refactor

As commented, if you let cuda be the one to detect double pin/unpinning
it actually creates an asyc GPU error.

.ci/windows_amd_base_files/README_VERY_IMPORTANT.txt

@@ -0,0 +1,27 @@
+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_amd_base_files/run_amd_gpu_disable_smart_memory.bat

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

.ci/windows_nvidia_base_files/advanced/run_nvidia_gpu_disable_api_nodes.bat

@@ -0,0 +1,3 @@
+..\python_embeded\python.exe -s ..\ComfyUI\main.py --windows-standalone-build --disable-api-nodes
+echo If you see this and ComfyUI did not start try updating your Nvidia Drivers to the latest.
+pause

.ci/windows_nvidia_base_files/run_nvidia_gpu.bat

@@ -0,0 +1,3 @@
+.\python_embeded\python.exe -s ComfyUI\main.py --windows-standalone-build
+echo If you see this and ComfyUI did not start try updating your Nvidia Drivers to the latest.
+pause

.ci/windows_nvidia_base_files/run_nvidia_gpu_fast_fp16_accumulation.bat

@@ -0,0 +1,3 @@
+.\python_embeded\python.exe -s ComfyUI\main.py --windows-standalone-build --fast fp16_accumulation
+echo If you see this and ComfyUI did not start try updating your Nvidia Drivers to the latest.
+pause

.github/ISSUE_TEMPLATE/bug-report.yml

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

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

@@ -0,0 +1,61 @@
+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 (cu129)"
+    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_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

@@ -21,3 +21,28 @@ 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

@@ -2,17 +2,17 @@
 name: "Release Stable Version"
 
 on:
-  workflow_dispatch:
+  workflow_call:
     inputs:
       git_tag:
         description: 'Git tag'
         required: true
         type: string
-      cu:
-        description: 'CUDA version'
+      cache_tag:
+        description: 'Cached dependencies tag'
         required: true
         type: string
-        default: "129"
+        default: "cu129"
       python_minor:
         description: 'Python minor version'
         required: true
@@ -23,7 +23,57 @@ 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
+  workflow_dispatch:
+    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
 
 jobs:
   package_comfy_windows:
@@ -42,15 +92,15 @@ jobs:
         id: cache
         with:
           path: |
-            cu${{ inputs.cu }}_python_deps.tar
+            ${{ inputs.cache_tag }}_python_deps.tar
             update_comfyui_and_python_dependencies.bat
-          key: ${{ runner.os }}-build-cu${{ inputs.cu }}-${{ inputs.python_minor }}
+          key: ${{ runner.os }}-build-${{ inputs.cache_tag }}-${{ inputs.python_minor }}
       - shell: bash
         run: |
-          mv cu${{ inputs.cu }}_python_deps.tar ../
+          mv ${{ inputs.cache_tag }}_python_deps.tar ../
           mv update_comfyui_and_python_dependencies.bat ../
           cd ..
-          tar xf cu${{ inputs.cu }}_python_deps.tar
+          tar xf ${{ inputs.cache_tag }}_python_deps.tar
           pwd
           ls
 
@@ -65,12 +115,19 @@ jobs:
           echo 'import site' >> ./python3${{ inputs.python_minor }}._pth
           curl https://bootstrap.pypa.io/get-pip.py -o get-pip.py
           ./python.exe get-pip.py
-          ./python.exe -s -m pip install ../cu${{ inputs.cu }}_python_deps/*
+          ./python.exe -s -m pip install ../${{ inputs.cache_tag }}_python_deps/*
+
+          grep comfyui ../ComfyUI/requirements.txt > ./requirements_comfyui.txt
+          ./python.exe -s -m pip install -r requirements_comfyui.txt
+          rm requirements_comfyui.txt
+
           sed -i '1i../ComfyUI' ./python3${{ inputs.python_minor }}._pth
 
-          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
+          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
 
           cd ..
 
@@ -85,14 +142,18 @@ jobs:
 
           mkdir update
           cp -r ComfyUI/.ci/update_windows/* ./update/
-          cp -r ComfyUI/.ci/windows_base_files/* ./
+          cp -r ComfyUI/.ci/windows_${{ inputs.rel_name }}_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_nvidia.7z
+          mv ComfyUI_windows_portable.7z ComfyUI/ComfyUI_windows_portable_${{ inputs.rel_name }}${{ inputs.rel_extra_name }}.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
 
@@ -101,10 +162,9 @@ jobs:
           ls
 
       - name: Upload binaries to release
-        uses: svenstaro/upload-release-action@v2
+        uses: softprops/action-gh-release@v2
         with:
-          repo_token: ${{ secrets.GITHUB_TOKEN }}
-          file: ComfyUI_windows_portable_nvidia.7z
-          tag: ${{ inputs.git_tag }}
-          overwrite: true
+          files: ComfyUI_windows_portable_${{ inputs.rel_name }}${{ inputs.rel_extra_name }}.7z
+          tag_name: ${{ inputs.git_tag }}
           draft: true
+          overwrite_files: true

.github/workflows/test-assets.yml

@@ -1,173 +0,0 @@
-name: Asset System Tests
-
-on:
-  push:
-    paths:
-      - 'app/**'
-      - 'tests-assets/**'
-      - '.github/workflows/test-assets.yml'
-      - 'requirements.txt'
-  pull_request:
-    branches: [master]
-  workflow_dispatch:
-
-permissions:
-  contents: read
-
-env:
-  PIP_DISABLE_PIP_VERSION_CHECK: '1'
-  PYTHONUNBUFFERED: '1'
-
-jobs:
-  sqlite:
-    name: SQLite (${{ matrix.sqlite_mode }}) • Python ${{ matrix.python }}
-    runs-on: ubuntu-latest
-    timeout-minutes: 40
-    strategy:
-      fail-fast: false
-      matrix:
-        python: ['3.9', '3.12']
-        sqlite_mode: ['memory', 'file']
-
-    steps:
-      - uses: actions/checkout@v4
-      - name: Set up Python
-        uses: actions/setup-python@v5
-        with:
-          python-version: ${{ matrix.python }}
-
-      - name: Install dependencies
-        run: |
-          python -m pip install -U pip wheel
-          pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cpu
-          pip install -r requirements.txt
-          pip install pytest pytest-aiohttp pytest-asyncio
-
-      - name: Set deterministic test base dir
-        id: basedir
-        shell: bash
-        run: |
-          BASE="$RUNNER_TEMP/comfyui-assets-tests-${{ matrix.python }}-${{ matrix.sqlite_mode }}-${{ github.run_id }}-${{ github.run_attempt }}"
-          echo "ASSETS_TEST_BASE_DIR=$BASE" >> "$GITHUB_ENV"
-          echo "ASSETS_TEST_LOGS=$BASE/logs" >> "$GITHUB_ENV"
-          mkdir -p "$BASE/logs"
-          echo "ASSETS_TEST_BASE_DIR=$BASE"
-
-      - name: Set DB URL for SQLite
-        id: setdb
-        shell: bash
-        run: |
-          if [ "${{ matrix.sqlite_mode }}" = "memory" ]; then
-            echo "ASSETS_TEST_DB_URL=sqlite+aiosqlite:///:memory:" >> "$GITHUB_ENV"
-          else
-            DBFILE="$RUNNER_TEMP/assets-tests.sqlite"
-            mkdir -p "$(dirname "$DBFILE")"
-            echo "ASSETS_TEST_DB_URL=sqlite+aiosqlite:///$DBFILE" >> "$GITHUB_ENV"
-          fi
-
-      - name: Run tests
-        run: python -m pytest tests-assets
-
-      - name: Show ComfyUI logs
-        if: always()
-        shell: bash
-        run: |
-          echo "==== ASSETS_TEST_BASE_DIR: $ASSETS_TEST_BASE_DIR ===="
-          echo "==== ASSETS_TEST_LOGS: $ASSETS_TEST_LOGS ===="
-          ls -la "$ASSETS_TEST_LOGS" || true
-          for f in "$ASSETS_TEST_LOGS"/stdout.log "$ASSETS_TEST_LOGS"/stderr.log; do
-            if [ -f "$f" ]; then
-              echo "----- BEGIN $f -----"
-              sed -n '1,400p' "$f"
-              echo "----- END $f -----"
-            fi
-          done
-
-      - name: Upload ComfyUI logs
-        if: always()
-        uses: actions/upload-artifact@v4
-        with:
-          name: asset-logs-sqlite-${{ matrix.sqlite_mode }}-py${{ matrix.python }}
-          path: ${{ env.ASSETS_TEST_LOGS }}/*.log
-          if-no-files-found: warn
-
-  postgres:
-    name: PostgreSQL ${{ matrix.pgsql }} • Python ${{ matrix.python }}
-    runs-on: ubuntu-latest
-    timeout-minutes: 40
-    strategy:
-      fail-fast: false
-      matrix:
-        python: ['3.9', '3.12']
-        pgsql: ['16', '18']
-
-    services:
-      postgres:
-        image: postgres:${{ matrix.pgsql }}
-        env:
-          POSTGRES_DB: assets
-          POSTGRES_USER: postgres
-          POSTGRES_PASSWORD: postgres
-        ports:
-          - 5432:5432
-        options: >-
-          --health-cmd "pg_isready -U postgres -d assets"
-          --health-interval 10s
-          --health-timeout 5s
-          --health-retries 12
-
-    steps:
-      - uses: actions/checkout@v4
-      - name: Set up Python
-        uses: actions/setup-python@v5
-        with:
-          python-version: ${{ matrix.python }}
-
-      - name: Install dependencies
-        run: |
-          python -m pip install -U pip wheel
-          pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cpu
-          pip install -r requirements.txt
-          pip install pytest pytest-aiohttp pytest-asyncio
-          pip install greenlet psycopg
-
-      - name: Set deterministic test base dir
-        id: basedir
-        shell: bash
-        run: |
-          BASE="$RUNNER_TEMP/comfyui-assets-tests-${{ matrix.python }}-${{ matrix.sqlite_mode }}-${{ github.run_id }}-${{ github.run_attempt }}"
-          echo "ASSETS_TEST_BASE_DIR=$BASE" >> "$GITHUB_ENV"
-          echo "ASSETS_TEST_LOGS=$BASE/logs" >> "$GITHUB_ENV"
-          mkdir -p "$BASE/logs"
-          echo "ASSETS_TEST_BASE_DIR=$BASE"
-
-      - name: Set DB URL for PostgreSQL
-        shell: bash
-        run: |
-          echo "ASSETS_TEST_DB_URL=postgresql+psycopg://postgres:postgres@localhost:5432/assets" >> "$GITHUB_ENV"
-
-      - name: Run tests
-        run: python -m pytest tests-assets
-
-      - name: Show ComfyUI logs
-        if: always()
-        shell: bash
-        run: |
-          echo "==== ASSETS_TEST_BASE_DIR: $ASSETS_TEST_BASE_DIR ===="
-          echo "==== ASSETS_TEST_LOGS: $ASSETS_TEST_LOGS ===="
-          ls -la "$ASSETS_TEST_LOGS" || true
-          for f in "$ASSETS_TEST_LOGS"/stdout.log "$ASSETS_TEST_LOGS"/stderr.log; do
-            if [ -f "$f" ]; then
-              echo "----- BEGIN $f -----"
-              sed -n '1,400p' "$f"
-              echo "----- END $f -----"
-            fi
-          done
-
-      - name: Upload ComfyUI logs
-        if: always()
-        uses: actions/upload-artifact@v4
-        with:
-          name: asset-logs-pgsql-${{ matrix.pgsql }}-py${{ matrix.python }}
-          path: ${{ env.ASSETS_TEST_LOGS }}/*.log
-          if-no-files-found: warn

.github/workflows/windows_release_dependencies.yml

@@ -17,7 +17,7 @@ on:
         description: 'cuda version'
         required: true
         type: string
-        default: "129"
+        default: "130"
 
       python_minor:
         description: 'python minor version'
@@ -29,7 +29,7 @@ on:
         description: 'python patch version'
         required: true
         type: string
-        default: "6"
+        default: "9"
 #  push:
 #    branches:
 #      - master
@@ -56,7 +56,8 @@ 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
 
-            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
+            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 install --no-cache-dir ./temp_wheel_dir/*
             echo installed basic
             ls -lah temp_wheel_dir

.github/workflows/windows_release_dependencies_manual.yml

@@ -0,0 +1,64 @@
+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_base_files/* ./
+            cp -r ComfyUI/.ci/windows_nvidia_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_base_files/* ./
+            cp -r ComfyUI/.ci/windows_nvidia_base_files/* ./
             cp ../update_comfyui_and_python_dependencies.bat ./update/
 
             cd ..

README.md

@@ -112,10 +112,11 @@ Workflow examples can be found on the [Examples page](https://comfyanonymous.git
 
 ## Release Process
 
-ComfyUI follows a weekly release cycle targeting Friday but this regularly changes because of model releases or large changes to the codebase. There are three interconnected repositories:
+ComfyUI follows a weekly release cycle targeting Monday but this regularly changes because of model releases or large changes to the codebase. There are three interconnected repositories:
 
 1. **[ComfyUI Core](https://github.com/comfyanonymous/ComfyUI)**
-   - Releases a new stable version (e.g., v0.7.0)
+   - Releases a new stable version (e.g., v0.7.0) roughly every week.
+   - Commits outside of the stable release tags may be very unstable and break many custom nodes.
    - Serves as the foundation for the desktop release
 
 2. **[ComfyUI Desktop](https://github.com/Comfy-Org/desktop)**
@@ -176,6 +177,14 @@ Simply download, extract with [7-Zip](https://7-zip.org) and run. Make sure you
 
 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) (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.
@@ -191,7 +200,11 @@ comfy install
 
 ## Manual Install (Windows, Linux)
 
-Python 3.13 is very well supported. If you have trouble with some custom node dependencies you can try 3.12
+Python 3.14 will work if you comment out the `kornia` dependency in the requirements.txt file (breaks the canny node) but it is not recommended.
+
+Python 3.13 is very well supported. If you have trouble with some custom node dependencies on 3.13 you can try 3.12
+
+### Instructions:
 
 Git clone this repo.
 
