    Returns:
+        Contiguous numpy array of the audio waveform. If the audio was batched,
+            the first item is taken.
+    """
+    if waveform.ndim != 3 or waveform.shape[0] != 1:
+        raise ValueError("Expected waveform tensor shape (1, channels, samples)")
+
+    # If batch is > 1, take first item
+    if waveform.shape[0] > 1:
+        waveform = waveform[0]
+
+    # Prepare for av: remove batch dim, move to CPU, make contiguous, convert to numpy array
+    audio_data_np = waveform.squeeze(0).cpu().contiguous().numpy()
+    if audio_data_np.dtype != np.float32:
+        audio_data_np = audio_data_np.astype(np.float32)
+
+    return audio_data_np
+
+
+def audio_ndarray_to_bytesio(
+    audio_data_np: np.ndarray,
+    sample_rate: int,
+    container_format: str = "mp4",
+    codec_name: str = "aac",
+) -> BytesIO:
+    """
+    Encodes a numpy array of audio data into a BytesIO object.
+    """
+    audio_bytes_io = io.BytesIO()
+    with av.open(audio_bytes_io, mode="w", format=container_format) as output_container:
+        audio_stream = output_container.add_stream(codec_name, rate=sample_rate)
+        frame = av.AudioFrame.from_ndarray(
+            audio_data_np,
+            format="fltp",
+            layout="stereo" if audio_data_np.shape[0] > 1 else "mono",
+        )
+        frame.sample_rate = sample_rate
+        frame.pts = 0
+
+        for packet in audio_stream.encode(frame):
+            output_container.mux(packet)
+
+        # Flush stream
+        for packet in audio_stream.encode(None):
+            output_container.mux(packet)
+
+    audio_bytes_io.seek(0)
+    return audio_bytes_io
+
+
+async def upload_audio_to_comfyapi(
+    audio: AudioInput,
+    auth_kwargs: Optional[dict[str, str]] = None,
+    container_format: str = "mp4",
+    codec_name: str = "aac",
+    mime_type: str = "audio/mp4",
+    filename: str = "uploaded_audio.mp4",
+) -> str:
+    """
+    Uploads a single audio input to ComfyUI API and returns its download URL.
+    Encodes the raw waveform into the specified format before uploading.
+
+    Args:
+        audio: a Comfy `AUDIO` type (contains waveform tensor and sample_rate)
+        auth_kwargs: Optional authentication token(s).
+
+    Returns:
+        The download URL for the uploaded audio file.
+    """
+    sample_rate: int = audio["sample_rate"]
+    waveform: torch.Tensor = audio["waveform"]
+    audio_data_np = audio_tensor_to_contiguous_ndarray(waveform)
+    audio_bytes_io = audio_ndarray_to_bytesio(
+        audio_data_np, sample_rate, container_format, codec_name
+    )
+
+    return await upload_file_to_comfyapi(audio_bytes_io, filename, mime_type, auth_kwargs)
+
+
+def f32_pcm(wav: torch.Tensor) -> torch.Tensor:
+    """Convert audio to float 32 bits PCM format. Copy-paste from nodes_audio.py file."""
+    if wav.dtype.is_floating_point:
+        return wav
+    elif wav.dtype == torch.int16:
+        return wav.float() / (2 ** 15)
+    elif wav.dtype == torch.int32:
+        return wav.float() / (2 ** 31)
+    raise ValueError(f"Unsupported wav dtype: {wav.dtype}")
+
+
+def audio_bytes_to_audio_input(audio_bytes: bytes,) -> dict:
+    """
+    Decode any common audio container from bytes using PyAV and return
+    a Comfy AUDIO dict: {"waveform": [1, C, T] float32, "sample_rate": int}.
+    """
+    with av.open(io.BytesIO(audio_bytes)) as af:
+        if not af.streams.audio:
+            raise ValueError("No audio stream found in response.")
+        stream = af.streams.audio[0]
+
+        in_sr = int(stream.codec_context.sample_rate)
+        out_sr = in_sr
+
+        frames: list[torch.Tensor] = []
+        n_channels = stream.channels or 1
+
+        for frame in af.decode(streams=stream.index):
+            arr = frame.to_ndarray()  # shape can be [C, T] or [T, C] or [T]
+            buf = torch.from_numpy(arr)
+            if buf.ndim == 1:
+                buf = buf.unsqueeze(0)  # [T] -> [1, T]
+            elif buf.shape[0] != n_channels and buf.shape[-1] == n_channels:
+                buf = buf.transpose(0, 1).contiguous()  # [T, C] -> [C, T]
+            elif buf.shape[0] != n_channels:
+                buf = buf.reshape(-1, n_channels).t().contiguous()  # fallback to [C, T]
+            frames.append(buf)
+
+    if not frames:
+        raise ValueError("Decoded zero audio frames.")
+
+    wav = torch.cat(frames, dim=1)  # [C, T]
+    wav = f32_pcm(wav)
+    return {"waveform": wav.unsqueeze(0).contiguous(), "sample_rate": out_sr}
+
+
+def audio_input_to_mp3(audio: AudioInput) -> io.BytesIO:
+    waveform = audio["waveform"].cpu()
+
+    output_buffer = io.BytesIO()
+    output_container = av.open(output_buffer, mode='w', format="mp3")
+
+    out_stream = output_container.add_stream("libmp3lame", rate=audio["sample_rate"])
+    out_stream.bit_rate = 320000
+
+    frame = av.AudioFrame.from_ndarray(waveform.movedim(0, 1).reshape(1, -1).float().numpy(), format='flt', layout='mono' if waveform.shape[0] == 1 else 'stereo')
+    frame.sample_rate = audio["sample_rate"]
+    frame.pts = 0
+    output_container.mux(out_stream.encode(frame))
+    output_container.mux(out_stream.encode(None))
+    output_container.close()
+    output_buffer.seek(0)
+    return output_buffer
+
+
+def audio_to_base64_string(
+    audio: AudioInput, container_format: str = "mp4", codec_name: str = "aac"
+) -> str:
+    """Converts an audio input to a base64 string."""
+    sample_rate: int = audio["sample_rate"]
+    waveform: torch.Tensor = audio["waveform"]
+    audio_data_np = audio_tensor_to_contiguous_ndarray(waveform)
+    audio_bytes_io = audio_ndarray_to_bytesio(
+        audio_data_np, sample_rate, container_format, codec_name
+    )
+    audio_bytes = audio_bytes_io.getvalue()
+    return base64.b64encode(audio_bytes).decode("utf-8")
+
+
+async def upload_images_to_comfyapi(
+    image: torch.Tensor,
+    max_images=8,
+    auth_kwargs: Optional[dict[str, str]] = None,
+    mime_type: Optional[str] = None,
+) -> list[str]:
+    """
+    Uploads images to ComfyUI API and returns download URLs.
+    To upload multiple images, stack them in the batch dimension first.
+
+    Args:
+        image: Input torch.Tensor image.
+        max_images: Maximum number of images to upload.
+        auth_kwargs: Optional authentication token(s).
+        mime_type: Optional MIME type for the image.
+    """
+    # if batch, try to upload each file if max_images is greater than 0
+    download_urls: list[str] = []
+    is_batch = len(image.shape) > 3
+    batch_len = image.shape[0] if is_batch else 1
+
+    for idx in range(min(batch_len, max_images)):
+        tensor = image[idx] if is_batch else image
+        img_io = tensor_to_bytesio(tensor, mime_type=mime_type)
+        url = await upload_file_to_comfyapi(img_io, img_io.name, mime_type, auth_kwargs)
+        download_urls.append(url)
+    return download_urls
+
+
+def resize_mask_to_image(
+    mask: torch.Tensor,
+    image: torch.Tensor,
+    upscale_method="nearest-exact",
+    crop="disabled",
+    allow_gradient=True,
+    add_channel_dim=False,
+):
+    """
+    Resize mask to be the same dimensions as an image, while maintaining proper format for API calls.
+    """
+    _, H, W, _ = image.shape
+    mask = mask.unsqueeze(-1)
+    mask = mask.movedim(-1, 1)
+    mask = common_upscale(
+        mask, width=W, height=H, upscale_method=upscale_method, crop=crop
+    )
+    mask = mask.movedim(1, -1)
+    if not add_channel_dim:
+        mask = mask.squeeze(-1)
+    if not allow_gradient:
+        mask = (mask > 0.5).float()
+    return mask
+
+
+def validate_string(
+    string: str,
+    strip_whitespace=True,
+    field_name="prompt",
+    min_length=None,
+    max_length=None,
+):
+    if string is None:
+        raise Exception(f"Field '{field_name}' cannot be empty.")
+    if strip_whitespace:
+        string = string.strip()
+    if min_length and len(string) < min_length:
+        raise Exception(
+            f"Field '{field_name}' cannot be shorter than {min_length} characters; was {len(string)} characters long."
+        )
+    if max_length and len(string) > max_length:
+        raise Exception(
+            f" Field '{field_name} cannot be longer than {max_length} characters; was {len(string)} characters long."
+        )
+
+
+def image_tensor_pair_to_batch(
+    image1: torch.Tensor, image2: torch.Tensor
+) -> torch.Tensor:
+    """
+    Converts a pair of image tensors to a batch tensor.
+    If the images are not the same size, the smaller image is resized to
+    match the larger image.
+    """
+    if image1.shape[1:] != image2.shape[1:]:
+        image2 = common_upscale(
+            image2.movedim(-1, 1),
+            image1.shape[2],
+            image1.shape[1],
+            "bilinear",
+            "center",
+        ).movedim(1, -1)
+    return torch.cat((image1, image2), dim=0)

comfy_api_nodes/apis/PixverseController.py

@@ -0,0 +1,17 @@
+# generated by datamodel-codegen:
+#   filename:  filtered-openapi.yaml
+#   timestamp: 2025-04-29T23:44:54+00:00
+
+from __future__ import annotations
+
+from typing import Optional
+
+from pydantic import BaseModel
+
+from . import PixverseDto
+
+
+class ResponseData(BaseModel):
+    ErrCode: Optional[int] = None
+    ErrMsg: Optional[str] = None
+    Resp: Optional[PixverseDto.V2OpenAPII2VResp] = None

comfy_api_nodes/apis/PixverseDto.py

@@ -0,0 +1,57 @@
+# generated by datamodel-codegen:
+#   filename:  filtered-openapi.yaml
+#   timestamp: 2025-04-29T23:44:54+00:00
+
+from __future__ import annotations
+
+from typing import Optional
+
+from pydantic import BaseModel, Field
+
+
+class V2OpenAPII2VResp(BaseModel):
+    video_id: Optional[int] = Field(None, description='Video_id')
+
+
+class V2OpenAPIT2VReq(BaseModel):
+    aspect_ratio: str = Field(
+        ..., description='Aspect ratio (16:9, 4:3, 1:1, 3:4, 9:16)', examples=['16:9']
+    )
+    duration: int = Field(
+        ...,
+        description='Video duration (5, 8 seconds, --model=v3.5 only allows 5,8; --quality=1080p does not support 8s)',
+        examples=[5],
+    )
+    model: str = Field(
+        ..., description='Model version (only supports v3.5)', examples=['v3.5']
+    )
+    motion_mode: Optional[str] = Field(
+        'normal',
+        description='Motion mode (normal, fast, --fast only available when duration=5; --quality=1080p does not support fast)',
+        examples=['normal'],
+    )
+    negative_prompt: Optional[str] = Field(
+        None, description='Negative prompt\n', max_length=2048
+    )
+    prompt: str = Field(..., description='Prompt', max_length=2048)
+    quality: str = Field(
+        ...,
+        description='Video quality ("360p"(Turbo model), "540p", "720p", "1080p")',
+        examples=['540p'],
+    )
+    seed: Optional[int] = Field(None, description='Random seed, range: 0 - 2147483647')
+    style: Optional[str] = Field(
+        None,
+        description='Style (effective when model=v3.5, "anime", "3d_animation", "clay", "comic", "cyberpunk") Do not include style parameter unless needed',
+        examples=['anime'],
+    )
+    template_id: Optional[int] = Field(
+        None,
+        description='Template ID (template_id must be activated before use)',
+        examples=[302325299692608],
+    )
+    water_mark: Optional[bool] = Field(
+        False,
+        description='Watermark (true: add watermark, false: no watermark)',
+        examples=[False],
+    )

comfy_api_nodes/apis/__init__.py

@@ -2,7 +2,6 @@
 #   filename:  filtered-openapi.yaml
 #   timestamp: 2025-07-30T08:54:00+00:00
 
-# pylint: disable
 from __future__ import annotations
 
 from datetime import date, datetime
@@ -1321,7 +1320,6 @@ class KlingTextToVideoModelName(str, Enum):
     kling_v1 = 'kling-v1'
     kling_v1_6 = 'kling-v1-6'
     kling_v2_1_master = 'kling-v2-1-master'
-    kling_v2_5_turbo = 'kling-v2-5-turbo'
 
 
 class KlingVideoGenAspectRatio(str, Enum):
@@ -1356,7 +1354,6 @@ class KlingVideoGenModelName(str, Enum):
     kling_v2_master = 'kling-v2-master'
     kling_v2_1 = 'kling-v2-1'
     kling_v2_1_master = 'kling-v2-1-master'
-    kling_v2_5_turbo = 'kling-v2-5-turbo'
 
 
 class KlingVideoResult(BaseModel):

comfy_api_nodes/apis/bfl_api.py

@@ -50,6 +50,44 @@ class BFLFluxFillImageRequest(BaseModel):
     mask: str = Field(None, description='A Base64-encoded string representing the mask of the areas you with to modify.')
 