@@ -200,14 +213,32 @@ Put your SD checkpoints (the huge ckpt/safetensors files) in: models/checkpoints
 Put your VAE in: models/vae
 
 
-### AMD GPUs (Linux only)
+### AMD GPUs (Linux)
+
 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 6.4 which might have some performance improvements:
+This is the command to install the nightly with ROCm 7.0 which might have some performance improvements:
 
-```pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/rocm6.4```
+```pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/rocm7.0```
+
+
+### 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/```
 
 ### Intel GPUs (Windows and Linux)
 
@@ -229,11 +260,11 @@ This is the command to install the Pytorch xpu nightly which might have some per
 
 Nvidia users should install stable pytorch using this command:
 
-```pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu129```
+```pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu130```
 
 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/cu129```
+```pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu130```
 
 #### Troubleshooting
 
@@ -264,12 +295,6 @@ 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:

alembic.ini

@@ -3,7 +3,7 @@
 [alembic]
 # path to migration scripts
 # Use forward slashes (/) also on windows to provide an os agnostic path
-script_location = app/alembic_db
+script_location = alembic_db
 
 # template used to generate migration file names; The default value is %%(rev)s_%%(slug)s
 # Uncomment the line below if you want the files to be prepended with date and time

alembic_db/env.py

@@ -2,12 +2,13 @@ from sqlalchemy import engine_from_config
 from sqlalchemy import pool
 
 from alembic import context
-from app.assets.database.models import Base
 
 # this is the Alembic Config object, which provides
 # access to the values within the .ini file in use.
 config = context.config
 
+
+from app.database.models import Base
 target_metadata = Base.metadata
 
 # other values from the config, defined by the needs of env.py,

app/alembic_db/versions/0001_assets.py

@@ -1,175 +0,0 @@
-"""initial assets schema
-
-Revision ID: 0001_assets
-Revises:
-Create Date: 2025-08-20 00:00:00
-"""
-
-from alembic import op
-import sqlalchemy as sa
-from sqlalchemy.dialects import postgresql
-
-revision = "0001_assets"
-down_revision = None
-branch_labels = None
-depends_on = None
-
-
-def upgrade() -> None:
-    # ASSETS: content identity
-    op.create_table(
-        "assets",
-        sa.Column("id", sa.String(length=36), primary_key=True),
-        sa.Column("hash", sa.String(length=256), nullable=True),
-        sa.Column("size_bytes", sa.BigInteger(), nullable=False, server_default="0"),
-        sa.Column("mime_type", sa.String(length=255), nullable=True),
-        sa.Column("created_at", sa.DateTime(timezone=False), nullable=False),
-        sa.CheckConstraint("size_bytes >= 0", name="ck_assets_size_nonneg"),
-    )
-    op.create_index("uq_assets_hash", "assets", ["hash"], unique=True)
-    op.create_index("ix_assets_mime_type", "assets", ["mime_type"])
-
-    # ASSETS_INFO: user-visible references
-    op.create_table(
-        "assets_info",
-        sa.Column("id", sa.String(length=36), primary_key=True),
-        sa.Column("owner_id", sa.String(length=128), nullable=False, server_default=""),
-        sa.Column("name", sa.String(length=512), nullable=False),
-        sa.Column("asset_id", sa.String(length=36), sa.ForeignKey("assets.id", ondelete="RESTRICT"), nullable=False),
-        sa.Column("preview_id", sa.String(length=36), sa.ForeignKey("assets.id", ondelete="SET NULL"), nullable=True),
-        sa.Column("user_metadata", sa.JSON(), nullable=True),
-        sa.Column("created_at", sa.DateTime(timezone=False), nullable=False),
-        sa.Column("updated_at", sa.DateTime(timezone=False), nullable=False),
-        sa.Column("last_access_time", sa.DateTime(timezone=False), nullable=False),
-        sa.UniqueConstraint("asset_id", "owner_id", "name", name="uq_assets_info_asset_owner_name"),
-    )
-    op.create_index("ix_assets_info_owner_id", "assets_info", ["owner_id"])
-    op.create_index("ix_assets_info_asset_id", "assets_info", ["asset_id"])
-    op.create_index("ix_assets_info_name", "assets_info", ["name"])
-    op.create_index("ix_assets_info_created_at", "assets_info", ["created_at"])
-    op.create_index("ix_assets_info_last_access_time", "assets_info", ["last_access_time"])
-    op.create_index("ix_assets_info_owner_name", "assets_info", ["owner_id", "name"])
-
-    # TAGS: normalized tag vocabulary
-    op.create_table(
-        "tags",
-        sa.Column("name", sa.String(length=512), primary_key=True),
-        sa.Column("tag_type", sa.String(length=32), nullable=False, server_default="user"),
-        sa.CheckConstraint("name = lower(name)", name="ck_tags_lowercase"),
-    )
-    op.create_index("ix_tags_tag_type", "tags", ["tag_type"])
-
-    # ASSET_INFO_TAGS: many-to-many for tags on AssetInfo
-    op.create_table(
-        "asset_info_tags",
-        sa.Column("asset_info_id", sa.String(length=36), sa.ForeignKey("assets_info.id", ondelete="CASCADE"), nullable=False),
-        sa.Column("tag_name", sa.String(length=512), sa.ForeignKey("tags.name", ondelete="RESTRICT"), nullable=False),
-        sa.Column("origin", sa.String(length=32), nullable=False, server_default="manual"),
-        sa.Column("added_at", sa.DateTime(timezone=False), nullable=False),
-        sa.PrimaryKeyConstraint("asset_info_id", "tag_name", name="pk_asset_info_tags"),
-    )
-    op.create_index("ix_asset_info_tags_tag_name", "asset_info_tags", ["tag_name"])
-    op.create_index("ix_asset_info_tags_asset_info_id", "asset_info_tags", ["asset_info_id"])
-
-    # ASSET_CACHE_STATE: N:1 local cache rows per Asset
-    op.create_table(
-        "asset_cache_state",
-        sa.Column("id", sa.Integer(), primary_key=True, autoincrement=True),
-        sa.Column("asset_id", sa.String(length=36), sa.ForeignKey("assets.id", ondelete="CASCADE"), nullable=False),
-        sa.Column("file_path", sa.Text(), nullable=False),  # absolute local path to cached file
-        sa.Column("mtime_ns", sa.BigInteger(), nullable=True),
-        sa.Column("needs_verify", sa.Boolean(), nullable=False, server_default=sa.text("false")),
-        sa.CheckConstraint("(mtime_ns IS NULL) OR (mtime_ns >= 0)", name="ck_acs_mtime_nonneg"),
-        sa.UniqueConstraint("file_path", name="uq_asset_cache_state_file_path"),
-    )
-    op.create_index("ix_asset_cache_state_file_path", "asset_cache_state", ["file_path"])
-    op.create_index("ix_asset_cache_state_asset_id", "asset_cache_state", ["asset_id"])
-
-    # ASSET_INFO_META: typed KV projection of user_metadata for filtering/sorting
-    op.create_table(
-        "asset_info_meta",
-        sa.Column("asset_info_id", sa.String(length=36), sa.ForeignKey("assets_info.id", ondelete="CASCADE"), nullable=False),
-        sa.Column("key", sa.String(length=256), nullable=False),
-        sa.Column("ordinal", sa.Integer(), nullable=False, server_default="0"),
-        sa.Column("val_str", sa.String(length=2048), nullable=True),
-        sa.Column("val_num", sa.Numeric(38, 10), nullable=True),
-        sa.Column("val_bool", sa.Boolean(), nullable=True),
-        sa.Column("val_json", sa.JSON().with_variant(postgresql.JSONB(), 'postgresql'), nullable=True),
-        sa.PrimaryKeyConstraint("asset_info_id", "key", "ordinal", name="pk_asset_info_meta"),
-    )
-    op.create_index("ix_asset_info_meta_key", "asset_info_meta", ["key"])
-    op.create_index("ix_asset_info_meta_key_val_str", "asset_info_meta", ["key", "val_str"])
-    op.create_index("ix_asset_info_meta_key_val_num", "asset_info_meta", ["key", "val_num"])
-    op.create_index("ix_asset_info_meta_key_val_bool", "asset_info_meta", ["key", "val_bool"])
-
-    # Tags vocabulary
-    tags_table = sa.table(
-        "tags",
-        sa.column("name", sa.String(length=512)),
-        sa.column("tag_type", sa.String()),
-    )
-    op.bulk_insert(
-        tags_table,
-        [
-            {"name": "models", "tag_type": "system"},
-            {"name": "input", "tag_type": "system"},
-            {"name": "output", "tag_type": "system"},
-
-            {"name": "configs", "tag_type": "system"},
-            {"name": "checkpoints", "tag_type": "system"},
-            {"name": "loras", "tag_type": "system"},
-            {"name": "vae", "tag_type": "system"},
-            {"name": "text_encoders", "tag_type": "system"},
-            {"name": "diffusion_models", "tag_type": "system"},
-            {"name": "clip_vision", "tag_type": "system"},
-            {"name": "style_models", "tag_type": "system"},
-            {"name": "embeddings", "tag_type": "system"},
-            {"name": "diffusers", "tag_type": "system"},
-            {"name": "vae_approx", "tag_type": "system"},
-            {"name": "controlnet", "tag_type": "system"},
-            {"name": "gligen", "tag_type": "system"},
-            {"name": "upscale_models", "tag_type": "system"},
-            {"name": "hypernetworks", "tag_type": "system"},
-            {"name": "photomaker", "tag_type": "system"},
-            {"name": "classifiers", "tag_type": "system"},
-
-            {"name": "encoder", "tag_type": "system"},
-            {"name": "decoder", "tag_type": "system"},
-
-            {"name": "missing", "tag_type": "system"},
-            {"name": "rescan", "tag_type": "system"},
-        ],
-    )
-
-
-def downgrade() -> None:
-    op.drop_index("ix_asset_info_meta_key_val_bool", table_name="asset_info_meta")
-    op.drop_index("ix_asset_info_meta_key_val_num", table_name="asset_info_meta")
-    op.drop_index("ix_asset_info_meta_key_val_str", table_name="asset_info_meta")
-    op.drop_index("ix_asset_info_meta_key", table_name="asset_info_meta")
-    op.drop_table("asset_info_meta")
-
-    op.drop_index("ix_asset_cache_state_asset_id", table_name="asset_cache_state")
-    op.drop_index("ix_asset_cache_state_file_path", table_name="asset_cache_state")
-    op.drop_constraint("uq_asset_cache_state_file_path", table_name="asset_cache_state")
-    op.drop_table("asset_cache_state")
-
-    op.drop_index("ix_asset_info_tags_asset_info_id", table_name="asset_info_tags")
-    op.drop_index("ix_asset_info_tags_tag_name", table_name="asset_info_tags")
-    op.drop_table("asset_info_tags")
-
-    op.drop_index("ix_tags_tag_type", table_name="tags")
-    op.drop_table("tags")
-
-    op.drop_constraint("uq_assets_info_asset_owner_name", table_name="assets_info")
-    op.drop_index("ix_assets_info_owner_name", table_name="assets_info")
-    op.drop_index("ix_assets_info_last_access_time", table_name="assets_info")
-    op.drop_index("ix_assets_info_created_at", table_name="assets_info")
-    op.drop_index("ix_assets_info_name", table_name="assets_info")
-    op.drop_index("ix_assets_info_asset_id", table_name="assets_info")
-    op.drop_index("ix_assets_info_owner_id", table_name="assets_info")
-    op.drop_table("assets_info")
-
-    op.drop_index("uq_assets_hash", table_name="assets")
-    op.drop_index("ix_assets_mime_type", table_name="assets")
-    op.drop_table("assets")

app/assets/__init__.py

@@ -1,4 +0,0 @@
-from .api.routes import register_assets_system
-from .scanner import sync_seed_assets
-
-__all__ = ["sync_seed_assets", "register_assets_system"]

app/assets/_helpers.py

@@ -1,225 +0,0 @@
-import contextlib
-import os
-import uuid
-from datetime import datetime, timezone
-from pathlib import Path
-from typing import Literal, Optional, Sequence
-
-import folder_paths
-
-from .api import schemas_in
-
-
-def get_comfy_models_folders() -> list[tuple[str, list[str]]]:
-    """Build a list of (folder_name, base_paths[]) categories that are configured for model locations.
-
-    We trust `folder_paths.folder_names_and_paths` and include a category if
-    *any* of its base paths lies under the Comfy `models_dir`.
-    """
-    targets: list[tuple[str, list[str]]] = []
-    models_root = os.path.abspath(folder_paths.models_dir)
-    for name, (paths, _exts) in folder_paths.folder_names_and_paths.items():
-        if any(os.path.abspath(p).startswith(models_root + os.sep) for p in paths):
-            targets.append((name, paths))
-    return targets
-
-
-def get_relative_to_root_category_path_of_asset(file_path: str) -> tuple[Literal["input", "output", "models"], str]:
-    """Given an absolute or relative file path, determine which root category the path belongs to:
-      - 'input' if the file resides under `folder_paths.get_input_directory()`
-      - 'output' if the file resides under `folder_paths.get_output_directory()`
-      - 'models' if the file resides under any base path of categories returned by `get_comfy_models_folders()`
-
-    Returns:
-        (root_category, relative_path_inside_that_root)
-        For 'models', the relative path is prefixed with the category name:
-            e.g. ('models', 'vae/test/sub/ae.safetensors')
-
-    Raises:
-        ValueError: if the path does not belong to input, output, or configured model bases.
-    """
-    fp_abs = os.path.abspath(file_path)
-
-    def _is_within(child: str, parent: str) -> bool:
-        try:
-            return os.path.commonpath([child, parent]) == parent
-        except Exception:
-            return False
-
-    def _rel(child: str, parent: str) -> str:
-        return os.path.relpath(os.path.join(os.sep, os.path.relpath(child, parent)), os.sep)
-
-    # 1) input
-    input_base = os.path.abspath(folder_paths.get_input_directory())
-    if _is_within(fp_abs, input_base):
-        return "input", _rel(fp_abs, input_base)
-
-    # 2) output
-    output_base = os.path.abspath(folder_paths.get_output_directory())
-    if _is_within(fp_abs, output_base):
-        return "output", _rel(fp_abs, output_base)
-
-    # 3) models (check deepest matching base to avoid ambiguity)
-    best: Optional[tuple[int, str, str]] = None  # (base_len, bucket, rel_inside_bucket)
-    for bucket, bases in get_comfy_models_folders():
-        for b in bases:
-            base_abs = os.path.abspath(b)
-            if not _is_within(fp_abs, base_abs):
-                continue
-            cand = (len(base_abs), bucket, _rel(fp_abs, base_abs))
-            if best is None or cand[0] > best[0]:
-                best = cand
-
-    if best is not None:
-        _, bucket, rel_inside = best
-        combined = os.path.join(bucket, rel_inside)
-        return "models", os.path.relpath(os.path.join(os.sep, combined), os.sep)
-
-    raise ValueError(f"Path is not within input, output, or configured model bases: {file_path}")
-
-
-def get_name_and_tags_from_asset_path(file_path: str) -> tuple[str, list[str]]:
-    """Return a tuple (name, tags) derived from a filesystem path.
-
-    Semantics:
-      - Root category is determined by `get_relative_to_root_category_path_of_asset`.
-      - The returned `name` is the base filename with extension from the relative path.
-      - The returned `tags` are:
-            [root_category] + parent folders of the relative path (in order)
-        For 'models', this means:
-            file '/.../ModelsDir/vae/test_tag/ae.safetensors'
-            -> root_category='models', some_path='vae/test_tag/ae.safetensors'
-            -> name='ae.safetensors', tags=['models', 'vae', 'test_tag']
-
-    Raises:
-        ValueError: if the path does not belong to input, output, or configured model bases.
-    """
-    root_category, some_path = get_relative_to_root_category_path_of_asset(file_path)
-    p = Path(some_path)
-    parent_parts = [part for part in p.parent.parts if part not in (".", "..", p.anchor)]
-    return p.name, list(dict.fromkeys(normalize_tags([root_category, *parent_parts])))
-
-
-def normalize_tags(tags: Optional[Sequence[str]]) -> list[str]:
-    return [t.strip().lower() for t in (tags or []) if (t or "").strip()]
-
-
-def resolve_destination_from_tags(tags: list[str]) -> tuple[str, list[str]]:
-    """Validates and maps tags -> (base_dir, subdirs_for_fs)"""
-    root = tags[0]
-    if root == "models":
-        if len(tags) < 2:
-            raise ValueError("at least two tags required for model asset")
-        try:
-            bases = folder_paths.folder_names_and_paths[tags[1]][0]
-        except KeyError:
-            raise ValueError(f"unknown model category '{tags[1]}'")
-        if not bases:
-            raise ValueError(f"no base path configured for category '{tags[1]}'")
-        base_dir = os.path.abspath(bases[0])
-        raw_subdirs = tags[2:]
-    else:
-        base_dir = os.path.abspath(
-            folder_paths.get_input_directory() if root == "input" else folder_paths.get_output_directory()
-        )
-        raw_subdirs = tags[1:]
-    for i in raw_subdirs:
-        if i in (".", ".."):
-            raise ValueError("invalid path component in tags")
-
-    return base_dir, raw_subdirs if raw_subdirs else []
-
-
-def ensure_within_base(candidate: str, base: str) -> None:
-    cand_abs = os.path.abspath(candidate)
-    base_abs = os.path.abspath(base)
-    try:
-        if os.path.commonpath([cand_abs, base_abs]) != base_abs:
-            raise ValueError("destination escapes base directory")
-    except Exception:
-        raise ValueError("invalid destination path")
-
-
-def compute_relative_filename(file_path: str) -> Optional[str]:
-    """
-    Return the model's path relative to the last well-known folder (the model category),
-    using forward slashes, eg:
-      /.../models/checkpoints/flux/123/flux.safetensors -> "flux/123/flux.safetensors"
-      /.../models/text_encoders/clip_g.safetensors -> "clip_g.safetensors"
-
-    For non-model paths, returns None.
-    NOTE: this is a temporary helper, used only for initializing metadata["filename"] field.
-    """
-    try:
-        root_category, rel_path = get_relative_to_root_category_path_of_asset(file_path)
-    except ValueError:
-        return None
-
-    p = Path(rel_path)
-    parts = [seg for seg in p.parts if seg not in (".", "..", p.anchor)]
-    if not parts:
-        return None
-
-    if root_category == "models":
-        # parts[0] is the category ("checkpoints", "vae", etc) – drop it
-        inside = parts[1:] if len(parts) > 1 else [parts[0]]
-        return "/".join(inside)
-    return "/".join(parts)  # input/output: keep all parts
-
-
-def list_tree(base_dir: str) -> list[str]:
-    out: list[str] = []
-    base_abs = os.path.abspath(base_dir)
-    if not os.path.isdir(base_abs):
-        return out
-    for dirpath, _subdirs, filenames in os.walk(base_abs, topdown=True, followlinks=False):
-        for name in filenames:
-            out.append(os.path.abspath(os.path.join(dirpath, name)))
-    return out
-
-
-def prefixes_for_root(root: schemas_in.RootType) -> list[str]:
-    if root == "models":
-        bases: list[str] = []
-        for _bucket, paths in get_comfy_models_folders():
-            bases.extend(paths)
-        return [os.path.abspath(p) for p in bases]
-    if root == "input":
-        return [os.path.abspath(folder_paths.get_input_directory())]
-    if root == "output":
-        return [os.path.abspath(folder_paths.get_output_directory())]
-    return []
-
-
-def ts_to_iso(ts: Optional[float]) -> Optional[str]:
-    if ts is None:
-        return None
-    try:
-        return datetime.fromtimestamp(float(ts), tz=timezone.utc).replace(tzinfo=None).isoformat()
-    except Exception:
-        return None
-
-
-def new_scan_id(root: schemas_in.RootType) -> str:
-    return f"scan-{root}-{uuid.uuid4().hex[:8]}"
-
-
-def collect_models_files() -> list[str]:
-    out: list[str] = []
-    for folder_name, bases in get_comfy_models_folders():
-        rel_files = folder_paths.get_filename_list(folder_name) or []
-        for rel_path in rel_files:
-            abs_path = folder_paths.get_full_path(folder_name, rel_path)
-            if not abs_path:
-                continue
-            abs_path = os.path.abspath(abs_path)
-            allowed = False
-            for b in bases:
-                base_abs = os.path.abspath(b)
-                with contextlib.suppress(Exception):
-                    if os.path.commonpath([abs_path, base_abs]) == base_abs:
-                        allowed = True
-                        break
-            if allowed:
-                out.append(abs_path)
-    return out