 
+class BFLFluxCannyImageRequest(BaseModel):
+    prompt: str = Field(..., description='Text prompt for image generation')
+    prompt_upsampling: Optional[bool] = Field(
+        None, description='Whether to perform upsampling on the prompt. If active, automatically modifies the prompt for more creative generation.'
+    )
+    canny_low_threshold: Optional[int] = Field(None, description='Low threshold for Canny edge detection')
+    canny_high_threshold: Optional[int] = Field(None, description='High threshold for Canny edge detection')
+    seed: Optional[int] = Field(None, description='The seed value for reproducibility.')
+    steps: conint(ge=15, le=50) = Field(..., description='Number of steps for the image generation process')
+    guidance: confloat(ge=1, le=100) = Field(..., description='Guidance strength for the image generation process')
+    safety_tolerance: Optional[conint(ge=0, le=6)] = Field(
+        6, description='Tolerance level for input and output moderation. Between 0 and 6, 0 being most strict, 6 being least strict. Defaults to 2.'
+    )
+    output_format: Optional[BFLOutputFormat] = Field(
+        BFLOutputFormat.png, description="Output format for the generated image. Can be 'jpeg' or 'png'.", examples=['png']
+    )
+    control_image: Optional[str] = Field(None, description='Base64 encoded image to use as control input if no preprocessed image is provided')
+    preprocessed_image: Optional[str] = Field(None, description='Optional pre-processed image that will bypass the control preprocessing step')
+
+
+class BFLFluxDepthImageRequest(BaseModel):
+    prompt: str = Field(..., description='Text prompt for image generation')
+    prompt_upsampling: Optional[bool] = Field(
+        None, description='Whether to perform upsampling on the prompt. If active, automatically modifies the prompt for more creative generation.'
+    )
+    seed: Optional[int] = Field(None, description='The seed value for reproducibility.')
+    steps: conint(ge=15, le=50) = Field(..., description='Number of steps for the image generation process')
+    guidance: confloat(ge=1, le=100) = Field(..., description='Guidance strength for the image generation process')
+    safety_tolerance: Optional[conint(ge=0, le=6)] = Field(
+        6, description='Tolerance level for input and output moderation. Between 0 and 6, 0 being most strict, 6 being least strict. Defaults to 2.'
+    )
+    output_format: Optional[BFLOutputFormat] = Field(
+        BFLOutputFormat.png, description="Output format for the generated image. Can be 'jpeg' or 'png'.", examples=['png']
+    )
+    control_image: Optional[str] = Field(None, description='Base64 encoded image to use as control input if no preprocessed image is provided')
+    preprocessed_image: Optional[str] = Field(None, description='Optional pre-processed image that will bypass the control preprocessing step')
+
+
 class BFLFluxProGenerateRequest(BaseModel):
     prompt: str = Field(..., description='The text prompt for image generation.')
     prompt_upsampling: Optional[bool] = Field(
@@ -122,8 +160,15 @@ class BFLStatus(str, Enum):
     error = "Error"
 
 
-class BFLFluxStatusResponse(BaseModel):
+class BFLFluxProStatusResponse(BaseModel):
     id: str = Field(..., description="The unique identifier for the generation task.")
     status: BFLStatus = Field(..., description="The status of the task.")
-    result: Optional[Dict[str, Any]] = Field(None, description="The result of the task (null if not completed).")
-    progress: Optional[float] = Field(None, description="The progress of the task (0.0 to 1.0).", ge=0.0, le=1.0)
+    result: Optional[Dict[str, Any]] = Field(
+        None, description="The result of the task (null if not completed)."
+    )
+    progress: confloat(ge=0.0, le=1.0) = Field(
+        ..., description="The progress of the task (0.0 to 1.0)."
+    )
+    details: Optional[Dict[str, Any]] = Field(
+        None, description="Additional details about the task (null if not available)."
+    )