app/assets/api/routes.py

@@ -1,544 +0,0 @@
-import contextlib
-import logging
-import os
-import urllib.parse
-import uuid
-from typing import Optional
-
-from aiohttp import web
-from pydantic import ValidationError
-
-import folder_paths
-
-from ... import user_manager
-from .. import manager, scanner
-from . import schemas_in, schemas_out
-
-ROUTES = web.RouteTableDef()
-USER_MANAGER: Optional[user_manager.UserManager] = None
-LOGGER = logging.getLogger(__name__)
-
-# UUID regex (canonical hyphenated form, case-insensitive)
-UUID_RE = r"[0-9a-fA-F]{8}-[0-9a-fA-F]{4}-[0-9a-fA-F]{4}-[0-9a-fA-F]{4}-[0-9a-fA-F]{12}"
-
-
-@ROUTES.head("/api/assets/hash/{hash}")
-async def head_asset_by_hash(request: web.Request) -> web.Response:
-    hash_str = request.match_info.get("hash", "").strip().lower()
-    if not hash_str or ":" not in hash_str:
-        return _error_response(400, "INVALID_HASH", "hash must be like 'blake3:<hex>'")
-    algo, digest = hash_str.split(":", 1)
-    if algo != "blake3" or not digest or any(c for c in digest if c not in "0123456789abcdef"):
-        return _error_response(400, "INVALID_HASH", "hash must be like 'blake3:<hex>'")
-    exists = await manager.asset_exists(asset_hash=hash_str)
-    return web.Response(status=200 if exists else 404)
-
-
-@ROUTES.get("/api/assets")
-async def list_assets(request: web.Request) -> web.Response:
-    qp = request.rel_url.query
-    query_dict = {}
-    if "include_tags" in qp:
-        query_dict["include_tags"] = qp.getall("include_tags")
-    if "exclude_tags" in qp:
-        query_dict["exclude_tags"] = qp.getall("exclude_tags")
-    for k in ("name_contains", "metadata_filter", "limit", "offset", "sort", "order"):
-        v = qp.get(k)
-        if v is not None:
-            query_dict[k] = v
-
-    try:
-        q = schemas_in.ListAssetsQuery.model_validate(query_dict)
-    except ValidationError as ve:
-        return _validation_error_response("INVALID_QUERY", ve)
-
-    payload = await manager.list_assets(
-        include_tags=q.include_tags,
-        exclude_tags=q.exclude_tags,
-        name_contains=q.name_contains,
-        metadata_filter=q.metadata_filter,
-        limit=q.limit,
-        offset=q.offset,
-        sort=q.sort,
-        order=q.order,
-        owner_id=USER_MANAGER.get_request_user_id(request),
-    )
-    return web.json_response(payload.model_dump(mode="json"))
-
-
-@ROUTES.get(f"/api/assets/{{id:{UUID_RE}}}/content")
-async def download_asset_content(request: web.Request) -> web.Response:
-    disposition = request.query.get("disposition", "attachment").lower().strip()
-    if disposition not in {"inline", "attachment"}:
-        disposition = "attachment"
-
-    try:
-        abs_path, content_type, filename = await manager.resolve_asset_content_for_download(
-            asset_info_id=str(uuid.UUID(request.match_info["id"])),
-            owner_id=USER_MANAGER.get_request_user_id(request),
-        )
-    except ValueError as ve:
-        return _error_response(404, "ASSET_NOT_FOUND", str(ve))
-    except NotImplementedError as nie:
-        return _error_response(501, "BACKEND_UNSUPPORTED", str(nie))
-    except FileNotFoundError:
-        return _error_response(404, "FILE_NOT_FOUND", "Underlying file not found on disk.")
-
-    quoted = (filename or "").replace("\r", "").replace("\n", "").replace('"', "'")
-    cd = f'{disposition}; filename="{quoted}"; filename*=UTF-8\'\'{urllib.parse.quote(filename)}'
-
-    resp = web.FileResponse(abs_path)
-    resp.content_type = content_type
-    resp.headers["Content-Disposition"] = cd
-    return resp
-
-
-@ROUTES.post("/api/assets/from-hash")
-async def create_asset_from_hash(request: web.Request) -> web.Response:
-    try:
-        payload = await request.json()
-        body = schemas_in.CreateFromHashBody.model_validate(payload)
-    except ValidationError as ve:
-        return _validation_error_response("INVALID_BODY", ve)
-    except Exception:
-        return _error_response(400, "INVALID_JSON", "Request body must be valid JSON.")
-
-    result = await manager.create_asset_from_hash(
-        hash_str=body.hash,
-        name=body.name,
-        tags=body.tags,
-        user_metadata=body.user_metadata,
-        owner_id=USER_MANAGER.get_request_user_id(request),
-    )
-    if result is None:
-        return _error_response(404, "ASSET_NOT_FOUND", f"Asset content {body.hash} does not exist")
-    return web.json_response(result.model_dump(mode="json"), status=201)
-
-
-@ROUTES.post("/api/assets")
-async def upload_asset(request: web.Request) -> web.Response:
-    """Multipart/form-data endpoint for Asset uploads."""
-
-    if not (request.content_type or "").lower().startswith("multipart/"):
-        return _error_response(415, "UNSUPPORTED_MEDIA_TYPE", "Use multipart/form-data for uploads.")
-
-    reader = await request.multipart()
-
-    file_present = False
-    file_client_name: Optional[str] = None
-    tags_raw: list[str] = []
-    provided_name: Optional[str] = None
-    user_metadata_raw: Optional[str] = None
-    provided_hash: Optional[str] = None
-    provided_hash_exists: Optional[bool] = None
-
-    file_written = 0
-    tmp_path: Optional[str] = None
-    while True:
-        field = await reader.next()
-        if field is None:
-            break
-
-        fname = getattr(field, "name", "") or ""
-
-        if fname == "hash":
-            try:
-                s = ((await field.text()) or "").strip().lower()
-            except Exception:
-                return _error_response(400, "INVALID_HASH", "hash must be like 'blake3:<hex>'")
-
-            if s:
-                if ":" not in s:
-                    return _error_response(400, "INVALID_HASH", "hash must be like 'blake3:<hex>'")
-                algo, digest = s.split(":", 1)
-                if algo != "blake3" or not digest or any(c for c in digest if c not in "0123456789abcdef"):
-                    return _error_response(400, "INVALID_HASH", "hash must be like 'blake3:<hex>'")
-                provided_hash = f"{algo}:{digest}"
-                try:
-                    provided_hash_exists = await manager.asset_exists(asset_hash=provided_hash)
-                except Exception:
-                    provided_hash_exists = None  # do not fail the whole request here
-
-        elif fname == "file":
-            file_present = True
-            file_client_name = (field.filename or "").strip()
-
-            if provided_hash and provided_hash_exists is True:
-                # If client supplied a hash that we know exists, drain but do not write to disk
-                try:
-                    while True:
-                        chunk = await field.read_chunk(8 * 1024 * 1024)
-                        if not chunk:
-                            break
-                        file_written += len(chunk)
-                except Exception:
-                    return _error_response(500, "UPLOAD_IO_ERROR", "Failed to receive uploaded file.")
-                continue  # Do not create temp file; we will create AssetInfo from the existing content
-
-            # Otherwise, store to temp for hashing/ingest
-            uploads_root = os.path.join(folder_paths.get_temp_directory(), "uploads")
-            unique_dir = os.path.join(uploads_root, uuid.uuid4().hex)
-            os.makedirs(unique_dir, exist_ok=True)
-            tmp_path = os.path.join(unique_dir, ".upload.part")
-
-            try:
-                with open(tmp_path, "wb") as f:
-                    while True:
-                        chunk = await field.read_chunk(8 * 1024 * 1024)
-                        if not chunk:
-                            break
-                        f.write(chunk)
-                        file_written += len(chunk)
-            except Exception:
-                try:
-                    if os.path.exists(tmp_path or ""):
-                        os.remove(tmp_path)
-                finally:
-                    return _error_response(500, "UPLOAD_IO_ERROR", "Failed to receive and store uploaded file.")
-        elif fname == "tags":
-            tags_raw.append((await field.text()) or "")
-        elif fname == "name":
-            provided_name = (await field.text()) or None
-        elif fname == "user_metadata":
-            user_metadata_raw = (await field.text()) or None
-
-    # If client did not send file, and we are not doing a from-hash fast path -> error
-    if not file_present and not (provided_hash and provided_hash_exists):
-        return _error_response(400, "MISSING_FILE", "Form must include a 'file' part or a known 'hash'.")
-
-    if file_present and file_written == 0 and not (provided_hash and provided_hash_exists):
-        # Empty upload is only acceptable if we are fast-pathing from existing hash
-        try:
-            if tmp_path and os.path.exists(tmp_path):
-                os.remove(tmp_path)
-        finally:
-            return _error_response(400, "EMPTY_UPLOAD", "Uploaded file is empty.")
-
-    try:
-        spec = schemas_in.UploadAssetSpec.model_validate({
-            "tags": tags_raw,
-            "name": provided_name,
-            "user_metadata": user_metadata_raw,
-            "hash": provided_hash,
-        })
-    except ValidationError as ve:
-        try:
-            if tmp_path and os.path.exists(tmp_path):
-                os.remove(tmp_path)
-        finally:
-            return _validation_error_response("INVALID_BODY", ve)
-
-    # Validate models category against configured folders (consistent with previous behavior)
-    if spec.tags and spec.tags[0] == "models":
-        if len(spec.tags) < 2 or spec.tags[1] not in folder_paths.folder_names_and_paths:
-            if tmp_path and os.path.exists(tmp_path):
-                os.remove(tmp_path)
-            return _error_response(
-                400, "INVALID_BODY", f"unknown models category '{spec.tags[1] if len(spec.tags) >= 2 else ''}'"
-            )
-
-    owner_id = USER_MANAGER.get_request_user_id(request)
-
-    # Fast path: if a valid provided hash exists, create AssetInfo without writing anything
-    if spec.hash and provided_hash_exists is True:
-        try:
-            result = await manager.create_asset_from_hash(
-                hash_str=spec.hash,
-                name=spec.name or (spec.hash.split(":", 1)[1]),
-                tags=spec.tags,
-                user_metadata=spec.user_metadata or {},
-                owner_id=owner_id,
-            )
-        except Exception:
-            LOGGER.exception("create_asset_from_hash failed for hash=%s, owner_id=%s", spec.hash, owner_id)
-            return _error_response(500, "INTERNAL", "Unexpected server error.")
-
-        if result is None:
-            return _error_response(404, "ASSET_NOT_FOUND", f"Asset content {spec.hash} does not exist")
-
-        # Drain temp if we accidentally saved (e.g., hash field came after file)
-        if tmp_path and os.path.exists(tmp_path):
-            with contextlib.suppress(Exception):
-                os.remove(tmp_path)
-
-        status = 200 if (not result.created_new) else 201
-        return web.json_response(result.model_dump(mode="json"), status=status)
-
-    # Otherwise, we must have a temp file path to ingest
-    if not tmp_path or not os.path.exists(tmp_path):
-        # The only case we reach here without a temp file is: client sent a hash that does not exist and no file
-        return _error_response(404, "ASSET_NOT_FOUND", "Provided hash not found and no file uploaded.")
-
-    try:
-        created = await manager.upload_asset_from_temp_path(
-            spec,
-            temp_path=tmp_path,
-            client_filename=file_client_name,
-            owner_id=owner_id,
-            expected_asset_hash=spec.hash,
-        )
-        status = 201 if created.created_new else 200
-        return web.json_response(created.model_dump(mode="json"), status=status)
-    except ValueError as e:
-        if tmp_path and os.path.exists(tmp_path):
-            os.remove(tmp_path)
-        msg = str(e)
-        if "HASH_MISMATCH" in msg or msg.strip().upper() == "HASH_MISMATCH":
-            return _error_response(
-                400,
-                "HASH_MISMATCH",
-                "Uploaded file hash does not match provided hash.",
-            )
-        return _error_response(400, "BAD_REQUEST", "Invalid inputs.")
-    except Exception:
-        if tmp_path and os.path.exists(tmp_path):
-            os.remove(tmp_path)
-        LOGGER.exception("upload_asset_from_temp_path failed for tmp_path=%s, owner_id=%s", tmp_path, owner_id)
-        return _error_response(500, "INTERNAL", "Unexpected server error.")
-
-
-@ROUTES.get(f"/api/assets/{{id:{UUID_RE}}}")
-async def get_asset(request: web.Request) -> web.Response:
-    asset_info_id = str(uuid.UUID(request.match_info["id"]))
-    try:
-        result = await manager.get_asset(
-            asset_info_id=asset_info_id,
-            owner_id=USER_MANAGER.get_request_user_id(request),
-        )
-    except ValueError as ve:
-        return _error_response(404, "ASSET_NOT_FOUND", str(ve), {"id": asset_info_id})
-    except Exception:
-        LOGGER.exception(
-            "get_asset failed for asset_info_id=%s, owner_id=%s",
-            asset_info_id,
-            USER_MANAGER.get_request_user_id(request),
-        )
-        return _error_response(500, "INTERNAL", "Unexpected server error.")
-    return web.json_response(result.model_dump(mode="json"), status=200)
-
-
-@ROUTES.put(f"/api/assets/{{id:{UUID_RE}}}")
-async def update_asset(request: web.Request) -> web.Response:
-    asset_info_id = str(uuid.UUID(request.match_info["id"]))
-    try:
-        body = schemas_in.UpdateAssetBody.model_validate(await request.json())
-    except ValidationError as ve:
-        return _validation_error_response("INVALID_BODY", ve)
-    except Exception:
-        return _error_response(400, "INVALID_JSON", "Request body must be valid JSON.")
-
-    try:
-        result = await manager.update_asset(
-            asset_info_id=asset_info_id,
-            name=body.name,
-            tags=body.tags,
-            user_metadata=body.user_metadata,
-            owner_id=USER_MANAGER.get_request_user_id(request),
-        )
-    except (ValueError, PermissionError) as ve:
-        return _error_response(404, "ASSET_NOT_FOUND", str(ve), {"id": asset_info_id})
-    except Exception:
-        LOGGER.exception(
-            "update_asset failed for asset_info_id=%s, owner_id=%s",
-            asset_info_id,
-            USER_MANAGER.get_request_user_id(request),
-        )
-        return _error_response(500, "INTERNAL", "Unexpected server error.")
-    return web.json_response(result.model_dump(mode="json"), status=200)
-
-
-@ROUTES.put(f"/api/assets/{{id:{UUID_RE}}}/preview")
-async def set_asset_preview(request: web.Request) -> web.Response:
-    asset_info_id = str(uuid.UUID(request.match_info["id"]))
-    try:
-        body = schemas_in.SetPreviewBody.model_validate(await request.json())
-    except ValidationError as ve:
-        return _validation_error_response("INVALID_BODY", ve)
-    except Exception:
-        return _error_response(400, "INVALID_JSON", "Request body must be valid JSON.")
-
-    try:
-        result = await manager.set_asset_preview(
-            asset_info_id=asset_info_id,
-            preview_asset_id=body.preview_id,
-            owner_id=USER_MANAGER.get_request_user_id(request),
-        )
-    except (PermissionError, ValueError) as ve:
-        return _error_response(404, "ASSET_NOT_FOUND", str(ve), {"id": asset_info_id})
-    except Exception:
-        LOGGER.exception(
-            "set_asset_preview failed for asset_info_id=%s, owner_id=%s",
-            asset_info_id,
-            USER_MANAGER.get_request_user_id(request),
-        )
-        return _error_response(500, "INTERNAL", "Unexpected server error.")
-    return web.json_response(result.model_dump(mode="json"), status=200)
-
-
-@ROUTES.delete(f"/api/assets/{{id:{UUID_RE}}}")
-async def delete_asset(request: web.Request) -> web.Response:
-    asset_info_id = str(uuid.UUID(request.match_info["id"]))
-    delete_content = request.query.get("delete_content")
-    delete_content = True if delete_content is None else delete_content.lower() not in {"0", "false", "no"}
-
-    try:
-        deleted = await manager.delete_asset_reference(
-            asset_info_id=asset_info_id,
-            owner_id=USER_MANAGER.get_request_user_id(request),
-            delete_content_if_orphan=delete_content,
-        )
-    except Exception:
-        LOGGER.exception(
-            "delete_asset_reference failed for asset_info_id=%s, owner_id=%s",
-            asset_info_id,
-            USER_MANAGER.get_request_user_id(request),
-        )
-        return _error_response(500, "INTERNAL", "Unexpected server error.")
-
-    if not deleted:
-        return _error_response(404, "ASSET_NOT_FOUND", f"AssetInfo {asset_info_id} not found.")
-    return web.Response(status=204)
-
-
-@ROUTES.get("/api/tags")
-async def get_tags(request: web.Request) -> web.Response:
-    query_map = dict(request.rel_url.query)
-
-    try:
-        query = schemas_in.TagsListQuery.model_validate(query_map)
-    except ValidationError as ve:
-        return web.json_response(
-            {"error": {"code": "INVALID_QUERY", "message": "Invalid query parameters", "details": ve.errors()}},
-            status=400,
-        )
-
-    result = await manager.list_tags(
-        prefix=query.prefix,
-        limit=query.limit,
-        offset=query.offset,
-        order=query.order,
-        include_zero=query.include_zero,
-        owner_id=USER_MANAGER.get_request_user_id(request),
-    )
-    return web.json_response(result.model_dump(mode="json"))
-
-
-@ROUTES.post(f"/api/assets/{{id:{UUID_RE}}}/tags")
-async def add_asset_tags(request: web.Request) -> web.Response:
-    asset_info_id = str(uuid.UUID(request.match_info["id"]))
-    try:
-        payload = await request.json()
-        data = schemas_in.TagsAdd.model_validate(payload)
-    except ValidationError as ve:
-        return _error_response(400, "INVALID_BODY", "Invalid JSON body for tags add.", {"errors": ve.errors()})
-    except Exception:
-        return _error_response(400, "INVALID_JSON", "Request body must be valid JSON.")
-
-    try:
-        result = await manager.add_tags_to_asset(
-            asset_info_id=asset_info_id,
-            tags=data.tags,
-            origin="manual",
-            owner_id=USER_MANAGER.get_request_user_id(request),
-        )
-    except (ValueError, PermissionError) as ve:
-        return _error_response(404, "ASSET_NOT_FOUND", str(ve), {"id": asset_info_id})
-    except Exception:
-        LOGGER.exception(
-            "add_tags_to_asset failed for asset_info_id=%s, owner_id=%s",
-            asset_info_id,
-            USER_MANAGER.get_request_user_id(request),
-        )
-        return _error_response(500, "INTERNAL", "Unexpected server error.")
-
-    return web.json_response(result.model_dump(mode="json"), status=200)
-
-
-@ROUTES.delete(f"/api/assets/{{id:{UUID_RE}}}/tags")
-async def delete_asset_tags(request: web.Request) -> web.Response:
-    asset_info_id = str(uuid.UUID(request.match_info["id"]))
-    try:
-        payload = await request.json()
-        data = schemas_in.TagsRemove.model_validate(payload)
-    except ValidationError as ve:
-        return _error_response(400, "INVALID_BODY", "Invalid JSON body for tags remove.", {"errors": ve.errors()})
-    except Exception:
-        return _error_response(400, "INVALID_JSON", "Request body must be valid JSON.")
-
-    try:
-        result = await manager.remove_tags_from_asset(
-            asset_info_id=asset_info_id,
-            tags=data.tags,
-            owner_id=USER_MANAGER.get_request_user_id(request),
-        )
-    except ValueError as ve:
-        return _error_response(404, "ASSET_NOT_FOUND", str(ve), {"id": asset_info_id})
-    except Exception:
-        LOGGER.exception(
-            "remove_tags_from_asset failed for asset_info_id=%s, owner_id=%s",
-            asset_info_id,
-            USER_MANAGER.get_request_user_id(request),
-        )
-        return _error_response(500, "INTERNAL", "Unexpected server error.")
-
-    return web.json_response(result.model_dump(mode="json"), status=200)
-
-
-@ROUTES.post("/api/assets/scan/seed")
-async def seed_assets(request: web.Request) -> web.Response:
-    try:
-        payload = await request.json()
-    except Exception:
-        payload = {}
-
-    try:
-        body = schemas_in.ScheduleAssetScanBody.model_validate(payload)
-    except ValidationError as ve:
-        return _validation_error_response("INVALID_BODY", ve)
-
-    try:
-        await scanner.sync_seed_assets(body.roots)
-    except Exception:
-        LOGGER.exception("sync_seed_assets failed for roots=%s", body.roots)
-        return _error_response(500, "INTERNAL", "Unexpected server error.")
-    return web.json_response({"synced": True, "roots": body.roots}, status=200)
-
-
-@ROUTES.post("/api/assets/scan/schedule")
-async def schedule_asset_scan(request: web.Request) -> web.Response:
-    try:
-        payload = await request.json()
-    except Exception:
-        payload = {}
-
-    try:
-        body = schemas_in.ScheduleAssetScanBody.model_validate(payload)
-    except ValidationError as ve:
-        return _validation_error_response("INVALID_BODY", ve)
-
-    states = await scanner.schedule_scans(body.roots)
-    return web.json_response(states.model_dump(mode="json"), status=202)
-
-
-@ROUTES.get("/api/assets/scan")
-async def get_asset_scan_status(request: web.Request) -> web.Response:
-    root = request.query.get("root", "").strip().lower()
-    states = scanner.current_statuses()
-    if root in {"models", "input", "output"}:
-        states = [s for s in states.scans if s.root == root]  # type: ignore
-        states = schemas_out.AssetScanStatusResponse(scans=states)
-    return web.json_response(states.model_dump(mode="json"), status=200)
-
-
-def register_assets_system(app: web.Application, user_manager_instance: user_manager.UserManager) -> None:
-    global USER_MANAGER
-    USER_MANAGER = user_manager_instance
-    app.add_routes(ROUTES)
-
-
-def _error_response(status: int, code: str, message: str, details: Optional[dict] = None) -> web.Response:
-    return web.json_response({"error": {"code": code, "message": message, "details": details or {}}}, status=status)
-
-
-def _validation_error_response(code: str, ve: ValidationError) -> web.Response:
-    return _error_response(400, code, "Validation failed.", {"errors": ve.json()})