comfy_api_nodes/apis/client.py

@@ -0,0 +1,957 @@
+"""
+API Client Framework for api.comfy.org.
+
+This module provides a flexible framework for making API requests from ComfyUI nodes.
+It supports both synchronous and asynchronous API operations with proper type validation.
+
+Key Components:
+--------------
+1. ApiClient - Handles HTTP requests with authentication and error handling
+2. ApiEndpoint - Defines a single HTTP endpoint with its request/response models
+3. ApiOperation - Executes a single synchronous API operation
+
+Usage Examples:
+--------------
+
+# Example 1: Synchronous API Operation
+# ------------------------------------
+# For a simple API call that returns the result immediately:
+
+# 1. Create the API client
+api_client = ApiClient(
+    base_url="https://api.example.com",
+    auth_token="your_auth_token_here",
+    comfy_api_key="your_comfy_api_key_here",
+    timeout=30.0,
+    verify_ssl=True
+)
+
+# 2. Define the endpoint
+user_info_endpoint = ApiEndpoint(
+    path="/v1/users/me",
+    method=HttpMethod.GET,
+    request_model=EmptyRequest,  # No request body needed
+    response_model=UserProfile,   # Pydantic model for the response
+    query_params=None
+)
+
+# 3. Create the request object
+request = EmptyRequest()
+
+# 4. Create and execute the operation
+operation = ApiOperation(
+    endpoint=user_info_endpoint,
+    request=request
+)
+user_profile = await operation.execute(client=api_client)  # Returns immediately with the result
+
+
+# Example 2: Asynchronous API Operation with Polling
+# -------------------------------------------------
+# For an API that starts a task and requires polling for completion:
+
+# 1. Define the endpoints (initial request and polling)
+generate_image_endpoint = ApiEndpoint(
+    path="/v1/images/generate",
+    method=HttpMethod.POST,
+    request_model=ImageGenerationRequest,
+    response_model=TaskCreatedResponse,
+    query_params=None
+)
+
+check_task_endpoint = ApiEndpoint(
+    path="/v1/tasks/{task_id}",
+    method=HttpMethod.GET,
+    request_model=EmptyRequest,
+    response_model=ImageGenerationResult,
+    query_params=None
+)
+
+# 2. Create the request object
+request = ImageGenerationRequest(
+    prompt="a beautiful sunset over mountains",
+    width=1024,
+    height=1024,
+    num_images=1
+)
+
+# 3. Create and execute the polling operation
+operation = PollingOperation(
+    initial_endpoint=generate_image_endpoint,
+    initial_request=request,
+    poll_endpoint=check_task_endpoint,
+    task_id_field="task_id",
+    status_field="status",
+    completed_statuses=["completed"],
+    failed_statuses=["failed", "error"]
+)
+
+# This will make the initial request and then poll until completion
+result = await operation.execute(client=api_client)  # Returns the final ImageGenerationResult when done
+"""
+
+from __future__ import annotations
+import aiohttp
+import asyncio
+import logging
+import io
+import socket
+from aiohttp.client_exceptions import ClientError, ClientResponseError
+from typing import Dict, Type, Optional, Any, TypeVar, Generic, Callable, Tuple
+from enum import Enum
+import json
+from urllib.parse import urljoin, urlparse
+from pydantic import BaseModel, Field
+import uuid # For generating unique operation IDs
+
+from server import PromptServer
+from comfy.cli_args import args
+from comfy import utils
+from . import request_logger
+
+T = TypeVar("T", bound=BaseModel)
+R = TypeVar("R", bound=BaseModel)
+P = TypeVar("P", bound=BaseModel)  # For poll response
+
+PROGRESS_BAR_MAX = 100
+
+
+class NetworkError(Exception):
+    """Base exception for network-related errors with diagnostic information."""
+    pass
+
+
+class LocalNetworkError(NetworkError):
+    """Exception raised when local network connectivity issues are detected."""
+    pass
+
+
+class ApiServerError(NetworkError):
+    """Exception raised when the API server is unreachable but internet is working."""
+    pass
+
+
+class EmptyRequest(BaseModel):
+    """Base class for empty request bodies.
+    For GET requests, fields will be sent as query parameters."""
+
+    pass
+
+
+class UploadRequest(BaseModel):
+    file_name: str = Field(..., description="Filename to upload")
+    content_type: Optional[str] = Field(
+        None,
+        description="Mime type of the file. For example: image/png, image/jpeg, video/mp4, etc.",
+    )
+
+
+class UploadResponse(BaseModel):
+    download_url: str = Field(..., description="URL to GET uploaded file")
+    upload_url: str = Field(..., description="URL to PUT file to upload")
+
+
+class HttpMethod(str, Enum):
+    GET = "GET"
+    POST = "POST"
+    PUT = "PUT"
+    DELETE = "DELETE"
+    PATCH = "PATCH"
+
+
+class ApiClient:
+    """
+    Client for making HTTP requests to an API with authentication, error handling, and retry logic.
+    """
+
+    def __init__(
+        self,
+        base_url: str,
+        auth_token: Optional[str] = None,
+        comfy_api_key: Optional[str] = None,
+        timeout: float = 3600.0,
+        verify_ssl: bool = True,
+        max_retries: int = 3,
+        retry_delay: float = 1.0,
+        retry_backoff_factor: float = 2.0,
+        retry_status_codes: Optional[Tuple[int, ...]] = None,
+        session: Optional[aiohttp.ClientSession] = None,
+    ):
+        self.base_url = base_url
+        self.auth_token = auth_token
+        self.comfy_api_key = comfy_api_key
+        self.timeout = timeout
+        self.verify_ssl = verify_ssl
+        self.max_retries = max_retries
+        self.retry_delay = retry_delay
+        self.retry_backoff_factor = retry_backoff_factor
+        # Default retry status codes: 408 (Request Timeout), 429 (Too Many Requests),
+        # 500, 502, 503, 504 (Server Errors)
+        self.retry_status_codes = retry_status_codes or (408, 429, 500, 502, 503, 504)
+        self._session: Optional[aiohttp.ClientSession] = session
+        self._owns_session = session is None  # Track if we have to close it
+
+    @staticmethod
+    def _generate_operation_id(path: str) -> str:
+        """Generates a unique operation ID for logging."""
+        return f"{path.strip('/').replace('/', '_')}_{uuid.uuid4().hex[:8]}"
+
+    @staticmethod
+    def _create_json_payload_args(
+        data: Optional[Dict[str, Any]] = None,
+        headers: Optional[Dict[str, str]] = None,
+    ) -> Dict[str, Any]:
+        return {
+            "json": data,
+            "headers": headers,
+        }
+
+    def _create_form_data_args(
+        self,
+        data: Dict[str, Any] | None,
+        files: Dict[str, Any] | None,
+        headers: Optional[Dict[str, str]] = None,
+        multipart_parser: Callable | None = None,
+    ) -> Dict[str, Any]:
+        if headers and "Content-Type" in headers:
+            del headers["Content-Type"]
+
+        if multipart_parser and data:
+            data = multipart_parser(data)
+
+        form = aiohttp.FormData(default_to_multipart=True)
+        if data:  # regular text fields
+            for k, v in data.items():
+                if v is None:
+                    continue  # aiohttp fails to serialize "None" values
+                # aiohttp expects strings or bytes; convert enums etc.
+                form.add_field(k, str(v) if not isinstance(v, (bytes, bytearray)) else v)
+
+        if files:
+            file_iter = files if isinstance(files, list) else files.items()
+            for field_name, file_obj in file_iter:
+                if file_obj is None:
+                    continue  # aiohttp fails to serialize "None" values
+                # file_obj can be (filename, bytes/io.BytesIO, content_type) tuple
+                if isinstance(file_obj, tuple):
+                    filename, file_value, content_type = self._unpack_tuple(file_obj)
+                else:
+                    file_value = file_obj
+                    filename = getattr(file_obj, "name", field_name)
+                    content_type = "application/octet-stream"
+
+                form.add_field(
+                    name=field_name,
+                    value=file_value,
+                    filename=filename,
+                    content_type=content_type,
+                )
+        return {"data": form, "headers": headers or {}}
+
+    @staticmethod
+    def _create_urlencoded_form_data_args(
+        data: Dict[str, Any],
+        headers: Optional[Dict[str, str]] = None,
+    ) -> Dict[str, Any]:
+        headers = headers or {}
+        headers["Content-Type"] = "application/x-www-form-urlencoded"
+        return {
+            "data": data,
+            "headers": headers,
+        }
+
+    def get_headers(self) -> Dict[str, str]:
+        """Get headers for API requests, including authentication if available"""
+        headers = {"Content-Type": "application/json", "Accept": "application/json"}
+
+        if self.auth_token:
+            headers["Authorization"] = f"Bearer {self.auth_token}"
+        elif self.comfy_api_key:
+            headers["X-API-KEY"] = self.comfy_api_key
+
+        return headers
+
+    async def _check_connectivity(self, target_url: str) -> Dict[str, bool]:
+        """
+        Check connectivity to determine if network issues are local or server-related.
+
+        Args:
+            target_url: URL to check connectivity to
+
+        Returns:
+            Dictionary with connectivity status details
+        """
+        results = {
+            "internet_accessible": False,
+            "api_accessible": False,
+            "is_local_issue": False,
+            "is_api_issue": False,
+        }
+        timeout = aiohttp.ClientTimeout(total=5.0)
+        async with aiohttp.ClientSession(timeout=timeout) as session:
+            try:
+                async with session.get("https://www.google.com", ssl=self.verify_ssl) as resp:
+                    results["internet_accessible"] = resp.status < 500
+            except (ClientError, asyncio.TimeoutError, socket.gaierror):
+                results["is_local_issue"] = True
+                return results  # cannot reach the internet – early exit
+
+            # Now check API health endpoint
+            parsed = urlparse(target_url)
+            health_url = f"{parsed.scheme}://{parsed.netloc}/health"
+            try:
+                async with session.get(health_url, ssl=self.verify_ssl) as resp:
+                    results["api_accessible"] = resp.status < 500
+            except ClientError:
+                pass  # leave as False
+
+        results["is_api_issue"] = results["internet_accessible"] and not results["api_accessible"]
+        return results
+
+    async def request(
+        self,
+        method: str,
+        path: str,
+        params: Optional[Dict[str, Any]] = None,
+        data: Optional[Dict[str, Any]] = None,
+        files: Optional[Dict[str, Any] | list[tuple[str, Any]]] = None,
+        headers: Optional[Dict[str, str]] = None,
+        content_type: str = "application/json",
+        multipart_parser: Callable | None = None,
+        retry_count: int = 0,  # Used internally for tracking retries
+    ) -> Dict[str, Any]:
+        """
+        Make an HTTP request to the API with automatic retries for transient errors.
+
+        Args:
+            method: HTTP method (GET, POST, etc.)
+            path: API endpoint path (will be joined with base_url)
+            params: Query parameters
+            data: body data
+            files: Files to upload
+            headers: Additional headers
+            content_type: Content type of the request. Defaults to application/json.
+            retry_count: Internal parameter for tracking retries, do not set manually
+
+        Returns:
+            Parsed JSON response
+
+        Raises:
+            LocalNetworkError: If local network connectivity issues are detected
+            ApiServerError: If the API server is unreachable but internet is working
+            Exception: For other request failures
+        """
+
+        # Build full URL and merge headers
+        relative_path = path.lstrip("/")
+        url = urljoin(self.base_url, relative_path)
+        self._check_auth(self.auth_token, self.comfy_api_key)
+
+        request_headers = self.get_headers()
+        if headers:
+            request_headers.update(headers)
+        if files:
+            request_headers.pop("Content-Type", None)
+        if params:
+            params = {k: v for k, v in params.items() if v is not None}  # aiohttp fails to serialize None values
+
+        logging.debug(f"[DEBUG] Request Headers: {request_headers}")
+        logging.debug(f"[DEBUG] Files: {files}")
+        logging.debug(f"[DEBUG] Params: {params}")
+        logging.debug(f"[DEBUG] Data: {data}")
+
+        if content_type == "application/x-www-form-urlencoded":
+            payload_args = self._create_urlencoded_form_data_args(data or {}, request_headers)
+        elif content_type == "multipart/form-data":
+            payload_args = self._create_form_data_args(data, files, request_headers, multipart_parser)
+        else:
+            payload_args = self._create_json_payload_args(data, request_headers)
+
+        operation_id = self._generate_operation_id(path)
+        request_logger.log_request_response(
+            operation_id=operation_id,
+            request_method=method,
+            request_url=url,
+            request_headers=request_headers,
+            request_params=params,
+            request_data=data if content_type == "application/json" else "[form-data or other]",
+        )
+
+        session = await self._get_session()
+        try:
+            async with session.request(
+                method,
+                url,
+                params=params,
+                ssl=self.verify_ssl,
+                **payload_args,
+            ) as resp:
+                if resp.status >= 400:
+                    try:
+                        error_data = await resp.json()
+                    except (aiohttp.ContentTypeError, json.JSONDecodeError):
+                        error_data = await resp.text()
+
+                    return await self._handle_http_error(
+                        ClientResponseError(resp.request_info, resp.history, status=resp.status, message=error_data),
+                        operation_id,
+                        method,
+                        url,
+                        params,
+                        data,
+                        files,
+                        headers,
+                        content_type,
+                        multipart_parser,
+                        retry_count=retry_count,
+                        response_content=error_data,
+                    )
+
+                # Success – parse JSON (safely) and log
+                try:
+                    payload = await resp.json()
+                    response_content_to_log = payload
+                except (aiohttp.ContentTypeError, json.JSONDecodeError):
+                    payload = {}
+                    response_content_to_log = await resp.text()
+
+                request_logger.log_request_response(
+                    operation_id=operation_id,
+                    request_method=method,
+                    request_url=url,
+                    response_status_code=resp.status,
+                    response_headers=dict(resp.headers),
+                    response_content=response_content_to_log,
+                )
+                return payload
+
+        except (ClientError, asyncio.TimeoutError, socket.gaierror) as e:
+            # Treat as *connection* problem – optionally retry, else escalate
+            if retry_count < self.max_retries:
+                delay = self.retry_delay * (self.retry_backoff_factor ** retry_count)
+                logging.warning("Connection error. Retrying in %.2fs (%s/%s): %s", delay, retry_count + 1,
+                                self.max_retries, str(e))
+                await asyncio.sleep(delay)
+                return await self.request(
+                    method,
+                    path,
+                    params=params,
+                    data=data,
+                    files=files,
+                    headers=headers,
+                    content_type=content_type,
+                    multipart_parser=multipart_parser,
+                    retry_count=retry_count + 1,
+                )
+            # One final connectivity check for diagnostics
+            connectivity = await self._check_connectivity(self.base_url)
+            if connectivity["is_local_issue"]:
+                raise LocalNetworkError(
+                    "Unable to connect to the API server due to local network issues. "
+                    "Please check your internet connection and try again."
+                ) from e
+            raise ApiServerError(
+                f"The API server at {self.base_url} is currently unreachable. "
+                f"The service may be experiencing issues. Please try again later."
+            ) from e
+
+    @staticmethod
+    def _check_auth(auth_token, comfy_api_key):
+        """Verify that an auth token is present or comfy_api_key is present"""
+        if auth_token is None and comfy_api_key is None:
+            raise Exception("Unauthorized: Please login first to use this node.")
+        return auth_token or comfy_api_key
+
+    @staticmethod
+    async def upload_file(
+        upload_url: str,
+        file: io.BytesIO | str,
+        content_type: str | None = None,
+        max_retries: int = 3,
+        retry_delay: float = 1.0,
+        retry_backoff_factor: float = 2.0,
+    ) -> aiohttp.ClientResponse:
+        """Upload a file to the API with retry logic.
+
+        Args:
+            upload_url: The URL to upload to
+            file: Either a file path string, BytesIO object, or tuple of (file_path, filename)
+            content_type: Optional mime type to set for the upload
+            max_retries: Maximum number of retry attempts
+            retry_delay: Initial delay between retries in seconds
+            retry_backoff_factor: Multiplier for the delay after each retry
+        """
+        headers: Dict[str, str] = {}
+        skip_auto_headers: set[str] = set()
+        if content_type:
+            headers["Content-Type"] = content_type
+        else:
+            # tell aiohttp not to add Content-Type that will break the request signature and result in a 403 status.
+            skip_auto_headers.add("Content-Type")
+
+        # Extract file bytes
+        if isinstance(file, io.BytesIO):
+            file.seek(0)
+            data = file.read()
+        elif isinstance(file, str):
+            with open(file, "rb") as f:
+                data = f.read()
+        else:
+            raise ValueError("File must be BytesIO or str path")
+
+        operation_id = f"upload_{upload_url.split('/')[-1]}_{uuid.uuid4().hex[:8]}"
+        request_logger.log_request_response(
+            operation_id=operation_id,
+            request_method="PUT",
+            request_url=upload_url,
+            request_headers=headers,
+            request_data=f"[File data {len(data)} bytes]",
+        )
+
+        delay = retry_delay
+        for attempt in range(max_retries + 1):
+            try:
+                timeout = aiohttp.ClientTimeout(total=None)  # honour server side timeouts
+                async with aiohttp.ClientSession(timeout=timeout) as session:
+                    async with session.put(
+                        upload_url, data=data, headers=headers, skip_auto_headers=skip_auto_headers,
+                    ) as resp:
+                        resp.raise_for_status()
+                        request_logger.log_request_response(
+                            operation_id=operation_id,
+                            request_method="PUT",
+                            request_url=upload_url,
+                            response_status_code=resp.status,
+                            response_headers=dict(resp.headers),
+                            response_content="File uploaded successfully.",
+                        )
+                        return resp
+            except (ClientError, asyncio.TimeoutError) as e:
+                request_logger.log_request_response(
+                    operation_id=operation_id,
+                    request_method="PUT",
+                    request_url=upload_url,
+                    response_status_code=e.status if hasattr(e, "status") else None,
+                    response_headers=dict(e.headers) if getattr(e, "headers") else None,
+                    response_content=None,
+                    error_message=f"{type(e).__name__}: {str(e)}",
+                )
+                if attempt < max_retries:
+                    logging.warning(
+                        "Upload failed (%s/%s). Retrying in %.2fs. %s", attempt + 1, max_retries, delay, str(e)
+                    )
+                    await asyncio.sleep(delay)
+                    delay *= retry_backoff_factor
+                else:
+                    raise NetworkError(f"Failed to upload file after {max_retries + 1} attempts: {e}") from e
+
+    async def _handle_http_error(
+        self,
+        exc: ClientResponseError,
+        operation_id: str,
+        *req_meta,
+        retry_count: int,
+        response_content: dict | str = "",
+    ) -> Dict[str, Any]:
+        status_code = exc.status
+        if status_code == 401:
+            user_friendly = "Unauthorized: Please login first to use this node."
+        elif status_code == 402:
+            user_friendly = "Payment Required: Please add credits to your account to use this node."
+        elif status_code == 409:
+            user_friendly = "There is a problem with your account. Please contact support@comfy.org."
+        elif status_code == 429:
+            user_friendly = "Rate Limit Exceeded: Please try again later."
+        else:
+            if isinstance(response_content, dict):
+                if "error" in response_content and "message" in response_content["error"]:
+                    user_friendly = f"API Error: {response_content['error']['message']}"
+                    if "type" in response_content["error"]:
+                        user_friendly += f" (Type: {response_content['error']['type']})"
+                else: # Handle cases where error is just a JSON dict with unknown format
+                    user_friendly = f"API Error: {json.dumps(response_content)}"
+            else:
+                if len(response_content) < 200:  # Arbitrary limit for display
+                    user_friendly = f"API Error (raw): {response_content}"
+                else:
+                    user_friendly = f"API Error (raw, status {response_content})"
+
+        request_logger.log_request_response(
+            operation_id=operation_id,
+            request_method=req_meta[0],
+            request_url=req_meta[1],
+            response_status_code=exc.status,
+            response_headers=dict(req_meta[5]) if req_meta[5] else None,
+            response_content=response_content,
+            error_message=f"HTTP Error {exc.status}",
+        )
+
+        logging.debug(f"[DEBUG] API Error: {user_friendly} (Status: {status_code})")
+        if response_content:
+            logging.debug(f"[DEBUG] Response content: {response_content}")
+
+        # Retry if eligible
+        if status_code in self.retry_status_codes and retry_count < self.max_retries:
+            delay = self.retry_delay * (self.retry_backoff_factor ** retry_count)
+            logging.warning(
+                "HTTP error %s. Retrying in %.2fs (%s/%s)",
+                status_code,
+                delay,
+                retry_count + 1,
+                self.max_retries,
+            )
+            await asyncio.sleep(delay)
+            return await self.request(
+                req_meta[0],  # method
+                req_meta[1].replace(self.base_url, ""),  # path
+                params=req_meta[2],
+                data=req_meta[3],
+                files=req_meta[4],
+                headers=req_meta[5],
+                content_type=req_meta[6],
+                multipart_parser=req_meta[7],
+                retry_count=retry_count + 1,
+            )
+
+        raise Exception(user_friendly) from exc
+
+    @staticmethod
+    def _unpack_tuple(t):
+        """Helper to normalise (filename, file, content_type) tuples."""
+        if len(t) == 3:
+            return t
+        elif len(t) == 2:
+            return t[0], t[1], "application/octet-stream"
+        else:
+            raise ValueError("files tuple must be (filename, file[, content_type])")
+
+    async def _get_session(self) -> aiohttp.ClientSession:
+        if self._session is None or self._session.closed:
+            timeout = aiohttp.ClientTimeout(total=self.timeout)
+            self._session = aiohttp.ClientSession(timeout=timeout)
+            self._owns_session = True
+        return self._session
+
+    async def close(self) -> None:
+        if self._owns_session and self._session and not self._session.closed:
+            await self._session.close()
+
+    async def __aenter__(self) -> "ApiClient":
+        """Allow usage as async‑context‑manager – ensures clean teardown"""
+        return self
+
+    async def __aexit__(self, exc_type, exc, tb):
+        await self.close()
+
+
+class ApiEndpoint(Generic[T, R]):
+    """Defines an API endpoint with its request and response types"""
+
+    def __init__(
+        self,
+        path: str,
+        method: HttpMethod,
+        request_model: Type[T],
+        response_model: Type[R],
+        query_params: Optional[Dict[str, Any]] = None,
+    ):
+        """Initialize an API endpoint definition.
+
+        Args:
+            path: The URL path for this endpoint, can include placeholders like {id}
+            method: The HTTP method to use (GET, POST, etc.)
+            request_model: Pydantic model class that defines the structure and validation rules for API requests to this endpoint
+            response_model: Pydantic model class that defines the structure and validation rules for API responses from this endpoint
+            query_params: Optional dictionary of query parameters to include in the request
+        """
+        self.path = path
+        self.method = method
+        self.request_model = request_model
+        self.response_model = response_model
+        self.query_params = query_params or {}
+
+
+class SynchronousOperation(Generic[T, R]):
+    """Represents a single synchronous API operation."""
+
+    def __init__(
+        self,
+        endpoint: ApiEndpoint[T, R],
+        request: T,
+        files: Optional[Dict[str, Any] | list[tuple[str, Any]]] = None,
+        api_base: str | None = None,
+        auth_token: Optional[str] = None,
+        comfy_api_key: Optional[str] = None,
+        auth_kwargs: Optional[Dict[str, str]] = None,
+        timeout: float = 7200.0,
+        verify_ssl: bool = True,
+        content_type: str = "application/json",
+        multipart_parser: Callable | None = None,
+        max_retries: int = 3,
+        retry_delay: float = 1.0,
+        retry_backoff_factor: float = 2.0,
+    ) -> None:
+        self.endpoint = endpoint
+        self.request = request
+        self.files = files
+        self.api_base: str = api_base or args.comfy_api_base
+        self.auth_token = auth_token
+        self.comfy_api_key = comfy_api_key
+        if auth_kwargs is not None:
+            self.auth_token = auth_kwargs.get("auth_token", self.auth_token)
+            self.comfy_api_key = auth_kwargs.get("comfy_api_key", self.comfy_api_key)
+        self.timeout = timeout
+        self.verify_ssl = verify_ssl
+        self.content_type = content_type
+        self.multipart_parser = multipart_parser
+        self.max_retries = max_retries
+        self.retry_delay = retry_delay
+        self.retry_backoff_factor = retry_backoff_factor
+
+    async def execute(self, client: Optional[ApiClient] = None) -> R:
+        owns_client = client is None
+        if owns_client:
+            client = ApiClient(
+                base_url=self.api_base,
+                auth_token=self.auth_token,
+                comfy_api_key=self.comfy_api_key,
+                timeout=self.timeout,
+                verify_ssl=self.verify_ssl,
+                max_retries=self.max_retries,
+                retry_delay=self.retry_delay,
+                retry_backoff_factor=self.retry_backoff_factor,
+            )
+
+        try:
+            request_dict: Optional[Dict[str, Any]]
+            if isinstance(self.request, EmptyRequest):
+                request_dict = None
+            else:
+                request_dict = self.request.model_dump(exclude_none=True)
+                for k, v in list(request_dict.items()):
+                    if isinstance(v, Enum):
+                        request_dict[k] = v.value
+
+            logging.debug(
+                f"[DEBUG] API Request: {self.endpoint.method.value} {self.endpoint.path}"
+            )
+            logging.debug(f"[DEBUG] Request Data: {json.dumps(request_dict, indent=2)}")
+            logging.debug(f"[DEBUG] Query Params: {self.endpoint.query_params}")
+
+            response_json = await client.request(
+                self.endpoint.method.value,
+                self.endpoint.path,
+                params=self.endpoint.query_params,
+                data=request_dict,
+                files=self.files,
+                content_type=self.content_type,
+                multipart_parser=self.multipart_parser,
+            )
+
+            logging.debug("=" * 50)
+            logging.debug("[DEBUG] RESPONSE DETAILS:")
+            logging.debug("[DEBUG] Status Code: 200 (Success)")
+            logging.debug(f"[DEBUG] Response Body: {json.dumps(response_json, indent=2)}")
+            logging.debug("=" * 50)
+
+            parsed_response = self.endpoint.response_model.model_validate(response_json)
+            logging.debug(f"[DEBUG] Parsed Response: {parsed_response}")
+            return parsed_response
+        finally:
+            if owns_client:
+                await client.close()
+
+
+class TaskStatus(str, Enum):
+    """Enum for task status values"""
+
+    COMPLETED = "completed"
+    FAILED = "failed"
+    PENDING = "pending"
+
+
+class PollingOperation(Generic[T, R]):
+    """Represents an asynchronous API operation that requires polling for completion."""
+
+    def __init__(
+        self,
+        poll_endpoint: ApiEndpoint[EmptyRequest, R],
+        completed_statuses: list[str],
+        failed_statuses: list[str],
+        status_extractor: Callable[[R], str],
+        progress_extractor: Callable[[R], float] | None = None,
+        result_url_extractor: Callable[[R], str] | None = None,
+        request: Optional[T] = None,
+        api_base: str | None = None,
+        auth_token: Optional[str] = None,
+        comfy_api_key: Optional[str] = None,
+        auth_kwargs: Optional[Dict[str, str]] = None,
+        poll_interval: float = 5.0,
+        max_poll_attempts: int = 120,  # Default max polling attempts (10 minutes with 5s interval)
+        max_retries: int = 3,  # Max retries per individual API call
+        retry_delay: float = 1.0,
+        retry_backoff_factor: float = 2.0,
+        estimated_duration: Optional[float] = None,
+        node_id: Optional[str] = None,
+    ) -> None:
+        self.poll_endpoint = poll_endpoint
+        self.request = request
+        self.api_base: str = api_base or args.comfy_api_base
+        self.auth_token = auth_token
+        self.comfy_api_key = comfy_api_key
+        if auth_kwargs is not None:
+            self.auth_token = auth_kwargs.get("auth_token", self.auth_token)
+            self.comfy_api_key = auth_kwargs.get("comfy_api_key", self.comfy_api_key)
+        self.poll_interval = poll_interval
+        self.max_poll_attempts = max_poll_attempts
+        self.max_retries = max_retries
+        self.retry_delay = retry_delay
+        self.retry_backoff_factor = retry_backoff_factor
+        self.estimated_duration = estimated_duration
+        self.status_extractor = status_extractor or (lambda x: getattr(x, "status", None))
+        self.progress_extractor = progress_extractor
+        self.result_url_extractor = result_url_extractor
+        self.node_id = node_id
+        self.completed_statuses = completed_statuses
+        self.failed_statuses = failed_statuses
+        self.final_response: Optional[R] = None
+
+    async def execute(self, client: Optional[ApiClient] = None) -> R:
+        owns_client = client is None
+        if owns_client:
+            client = ApiClient(
+                base_url=self.api_base,
+                auth_token=self.auth_token,
+                comfy_api_key=self.comfy_api_key,
+                max_retries=self.max_retries,
+                retry_delay=self.retry_delay,
+                retry_backoff_factor=self.retry_backoff_factor,
+            )
+        try:
+            return await self._poll_until_complete(client)
+        finally:
+            if owns_client:
+                await client.close()
+
+    def _display_text_on_node(self, text: str):
+        if not self.node_id:
+            return
+        PromptServer.instance.send_progress_text(text, self.node_id)
+
+    def _display_time_progress_on_node(self, time_completed: int | float):
+        if not self.node_id:
+            return
+        if self.estimated_duration is not None:
+            remaining = max(0, int(self.estimated_duration) - time_completed)
+            message = f"Task in progress: {time_completed}s (~{remaining}s remaining)"
+        else:
+            message = f"Task in progress: {time_completed}s"
+        self._display_text_on_node(message)
+
+    def _check_task_status(self, response: R) -> TaskStatus:
+        try:
+            status = self.status_extractor(response)
+            if status in self.completed_statuses:
+                return TaskStatus.COMPLETED
+            if status in self.failed_statuses:
+                return TaskStatus.FAILED
+            return TaskStatus.PENDING
+        except Exception as e:
+            logging.error("Error extracting status: %s", e)
+            return TaskStatus.PENDING
+
+    async def _poll_until_complete(self, client: ApiClient) -> R:
+        """Poll until the task is complete"""
+        consecutive_errors = 0
+        max_consecutive_errors = min(5, self.max_retries * 2)  # Limit consecutive errors
+
+        if self.progress_extractor:
+            progress = utils.ProgressBar(PROGRESS_BAR_MAX)
+
+        status = TaskStatus.PENDING
+        for poll_count in range(1, self.max_poll_attempts + 1):
+            try:
+                logging.debug(f"[DEBUG] Polling attempt #{poll_count}")
+
+                request_dict = (
+                    None if self.request is None else self.request.model_dump(exclude_none=True)
+                )
+
+                if poll_count == 1:
+                    logging.debug(
+                        f"[DEBUG] Poll Request: {self.poll_endpoint.method.value} {self.poll_endpoint.path}"
+                    )
+                    logging.debug(
+                        f"[DEBUG] Poll Request Data: {json.dumps(request_dict, indent=2) if request_dict else 'None'}"
+                    )
+
+                # Query task status
+                resp = await client.request(
+                    self.poll_endpoint.method.value,
+                    self.poll_endpoint.path,
+                    params=self.poll_endpoint.query_params,
+                    data=request_dict,
+                )
+                consecutive_errors = 0  # reset on success
+                response_obj: R = self.poll_endpoint.response_model.model_validate(resp)
+
+                # Check if task is complete
+                status = self._check_task_status(response_obj)
+                logging.debug(f"[DEBUG] Task Status: {status}")
+
+                # If progress extractor is provided, extract progress
+                if self.progress_extractor:
+                    new_progress = self.progress_extractor(response_obj)
+                    if new_progress is not None:
+                        progress.update_absolute(new_progress, total=PROGRESS_BAR_MAX)
+
+                if status == TaskStatus.COMPLETED:
+                    message = "Task completed successfully"
+                    if self.result_url_extractor:
+                        result_url = self.result_url_extractor(response_obj)
+                        if result_url:
+                            message = f"Result URL: {result_url}"
+                    logging.debug(f"[DEBUG] {message}")
+                    self._display_text_on_node(message)
+                    self.final_response = response_obj
+                    if self.progress_extractor:
+                        progress.update(100)
+                    return self.final_response
+                if status == TaskStatus.FAILED:
+                    message = f"Task failed: {json.dumps(resp)}"
+                    logging.error(f"[DEBUG] {message}")
+                    raise Exception(message)
+                logging.debug("[DEBUG] Task still pending, continuing to poll...")
+                # Task pending – wait
+                for i in range(int(self.poll_interval)):
+                    self._display_time_progress_on_node((poll_count - 1) * self.poll_interval + i)
+                    await asyncio.sleep(1)
+
+            except (LocalNetworkError, ApiServerError, NetworkError) as e:
+                consecutive_errors += 1
+                if consecutive_errors >= max_consecutive_errors:
+                    raise Exception(
+                        f"Polling aborted after {consecutive_errors} network errors: {str(e)}"
+                    ) from e
+                logging.warning("Network error (%s/%s): %s", consecutive_errors, max_consecutive_errors, str(e))
+                await asyncio.sleep(self.poll_interval)
+            except Exception as e:
+                # For other errors, increment count and potentially abort
+                consecutive_errors += 1
+                if consecutive_errors >= max_consecutive_errors or status == TaskStatus.FAILED:
+                    raise Exception(
+                        f"Polling aborted after {consecutive_errors} consecutive errors: {str(e)}"
+                    ) from e
+
+                logging.error(f"[DEBUG] Polling error: {str(e)}")
+                logging.warning(
+                    f"Error during polling (attempt {poll_count}/{self.max_poll_attempts}): {str(e)}. "
+                    f"Will retry in {self.poll_interval} seconds."
+                )
+                await asyncio.sleep(self.poll_interval)
+
+        # If we've exhausted all polling attempts
+        raise Exception(
+            f"Polling timed out after {self.max_poll_attempts} attempts (" f"{self.max_poll_attempts * self.poll_interval} seconds). "
+            "The operation may still be running on the server but is taking longer than expected."
+        )