app/assets/api/schemas_in.py

@@ -1,297 +0,0 @@
-import json
-import uuid
-from typing import Any, Literal, Optional
-
-from pydantic import (
-    BaseModel,
-    ConfigDict,
-    Field,
-    conint,
-    field_validator,
-    model_validator,
-)
-
-
-class ListAssetsQuery(BaseModel):
-    include_tags: list[str] = Field(default_factory=list)
-    exclude_tags: list[str] = Field(default_factory=list)
-    name_contains: Optional[str] = None
-
-    # Accept either a JSON string (query param) or a dict
-    metadata_filter: Optional[dict[str, Any]] = None
-
-    limit: conint(ge=1, le=500) = 20
-    offset: conint(ge=0) = 0
-
-    sort: Literal["name", "created_at", "updated_at", "size", "last_access_time"] = "created_at"
-    order: Literal["asc", "desc"] = "desc"
-
-    @field_validator("include_tags", "exclude_tags", mode="before")
-    @classmethod
-    def _split_csv_tags(cls, v):
-        # Accept "a,b,c" or ["a","b"] (we are liberal in what we accept)
-        if v is None:
-            return []
-        if isinstance(v, str):
-            return [t.strip() for t in v.split(",") if t.strip()]
-        if isinstance(v, list):
-            out: list[str] = []
-            for item in v:
-                if isinstance(item, str):
-                    out.extend([t.strip() for t in item.split(",") if t.strip()])
-            return out
-        return v
-
-    @field_validator("metadata_filter", mode="before")
-    @classmethod
-    def _parse_metadata_json(cls, v):
-        if v is None or isinstance(v, dict):
-            return v
-        if isinstance(v, str) and v.strip():
-            try:
-                parsed = json.loads(v)
-            except Exception as e:
-                raise ValueError(f"metadata_filter must be JSON: {e}") from e
-            if not isinstance(parsed, dict):
-                raise ValueError("metadata_filter must be a JSON object")
-            return parsed
-        return None
-
-
-class UpdateAssetBody(BaseModel):
-    name: Optional[str] = None
-    tags: Optional[list[str]] = None
-    user_metadata: Optional[dict[str, Any]] = None
-
-    @model_validator(mode="after")
-    def _at_least_one(self):
-        if self.name is None and self.tags is None and self.user_metadata is None:
-            raise ValueError("Provide at least one of: name, tags, user_metadata.")
-        if self.tags is not None:
-            if not isinstance(self.tags, list) or not all(isinstance(t, str) for t in self.tags):
-                raise ValueError("Field 'tags' must be an array of strings.")
-        return self
-
-
-class CreateFromHashBody(BaseModel):
-    model_config = ConfigDict(extra="ignore", str_strip_whitespace=True)
-
-    hash: str
-    name: str
-    tags: list[str] = Field(default_factory=list)
-    user_metadata: dict[str, Any] = Field(default_factory=dict)
-
-    @field_validator("hash")
-    @classmethod
-    def _require_blake3(cls, v):
-        s = (v or "").strip().lower()
-        if ":" not in s:
-            raise ValueError("hash must be 'blake3:<hex>'")
-        algo, digest = s.split(":", 1)
-        if algo != "blake3":
-            raise ValueError("only canonical 'blake3:<hex>' is accepted here")
-        if not digest or any(c for c in digest if c not in "0123456789abcdef"):
-            raise ValueError("hash digest must be lowercase hex")
-        return s
-
-    @field_validator("tags", mode="before")
-    @classmethod
-    def _tags_norm(cls, v):
-        if v is None:
-            return []
-        if isinstance(v, list):
-            out = [str(t).strip().lower() for t in v if str(t).strip()]
-            seen = set()
-            dedup = []
-            for t in out:
-                if t not in seen:
-                    seen.add(t)
-                    dedup.append(t)
-            return dedup
-        if isinstance(v, str):
-            return [t.strip().lower() for t in v.split(",") if t.strip()]
-        return []
-
-
-class TagsListQuery(BaseModel):
-    model_config = ConfigDict(extra="ignore", str_strip_whitespace=True)
-
-    prefix: Optional[str] = Field(None, min_length=1, max_length=256)
-    limit: int = Field(100, ge=1, le=1000)
-    offset: int = Field(0, ge=0, le=10_000_000)
-    order: Literal["count_desc", "name_asc"] = "count_desc"
-    include_zero: bool = True
-
-    @field_validator("prefix")
-    @classmethod
-    def normalize_prefix(cls, v: Optional[str]) -> Optional[str]:
-        if v is None:
-            return v
-        v = v.strip()
-        return v.lower() or None
-
-
-class TagsAdd(BaseModel):
-    model_config = ConfigDict(extra="ignore")
-    tags: list[str] = Field(..., min_length=1)
-
-    @field_validator("tags")
-    @classmethod
-    def normalize_tags(cls, v: list[str]) -> list[str]:
-        out = []
-        for t in v:
-            if not isinstance(t, str):
-                raise TypeError("tags must be strings")
-            tnorm = t.strip().lower()
-            if tnorm:
-                out.append(tnorm)
-        seen = set()
-        deduplicated = []
-        for x in out:
-            if x not in seen:
-                seen.add(x)
-                deduplicated.append(x)
-        return deduplicated
-
-
-class TagsRemove(TagsAdd):
-    pass
-
-
-RootType = Literal["models", "input", "output"]
-ALLOWED_ROOTS: tuple[RootType, ...] = ("models", "input", "output")
-
-
-class ScheduleAssetScanBody(BaseModel):
-    roots: list[RootType] = Field(..., min_length=1)
-
-
-class UploadAssetSpec(BaseModel):
-    """Upload Asset operation.
-    - tags: ordered; first is root ('models'|'input'|'output');
-            if root == 'models', second must be a valid category from folder_paths.folder_names_and_paths
-    - name: display name
-    - user_metadata: arbitrary JSON object (optional)
-    - hash: optional canonical 'blake3:<hex>' provided by the client for validation / fast-path
-
-    Files created via this endpoint are stored on disk using the **content hash** as the filename stem
-    and the original extension is preserved when available.
-    """
-    model_config = ConfigDict(extra="ignore", str_strip_whitespace=True)
-
-    tags: list[str] = Field(..., min_length=1)
-    name: Optional[str] = Field(default=None, max_length=512, description="Display Name")
-    user_metadata: dict[str, Any] = Field(default_factory=dict)
-    hash: Optional[str] = Field(default=None)
-
-    @field_validator("hash", mode="before")
-    @classmethod
-    def _parse_hash(cls, v):
-        if v is None:
-            return None
-        s = str(v).strip().lower()
-        if not s:
-            return None
-        if ":" not in s:
-            raise ValueError("hash must be 'blake3:<hex>'")
-        algo, digest = s.split(":", 1)
-        if algo != "blake3":
-            raise ValueError("only canonical 'blake3:<hex>' is accepted here")
-        if not digest or any(c for c in digest if c not in "0123456789abcdef"):
-            raise ValueError("hash digest must be lowercase hex")
-        return f"{algo}:{digest}"
-
-    @field_validator("tags", mode="before")
-    @classmethod
-    def _parse_tags(cls, v):
-        """
-        Accepts a list of strings (possibly multiple form fields),
-        where each string can be:
-          - JSON array (e.g., '["models","loras","foo"]')
-          - comma-separated ('models, loras, foo')
-          - single token ('models')
-        Returns a normalized, deduplicated, ordered list.
-        """
-        items: list[str] = []
-        if v is None:
-            return []
-        if isinstance(v, str):
-            v = [v]
-
-        if isinstance(v, list):
-            for item in v:
-                if item is None:
-                    continue
-                s = str(item).strip()
-                if not s:
-                    continue
-                if s.startswith("["):
-                    try:
-                        arr = json.loads(s)
-                        if isinstance(arr, list):
-                            items.extend(str(x) for x in arr)
-                            continue
-                    except Exception:
-                        pass  # fallback to CSV parse below
-                items.extend([p for p in s.split(",") if p.strip()])
-        else:
-            return []
-
-        # normalize + dedupe
-        norm = []
-        seen = set()
-        for t in items:
-            tnorm = str(t).strip().lower()
-            if tnorm and tnorm not in seen:
-                seen.add(tnorm)
-                norm.append(tnorm)
-        return norm
-
-    @field_validator("user_metadata", mode="before")
-    @classmethod
-    def _parse_metadata_json(cls, v):
-        if v is None or isinstance(v, dict):
-            return v or {}
-        if isinstance(v, str):
-            s = v.strip()
-            if not s:
-                return {}
-            try:
-                parsed = json.loads(s)
-            except Exception as e:
-                raise ValueError(f"user_metadata must be JSON: {e}") from e
-            if not isinstance(parsed, dict):
-                raise ValueError("user_metadata must be a JSON object")
-            return parsed
-        return {}
-
-    @model_validator(mode="after")
-    def _validate_order(self):
-        if not self.tags:
-            raise ValueError("tags must be provided and non-empty")
-        root = self.tags[0]
-        if root not in {"models", "input", "output"}:
-            raise ValueError("first tag must be one of: models, input, output")
-        if root == "models":
-            if len(self.tags) < 2:
-                raise ValueError("models uploads require a category tag as the second tag")
-        return self
-
-
-class SetPreviewBody(BaseModel):
-    """Set or clear the preview for an AssetInfo. Provide an Asset.id or null."""
-    preview_id: Optional[str] = None
-
-    @field_validator("preview_id", mode="before")
-    @classmethod
-    def _norm_uuid(cls, v):
-        if v is None:
-            return None
-        s = str(v).strip()
-        if not s:
-            return None
-        try:
-            uuid.UUID(s)
-        except Exception:
-            raise ValueError("preview_id must be a UUID")
-        return s

app/assets/api/schemas_out.py

@@ -1,115 +0,0 @@
-from datetime import datetime
-from typing import Any, Literal, Optional
-
-from pydantic import BaseModel, ConfigDict, Field, field_serializer
-
-
-class AssetSummary(BaseModel):
-    id: str
-    name: str
-    asset_hash: Optional[str]
-    size: Optional[int] = None
-    mime_type: Optional[str] = None
-    tags: list[str] = Field(default_factory=list)
-    preview_url: Optional[str] = None
-    created_at: Optional[datetime] = None
-    updated_at: Optional[datetime] = None
-    last_access_time: Optional[datetime] = None
-
-    model_config = ConfigDict(from_attributes=True)
-
-    @field_serializer("created_at", "updated_at", "last_access_time")
-    def _ser_dt(self, v: Optional[datetime], _info):
-        return v.isoformat() if v else None
-
-
-class AssetsList(BaseModel):
-    assets: list[AssetSummary]
-    total: int
-    has_more: bool
-
-
-class AssetUpdated(BaseModel):
-    id: str
-    name: str
-    asset_hash: Optional[str]
-    tags: list[str] = Field(default_factory=list)
-    user_metadata: dict[str, Any] = Field(default_factory=dict)
-    updated_at: Optional[datetime] = None
-
-    model_config = ConfigDict(from_attributes=True)
-
-    @field_serializer("updated_at")
-    def _ser_updated(self, v: Optional[datetime], _info):
-        return v.isoformat() if v else None
-
-
-class AssetDetail(BaseModel):
-    id: str
-    name: str
-    asset_hash: Optional[str]
-    size: Optional[int] = None
-    mime_type: Optional[str] = None
-    tags: list[str] = Field(default_factory=list)
-    user_metadata: dict[str, Any] = Field(default_factory=dict)
-    preview_id: Optional[str] = None
-    created_at: Optional[datetime] = None
-    last_access_time: Optional[datetime] = None
-
-    model_config = ConfigDict(from_attributes=True)
-
-    @field_serializer("created_at", "last_access_time")
-    def _ser_dt(self, v: Optional[datetime], _info):
-        return v.isoformat() if v else None
-
-
-class AssetCreated(AssetDetail):
-    created_new: bool
-
-
-class TagUsage(BaseModel):
-    name: str
-    count: int
-    type: str
-
-
-class TagsList(BaseModel):
-    tags: list[TagUsage] = Field(default_factory=list)
-    total: int
-    has_more: bool
-
-
-class TagsAdd(BaseModel):
-    model_config = ConfigDict(str_strip_whitespace=True)
-    added: list[str] = Field(default_factory=list)
-    already_present: list[str] = Field(default_factory=list)
-    total_tags: list[str] = Field(default_factory=list)
-
-
-class TagsRemove(BaseModel):
-    model_config = ConfigDict(str_strip_whitespace=True)
-    removed: list[str] = Field(default_factory=list)
-    not_present: list[str] = Field(default_factory=list)
-    total_tags: list[str] = Field(default_factory=list)
-
-
-class AssetScanError(BaseModel):
-    path: str
-    message: str
-    at: Optional[str] = Field(None, description="ISO timestamp")
-
-
-class AssetScanStatus(BaseModel):
-    scan_id: str
-    root: Literal["models", "input", "output"]
-    status: Literal["scheduled", "running", "completed", "failed", "cancelled"]
-    scheduled_at: Optional[str] = None
-    started_at: Optional[str] = None
-    finished_at: Optional[str] = None
-    discovered: int = 0
-    processed: int = 0
-    file_errors: list[AssetScanError] = Field(default_factory=list)
-
-
-class AssetScanStatusResponse(BaseModel):
-    scans: list[AssetScanStatus] = Field(default_factory=list)

app/assets/database/helpers/__init__.py

@@ -1,25 +0,0 @@
-from .bulk_ops import seed_from_paths_batch
-from .escape_like import escape_like_prefix
-from .fast_check import fast_asset_file_check
-from .filters import apply_metadata_filter, apply_tag_filters
-from .ownership import visible_owner_clause
-from .projection import is_scalar, project_kv
-from .tags import (
-    add_missing_tag_for_asset_id,
-    ensure_tags_exist,
-    remove_missing_tag_for_asset_id,
-)
-
-__all__ = [
-    "apply_tag_filters",
-    "apply_metadata_filter",
-    "escape_like_prefix",
-    "fast_asset_file_check",
-    "is_scalar",
-    "project_kv",
-    "ensure_tags_exist",
-    "add_missing_tag_for_asset_id",
-    "remove_missing_tag_for_asset_id",
-    "seed_from_paths_batch",
-    "visible_owner_clause",
-]

app/assets/database/helpers/bulk_ops.py

@@ -1,230 +0,0 @@
-import os
-import uuid
-from typing import Iterable, Sequence
-
-import sqlalchemy as sa
-from sqlalchemy.dialects import postgresql as d_pg
-from sqlalchemy.dialects import sqlite as d_sqlite
-from sqlalchemy.ext.asyncio import AsyncSession
-
-from ..models import Asset, AssetCacheState, AssetInfo, AssetInfoMeta, AssetInfoTag
-from ..timeutil import utcnow
-
-MAX_BIND_PARAMS = 800
-
-
-async def seed_from_paths_batch(
-    session: AsyncSession,
-    *,
-    specs: Sequence[dict],
-    owner_id: str = "",
-) -> dict:
-    """Each spec is a dict with keys:
-      - abs_path: str
-      - size_bytes: int
-      - mtime_ns: int
-      - info_name: str
-      - tags: list[str]
-      - fname: Optional[str]
-    """
-    if not specs:
-        return {"inserted_infos": 0, "won_states": 0, "lost_states": 0}
-
-    now = utcnow()
-    dialect = session.bind.dialect.name
-    if dialect not in ("sqlite", "postgresql"):
-        raise NotImplementedError(f"Unsupported database dialect: {dialect}")
-
-    asset_rows: list[dict] = []
-    state_rows: list[dict] = []
-    path_to_asset: dict[str, str] = {}
-    asset_to_info: dict[str, dict] = {}  # asset_id -> prepared info row
-    path_list: list[str] = []
-
-    for sp in specs:
-        ap = os.path.abspath(sp["abs_path"])
-        aid = str(uuid.uuid4())
-        iid = str(uuid.uuid4())
-        path_list.append(ap)
-        path_to_asset[ap] = aid
-
-        asset_rows.append(
-            {
-                "id": aid,
-                "hash": None,
-                "size_bytes": sp["size_bytes"],
-                "mime_type": None,
-                "created_at": now,
-            }
-        )
-        state_rows.append(
-            {
-                "asset_id": aid,
-                "file_path": ap,
-                "mtime_ns": sp["mtime_ns"],
-            }
-        )
-        asset_to_info[aid] = {
-            "id": iid,
-            "owner_id": owner_id,
-            "name": sp["info_name"],
-            "asset_id": aid,
-            "preview_id": None,
-            "user_metadata": {"filename": sp["fname"]} if sp["fname"] else None,
-            "created_at": now,
-            "updated_at": now,
-            "last_access_time": now,
-            "_tags": sp["tags"],
-            "_filename": sp["fname"],
-        }
-
-    # insert all seed Assets (hash=NULL)
-    ins_asset = d_sqlite.insert(Asset) if dialect == "sqlite" else d_pg.insert(Asset)
-    for chunk in _iter_chunks(asset_rows, _rows_per_stmt(5)):
-        await session.execute(ins_asset, chunk)
-
-    # try to claim AssetCacheState (file_path)
-    winners_by_path: set[str] = set()
-    if dialect == "sqlite":
-        ins_state = (
-            d_sqlite.insert(AssetCacheState)
-            .on_conflict_do_nothing(index_elements=[AssetCacheState.file_path])
-            .returning(AssetCacheState.file_path)
-        )
-    else:
-        ins_state = (
-            d_pg.insert(AssetCacheState)
-            .on_conflict_do_nothing(index_elements=[AssetCacheState.file_path])
-            .returning(AssetCacheState.file_path)
-        )
-    for chunk in _iter_chunks(state_rows, _rows_per_stmt(3)):
-        winners_by_path.update((await session.execute(ins_state, chunk)).scalars().all())
-
-    all_paths_set = set(path_list)
-    losers_by_path = all_paths_set - winners_by_path
-    lost_assets = [path_to_asset[p] for p in losers_by_path]
-    if lost_assets:  # losers get their Asset removed
-        for id_chunk in _iter_chunks(lost_assets, MAX_BIND_PARAMS):
-            await session.execute(sa.delete(Asset).where(Asset.id.in_(id_chunk)))
-
-    if not winners_by_path:
-        return {"inserted_infos": 0, "won_states": 0, "lost_states": len(losers_by_path)}
-
-    # insert AssetInfo only for winners
-    winner_info_rows = [asset_to_info[path_to_asset[p]] for p in winners_by_path]
-    if dialect == "sqlite":
-        ins_info = (
-            d_sqlite.insert(AssetInfo)
-            .on_conflict_do_nothing(index_elements=[AssetInfo.asset_id, AssetInfo.owner_id, AssetInfo.name])
-            .returning(AssetInfo.id)
-        )
-    else:
-        ins_info = (
-            d_pg.insert(AssetInfo)
-            .on_conflict_do_nothing(index_elements=[AssetInfo.asset_id, AssetInfo.owner_id, AssetInfo.name])
-            .returning(AssetInfo.id)
-        )
-
-    inserted_info_ids: set[str] = set()
-    for chunk in _iter_chunks(winner_info_rows, _rows_per_stmt(9)):
-        inserted_info_ids.update((await session.execute(ins_info, chunk)).scalars().all())
-
-    # build and insert tag + meta rows for the AssetInfo
-    tag_rows: list[dict] = []
-    meta_rows: list[dict] = []
-    if inserted_info_ids:
-        for row in winner_info_rows:
-            iid = row["id"]
-            if iid not in inserted_info_ids:
-                continue
-            for t in row["_tags"]:
-                tag_rows.append({
-                    "asset_info_id": iid,
-                    "tag_name": t,
-                    "origin": "automatic",
-                    "added_at": now,
-                })
-            if row["_filename"]:
-                meta_rows.append(
-                    {
-                        "asset_info_id": iid,
-                        "key": "filename",
-                        "ordinal": 0,
-                        "val_str": row["_filename"],
-                        "val_num": None,
-                        "val_bool": None,
-                        "val_json": None,
-                    }
-                )
-
-    await bulk_insert_tags_and_meta(session, tag_rows=tag_rows, meta_rows=meta_rows, max_bind_params=MAX_BIND_PARAMS)
-    return {
-        "inserted_infos": len(inserted_info_ids),
-        "won_states": len(winners_by_path),
-        "lost_states": len(losers_by_path),
-    }
-
-
-async def bulk_insert_tags_and_meta(
-    session: AsyncSession,
-    *,
-    tag_rows: list[dict],
-    meta_rows: list[dict],
-    max_bind_params: int,
-) -> None:
-    """Batch insert into asset_info_tags and asset_info_meta with ON CONFLICT DO NOTHING.
-    - tag_rows keys: asset_info_id, tag_name, origin, added_at
-    - meta_rows keys: asset_info_id, key, ordinal, val_str, val_num, val_bool, val_json
-    """
-    dialect = session.bind.dialect.name
-    if tag_rows:
-        if dialect == "sqlite":
-            ins_links = (
-                d_sqlite.insert(AssetInfoTag)
-                .on_conflict_do_nothing(index_elements=[AssetInfoTag.asset_info_id, AssetInfoTag.tag_name])
-            )
-        elif dialect == "postgresql":
-            ins_links = (
-                d_pg.insert(AssetInfoTag)
-                .on_conflict_do_nothing(index_elements=[AssetInfoTag.asset_info_id, AssetInfoTag.tag_name])
-            )
-        else:
-            raise NotImplementedError(f"Unsupported database dialect: {dialect}")
-        for chunk in _chunk_rows(tag_rows, cols_per_row=4, max_bind_params=max_bind_params):
-            await session.execute(ins_links, chunk)
-    if meta_rows:
-        if dialect == "sqlite":
-            ins_meta = (
-                d_sqlite.insert(AssetInfoMeta)
-                .on_conflict_do_nothing(
-                    index_elements=[AssetInfoMeta.asset_info_id, AssetInfoMeta.key, AssetInfoMeta.ordinal]
-                )
-            )
-        elif dialect == "postgresql":
-            ins_meta = (
-                d_pg.insert(AssetInfoMeta)
-                .on_conflict_do_nothing(
-                    index_elements=[AssetInfoMeta.asset_info_id, AssetInfoMeta.key, AssetInfoMeta.ordinal]
-                )
-            )
-        else:
-            raise NotImplementedError(f"Unsupported database dialect: {dialect}")
-        for chunk in _chunk_rows(meta_rows, cols_per_row=7, max_bind_params=max_bind_params):
-            await session.execute(ins_meta, chunk)
-
-
-def _chunk_rows(rows: list[dict], cols_per_row: int, max_bind_params: int) -> Iterable[list[dict]]:
-    if not rows:
-        return []
-    rows_per_stmt = max(1, max_bind_params // max(1, cols_per_row))
-    for i in range(0, len(rows), rows_per_stmt):
-        yield rows[i:i + rows_per_stmt]
-
-
-def _iter_chunks(seq, n: int):
-    for i in range(0, len(seq), n):
-        yield seq[i:i + n]
-
-
-def _rows_per_stmt(cols: int) -> int:
-    return max(1, MAX_BIND_PARAMS // max(1, cols))