comfy_api_nodes/apis/gemini_api.py

@@ -1,230 +1,19 @@
-from datetime import date
-from enum import Enum
-from typing import Any
+from __future__ import annotations
 
-from pydantic import BaseModel, Field
+from typing import List, Optional
 
-
-class GeminiSafetyCategory(str, Enum):
-    HARM_CATEGORY_SEXUALLY_EXPLICIT = "HARM_CATEGORY_SEXUALLY_EXPLICIT"
-    HARM_CATEGORY_HATE_SPEECH = "HARM_CATEGORY_HATE_SPEECH"
-    HARM_CATEGORY_HARASSMENT = "HARM_CATEGORY_HARASSMENT"
-    HARM_CATEGORY_DANGEROUS_CONTENT = "HARM_CATEGORY_DANGEROUS_CONTENT"
-
-
-class GeminiSafetyThreshold(str, Enum):
-    OFF = "OFF"
-    BLOCK_NONE = "BLOCK_NONE"
-    BLOCK_LOW_AND_ABOVE = "BLOCK_LOW_AND_ABOVE"
-    BLOCK_MEDIUM_AND_ABOVE = "BLOCK_MEDIUM_AND_ABOVE"
-    BLOCK_ONLY_HIGH = "BLOCK_ONLY_HIGH"
-
-
-class GeminiSafetySetting(BaseModel):
-    category: GeminiSafetyCategory
-    threshold: GeminiSafetyThreshold
-
-
-class GeminiRole(str, Enum):
-    user = "user"
-    model = "model"
-
-
-class GeminiMimeType(str, Enum):
-    application_pdf = "application/pdf"
-    audio_mpeg = "audio/mpeg"
-    audio_mp3 = "audio/mp3"
-    audio_wav = "audio/wav"
-    image_png = "image/png"
-    image_jpeg = "image/jpeg"
-    image_webp = "image/webp"
-    text_plain = "text/plain"
-    video_mov = "video/mov"
-    video_mpeg = "video/mpeg"
-    video_mp4 = "video/mp4"
-    video_mpg = "video/mpg"
-    video_avi = "video/avi"
-    video_wmv = "video/wmv"
-    video_mpegps = "video/mpegps"
-    video_flv = "video/flv"
-
-
-class GeminiInlineData(BaseModel):
-    data: str | None = Field(
-        None,
-        description="The base64 encoding of the image, PDF, or video to include inline in the prompt. "
-        "When including media inline, you must also specify the media type (mimeType) of the data. Size limit: 20MB",
-    )
-    mimeType: GeminiMimeType | None = Field(None)
-
-
-class GeminiPart(BaseModel):
-    inlineData: GeminiInlineData | None = Field(None)
-    text: str | None = Field(None)
-
-
-class GeminiTextPart(BaseModel):
-    text: str | None = Field(None)
-
-
-class GeminiContent(BaseModel):
-    parts: list[GeminiPart] = Field([])
-    role: GeminiRole = Field(..., examples=["user"])
-
-
-class GeminiSystemInstructionContent(BaseModel):
-    parts: list[GeminiTextPart] = Field(
-        ...,
-        description="A list of ordered parts that make up a single message. "
-        "Different parts may have different IANA MIME types.",
-    )
-    role: GeminiRole = Field(
-        ...,
-        description="The identity of the entity that creates the message. "
-        "The following values are supported: "
-        "user: This indicates that the message is sent by a real person, typically a user-generated message. "
-        "model: This indicates that the message is generated by the model. "
-        "The model value is used to insert messages from model into the conversation during multi-turn conversations. "
-        "For non-multi-turn conversations, this field can be left blank or unset.",
-    )
-
-
-class GeminiFunctionDeclaration(BaseModel):
-    description: str | None = Field(None)
-    name: str = Field(...)
-    parameters: dict[str, Any] = Field(..., description="JSON schema for the function parameters")
-
-
-class GeminiTool(BaseModel):
-    functionDeclarations: list[GeminiFunctionDeclaration] | None = Field(None)
-
-
-class GeminiOffset(BaseModel):
-    nanos: int | None = Field(None, ge=0, le=999999999)
-    seconds: int | None = Field(None, ge=-315576000000, le=315576000000)
-
-
-class GeminiVideoMetadata(BaseModel):
-    endOffset: GeminiOffset | None = Field(None)
-    startOffset: GeminiOffset | None = Field(None)
-
-
-class GeminiGenerationConfig(BaseModel):
-    maxOutputTokens: int | None = Field(None, ge=16, le=8192)
-    seed: int | None = Field(None)
-    stopSequences: list[str] | None = Field(None)
-    temperature: float | None = Field(1, ge=0.0, le=2.0)
-    topK: int | None = Field(40, ge=1)
-    topP: float | None = Field(0.95, ge=0.0, le=1.0)
-
-
-class GeminiImageConfig(BaseModel):
-    aspectRatio: str | None = Field(None)
-    imageSize: str | None = Field(None)
+from comfy_api_nodes.apis import GeminiGenerationConfig, GeminiContent, GeminiSafetySetting, GeminiSystemInstructionContent, GeminiTool, GeminiVideoMetadata
+from pydantic import BaseModel
 