app/assets/database/helpers/escape_like.py

@@ -1,7 +0,0 @@
-def escape_like_prefix(s: str, escape: str = "!") -> tuple[str, str]:
-    """Escapes %, _ and the escape char itself in a LIKE prefix.
-    Returns (escaped_prefix, escape_char). Caller should append '%' and pass escape=escape_char to .like().
-    """
-    s = s.replace(escape, escape + escape)  # escape the escape char first
-    s = s.replace("%", escape + "%").replace("_", escape + "_")  # escape LIKE wildcards
-    return s, escape

app/assets/database/helpers/fast_check.py

@@ -1,19 +0,0 @@
-import os
-from typing import Optional
-
-
-def fast_asset_file_check(
-    *,
-    mtime_db: Optional[int],
-    size_db: Optional[int],
-    stat_result: os.stat_result,
-) -> bool:
-    if mtime_db is None:
-        return False
-    actual_mtime_ns = getattr(stat_result, "st_mtime_ns", int(stat_result.st_mtime * 1_000_000_000))
-    if int(mtime_db) != int(actual_mtime_ns):
-        return False
-    sz = int(size_db or 0)
-    if sz > 0:
-        return int(stat_result.st_size) == sz
-    return True

app/assets/database/helpers/filters.py

@@ -1,87 +0,0 @@
-from typing import Optional, Sequence
-
-import sqlalchemy as sa
-from sqlalchemy import exists
-
-from ..._helpers import normalize_tags
-from ..models import AssetInfo, AssetInfoMeta, AssetInfoTag
-
-
-def apply_tag_filters(
-    stmt: sa.sql.Select,
-    include_tags: Optional[Sequence[str]],
-    exclude_tags: Optional[Sequence[str]],
-) -> sa.sql.Select:
-    """include_tags: every tag must be present; exclude_tags: none may be present."""
-    include_tags = normalize_tags(include_tags)
-    exclude_tags = normalize_tags(exclude_tags)
-
-    if include_tags:
-        for tag_name in include_tags:
-            stmt = stmt.where(
-                exists().where(
-                    (AssetInfoTag.asset_info_id == AssetInfo.id)
-                    & (AssetInfoTag.tag_name == tag_name)
-                )
-            )
-
-    if exclude_tags:
-        stmt = stmt.where(
-            ~exists().where(
-                (AssetInfoTag.asset_info_id == AssetInfo.id)
-                & (AssetInfoTag.tag_name.in_(exclude_tags))
-            )
-        )
-    return stmt
-
-
-def apply_metadata_filter(
-    stmt: sa.sql.Select,
-    metadata_filter: Optional[dict],
-) -> sa.sql.Select:
-    """Apply filters using asset_info_meta projection table."""
-    if not metadata_filter:
-        return stmt
-
-    def _exists_for_pred(key: str, *preds) -> sa.sql.ClauseElement:
-        return sa.exists().where(
-            AssetInfoMeta.asset_info_id == AssetInfo.id,
-            AssetInfoMeta.key == key,
-            *preds,
-        )
-
-    def _exists_clause_for_value(key: str, value) -> sa.sql.ClauseElement:
-        if value is None:
-            no_row_for_key = sa.not_(
-                sa.exists().where(
-                    AssetInfoMeta.asset_info_id == AssetInfo.id,
-                    AssetInfoMeta.key == key,
-                )
-            )
-            null_row = _exists_for_pred(
-                key,
-                AssetInfoMeta.val_json.is_(None),
-                AssetInfoMeta.val_str.is_(None),
-                AssetInfoMeta.val_num.is_(None),
-                AssetInfoMeta.val_bool.is_(None),
-            )
-            return sa.or_(no_row_for_key, null_row)
-
-        if isinstance(value, bool):
-            return _exists_for_pred(key, AssetInfoMeta.val_bool == bool(value))
-        if isinstance(value, (int, float)):
-            from decimal import Decimal
-            num = value if isinstance(value, Decimal) else Decimal(str(value))
-            return _exists_for_pred(key, AssetInfoMeta.val_num == num)
-        if isinstance(value, str):
-            return _exists_for_pred(key, AssetInfoMeta.val_str == value)
-        return _exists_for_pred(key, AssetInfoMeta.val_json == value)
-
-    for k, v in metadata_filter.items():
-        if isinstance(v, list):
-            ors = [_exists_clause_for_value(k, elem) for elem in v]
-            if ors:
-                stmt = stmt.where(sa.or_(*ors))
-        else:
-            stmt = stmt.where(_exists_clause_for_value(k, v))
-    return stmt

app/assets/database/helpers/ownership.py

@@ -1,12 +0,0 @@
-import sqlalchemy as sa
-
-from ..models import AssetInfo
-
-
-def visible_owner_clause(owner_id: str) -> sa.sql.ClauseElement:
-    """Build owner visibility predicate for reads. Owner-less rows are visible to everyone."""
-
-    owner_id = (owner_id or "").strip()
-    if owner_id == "":
-        return AssetInfo.owner_id == ""
-    return AssetInfo.owner_id.in_(["", owner_id])

app/assets/database/helpers/projection.py

@@ -1,64 +0,0 @@
-from decimal import Decimal
-
-
-def is_scalar(v):
-    if v is None:
-        return True
-    if isinstance(v, bool):
-        return True
-    if isinstance(v, (int, float, Decimal, str)):
-        return True
-    return False
-
-
-def project_kv(key: str, value):
-    """
-    Turn a metadata key/value into typed projection rows.
-    Returns list[dict] with keys:
-      key, ordinal, and one of val_str / val_num / val_bool / val_json (others None)
-    """
-    rows: list[dict] = []
-
-    def _null_row(ordinal: int) -> dict:
-        return {
-            "key": key, "ordinal": ordinal,
-            "val_str": None, "val_num": None, "val_bool": None, "val_json": None
-        }
-
-    if value is None:
-        rows.append(_null_row(0))
-        return rows
-
-    if is_scalar(value):
-        if isinstance(value, bool):
-            rows.append({"key": key, "ordinal": 0, "val_bool": bool(value)})
-        elif isinstance(value, (int, float, Decimal)):
-            num = value if isinstance(value, Decimal) else Decimal(str(value))
-            rows.append({"key": key, "ordinal": 0, "val_num": num})
-        elif isinstance(value, str):
-            rows.append({"key": key, "ordinal": 0, "val_str": value})
-        else:
-            rows.append({"key": key, "ordinal": 0, "val_json": value})
-        return rows
-
-    if isinstance(value, list):
-        if all(is_scalar(x) for x in value):
-            for i, x in enumerate(value):
-                if x is None:
-                    rows.append(_null_row(i))
-                elif isinstance(x, bool):
-                    rows.append({"key": key, "ordinal": i, "val_bool": bool(x)})
-                elif isinstance(x, (int, float, Decimal)):
-                    num = x if isinstance(x, Decimal) else Decimal(str(x))
-                    rows.append({"key": key, "ordinal": i, "val_num": num})
-                elif isinstance(x, str):
-                    rows.append({"key": key, "ordinal": i, "val_str": x})
-                else:
-                    rows.append({"key": key, "ordinal": i, "val_json": x})
-            return rows
-        for i, x in enumerate(value):
-            rows.append({"key": key, "ordinal": i, "val_json": x})
-        return rows
-
-    rows.append({"key": key, "ordinal": 0, "val_json": value})
-    return rows

app/assets/database/helpers/tags.py

@@ -1,90 +0,0 @@
-from typing import Iterable
-
-import sqlalchemy as sa
-from sqlalchemy.dialects import postgresql as d_pg
-from sqlalchemy.dialects import sqlite as d_sqlite
-from sqlalchemy.ext.asyncio import AsyncSession
-
-from ..._helpers import normalize_tags
-from ..models import AssetInfo, AssetInfoTag, Tag
-from ..timeutil import utcnow
-
-
-async def ensure_tags_exist(session: AsyncSession, names: Iterable[str], tag_type: str = "user") -> None:
-    wanted = normalize_tags(list(names))
-    if not wanted:
-        return
-    rows = [{"name": n, "tag_type": tag_type} for n in list(dict.fromkeys(wanted))]
-    dialect = session.bind.dialect.name
-    if dialect == "sqlite":
-        ins = (
-            d_sqlite.insert(Tag)
-            .values(rows)
-            .on_conflict_do_nothing(index_elements=[Tag.name])
-        )
-    elif dialect == "postgresql":
-        ins = (
-            d_pg.insert(Tag)
-            .values(rows)
-            .on_conflict_do_nothing(index_elements=[Tag.name])
-        )
-    else:
-        raise NotImplementedError(f"Unsupported database dialect: {dialect}")
-    await session.execute(ins)
-
-
-async def add_missing_tag_for_asset_id(
-    session: AsyncSession,
-    *,
-    asset_id: str,
-    origin: str = "automatic",
-) -> None:
-    select_rows = (
-        sa.select(
-            AssetInfo.id.label("asset_info_id"),
-            sa.literal("missing").label("tag_name"),
-            sa.literal(origin).label("origin"),
-            sa.literal(utcnow()).label("added_at"),
-        )
-        .where(AssetInfo.asset_id == asset_id)
-        .where(
-            sa.not_(
-                sa.exists().where((AssetInfoTag.asset_info_id == AssetInfo.id) & (AssetInfoTag.tag_name == "missing"))
-            )
-        )
-    )
-    dialect = session.bind.dialect.name
-    if dialect == "sqlite":
-        ins = (
-            d_sqlite.insert(AssetInfoTag)
-            .from_select(
-                ["asset_info_id", "tag_name", "origin", "added_at"],
-                select_rows,
-            )
-            .on_conflict_do_nothing(index_elements=[AssetInfoTag.asset_info_id, AssetInfoTag.tag_name])
-        )
-    elif dialect == "postgresql":
-        ins = (
-            d_pg.insert(AssetInfoTag)
-            .from_select(
-                ["asset_info_id", "tag_name", "origin", "added_at"],
-                select_rows,
-            )
-            .on_conflict_do_nothing(index_elements=[AssetInfoTag.asset_info_id, AssetInfoTag.tag_name])
-        )
-    else:
-        raise NotImplementedError(f"Unsupported database dialect: {dialect}")
-    await session.execute(ins)
-
-
-async def remove_missing_tag_for_asset_id(
-    session: AsyncSession,
-    *,
-    asset_id: str,
-) -> None:
-    await session.execute(
-        sa.delete(AssetInfoTag).where(
-            AssetInfoTag.asset_info_id.in_(sa.select(AssetInfo.id).where(AssetInfo.asset_id == asset_id)),
-            AssetInfoTag.tag_name == "missing",
-        )
-    )