 
 class GeminiImageGenerationConfig(GeminiGenerationConfig):
-    responseModalities: list[str] | None = Field(None)
-    imageConfig: GeminiImageConfig | None = Field(None)
+    responseModalities: Optional[List[str]] = None
 
 
 class GeminiImageGenerateContentRequest(BaseModel):
-    contents: list[GeminiContent] = Field(...)
-    generationConfig: GeminiImageGenerationConfig | None = Field(None)
-    safetySettings: list[GeminiSafetySetting] | None = Field(None)
-    systemInstruction: GeminiSystemInstructionContent | None = Field(None)
-    tools: list[GeminiTool] | None = Field(None)
-    videoMetadata: GeminiVideoMetadata | None = Field(None)
-
-
-class GeminiGenerateContentRequest(BaseModel):
-    contents: list[GeminiContent] = Field(...)
-    generationConfig: GeminiGenerationConfig | None = Field(None)
-    safetySettings: list[GeminiSafetySetting] | None = Field(None)
-    systemInstruction: GeminiSystemInstructionContent | None = Field(None)
-    tools: list[GeminiTool] | None = Field(None)
-    videoMetadata: GeminiVideoMetadata | None = Field(None)
-
-
-class Modality(str, Enum):
-    MODALITY_UNSPECIFIED = "MODALITY_UNSPECIFIED"
-    TEXT = "TEXT"
-    IMAGE = "IMAGE"
-    VIDEO = "VIDEO"
-    AUDIO = "AUDIO"
-    DOCUMENT = "DOCUMENT"
-
-
-class ModalityTokenCount(BaseModel):
-    modality: Modality | None = None
-    tokenCount: int | None = Field(None, description="Number of tokens for the given modality.")
-
-
-class Probability(str, Enum):
-    NEGLIGIBLE = "NEGLIGIBLE"
-    LOW = "LOW"
-    MEDIUM = "MEDIUM"
-    HIGH = "HIGH"
-    UNKNOWN = "UNKNOWN"
-
-
-class GeminiSafetyRating(BaseModel):
-    category: GeminiSafetyCategory | None = None
-    probability: Probability | None = Field(
-        None,
-        description="The probability that the content violates the specified safety category",
-    )
-
-
-class GeminiCitation(BaseModel):
-    authors: list[str] | None = None
-    endIndex: int | None = None
-    license: str | None = None
-    publicationDate: date | None = None
-    startIndex: int | None = None
-    title: str | None = None
-    uri: str | None = None
-
-
-class GeminiCitationMetadata(BaseModel):
-    citations: list[GeminiCitation] | None = None
-
-
-class GeminiCandidate(BaseModel):
-    citationMetadata: GeminiCitationMetadata | None = None
-    content: GeminiContent | None = None
-    finishReason: str | None = None
-    safetyRatings: list[GeminiSafetyRating] | None = None
-
-
-class GeminiPromptFeedback(BaseModel):
-    blockReason: str | None = None
-    blockReasonMessage: str | None = None
-    safetyRatings: list[GeminiSafetyRating] | None = None
-
-
-class GeminiUsageMetadata(BaseModel):
-    cachedContentTokenCount: int | None = Field(
-        None,
-        description="Output only. Number of tokens in the cached part in the input (the cached content).",
-    )
-    candidatesTokenCount: int | None = Field(None, description="Number of tokens in the response(s).")
-    candidatesTokensDetails: list[ModalityTokenCount] | None = Field(
-        None, description="Breakdown of candidate tokens by modality."
-    )
-    promptTokenCount: int | None = Field(
-        None,
-        description="Number of tokens in the request. When cachedContent is set, this is still the total effective prompt size meaning this includes the number of tokens in the cached content.",
-    )
-    promptTokensDetails: list[ModalityTokenCount] | None = Field(
-        None, description="Breakdown of prompt tokens by modality."
-    )
-    thoughtsTokenCount: int | None = Field(None, description="Number of tokens present in thoughts output.")
-    toolUsePromptTokenCount: int | None = Field(None, description="Number of tokens present in tool-use prompt(s).")
-
-
-class GeminiGenerateContentResponse(BaseModel):
-    candidates: list[GeminiCandidate] | None = Field(None)
-    promptFeedback: GeminiPromptFeedback | None = Field(None)
-    usageMetadata: GeminiUsageMetadata | None = Field(None)
-    modelVersion: str | None = Field(None)
+    contents: List[GeminiContent]
+    generationConfig: Optional[GeminiImageGenerationConfig] = None
+    safetySettings: Optional[List[GeminiSafetySetting]] = None
+    systemInstruction: Optional[GeminiSystemInstructionContent] = None
+    tools: Optional[List[GeminiTool]] = None
+    videoMetadata: Optional[GeminiVideoMetadata] = None

comfy_api_nodes/apis/minimax_api.py

@@ -1,120 +0,0 @@
-from enum import Enum
-from typing import Optional
-
-from pydantic import BaseModel, Field
-
-
-class MinimaxBaseResponse(BaseModel):
-    status_code: int = Field(
-        ...,
-        description='Status code. 0 indicates success, other values indicate errors.',
-    )
-    status_msg: str = Field(
-        ..., description='Specific error details or success message.'
-    )
-
-
-class File(BaseModel):
-    bytes: Optional[int] = Field(None, description='File size in bytes')
-    created_at: Optional[int] = Field(
-        None, description='Unix timestamp when the file was created, in seconds'
-    )
-    download_url: Optional[str] = Field(
-        None, description='The URL to download the video'
-    )
-    backup_download_url: Optional[str] = Field(
-        None, description='The backup URL to download the video'
-    )
-
-    file_id: Optional[int] = Field(None, description='Unique identifier for the file')
-    filename: Optional[str] = Field(None, description='The name of the file')
-    purpose: Optional[str] = Field(None, description='The purpose of using the file')
-
-
-class MinimaxFileRetrieveResponse(BaseModel):
-    base_resp: MinimaxBaseResponse
-    file: File
-
-
-class MiniMaxModel(str, Enum):
-    T2V_01_Director = 'T2V-01-Director'
-    I2V_01_Director = 'I2V-01-Director'
-    S2V_01 = 'S2V-01'
-    I2V_01 = 'I2V-01'
-    I2V_01_live = 'I2V-01-live'
-    T2V_01 = 'T2V-01'
-    Hailuo_02 = 'MiniMax-Hailuo-02'
-
-
-class Status6(str, Enum):
-    Queueing = 'Queueing'
-    Preparing = 'Preparing'
-    Processing = 'Processing'
-    Success = 'Success'
-    Fail = 'Fail'
-
-
-class MinimaxTaskResultResponse(BaseModel):
-    base_resp: MinimaxBaseResponse
-    file_id: Optional[str] = Field(
-        None,
-        description='After the task status changes to Success, this field returns the file ID corresponding to the generated video.',
-    )
-    status: Status6 = Field(
-        ...,
-        description="Task status: 'Queueing' (in queue), 'Preparing' (task is preparing), 'Processing' (generating), 'Success' (task completed successfully), or 'Fail' (task failed).",
-    )
-    task_id: str = Field(..., description='The task ID being queried.')
-
-
-class SubjectReferenceItem(BaseModel):
-    image: Optional[str] = Field(
-        None, description='URL or base64 encoding of the subject reference image.'
-    )
-    mask: Optional[str] = Field(
-        None,
-        description='URL or base64 encoding of the mask for the subject reference image.',
-    )
-
-
-class MinimaxVideoGenerationRequest(BaseModel):
-    callback_url: Optional[str] = Field(
-        None,
-        description='Optional. URL to receive real-time status updates about the video generation task.',
-    )
-    first_frame_image: Optional[str] = Field(
-        None,
-        description='URL or base64 encoding of the first frame image. Required when model is I2V-01, I2V-01-Director, or I2V-01-live.',
-    )
-    model: MiniMaxModel = Field(
-        ...,
-        description='Required. ID of model. Options: T2V-01-Director, I2V-01-Director, S2V-01, I2V-01, I2V-01-live, T2V-01',
-    )
-    prompt: Optional[str] = Field(
-        None,
-        description='Description of the video. Should be less than 2000 characters. Supports camera movement instructions in [brackets].',
-        max_length=2000,
-    )
-    prompt_optimizer: Optional[bool] = Field(
-        True,
-        description='If true (default), the model will automatically optimize the prompt. Set to false for more precise control.',
-    )
-    subject_reference: Optional[list[SubjectReferenceItem]] = Field(
-        None,
-        description='Only available when model is S2V-01. The model will generate a video based on the subject uploaded through this parameter.',
-    )
-    duration: Optional[int] = Field(
-        None,
-        description="The length of the output video in seconds."
-    )
-    resolution: Optional[str] = Field(
-        None,
-        description="The dimensions of the video display. 1080p corresponds to 1920 x 1080 pixels, 768p corresponds to 1366 x 768 pixels."
-    )
-
-
-class MinimaxVideoGenerationResponse(BaseModel):
-    base_resp: MinimaxBaseResponse
-    task_id: str = Field(
-        ..., description='The task ID for the asynchronous video generation task.'
-    )

comfy_api_nodes/apis/pika_api.py

@@ -1,100 +0,0 @@
-from typing import Optional
-from enum import Enum
-from pydantic import BaseModel, Field
-
-
-class Pikaffect(str, Enum):
-    Cake_ify = "Cake-ify"
-    Crumble = "Crumble"
-    Crush = "Crush"
-    Decapitate = "Decapitate"
-    Deflate = "Deflate"
-    Dissolve = "Dissolve"
-    Explode = "Explode"
-    Eye_pop = "Eye-pop"
-    Inflate = "Inflate"
-    Levitate = "Levitate"
-    Melt = "Melt"
-    Peel = "Peel"
-    Poke = "Poke"
-    Squish = "Squish"
-    Ta_da = "Ta-da"
-    Tear = "Tear"
-
-
-class PikaBodyGenerate22C2vGenerate22PikascenesPost(BaseModel):
-    aspectRatio: Optional[float] = Field(None, description='Aspect ratio (width / height)')
-    duration: Optional[int] = Field(5)
-    ingredientsMode: str = Field(...)
-    negativePrompt: Optional[str] = Field(None)
-    promptText: Optional[str] = Field(None)
-    resolution: Optional[str] = Field('1080p')
-    seed: Optional[int] = Field(None)
-
-
-class PikaGenerateResponse(BaseModel):
-    video_id: str = Field(...)
-
-
-class PikaBodyGenerate22I2vGenerate22I2vPost(BaseModel):
-    duration: Optional[int] = 5
-    negativePrompt: Optional[str] = Field(None)
-    promptText: Optional[str] = Field(None)
-    resolution: Optional[str] = '1080p'
-    seed: Optional[int] = Field(None)
-
-
-class PikaBodyGenerate22KeyframeGenerate22PikaframesPost(BaseModel):
-    duration: Optional[int] = Field(None, ge=5, le=10)
-    negativePrompt: Optional[str] = Field(None)
-    promptText: str = Field(...)
-    resolution: Optional[str] = '1080p'
-    seed: Optional[int] = Field(None)
-
-
-class PikaBodyGenerate22T2vGenerate22T2vPost(BaseModel):
-    aspectRatio: Optional[float] = Field(
-        1.7777777777777777,
-        description='Aspect ratio (width / height)',
-        ge=0.4,
-        le=2.5,
-    )
-    duration: Optional[int] = 5
-    negativePrompt: Optional[str] = Field(None)
-    promptText: str = Field(...)
-    resolution: Optional[str] = '1080p'
-    seed: Optional[int] = Field(None)
-
-
-class PikaBodyGeneratePikadditionsGeneratePikadditionsPost(BaseModel):
-    negativePrompt: Optional[str] = Field(None)
-    promptText: Optional[str] = Field(None)
-    seed: Optional[int] = Field(None)
-
-
-class PikaBodyGeneratePikaffectsGeneratePikaffectsPost(BaseModel):
-    negativePrompt: Optional[str] = Field(None)
-    pikaffect: Optional[str] = None
-    promptText: Optional[str] = Field(None)
-    seed: Optional[int] = Field(None)
-
-
-class PikaBodyGeneratePikaswapsGeneratePikaswapsPost(BaseModel):
-    negativePrompt: Optional[str] = Field(None)
-    promptText: Optional[str] = Field(None)
-    seed: Optional[int] = Field(None)
-    modifyRegionRoi: Optional[str] = Field(None)
-
-
-class PikaStatusEnum(str, Enum):
-    queued = "queued"
-    started = "started"
-    finished = "finished"
-    failed = "failed"
-
-
-class PikaVideoResponse(BaseModel):
-    id: str = Field(...)
-    progress: Optional[int] = Field(None)
-    status: PikaStatusEnum
-    url: Optional[str] = Field(None)

comfy_api_nodes/apis/request_logger.py

@@ -1,102 +1,65 @@
 from __future__ import annotations
 
+import os
 import datetime
-import hashlib
 import json
 import logging
-import os
-import re
-from typing import Any
-
 import folder_paths
 
 # Get the logger instance
 logger = logging.getLogger(__name__)
 
-
 def get_log_directory():
-    """Ensures the API log directory exists within ComfyUI's temp directory and returns its path."""
+    """
+    Ensures the API log directory exists within ComfyUI's temp directory
+    and returns its path.
+    """
     base_temp_dir = folder_paths.get_temp_directory()
     log_dir = os.path.join(base_temp_dir, "api_logs")
     try:
         os.makedirs(log_dir, exist_ok=True)
     except Exception as e:
-        logger.error("Error creating API log directory %s: %s", log_dir, str(e))
+        logger.error(f"Error creating API log directory {log_dir}: {e}")
         # Fallback to base temp directory if sub-directory creation fails
         return base_temp_dir
     return log_dir
 