app/assets/database/models.py

@@ -1,251 +0,0 @@
-import uuid
-from datetime import datetime
-from typing import Any, Optional
-
-from sqlalchemy import (
-    JSON,
-    BigInteger,
-    Boolean,
-    CheckConstraint,
-    DateTime,
-    ForeignKey,
-    Index,
-    Integer,
-    Numeric,
-    String,
-    Text,
-    UniqueConstraint,
-)
-from sqlalchemy.dialects.postgresql import JSONB
-from sqlalchemy.orm import DeclarativeBase, Mapped, foreign, mapped_column, relationship
-
-from .timeutil import utcnow
-
-JSONB_V = JSON(none_as_null=True).with_variant(JSONB(none_as_null=True), 'postgresql')
-
-
-class Base(DeclarativeBase):
-    pass
-
-
-def to_dict(obj: Any, include_none: bool = False) -> dict[str, Any]:
-    fields = obj.__table__.columns.keys()
-    out: dict[str, Any] = {}
-    for field in fields:
-        val = getattr(obj, field)
-        if val is None and not include_none:
-            continue
-        if isinstance(val, datetime):
-            out[field] = val.isoformat()
-        else:
-            out[field] = val
-    return out
-
-
-class Asset(Base):
-    __tablename__ = "assets"
-
-    id: Mapped[str] = mapped_column(String(36), primary_key=True, default=lambda: str(uuid.uuid4()))
-    hash: Mapped[Optional[str]] = mapped_column(String(256), nullable=True)
-    size_bytes: Mapped[int] = mapped_column(BigInteger, nullable=False, default=0)
-    mime_type: Mapped[Optional[str]] = mapped_column(String(255))
-    created_at: Mapped[datetime] = mapped_column(
-        DateTime(timezone=False), nullable=False, default=utcnow
-    )
-
-    infos: Mapped[list["AssetInfo"]] = relationship(
-        "AssetInfo",
-        back_populates="asset",
-        primaryjoin=lambda: Asset.id == foreign(AssetInfo.asset_id),
-        foreign_keys=lambda: [AssetInfo.asset_id],
-        cascade="all,delete-orphan",
-        passive_deletes=True,
-    )
-
-    preview_of: Mapped[list["AssetInfo"]] = relationship(
-        "AssetInfo",
-        back_populates="preview_asset",
-        primaryjoin=lambda: Asset.id == foreign(AssetInfo.preview_id),
-        foreign_keys=lambda: [AssetInfo.preview_id],
-        viewonly=True,
-    )
-
-    cache_states: Mapped[list["AssetCacheState"]] = relationship(
-        back_populates="asset",
-        cascade="all, delete-orphan",
-        passive_deletes=True,
-    )
-
-    __table_args__ = (
-        Index("uq_assets_hash", "hash", unique=True),
-        Index("ix_assets_mime_type", "mime_type"),
-        CheckConstraint("size_bytes >= 0", name="ck_assets_size_nonneg"),
-    )
-
-    def to_dict(self, include_none: bool = False) -> dict[str, Any]:
-        return to_dict(self, include_none=include_none)
-
-    def __repr__(self) -> str:
-        return f"<Asset id={self.id} hash={(self.hash or '')[:12]}>"
-
-
-class AssetCacheState(Base):
-    __tablename__ = "asset_cache_state"
-
-    id: Mapped[int] = mapped_column(Integer, primary_key=True, autoincrement=True)
-    asset_id: Mapped[str] = mapped_column(String(36), ForeignKey("assets.id", ondelete="CASCADE"), nullable=False)
-    file_path: Mapped[str] = mapped_column(Text, nullable=False)
-    mtime_ns: Mapped[Optional[int]] = mapped_column(BigInteger, nullable=True)
-    needs_verify: Mapped[bool] = mapped_column(Boolean, nullable=False, default=False)
-
-    asset: Mapped["Asset"] = relationship(back_populates="cache_states")
-
-    __table_args__ = (
-        Index("ix_asset_cache_state_file_path", "file_path"),
-        Index("ix_asset_cache_state_asset_id", "asset_id"),
-        CheckConstraint("(mtime_ns IS NULL) OR (mtime_ns >= 0)", name="ck_acs_mtime_nonneg"),
-        UniqueConstraint("file_path", name="uq_asset_cache_state_file_path"),
-    )
-
-    def to_dict(self, include_none: bool = False) -> dict[str, Any]:
-        return to_dict(self, include_none=include_none)
-
-    def __repr__(self) -> str:
-        return f"<AssetCacheState id={self.id} asset_id={self.asset_id} path={self.file_path!r}>"
-
-
-class AssetInfo(Base):
-    __tablename__ = "assets_info"
-
-    id: Mapped[str] = mapped_column(String(36), primary_key=True, default=lambda: str(uuid.uuid4()))
-    owner_id: Mapped[str] = mapped_column(String(128), nullable=False, default="")
-    name: Mapped[str] = mapped_column(String(512), nullable=False)
-    asset_id: Mapped[str] = mapped_column(String(36), ForeignKey("assets.id", ondelete="RESTRICT"), nullable=False)
-    preview_id: Mapped[Optional[str]] = mapped_column(String(36), ForeignKey("assets.id", ondelete="SET NULL"))
-    user_metadata: Mapped[Optional[dict[str, Any]]] = mapped_column(JSON(none_as_null=True))
-    created_at: Mapped[datetime] = mapped_column(DateTime(timezone=False), nullable=False, default=utcnow)
-    updated_at: Mapped[datetime] = mapped_column(DateTime(timezone=False), nullable=False, default=utcnow)
-    last_access_time: Mapped[datetime] = mapped_column(DateTime(timezone=False), nullable=False, default=utcnow)
-
-    asset: Mapped[Asset] = relationship(
-        "Asset",
-        back_populates="infos",
-        foreign_keys=[asset_id],
-        lazy="selectin",
-    )
-    preview_asset: Mapped[Optional[Asset]] = relationship(
-        "Asset",
-        back_populates="preview_of",
-        foreign_keys=[preview_id],
-    )
-
-    metadata_entries: Mapped[list["AssetInfoMeta"]] = relationship(
-        back_populates="asset_info",
-        cascade="all,delete-orphan",
-        passive_deletes=True,
-    )
-
-    tag_links: Mapped[list["AssetInfoTag"]] = relationship(
-        back_populates="asset_info",
-        cascade="all,delete-orphan",
-        passive_deletes=True,
-        overlaps="tags,asset_infos",
-    )
-
-    tags: Mapped[list["Tag"]] = relationship(
-        secondary="asset_info_tags",
-        back_populates="asset_infos",
-        lazy="selectin",
-        viewonly=True,
-        overlaps="tag_links,asset_info_links,asset_infos,tag",
-    )
-
-    __table_args__ = (
-        UniqueConstraint("asset_id", "owner_id", "name", name="uq_assets_info_asset_owner_name"),
-        Index("ix_assets_info_owner_name", "owner_id", "name"),
-        Index("ix_assets_info_owner_id", "owner_id"),
-        Index("ix_assets_info_asset_id", "asset_id"),
-        Index("ix_assets_info_name", "name"),
-        Index("ix_assets_info_created_at", "created_at"),
-        Index("ix_assets_info_last_access_time", "last_access_time"),
-    )
-
-    def to_dict(self, include_none: bool = False) -> dict[str, Any]:
-        data = to_dict(self, include_none=include_none)
-        data["tags"] = [t.name for t in self.tags]
-        return data
-
-    def __repr__(self) -> str:
-        return f"<AssetInfo id={self.id} name={self.name!r} asset_id={self.asset_id}>"
-
-
-class AssetInfoMeta(Base):
-    __tablename__ = "asset_info_meta"
-
-    asset_info_id: Mapped[str] = mapped_column(
-        String(36), ForeignKey("assets_info.id", ondelete="CASCADE"), primary_key=True
-    )
-    key: Mapped[str] = mapped_column(String(256), primary_key=True)
-    ordinal: Mapped[int] = mapped_column(Integer, primary_key=True, default=0)
-
-    val_str: Mapped[Optional[str]] = mapped_column(String(2048), nullable=True)
-    val_num: Mapped[Optional[float]] = mapped_column(Numeric(38, 10), nullable=True)
-    val_bool: Mapped[Optional[bool]] = mapped_column(Boolean, nullable=True)
-    val_json: Mapped[Optional[Any]] = mapped_column(JSONB_V, nullable=True)
-
-    asset_info: Mapped["AssetInfo"] = relationship(back_populates="metadata_entries")
-
-    __table_args__ = (
-        Index("ix_asset_info_meta_key", "key"),
-        Index("ix_asset_info_meta_key_val_str", "key", "val_str"),
-        Index("ix_asset_info_meta_key_val_num", "key", "val_num"),
-        Index("ix_asset_info_meta_key_val_bool", "key", "val_bool"),
-    )
-
-
-class AssetInfoTag(Base):
-    __tablename__ = "asset_info_tags"
-
-    asset_info_id: Mapped[str] = mapped_column(
-        String(36), ForeignKey("assets_info.id", ondelete="CASCADE"), primary_key=True
-    )
-    tag_name: Mapped[str] = mapped_column(
-        String(512), ForeignKey("tags.name", ondelete="RESTRICT"), primary_key=True
-    )
-    origin: Mapped[str] = mapped_column(String(32), nullable=False, default="manual")
-    added_at: Mapped[datetime] = mapped_column(
-        DateTime(timezone=False), nullable=False, default=utcnow
-    )
-
-    asset_info: Mapped["AssetInfo"] = relationship(back_populates="tag_links")
-    tag: Mapped["Tag"] = relationship(back_populates="asset_info_links")
-
-    __table_args__ = (
-        Index("ix_asset_info_tags_tag_name", "tag_name"),
-        Index("ix_asset_info_tags_asset_info_id", "asset_info_id"),
-    )
-
-
-class Tag(Base):
-    __tablename__ = "tags"
-
-    name: Mapped[str] = mapped_column(String(512), primary_key=True)
-    tag_type: Mapped[str] = mapped_column(String(32), nullable=False, default="user")
-
-    asset_info_links: Mapped[list["AssetInfoTag"]] = relationship(
-        back_populates="tag",
-        overlaps="asset_infos,tags",
-    )
-    asset_infos: Mapped[list["AssetInfo"]] = relationship(
-        secondary="asset_info_tags",
-        back_populates="tags",
-        viewonly=True,
-        overlaps="asset_info_links,tag_links,tags,asset_info",
-    )
-
-    __table_args__ = (
-        Index("ix_tags_tag_type", "tag_type"),
-    )
-
-    def __repr__(self) -> str:
-        return f"<Tag {self.name}>"

app/assets/database/services/__init__.py

@@ -1,57 +0,0 @@
-from .content import (
-    check_fs_asset_exists_quick,
-    compute_hash_and_dedup_for_cache_state,
-    ingest_fs_asset,
-    list_cache_states_with_asset_under_prefixes,
-    list_unhashed_candidates_under_prefixes,
-    list_verify_candidates_under_prefixes,
-    redirect_all_references_then_delete_asset,
-    touch_asset_infos_by_fs_path,
-)
-from .info import (
-    add_tags_to_asset_info,
-    create_asset_info_for_existing_asset,
-    delete_asset_info_by_id,
-    fetch_asset_info_and_asset,
-    fetch_asset_info_asset_and_tags,
-    get_asset_tags,
-    list_asset_infos_page,
-    list_tags_with_usage,
-    remove_tags_from_asset_info,
-    replace_asset_info_metadata_projection,
-    set_asset_info_preview,
-    set_asset_info_tags,
-    touch_asset_info_by_id,
-    update_asset_info_full,
-)
-from .queries import (
-    asset_exists_by_hash,
-    asset_info_exists_for_asset_id,
-    get_asset_by_hash,
-    get_asset_info_by_id,
-    get_cache_state_by_asset_id,
-    list_cache_states_by_asset_id,
-    pick_best_live_path,
-)
-
-__all__ = [
-    # queries
-    "asset_exists_by_hash", "get_asset_by_hash", "get_asset_info_by_id", "asset_info_exists_for_asset_id",
-    "get_cache_state_by_asset_id",
-    "list_cache_states_by_asset_id",
-    "pick_best_live_path",
-    # info
-    "list_asset_infos_page", "create_asset_info_for_existing_asset", "set_asset_info_tags",
-    "update_asset_info_full", "replace_asset_info_metadata_projection",
-    "touch_asset_info_by_id", "delete_asset_info_by_id",
-    "add_tags_to_asset_info", "remove_tags_from_asset_info",
-    "get_asset_tags", "list_tags_with_usage", "set_asset_info_preview",
-    "fetch_asset_info_and_asset", "fetch_asset_info_asset_and_tags",
-    # content
-    "check_fs_asset_exists_quick",
-    "redirect_all_references_then_delete_asset",
-    "compute_hash_and_dedup_for_cache_state",
-    "list_unhashed_candidates_under_prefixes", "list_verify_candidates_under_prefixes",
-    "ingest_fs_asset", "touch_asset_infos_by_fs_path",
-    "list_cache_states_with_asset_under_prefixes",
-]