-
-def _sanitize_filename_component(name: str) -> str:
-    if not name:
-        return "log"
-    sanitized = re.sub(r"[^A-Za-z0-9._-]+", "_", name)  # Replace disallowed characters with underscore
-    sanitized = sanitized.strip(" ._")  # Windows: trailing dots or spaces are not allowed
-    if not sanitized:
-        sanitized = "log"
-    return sanitized
-
-
-def _short_hash(*parts: str, length: int = 10) -> str:
-    return hashlib.sha1(("|".join(parts)).encode("utf-8")).hexdigest()[:length]
-
-
-def _build_log_filepath(log_dir: str, operation_id: str, request_url: str) -> str:
-    """Build log filepath. We keep it well under common path length limits aiming for <= 240 characters total."""
-    timestamp = datetime.datetime.now().strftime("%Y%m%d_%H%M%S_%f")
-    slug = _sanitize_filename_component(operation_id)  # Best-effort human-readable slug from operation_id
-    h = _short_hash(operation_id or "", request_url or "")  # Short hash ties log to the full operation and URL
-
-    # Compute how much room we have for the slug given the directory length
-    # Keep total path length reasonably below ~260 on Windows.
-    max_total_path = 240
-    prefix = f"{timestamp}_"
-    suffix = f"_{h}.log"
-    if not slug:
-        slug = "op"
-    max_filename_len = max(60, max_total_path - len(log_dir) - 1)
-    max_slug_len = max(8, max_filename_len - len(prefix) - len(suffix))
-    if len(slug) > max_slug_len:
-        slug = slug[:max_slug_len].rstrip(" ._-")
-    return os.path.join(log_dir, f"{prefix}{slug}{suffix}")
-
-
-def _format_data_for_logging(data: Any) -> str:
+def _format_data_for_logging(data):
     """Helper to format data (dict, str, bytes) for logging."""
     if isinstance(data, bytes):
         try:
-            return data.decode("utf-8")  # Try to decode as text
+            return data.decode('utf-8')  # Try to decode as text
         except UnicodeDecodeError:
             return f"[Binary data of length {len(data)} bytes]"
     elif isinstance(data, (dict, list)):
         try:
             return json.dumps(data, indent=2, ensure_ascii=False)
         except TypeError:
-            return str(data)  # Fallback for non-serializable objects
+            return str(data) # Fallback for non-serializable objects
     return str(data)
 
-
 def log_request_response(
     operation_id: str,
     request_method: str,
     request_url: str,
     request_headers: dict | None = None,
     request_params: dict | None = None,
-    request_data: Any = None,
+    request_data: any = None,
     response_status_code: int | None = None,
     response_headers: dict | None = None,
-    response_content: Any = None,
-    error_message: str | None = None,
+    response_content: any = None,
+    error_message: str | None = None
 ):
     """
     Logs API request and response details to a file in the temp/api_logs directory.
-    Filenames are sanitized and length-limited for cross-platform safety.
-    If we still fail to write, we fall back to appending into api.log.
     """
     log_dir = get_log_directory()
-    filepath = _build_log_filepath(log_dir, operation_id, request_url)
+    timestamp = datetime.datetime.now().strftime("%Y%m%d_%H%M%S_%f")
+    filename = f"{timestamp}_{operation_id.replace('/', '_').replace(':', '_')}.log"
+    filepath = os.path.join(log_dir, filename)
+
+    log_content = []
 
-    log_content: list[str] = []
     log_content.append(f"Timestamp: {datetime.datetime.now().isoformat()}")
     log_content.append(f"Operation ID: {operation_id}")
     log_content.append("-" * 30 + " REQUEST " + "-" * 30)
@@ -106,7 +69,7 @@ def log_request_response(
         log_content.append(f"Headers:\n{_format_data_for_logging(request_headers)}")
     if request_params:
         log_content.append(f"Params:\n{_format_data_for_logging(request_params)}")
-    if request_data is not None:
+    if request_data:
         log_content.append(f"Data/Body:\n{_format_data_for_logging(request_data)}")
 
     log_content.append("\n" + "-" * 30 + " RESPONSE " + "-" * 30)
@@ -114,7 +77,7 @@ def log_request_response(
         log_content.append(f"Status Code: {response_status_code}")
     if response_headers:
         log_content.append(f"Headers:\n{_format_data_for_logging(response_headers)}")
-    if response_content is not None:
+    if response_content:
         log_content.append(f"Content:\n{_format_data_for_logging(response_content)}")
     if error_message:
         log_content.append(f"Error:\n{error_message}")
@@ -122,10 +85,9 @@ def log_request_response(
     try:
         with open(filepath, "w", encoding="utf-8") as f:
             f.write("\n".join(log_content))
-        logger.debug("API log saved to: %s", filepath)
+        logger.debug(f"API log saved to: {filepath}")
     except Exception as e:
-        logger.error("Error writing API log to %s: %s", filepath, str(e))
-
+        logger.error(f"Error writing API log to {filepath}: {e}")
 
 if __name__ == '__main__':
     # Example usage (for testing the logger directly)

comfy_api_nodes/apis/rodin_api.py

@@ -9,9 +9,8 @@ class Rodin3DGenerateRequest(BaseModel):
     seed: int = Field(..., description="seed_")
     tier: str = Field(..., description="Tier of generation.")
     material: str = Field(..., description="The material type.")
-    quality_override: int = Field(..., description="The poly count of the mesh.")
+    quality: str = Field(..., description="The generation quality of the mesh.")
     mesh_mode: str = Field(..., description="It controls the type of faces of generated models.")
-    TAPose: Optional[bool] = Field(None, description="")
 
 class GenerateJobsData(BaseModel):
     uuids: List[str] = Field(..., description="str LIST")
@@ -52,3 +51,7 @@ class RodinResourceItem(BaseModel):
 
 class Rodin3DDownloadResponse(BaseModel):
     list: List[RodinResourceItem] = Field(..., description="Source List")
+
+
+
+

comfy_api_nodes/apis/topaz_api.py

@@ -1,133 +0,0 @@
-from typing import Optional, Union
-
-from pydantic import BaseModel, Field
-
-
-class ImageEnhanceRequest(BaseModel):
-    model: str = Field("Reimagine")
-    output_format: str = Field("jpeg")
-    subject_detection: str = Field("All")
-    face_enhancement: bool = Field(True)
-    face_enhancement_creativity: float = Field(0, description="Is ignored if face_enhancement is false")
-    face_enhancement_strength: float = Field(0.8, description="Is ignored if face_enhancement is false")
-    source_url: str = Field(...)
-    output_width: Optional[int] = Field(None)
-    output_height: Optional[int] = Field(None)
-    crop_to_fill: bool = Field(False)
-    prompt: Optional[str] = Field(None, description="Text prompt for creative upscaling guidance")
-    creativity: int = Field(3, description="Creativity settings range from 1 to 9")
-    face_preservation: str = Field("true", description="To preserve the identity of characters")
-    color_preservation: str = Field("true", description="To preserve the original color")
-
-
-class ImageAsyncTaskResponse(BaseModel):
-    process_id: str = Field(...)
-
-
-class ImageStatusResponse(BaseModel):
-    process_id: str = Field(...)
-    status: str = Field(...)
-    progress: Optional[int] = Field(None)
-    credits: int = Field(...)
-
-
-class ImageDownloadResponse(BaseModel):
-    download_url: str = Field(...)
-    expiry: int = Field(...)
-
-
-class Resolution(BaseModel):
-    width: int = Field(...)
-    height: int = Field(...)
-
-
-class CreateCreateVideoRequestSource(BaseModel):
-    container: str = Field(...)
-    size: int = Field(..., description="Size of the video file in bytes")
-    duration: int = Field(..., description="Duration of the video file in seconds")
-    frameCount: int = Field(..., description="Total number of frames in the video")
-    frameRate: int = Field(...)
-    resolution: Resolution = Field(...)
-
-
-class VideoFrameInterpolationFilter(BaseModel):
-    model: str = Field(...)
-    slowmo: Optional[int] = Field(None)
-    fps: int = Field(...)
-    duplicate: bool = Field(...)
-    duplicate_threshold: float = Field(...)
-
-
-class VideoEnhancementFilter(BaseModel):
-    model: str = Field(...)
-    auto: Optional[str] = Field(None, description="Auto, Manual, Relative")
-    focusFixLevel: Optional[str] = Field(None, description="Downscales video input for correction of blurred subjects")
-    compression: Optional[float] = Field(None, description="Strength of compression recovery")
-    details: Optional[float] = Field(None, description="Amount of detail reconstruction")
-    prenoise: Optional[float] = Field(None, description="Amount of noise to add to input to reduce over-smoothing")
-    noise: Optional[float] = Field(None, description="Amount of noise reduction")
-    halo: Optional[float] = Field(None, description="Amount of halo reduction")
-    preblur: Optional[float] = Field(None, description="Anti-aliasing and deblurring strength")
-    blur: Optional[float] = Field(None, description="Amount of sharpness applied")
-    grain: Optional[float] = Field(None, description="Grain after AI model processing")
-    grainSize: Optional[float] = Field(None, description="Size of generated grain")
-    recoverOriginalDetailValue: Optional[float] = Field(None, description="Source details into the output video")
-    creativity: Optional[str] = Field(None, description="Creativity level(high, low) for slc-1 only")
-    isOptimizedMode: Optional[bool] = Field(None, description="Set to true for Starlight Creative (slc-1) only")
-
-
-class OutputInformationVideo(BaseModel):
-    resolution: Resolution = Field(...)
-    frameRate: int = Field(...)
-    audioCodec: Optional[str] = Field(..., description="Required if audioTransfer is Copy or Convert")
-    audioTransfer: str = Field(..., description="Copy, Convert, None")
-    dynamicCompressionLevel: str = Field(..., description="Low, Mid, High")
-
-
-class Overrides(BaseModel):
-    isPaidDiffusion: bool = Field(True)
-
-
-class CreateVideoRequest(BaseModel):
-    source: CreateCreateVideoRequestSource = Field(...)
-    filters: list[Union[VideoFrameInterpolationFilter, VideoEnhancementFilter]] = Field(...)
-    output: OutputInformationVideo = Field(...)
-    overrides: Overrides = Field(Overrides(isPaidDiffusion=True))
-
-
-class CreateVideoResponse(BaseModel):
-    requestId: str = Field(...)
-
-
-class VideoAcceptResponse(BaseModel):
-    uploadId: str = Field(...)
-    urls: list[str] = Field(...)
-
-
-class VideoCompleteUploadRequestPart(BaseModel):
-    partNum: int = Field(...)
-    eTag: str = Field(...)
-
-
-class VideoCompleteUploadRequest(BaseModel):
-    uploadResults: list[VideoCompleteUploadRequestPart] = Field(...)
-
-
-class VideoCompleteUploadResponse(BaseModel):
-    message: str = Field(..., description="Confirmation message")
-
-
-class VideoStatusResponseEstimates(BaseModel):
-    cost: list[int] = Field(...)
-
-
-class VideoStatusResponseDownloadUrl(BaseModel):
-    url: str = Field(...)
-
-
-class VideoStatusResponse(BaseModel):
-    status: str = Field(...)
-    estimates: Optional[VideoStatusResponseEstimates] = Field(None)
-    progress: Optional[float] = Field(None)
-    message: Optional[str] = Field("")
-    download: Optional[VideoStatusResponseDownloadUrl] = Field(None)

comfy_api_nodes/apis/tripo_api.py

@@ -1,20 +1,13 @@
 from __future__ import annotations
+from comfy_api_nodes.apis import (
+    TripoModelVersion,
+    TripoTextureQuality,
+)
 from enum import Enum
 from typing import Optional, List, Dict, Any, Union
 
 from pydantic import BaseModel, Field, RootModel
 
-class TripoModelVersion(str, Enum):
-    v2_5_20250123 = 'v2.5-20250123'
-    v2_0_20240919 = 'v2.0-20240919'
-    v1_4_20240625 = 'v1.4-20240625'
-
-
-class TripoTextureQuality(str, Enum):
-    standard = 'standard'
-    detailed = 'detailed'
-
-
 class TripoStyle(str, Enum):
     PERSON_TO_CARTOON = "person:person2cartoon"
     ANIMAL_VENOM = "animal:venom"

comfy_api_nodes/apis/veo_api.py

@@ -1,111 +0,0 @@
-from typing import Optional, Union
-from enum import Enum
-
-from pydantic import BaseModel, Field
-
-
-class Image2(BaseModel):
-    bytesBase64Encoded: str
-    gcsUri: Optional[str] = None
-    mimeType: Optional[str] = None
-
-
-class Image3(BaseModel):
-    bytesBase64Encoded: Optional[str] = None
-    gcsUri: str
-    mimeType: Optional[str] = None
-
-
-class Instance1(BaseModel):
-    image: Optional[Union[Image2, Image3]] = Field(
-        None, description='Optional image to guide video generation'
-    )
-    prompt: str = Field(..., description='Text description of the video')
-
-
-class PersonGeneration1(str, Enum):
-    ALLOW = 'ALLOW'
-    BLOCK = 'BLOCK'
-
-
-class Parameters1(BaseModel):
-    aspectRatio: Optional[str] = Field(None, examples=['16:9'])
-    durationSeconds: Optional[int] = None
-    enhancePrompt: Optional[bool] = None
-    generateAudio: Optional[bool] = Field(
-        None,
-        description='Generate audio for the video. Only supported by veo 3 models.',
-    )
-    negativePrompt: Optional[str] = None
-    personGeneration: Optional[PersonGeneration1] = None
-    sampleCount: Optional[int] = None
-    seed: Optional[int] = None
-    storageUri: Optional[str] = Field(
-        None, description='Optional Cloud Storage URI to upload the video'
-    )
-
-
-class VeoGenVidRequest(BaseModel):
-    instances: Optional[list[Instance1]] = None
-    parameters: Optional[Parameters1] = None
-
-
-class VeoGenVidResponse(BaseModel):
-    name: str = Field(
-        ...,
-        description='Operation resource name',
-        examples=[
-            'projects/PROJECT_ID/locations/us-central1/publishers/google/models/MODEL_ID/operations/a1b07c8e-7b5a-4aba-bb34-3e1ccb8afcc8'
-        ],
-    )
-
-
-class VeoGenVidPollRequest(BaseModel):
-    operationName: str = Field(
-        ...,
-        description='Full operation name (from predict response)',
-        examples=[
-            'projects/PROJECT_ID/locations/us-central1/publishers/google/models/MODEL_ID/operations/OPERATION_ID'
-        ],
-    )
-
-
-class Video(BaseModel):
-    bytesBase64Encoded: Optional[str] = Field(
-        None, description='Base64-encoded video content'
-    )
-    gcsUri: Optional[str] = Field(None, description='Cloud Storage URI of the video')
-    mimeType: Optional[str] = Field(None, description='Video MIME type')
-
-
-class Error1(BaseModel):
-    code: Optional[int] = Field(None, description='Error code')
-    message: Optional[str] = Field(None, description='Error message')
-
-
-class Response1(BaseModel):
-    field_type: Optional[str] = Field(
-        None,
-        alias='@type',
-        examples=[
-            'type.googleapis.com/cloud.ai.large_models.vision.GenerateVideoResponse'
-        ],
-    )
-    raiMediaFilteredCount: Optional[int] = Field(
-        None, description='Count of media filtered by responsible AI policies'
-    )
-    raiMediaFilteredReasons: Optional[list[str]] = Field(
-        None, description='Reasons why media was filtered by responsible AI policies'
-    )
-    videos: Optional[list[Video]] = None
-
-
-class VeoGenVidPollResponse(BaseModel):
-    done: Optional[bool] = None
-    error: Optional[Error1] = Field(
-        None, description='Error details if operation failed'
-    )
-    name: Optional[str] = None
-    response: Optional[Response1] = Field(
-        None, description='The actual prediction response if done is true'
-    )

comfy_api_nodes/nodes_ideogram.py

@@ -1,6 +1,6 @@
 from io import BytesIO
 from typing_extensions import override
-from comfy_api.latest import IO, ComfyExtension
+from comfy_api.latest import ComfyExtension, io as comfy_io
 from PIL import Image
 import numpy as np
 import torch
@@ -11,14 +11,20 @@ from comfy_api_nodes.apis import (
     IdeogramV3Request,
     IdeogramV3EditRequest,
 )
-from comfy_api_nodes.util import (
+
+from comfy_api_nodes.apis.client import (
     ApiEndpoint,
-    bytesio_to_image_tensor,
-    download_url_as_bytesio,
-    resize_mask_to_image,
-    sync_op,
+    HttpMethod,
+    SynchronousOperation,
 )
 
+from comfy_api_nodes.apinode_utils import (
+    download_url_to_bytesio,
+    bytesio_to_image_tensor,
+    resize_mask_to_image,
+)
+from server import PromptServer
+
 V1_V1_RES_MAP = {
   "Auto":"AUTO",
   "512 x 1536":"RESOLUTION_512_1536",
@@ -214,7 +220,7 @@ async def download_and_process_images(image_urls):
 
     for image_url in image_urls:
         # Using functions from apinode_utils.py to handle downloading and processing
-        image_bytesio = await download_url_as_bytesio(image_url)  # Download image content to BytesIO
+        image_bytesio = await download_url_to_bytesio(image_url)  # Download image content to BytesIO
         img_tensor = bytesio_to_image_tensor(image_bytesio, mode="RGB")  # Convert to torch.Tensor with RGB mode
         image_tensors.append(img_tensor)
 
@@ -227,76 +233,89 @@ async def download_and_process_images(image_urls):
     return stacked_tensors
 
 
-class IdeogramV1(IO.ComfyNode):
+def display_image_urls_on_node(image_urls, node_id):
+    if node_id and image_urls:
+        if len(image_urls) == 1:
+            PromptServer.instance.send_progress_text(
+                f"Generated Image URL:\n{image_urls[0]}", node_id
+            )
+        else:
+            urls_text = "Generated Image URLs:\n" + "\n".join(
+                f"{i+1}. {url}" for i, url in enumerate(image_urls)
+            )
+            PromptServer.instance.send_progress_text(urls_text, node_id)
+
+
+class IdeogramV1(comfy_io.ComfyNode):
 
     @classmethod
     def define_schema(cls):
-        return IO.Schema(
+        return comfy_io.Schema(
             node_id="IdeogramV1",
             display_name="Ideogram V1",
             category="api node/image/Ideogram",
             description="Generates images using the Ideogram V1 model.",
             is_api_node=True,
             inputs=[
-                IO.String.Input(
+                comfy_io.String.Input(
                     "prompt",
                     multiline=True,
                     default="",
                     tooltip="Prompt for the image generation",
                 ),
-                IO.Boolean.Input(
+                comfy_io.Boolean.Input(
                     "turbo",
                     default=False,
                     tooltip="Whether to use turbo mode (faster generation, potentially lower quality)",
                 ),
-                IO.Combo.Input(
+                comfy_io.Combo.Input(
                     "aspect_ratio",
                     options=list(V1_V2_RATIO_MAP.keys()),
                     default="1:1",
                     tooltip="The aspect ratio for image generation.",
                     optional=True,
                 ),
-                IO.Combo.Input(
+                comfy_io.Combo.Input(
                     "magic_prompt_option",
                     options=["AUTO", "ON", "OFF"],
                     default="AUTO",
                     tooltip="Determine if MagicPrompt should be used in generation",
                     optional=True,
                 ),
-                IO.Int.Input(
+                comfy_io.Int.Input(
                     "seed",
                     default=0,
                     min=0,
                     max=2147483647,
                     step=1,
                     control_after_generate=True,
-                    display_mode=IO.NumberDisplay.number,
+                    display_mode=comfy_io.NumberDisplay.number,
                     optional=True,
                 ),
-                IO.String.Input(
+                comfy_io.String.Input(
                     "negative_prompt",
                     multiline=True,
                     default="",
                     tooltip="Description of what to exclude from the image",
                     optional=True,
                 ),
-                IO.Int.Input(
+                comfy_io.Int.Input(
                     "num_images",
                     default=1,
                     min=1,
                     max=8,
                     step=1,
-                    display_mode=IO.NumberDisplay.number,
+                    display_mode=comfy_io.NumberDisplay.number,
                     optional=True,
                 ),
             ],
             outputs=[
-                IO.Image.Output(),
+                comfy_io.Image.Output(),
             ],
             hidden=[
-                IO.Hidden.auth_token_comfy_org,
-                IO.Hidden.api_key_comfy_org,
-                IO.Hidden.unique_id,
+                comfy_io.Hidden.auth_token_comfy_org,
+                comfy_io.Hidden.api_key_comfy_org,
+                comfy_io.Hidden.unique_id,
             ],
         )
 
@@ -315,63 +334,77 @@ class IdeogramV1(IO.ComfyNode):
         aspect_ratio = V1_V2_RATIO_MAP.get(aspect_ratio, None)
         model = "V_1_TURBO" if turbo else "V_1"
 
-        response = await sync_op(
-            cls,
-            ApiEndpoint(path="/proxy/ideogram/generate", method="POST"),
-            response_model=IdeogramGenerateResponse,
-            data=IdeogramGenerateRequest(
+        auth = {
+            "auth_token": cls.hidden.auth_token_comfy_org,
+            "comfy_api_key": cls.hidden.api_key_comfy_org,
+        }
+        operation = SynchronousOperation(
+            endpoint=ApiEndpoint(
+                path="/proxy/ideogram/generate",
+                method=HttpMethod.POST,
+                request_model=IdeogramGenerateRequest,
+                response_model=IdeogramGenerateResponse,
+            ),
+            request=IdeogramGenerateRequest(
                 image_request=ImageRequest(
                     prompt=prompt,
                     model=model,
                     num_images=num_images,
                     seed=seed,
                     aspect_ratio=aspect_ratio if aspect_ratio != "ASPECT_1_1" else None,
-                    magic_prompt_option=(magic_prompt_option if magic_prompt_option != "AUTO" else None),
+                    magic_prompt_option=(
+                        magic_prompt_option if magic_prompt_option != "AUTO" else None
+                    ),
                     negative_prompt=negative_prompt if negative_prompt else None,
                 )
             ),
-            max_retries=1,
+            auth_kwargs=auth,
         )
 
+        response = await operation.execute()
+
         if not response.data or len(response.data) == 0:
             raise Exception("No images were generated in the response")
 
         image_urls = [image_data.url for image_data in response.data if image_data.url]
+
         if not image_urls:
             raise Exception("No image URLs were generated in the response")
-        return IO.NodeOutput(await download_and_process_images(image_urls))
+
+        display_image_urls_on_node(image_urls, cls.hidden.unique_id)
+        return comfy_io.NodeOutput(await download_and_process_images(image_urls))
 
 
-class IdeogramV2(IO.ComfyNode):
+class IdeogramV2(comfy_io.ComfyNode):
 
     @classmethod
     def define_schema(cls):
-        return IO.Schema(
+        return comfy_io.Schema(
             node_id="IdeogramV2",
             display_name="Ideogram V2",
             category="api node/image/Ideogram",
             description="Generates images using the Ideogram V2 model.",
             is_api_node=True,
             inputs=[
-                IO.String.Input(
+                comfy_io.String.Input(
                     "prompt",
                     multiline=True,
                     default="",
                     tooltip="Prompt for the image generation",
                 ),
-                IO.Boolean.Input(
+                comfy_io.Boolean.Input(
                     "turbo",
                     default=False,
                     tooltip="Whether to use turbo mode (faster generation, potentially lower quality)",
                 ),
-                IO.Combo.Input(
+                comfy_io.Combo.Input(
                     "aspect_ratio",
                     options=list(V1_V2_RATIO_MAP.keys()),
                     default="1:1",
                     tooltip="The aspect ratio for image generation. Ignored if resolution is not set to AUTO.",
                     optional=True,
                 ),
-                IO.Combo.Input(
+                comfy_io.Combo.Input(
                     "resolution",
                     options=list(V1_V1_RES_MAP.keys()),
                     default="Auto",
@@ -379,44 +412,44 @@ class IdeogramV2(IO.ComfyNode):
                             "If not set to AUTO, this overrides the aspect_ratio setting.",
                     optional=True,
                 ),
-                IO.Combo.Input(
+                comfy_io.Combo.Input(
                     "magic_prompt_option",
                     options=["AUTO", "ON", "OFF"],
                     default="AUTO",
                     tooltip="Determine if MagicPrompt should be used in generation",
                     optional=True,
                 ),
-                IO.Int.Input(
+                comfy_io.Int.Input(
                     "seed",
                     default=0,
                     min=0,
                     max=2147483647,
                     step=1,
                     control_after_generate=True,
-                    display_mode=IO.NumberDisplay.number,
+                    display_mode=comfy_io.NumberDisplay.number,
                     optional=True,
                 ),
-                IO.Combo.Input(
+                comfy_io.Combo.Input(
                     "style_type",
                     options=["AUTO", "GENERAL", "REALISTIC", "DESIGN", "RENDER_3D", "ANIME"],
                     default="NONE",
                     tooltip="Style type for generation (V2 only)",
                     optional=True,
                 ),
-                IO.String.Input(
+                comfy_io.String.Input(
                     "negative_prompt",
                     multiline=True,
                     default="",
                     tooltip="Description of what to exclude from the image",
                     optional=True,
                 ),
-                IO.Int.Input(
+                comfy_io.Int.Input(
                     "num_images",
                     default=1,
                     min=1,
                     max=8,
                     step=1,
-                    display_mode=IO.NumberDisplay.number,
+                    display_mode=comfy_io.NumberDisplay.number,
                     optional=True,
                 ),
                 #"color_palette": (
@@ -429,12 +462,12 @@ class IdeogramV2(IO.ComfyNode):
                 #),
             ],
             outputs=[
-                IO.Image.Output(),
+                comfy_io.Image.Output(),
             ],
             hidden=[
-                IO.Hidden.auth_token_comfy_org,
-                IO.Hidden.api_key_comfy_org,
-                IO.Hidden.unique_id,
+                comfy_io.Hidden.auth_token_comfy_org,
+                comfy_io.Hidden.api_key_comfy_org,
+                comfy_io.Hidden.unique_id,
             ],
         )
 
@@ -467,11 +500,18 @@ class IdeogramV2(IO.ComfyNode):
         else:
             final_aspect_ratio = aspect_ratio if aspect_ratio != "ASPECT_1_1" else None
 
-        response = await sync_op(
-            cls,
-            endpoint=ApiEndpoint(path="/proxy/ideogram/generate", method="POST"),
-            response_model=IdeogramGenerateResponse,
-            data=IdeogramGenerateRequest(
+        auth = {
+            "auth_token": cls.hidden.auth_token_comfy_org,
+            "comfy_api_key": cls.hidden.api_key_comfy_org,
+        }
+        operation = SynchronousOperation(
+            endpoint=ApiEndpoint(
+                path="/proxy/ideogram/generate",
+                method=HttpMethod.POST,
+                request_model=IdeogramGenerateRequest,
+                response_model=IdeogramGenerateResponse,
+            ),
+            request=IdeogramGenerateRequest(
                 image_request=ImageRequest(
                     prompt=prompt,
                     model=model,
@@ -479,28 +519,36 @@ class IdeogramV2(IO.ComfyNode):
                     seed=seed,
                     aspect_ratio=final_aspect_ratio,
                     resolution=final_resolution,
-                    magic_prompt_option=(magic_prompt_option if magic_prompt_option != "AUTO" else None),
+                    magic_prompt_option=(
+                        magic_prompt_option if magic_prompt_option != "AUTO" else None
+                    ),
                     style_type=style_type if style_type != "NONE" else None,
                     negative_prompt=negative_prompt if negative_prompt else None,
                     color_palette=color_palette if color_palette else None,
                 )
             ),
-            max_retries=1,
+            auth_kwargs=auth,
         )
+
+        response = await operation.execute()
+
         if not response.data or len(response.data) == 0:
             raise Exception("No images were generated in the response")
 
         image_urls = [image_data.url for image_data in response.data if image_data.url]
+
         if not image_urls:
             raise Exception("No image URLs were generated in the response")
-        return IO.NodeOutput(await download_and_process_images(image_urls))
+
+        display_image_urls_on_node(image_urls, cls.hidden.unique_id)
+        return comfy_io.NodeOutput(await download_and_process_images(image_urls))
 
 
-class IdeogramV3(IO.ComfyNode):
+class IdeogramV3(comfy_io.ComfyNode):
 
     @classmethod
     def define_schema(cls):
-        return IO.Schema(
+        return comfy_io.Schema(
             node_id="IdeogramV3",
             display_name="Ideogram V3",
             category="api node/image/Ideogram",
@@ -508,30 +556,30 @@ class IdeogramV3(IO.ComfyNode):
                         "Supports both regular image generation from text prompts and image editing with mask.",
             is_api_node=True,
             inputs=[
-                IO.String.Input(
+                comfy_io.String.Input(
                     "prompt",
                     multiline=True,
                     default="",
                     tooltip="Prompt for the image generation or editing",
                 ),
-                IO.Image.Input(
+                comfy_io.Image.Input(
                     "image",
                     tooltip="Optional reference image for image editing.",
                     optional=True,
                 ),
-                IO.Mask.Input(
+                comfy_io.Mask.Input(
                     "mask",
                     tooltip="Optional mask for inpainting (white areas will be replaced)",
                     optional=True,
                 ),
-                IO.Combo.Input(
+                comfy_io.Combo.Input(
                     "aspect_ratio",
                     options=list(V3_RATIO_MAP.keys()),
                     default="1:1",
                     tooltip="The aspect ratio for image generation. Ignored if resolution is not set to Auto.",
                     optional=True,
                 ),
-                IO.Combo.Input(
+                comfy_io.Combo.Input(
                     "resolution",
                     options=V3_RESOLUTIONS,
                     default="Auto",
@@ -539,57 +587,57 @@ class IdeogramV3(IO.ComfyNode):
                             "If not set to Auto, this overrides the aspect_ratio setting.",
                     optional=True,
                 ),
-                IO.Combo.Input(
+                comfy_io.Combo.Input(
                     "magic_prompt_option",
                     options=["AUTO", "ON", "OFF"],
                     default="AUTO",
                     tooltip="Determine if MagicPrompt should be used in generation",
                     optional=True,
                 ),
-                IO.Int.Input(
+                comfy_io.Int.Input(
                     "seed",
                     default=0,
                     min=0,
                     max=2147483647,
                     step=1,
                     control_after_generate=True,
-                    display_mode=IO.NumberDisplay.number,
+                    display_mode=comfy_io.NumberDisplay.number,
                     optional=True,
                 ),
-                IO.Int.Input(
+                comfy_io.Int.Input(
                     "num_images",
                     default=1,
                     min=1,
                     max=8,
                     step=1,
-                    display_mode=IO.NumberDisplay.number,
+                    display_mode=comfy_io.NumberDisplay.number,
                     optional=True,
                 ),
-                IO.Combo.Input(
+                comfy_io.Combo.Input(
                     "rendering_speed",
                     options=["DEFAULT", "TURBO", "QUALITY"],
                     default="DEFAULT",
                     tooltip="Controls the trade-off between generation speed and quality",
                     optional=True,
                 ),
-                IO.Image.Input(
+                comfy_io.Image.Input(
                     "character_image",
                     tooltip="Image to use as character reference.",
                     optional=True,
                 ),
-                IO.Mask.Input(
+                comfy_io.Mask.Input(
                     "character_mask",
                     tooltip="Optional mask for character reference image.",
                     optional=True,
                 ),
             ],
             outputs=[
-                IO.Image.Output(),
+                comfy_io.Image.Output(),
             ],
             hidden=[
-                IO.Hidden.auth_token_comfy_org,
-                IO.Hidden.api_key_comfy_org,
-                IO.Hidden.unique_id,
+                comfy_io.Hidden.auth_token_comfy_org,
+                comfy_io.Hidden.api_key_comfy_org,
+                comfy_io.Hidden.unique_id,
             ],
         )
 
@@ -608,6 +656,10 @@ class IdeogramV3(IO.ComfyNode):
         character_image=None,
         character_mask=None,
     ):
+        auth = {
+            "auth_token": cls.hidden.auth_token_comfy_org,
+            "comfy_api_key": cls.hidden.api_key_comfy_org,
+        }
         if rendering_speed == "BALANCED":  # for backward compatibility
             rendering_speed = "DEFAULT"
 
@@ -642,6 +694,9 @@ class IdeogramV3(IO.ComfyNode):
 
         # Check if both image and mask are provided for editing mode
         if image is not None and mask is not None:
+            # Edit mode
+            path = "/proxy/ideogram/ideogram-v3/edit"
+
             # Process image and mask
             input_tensor = image.squeeze().cpu()
             # Resize mask to match image dimension
@@ -694,20 +749,27 @@ class IdeogramV3(IO.ComfyNode):
             if character_mask_binary:
                 files["character_mask_binary"] = character_mask_binary
 
-            response = await sync_op(
-                cls,
-                ApiEndpoint(path="/proxy/ideogram/ideogram-v3/edit", method="POST"),
-                response_model=IdeogramGenerateResponse,
-                data=edit_request,
+            # Execute the operation for edit mode
+            operation = SynchronousOperation(
+                endpoint=ApiEndpoint(
+                    path=path,
+                    method=HttpMethod.POST,
+                    request_model=IdeogramV3EditRequest,
+                    response_model=IdeogramGenerateResponse,
+                ),
+                request=edit_request,
                 files=files,
                 content_type="multipart/form-data",
-                max_retries=1,
+                auth_kwargs=auth,
             )
 
         elif image is not None or mask is not None:
             # If only one of image or mask is provided, raise an error
             raise Exception("Ideogram V3 image editing requires both an image AND a mask")
         else:
+            # Generation mode
+            path = "/proxy/ideogram/ideogram-v3/generate"
+
             # Create generation request
             gen_request = IdeogramV3Request(
                 prompt=prompt,
@@ -738,34 +800,43 @@ class IdeogramV3(IO.ComfyNode):
             if files:
                 gen_request.style_type = "AUTO"
 
-            response = await sync_op(
-                cls,
-                endpoint=ApiEndpoint(path="/proxy/ideogram/ideogram-v3/generate", method="POST"),
-                response_model=IdeogramGenerateResponse,
-                data=gen_request,
+            # Execute the operation for generation mode
+            operation = SynchronousOperation(
+                endpoint=ApiEndpoint(
+                    path=path,
+                    method=HttpMethod.POST,
+                    request_model=IdeogramV3Request,
+                    response_model=IdeogramGenerateResponse,
+                ),
+                request=gen_request,
                 files=files if files else None,
                 content_type="multipart/form-data",
-                max_retries=1,
+                auth_kwargs=auth,
             )
 
+        # Execute the operation and process response
+        response = await operation.execute()
+
         if not response.data or len(response.data) == 0:
             raise Exception("No images were generated in the response")
 
         image_urls = [image_data.url for image_data in response.data if image_data.url]
+
         if not image_urls:
             raise Exception("No image URLs were generated in the response")
-        return IO.NodeOutput(await download_and_process_images(image_urls))
+
+        display_image_urls_on_node(image_urls, cls.hidden.unique_id)
+        return comfy_io.NodeOutput(await download_and_process_images(image_urls))
 
 
 class IdeogramExtension(ComfyExtension):
     @override
-    async def get_node_list(self) -> list[type[IO.ComfyNode]]:
+    async def get_node_list(self) -> list[type[comfy_io.ComfyNode]]:
         return [
             IdeogramV1,
             IdeogramV2,
             IdeogramV3,
         ]
 
-
 async def comfy_entrypoint() -> IdeogramExtension:
     return IdeogramExtension()

comfy_api_nodes/nodes_ltxv.py

@@ -1,199 +0,0 @@
-from io import BytesIO
-from typing import Optional
-
-import torch
-from pydantic import BaseModel, Field
-from typing_extensions import override
-
-from comfy_api.input_impl import VideoFromFile
-from comfy_api.latest import IO, ComfyExtension
-from comfy_api_nodes.util import (
-    ApiEndpoint,
-    get_number_of_images,
-    sync_op_raw,
-    upload_images_to_comfyapi,
-    validate_string,
-)
-
-MODELS_MAP = {
-    "LTX-2 (Pro)": "ltx-2-pro",
-    "LTX-2 (Fast)": "ltx-2-fast",
-}
-
-
-class ExecuteTaskRequest(BaseModel):
-    prompt: str = Field(...)
-    model: str = Field(...)
-    duration: int = Field(...)
-    resolution: str = Field(...)
-    fps: Optional[int] = Field(25)
-    generate_audio: Optional[bool] = Field(True)
-    image_uri: Optional[str] = Field(None)
-
-
-class TextToVideoNode(IO.ComfyNode):
-    @classmethod
-    def define_schema(cls):
-        return IO.Schema(
-            node_id="LtxvApiTextToVideo",
-            display_name="LTXV Text To Video",
-            category="api node/video/LTXV",
-            description="Professional-quality videos with customizable duration and resolution.",
-            inputs=[
-                IO.Combo.Input("model", options=list(MODELS_MAP.keys())),
-                IO.String.Input(
-                    "prompt",
-                    multiline=True,
-                    default="",
-                ),
-                IO.Combo.Input("duration", options=[6, 8, 10, 12, 14, 16, 18, 20], default=8),
-                IO.Combo.Input(
-                    "resolution",
-                    options=[
-                        "1920x1080",
-                        "2560x1440",
-                        "3840x2160",
-                    ],
-                ),
-                IO.Combo.Input("fps", options=[25, 50], default=25),
-                IO.Boolean.Input(
-                    "generate_audio",
-                    default=False,
-                    optional=True,
-                    tooltip="When true, the generated video will include AI-generated audio matching the scene.",
-                ),
-            ],
-            outputs=[
-                IO.Video.Output(),
-            ],
-            hidden=[
-                IO.Hidden.auth_token_comfy_org,
-                IO.Hidden.api_key_comfy_org,
-                IO.Hidden.unique_id,
-            ],
-            is_api_node=True,
-        )
-
-    @classmethod
-    async def execute(
-        cls,
-        model: str,
-        prompt: str,
-        duration: int,
-        resolution: str,
-        fps: int = 25,
-        generate_audio: bool = False,
-    ) -> IO.NodeOutput:
-        validate_string(prompt, min_length=1, max_length=10000)
-        if duration > 10 and (model != "LTX-2 (Fast)" or resolution != "1920x1080" or fps != 25):
-            raise ValueError(
-                "Durations over 10s are only available for the Fast model at 1920x1080 resolution and 25 FPS."
-            )
-        response = await sync_op_raw(
-            cls,
-            ApiEndpoint("/proxy/ltx/v1/text-to-video", "POST"),
-            data=ExecuteTaskRequest(
-                prompt=prompt,
-                model=MODELS_MAP[model],
-                duration=duration,
-                resolution=resolution,
-                fps=fps,
-                generate_audio=generate_audio,
-            ),
-            as_binary=True,
-            max_retries=1,
-        )
-        return IO.NodeOutput(VideoFromFile(BytesIO(response)))
-
-
-class ImageToVideoNode(IO.ComfyNode):
-    @classmethod
-    def define_schema(cls):
-        return IO.Schema(
-            node_id="LtxvApiImageToVideo",
-            display_name="LTXV Image To Video",
-            category="api node/video/LTXV",
-            description="Professional-quality videos with customizable duration and resolution based on start image.",
-            inputs=[
-                IO.Image.Input("image", tooltip="First frame to be used for the video."),
-                IO.Combo.Input("model", options=list(MODELS_MAP.keys())),
-                IO.String.Input(
-                    "prompt",
-                    multiline=True,
-                    default="",
-                ),
-                IO.Combo.Input("duration", options=[6, 8, 10, 12, 14, 16, 18, 20], default=8),
-                IO.Combo.Input(
-                    "resolution",
-                    options=[
-                        "1920x1080",
-                        "2560x1440",
-                        "3840x2160",
-                    ],
-                ),
-                IO.Combo.Input("fps", options=[25, 50], default=25),
-                IO.Boolean.Input(
-                    "generate_audio",
-                    default=False,
-                    optional=True,
-                    tooltip="When true, the generated video will include AI-generated audio matching the scene.",
-                ),
-            ],
-            outputs=[
-                IO.Video.Output(),
-            ],
-            hidden=[
-                IO.Hidden.auth_token_comfy_org,
-                IO.Hidden.api_key_comfy_org,
-                IO.Hidden.unique_id,
-            ],
-            is_api_node=True,
-        )
-
-    @classmethod
-    async def execute(
-        cls,
-        image: torch.Tensor,
-        model: str,
-        prompt: str,
-        duration: int,
-        resolution: str,
-        fps: int = 25,
-        generate_audio: bool = False,
-    ) -> IO.NodeOutput:
-        validate_string(prompt, min_length=1, max_length=10000)
-        if duration > 10 and (model != "LTX-2 (Fast)" or resolution != "1920x1080" or fps != 25):
-            raise ValueError(
-                "Durations over 10s are only available for the Fast model at 1920x1080 resolution and 25 FPS."
-            )
-        if get_number_of_images(image) != 1:
-            raise ValueError("Currently only one input image is supported.")
-        response = await sync_op_raw(
-            cls,
-            ApiEndpoint("/proxy/ltx/v1/image-to-video", "POST"),
-            data=ExecuteTaskRequest(
-                image_uri=(await upload_images_to_comfyapi(cls, image, max_images=1, mime_type="image/png"))[0],
-                prompt=prompt,
-                model=MODELS_MAP[model],
-                duration=duration,
-                resolution=resolution,
-                fps=fps,
-                generate_audio=generate_audio,
-            ),
-            as_binary=True,
-            max_retries=1,
-        )
-        return IO.NodeOutput(VideoFromFile(BytesIO(response)))
-
-
-class LtxvApiExtension(ComfyExtension):
-    @override
-    async def get_node_list(self) -> list[type[IO.ComfyNode]]:
-        return [
-            TextToVideoNode,
-            ImageToVideoNode,
-        ]
-
-
-async def comfy_entrypoint() -> LtxvApiExtension:
-    return LtxvApiExtension()