app/assets/database/services/content.py

@@ -1,721 +0,0 @@
-import contextlib
-import logging
-import os
-from datetime import datetime
-from typing import Any, Optional, Sequence, Union
-
-import sqlalchemy as sa
-from sqlalchemy import select
-from sqlalchemy.dialects import postgresql as d_pg
-from sqlalchemy.dialects import sqlite as d_sqlite
-from sqlalchemy.exc import IntegrityError
-from sqlalchemy.ext.asyncio import AsyncSession
-from sqlalchemy.orm import noload
-
-from ..._helpers import compute_relative_filename
-from ...storage import hashing as hashing_mod
-from ..helpers import (
-    ensure_tags_exist,
-    escape_like_prefix,
-    remove_missing_tag_for_asset_id,
-)
-from ..models import Asset, AssetCacheState, AssetInfo, AssetInfoTag, Tag
-from ..timeutil import utcnow
-from .info import replace_asset_info_metadata_projection
-from .queries import list_cache_states_by_asset_id, pick_best_live_path
-
-
-async def check_fs_asset_exists_quick(
-    session: AsyncSession,
-    *,
-    file_path: str,
-    size_bytes: Optional[int] = None,
-    mtime_ns: Optional[int] = None,
-) -> bool:
-    """Returns True if we already track this absolute path with a HASHED asset and the cached mtime/size match."""
-    locator = os.path.abspath(file_path)
-
-    stmt = (
-        sa.select(sa.literal(True))
-        .select_from(AssetCacheState)
-        .join(Asset, Asset.id == AssetCacheState.asset_id)
-        .where(
-            AssetCacheState.file_path == locator,
-            Asset.hash.isnot(None),
-            AssetCacheState.needs_verify.is_(False),
-        )
-        .limit(1)
-    )
-
-    conds = []
-    if mtime_ns is not None:
-        conds.append(AssetCacheState.mtime_ns == int(mtime_ns))
-    if size_bytes is not None:
-        conds.append(sa.or_(Asset.size_bytes == 0, Asset.size_bytes == int(size_bytes)))
-    if conds:
-        stmt = stmt.where(*conds)
-    return (await session.execute(stmt)).first() is not None
-
-
-async def redirect_all_references_then_delete_asset(
-    session: AsyncSession,
-    *,
-    duplicate_asset_id: str,
-    canonical_asset_id: str,
-) -> None:
-    """
-    Safely migrate all references from duplicate_asset_id to canonical_asset_id.
-
-    - If an AssetInfo for (owner_id, name) already exists on the canonical asset,
-      merge tags, metadata, times, and preview, then delete the duplicate AssetInfo.
-    - Otherwise, simply repoint the AssetInfo.asset_id.
-    - Always retarget AssetCacheState rows.
-    - Finally delete the duplicate Asset row.
-    """
-    if duplicate_asset_id == canonical_asset_id:
-        return
-
-    # 1) Migrate AssetInfo rows one-by-one to avoid UNIQUE conflicts.
-    dup_infos = (
-        await session.execute(
-            select(AssetInfo).options(noload(AssetInfo.tags)).where(AssetInfo.asset_id == duplicate_asset_id)
-        )
-    ).unique().scalars().all()
-
-    for info in dup_infos:
-        # Try to find an existing collision on canonical
-        existing = (
-            await session.execute(
-                select(AssetInfo)
-                .options(noload(AssetInfo.tags))
-                .where(
-                    AssetInfo.asset_id == canonical_asset_id,
-                    AssetInfo.owner_id == info.owner_id,
-                    AssetInfo.name == info.name,
-                )
-                .limit(1)
-            )
-        ).unique().scalars().first()
-
-        if existing:
-            merged_meta = dict(existing.user_metadata or {})
-            other_meta = info.user_metadata or {}
-            for k, v in other_meta.items():
-                if k not in merged_meta:
-                    merged_meta[k] = v
-            if merged_meta != (existing.user_metadata or {}):
-                await replace_asset_info_metadata_projection(
-                    session,
-                    asset_info_id=existing.id,
-                    user_metadata=merged_meta,
-                )
-
-            existing_tags = {
-                t for (t,) in (
-                    await session.execute(
-                        select(AssetInfoTag.tag_name).where(AssetInfoTag.asset_info_id == existing.id)
-                    )
-                ).all()
-            }
-            from_tags = {
-                t for (t,) in (
-                    await session.execute(
-                        select(AssetInfoTag.tag_name).where(AssetInfoTag.asset_info_id == info.id)
-                    )
-                ).all()
-            }
-            to_add = sorted(from_tags - existing_tags)
-            if to_add:
-                await ensure_tags_exist(session, to_add, tag_type="user")
-                now = utcnow()
-                session.add_all([
-                    AssetInfoTag(asset_info_id=existing.id, tag_name=t, origin="automatic", added_at=now)
-                    for t in to_add
-                ])
-                await session.flush()
-
-            if existing.preview_id is None and info.preview_id is not None:
-                existing.preview_id = info.preview_id
-            if info.last_access_time and (
-                existing.last_access_time is None or info.last_access_time > existing.last_access_time
-            ):
-                existing.last_access_time = info.last_access_time
-            existing.updated_at = utcnow()
-            await session.flush()
-
-            # Delete the duplicate AssetInfo (cascades will clean its tags/meta)
-            await session.delete(info)
-            await session.flush()
-        else:
-            # Simple retarget
-            info.asset_id = canonical_asset_id
-            info.updated_at = utcnow()
-            await session.flush()
-
-    # 2) Repoint cache states and previews
-    await session.execute(
-        sa.update(AssetCacheState)
-        .where(AssetCacheState.asset_id == duplicate_asset_id)
-        .values(asset_id=canonical_asset_id)
-    )
-    await session.execute(
-        sa.update(AssetInfo)
-        .where(AssetInfo.preview_id == duplicate_asset_id)
-        .values(preview_id=canonical_asset_id)
-    )
-
-    # 3) Remove duplicate Asset
-    dup = await session.get(Asset, duplicate_asset_id)
-    if dup:
-        await session.delete(dup)
-    await session.flush()
-
-
-async def compute_hash_and_dedup_for_cache_state(
-    session: AsyncSession,
-    *,
-    state_id: int,
-) -> Optional[str]:
-    """
-    Compute hash for the given cache state, deduplicate, and settle verify cases.
-
-    Returns the asset_id that this state ends up pointing to, or None if file disappeared.
-    """
-    state = await session.get(AssetCacheState, state_id)
-    if not state:
-        return None
-
-    path = state.file_path
-    try:
-        if not os.path.isfile(path):
-            # File vanished: drop the state. If the Asset has hash=NULL and has no other states, drop the Asset too.
-            asset = await session.get(Asset, state.asset_id)
-            await session.delete(state)
-            await session.flush()
-
-            if asset and asset.hash is None:
-                remaining = (
-                    await session.execute(
-                        sa.select(sa.func.count())
-                        .select_from(AssetCacheState)
-                        .where(AssetCacheState.asset_id == asset.id)
-                    )
-                ).scalar_one()
-                if int(remaining or 0) == 0:
-                    await session.delete(asset)
-                    await session.flush()
-                else:
-                    await _recompute_and_apply_filename_for_asset(session, asset_id=asset.id)
-            return None
-
-        digest = await hashing_mod.blake3_hash(path)
-        new_hash = f"blake3:{digest}"
-
-        st = os.stat(path, follow_symlinks=True)
-        new_size = int(st.st_size)
-        mtime_ns = getattr(st, "st_mtime_ns", int(st.st_mtime * 1_000_000_000))
-
-        # Current asset of this state
-        this_asset = await session.get(Asset, state.asset_id)
-
-        # If the state got orphaned somehow (race), just reattach appropriately.
-        if not this_asset:
-            canonical = (
-                await session.execute(sa.select(Asset).where(Asset.hash == new_hash).limit(1))
-            ).scalars().first()
-            if canonical:
-                state.asset_id = canonical.id
-            else:
-                now = utcnow()
-                new_asset = Asset(hash=new_hash, size_bytes=new_size, mime_type=None, created_at=now)
-                session.add(new_asset)
-                await session.flush()
-                state.asset_id = new_asset.id
-            state.mtime_ns = mtime_ns
-            state.needs_verify = False
-            with contextlib.suppress(Exception):
-                await remove_missing_tag_for_asset_id(session, asset_id=state.asset_id)
-            await session.flush()
-            return state.asset_id
-
-        # 1) Seed asset case (hash is NULL): claim or merge into canonical
-        if this_asset.hash is None:
-            canonical = (
-                await session.execute(sa.select(Asset).where(Asset.hash == new_hash).limit(1))
-            ).scalars().first()
-
-            if canonical and canonical.id != this_asset.id:
-                # Merge seed asset into canonical (safe, collision-aware)
-                await redirect_all_references_then_delete_asset(
-                    session,
-                    duplicate_asset_id=this_asset.id,
-                    canonical_asset_id=canonical.id,
-                )
-                state = await session.get(AssetCacheState, state_id)
-                if state:
-                    state.mtime_ns = mtime_ns
-                    state.needs_verify = False
-                    with contextlib.suppress(Exception):
-                        await remove_missing_tag_for_asset_id(session, asset_id=canonical.id)
-                    await _recompute_and_apply_filename_for_asset(session, asset_id=canonical.id)
-                    await session.flush()
-                return canonical.id
-
-            # No canonical: try to claim the hash; handle races with a SAVEPOINT
-            try:
-                async with session.begin_nested():
-                    this_asset.hash = new_hash
-                    if int(this_asset.size_bytes or 0) == 0 and new_size > 0:
-                        this_asset.size_bytes = new_size
-                    await session.flush()
-            except IntegrityError:
-                # Someone else claimed it concurrently; fetch canonical and merge
-                canonical = (
-                    await session.execute(sa.select(Asset).where(Asset.hash == new_hash).limit(1))
-                ).scalars().first()
-                if canonical and canonical.id != this_asset.id:
-                    await redirect_all_references_then_delete_asset(
-                        session,
-                        duplicate_asset_id=this_asset.id,
-                        canonical_asset_id=canonical.id,
-                    )
-                    state = await session.get(AssetCacheState, state_id)
-                    if state:
-                        state.mtime_ns = mtime_ns
-                        state.needs_verify = False
-                        with contextlib.suppress(Exception):
-                            await remove_missing_tag_for_asset_id(session, asset_id=canonical.id)
-                        await _recompute_and_apply_filename_for_asset(session, asset_id=canonical.id)
-                        await session.flush()
-                    return canonical.id
-                # If we got here, the integrity error was not about hash uniqueness
-                raise
-
-            # Claimed successfully
-            state.mtime_ns = mtime_ns
-            state.needs_verify = False
-            with contextlib.suppress(Exception):
-                await remove_missing_tag_for_asset_id(session, asset_id=this_asset.id)
-            await _recompute_and_apply_filename_for_asset(session, asset_id=this_asset.id)
-            await session.flush()
-            return this_asset.id
-
-        # 2) Verify case for hashed assets
-        if this_asset.hash == new_hash:
-            if int(this_asset.size_bytes or 0) == 0 and new_size > 0:
-                this_asset.size_bytes = new_size
-            state.mtime_ns = mtime_ns
-            state.needs_verify = False
-            with contextlib.suppress(Exception):
-                await remove_missing_tag_for_asset_id(session, asset_id=this_asset.id)
-            await _recompute_and_apply_filename_for_asset(session, asset_id=this_asset.id)
-            await session.flush()
-            return this_asset.id
-
-        # Content changed on this path only: retarget THIS state, do not move AssetInfo rows
-        canonical = (
-            await session.execute(sa.select(Asset).where(Asset.hash == new_hash).limit(1))
-        ).scalars().first()
-        if canonical:
-            target_id = canonical.id
-        else:
-            now = utcnow()
-            new_asset = Asset(hash=new_hash, size_bytes=new_size, mime_type=None, created_at=now)
-            session.add(new_asset)
-            await session.flush()
-            target_id = new_asset.id
-
-        state.asset_id = target_id
-        state.mtime_ns = mtime_ns
-        state.needs_verify = False
-        with contextlib.suppress(Exception):
-            await remove_missing_tag_for_asset_id(session, asset_id=target_id)
-        await _recompute_and_apply_filename_for_asset(session, asset_id=target_id)
-        await session.flush()
-        return target_id
-    except Exception:
-        raise
-
-
-async def list_unhashed_candidates_under_prefixes(session: AsyncSession, *, prefixes: list[str]) -> list[int]:
-    if not prefixes:
-        return []
-
-    conds = []
-    for p in prefixes:
-        base = os.path.abspath(p)
-        if not base.endswith(os.sep):
-            base += os.sep
-        escaped, esc = escape_like_prefix(base)
-        conds.append(AssetCacheState.file_path.like(escaped + "%", escape=esc))
-
-    path_filter = sa.or_(*conds) if len(conds) > 1 else conds[0]
-    if session.bind.dialect.name == "postgresql":
-        stmt = (
-            sa.select(AssetCacheState.id)
-            .join(Asset, Asset.id == AssetCacheState.asset_id)
-            .where(Asset.hash.is_(None), path_filter)
-            .order_by(AssetCacheState.asset_id.asc(), AssetCacheState.id.asc())
-            .distinct(AssetCacheState.asset_id)
-        )
-    else:
-        first_id = sa.func.min(AssetCacheState.id).label("first_id")
-        stmt = (
-            sa.select(first_id)
-            .join(Asset, Asset.id == AssetCacheState.asset_id)
-            .where(Asset.hash.is_(None), path_filter)
-            .group_by(AssetCacheState.asset_id)
-            .order_by(first_id.asc())
-        )
-    return [int(x) for x in (await session.execute(stmt)).scalars().all()]
-
-
-async def list_verify_candidates_under_prefixes(
-    session: AsyncSession, *, prefixes: Sequence[str]
-) -> Union[list[int], Sequence[int]]:
-    if not prefixes:
-        return []
-    conds = []
-    for p in prefixes:
-        base = os.path.abspath(p)
-        if not base.endswith(os.sep):
-            base += os.sep
-        escaped, esc = escape_like_prefix(base)
-        conds.append(AssetCacheState.file_path.like(escaped + "%", escape=esc))
-
-    return (
-        await session.execute(
-            sa.select(AssetCacheState.id)
-            .where(AssetCacheState.needs_verify.is_(True))
-            .where(sa.or_(*conds))
-            .order_by(AssetCacheState.id.asc())
-        )
-    ).scalars().all()
-
-
-async def ingest_fs_asset(
-    session: AsyncSession,
-    *,
-    asset_hash: str,
-    abs_path: str,
-    size_bytes: int,
-    mtime_ns: int,
-    mime_type: Optional[str] = None,
-    info_name: Optional[str] = None,
-    owner_id: str = "",
-    preview_id: Optional[str] = None,
-    user_metadata: Optional[dict] = None,
-    tags: Sequence[str] = (),
-    tag_origin: str = "manual",
-    require_existing_tags: bool = False,
-) -> dict:
-    """
-    Idempotently upsert:
-      - Asset by content hash (create if missing)
-      - AssetCacheState(file_path) pointing to asset_id
-      - Optionally AssetInfo + tag links and metadata projection
-    Returns flags and ids.
-    """
-    locator = os.path.abspath(abs_path)
-    now = utcnow()
-    dialect = session.bind.dialect.name
-
-    if preview_id:
-        if not await session.get(Asset, preview_id):
-            preview_id = None
-
-    out: dict[str, Any] = {
-        "asset_created": False,
-        "asset_updated": False,
-        "state_created": False,
-        "state_updated": False,
-        "asset_info_id": None,
-    }
-
-    # 1) Asset by hash
-    asset = (
-        await session.execute(select(Asset).where(Asset.hash == asset_hash).limit(1))
-    ).scalars().first()
-    if not asset:
-        vals = {
-            "hash": asset_hash,
-            "size_bytes": int(size_bytes),
-            "mime_type": mime_type,
-            "created_at": now,
-        }
-        if dialect == "sqlite":
-            res = await session.execute(
-                d_sqlite.insert(Asset)
-                .values(**vals)
-                .on_conflict_do_nothing(index_elements=[Asset.hash])
-            )
-            if int(res.rowcount or 0) > 0:
-                out["asset_created"] = True
-            asset = (
-                await session.execute(
-                    select(Asset).where(Asset.hash == asset_hash).limit(1)
-                )
-            ).scalars().first()
-        elif dialect == "postgresql":
-            res = await session.execute(
-                d_pg.insert(Asset)
-                .values(**vals)
-                .on_conflict_do_nothing(
-                    index_elements=[Asset.hash],
-                    index_where=Asset.__table__.c.hash.isnot(None),
-                )
-                .returning(Asset.id)
-            )
-            inserted_id = res.scalar_one_or_none()
-            if inserted_id:
-                out["asset_created"] = True
-                asset = await session.get(Asset, inserted_id)
-            else:
-                asset = (
-                    await session.execute(
-                        select(Asset).where(Asset.hash == asset_hash).limit(1)
-                    )
-                ).scalars().first()
-        else:
-            raise NotImplementedError(f"Unsupported database dialect: {dialect}")
-        if not asset:
-            raise RuntimeError("Asset row not found after upsert.")
-    else:
-        changed = False
-        if asset.size_bytes != int(size_bytes) and int(size_bytes) > 0:
-            asset.size_bytes = int(size_bytes)
-            changed = True
-        if mime_type and asset.mime_type != mime_type:
-            asset.mime_type = mime_type
-            changed = True
-        if changed:
-            out["asset_updated"] = True
-
-    # 2) AssetCacheState upsert by file_path (unique)
-    vals = {
-        "asset_id": asset.id,
-        "file_path": locator,
-        "mtime_ns": int(mtime_ns),
-    }
-    if dialect == "sqlite":
-        ins = (
-            d_sqlite.insert(AssetCacheState)
-            .values(**vals)
-            .on_conflict_do_nothing(index_elements=[AssetCacheState.file_path])
-        )
-    elif dialect == "postgresql":
-        ins = (
-            d_pg.insert(AssetCacheState)
-            .values(**vals)
-            .on_conflict_do_nothing(index_elements=[AssetCacheState.file_path])
-        )
-    else:
-        raise NotImplementedError(f"Unsupported database dialect: {dialect}")
-
-    res = await session.execute(ins)
-    if int(res.rowcount or 0) > 0:
-        out["state_created"] = True
-    else:
-        upd = (
-            sa.update(AssetCacheState)
-            .where(AssetCacheState.file_path == locator)
-            .where(
-                sa.or_(
-                    AssetCacheState.asset_id != asset.id,
-                    AssetCacheState.mtime_ns.is_(None),
-                    AssetCacheState.mtime_ns != int(mtime_ns),
-                )
-            )
-            .values(asset_id=asset.id, mtime_ns=int(mtime_ns))
-        )
-        res2 = await session.execute(upd)
-        if int(res2.rowcount or 0) > 0:
-            out["state_updated"] = True
-
-    # 3) Optional AssetInfo + tags + metadata
-    if info_name:
-        try:
-            async with session.begin_nested():
-                info = AssetInfo(
-                    owner_id=owner_id,
-                    name=info_name,
-                    asset_id=asset.id,
-                    preview_id=preview_id,
-                    created_at=now,
-                    updated_at=now,
-                    last_access_time=now,
-                )
-                session.add(info)
-                await session.flush()
-                out["asset_info_id"] = info.id
-        except IntegrityError:
-            pass
-
-        existing_info = (
-            await session.execute(
-                select(AssetInfo)
-                .where(
-                    AssetInfo.asset_id == asset.id,
-                    AssetInfo.name == info_name,
-                    (AssetInfo.owner_id == owner_id),
-                )
-                .limit(1)
-            )
-        ).unique().scalar_one_or_none()
-        if not existing_info:
-            raise RuntimeError("Failed to update or insert AssetInfo.")
-
-        if preview_id and existing_info.preview_id != preview_id:
-            existing_info.preview_id = preview_id
-
-        existing_info.updated_at = now
-        if existing_info.last_access_time < now:
-            existing_info.last_access_time = now
-        await session.flush()
-        out["asset_info_id"] = existing_info.id
-
-        norm = [t.strip().lower() for t in (tags or []) if (t or "").strip()]
-        if norm and out["asset_info_id"] is not None:
-            if not require_existing_tags:
-                await ensure_tags_exist(session, norm, tag_type="user")
-
-            existing_tag_names = set(
-                name for (name,) in (await session.execute(select(Tag.name).where(Tag.name.in_(norm)))).all()
-            )
-            missing = [t for t in norm if t not in existing_tag_names]
-            if missing and require_existing_tags:
-                raise ValueError(f"Unknown tags: {missing}")
-
-            existing_links = set(
-                tag_name
-                for (tag_name,) in (
-                    await session.execute(
-                        select(AssetInfoTag.tag_name).where(AssetInfoTag.asset_info_id == out["asset_info_id"])
-                    )
-                ).all()
-            )
-            to_add = [t for t in norm if t in existing_tag_names and t not in existing_links]
-            if to_add:
-                session.add_all(
-                    [
-                        AssetInfoTag(
-                            asset_info_id=out["asset_info_id"],
-                            tag_name=t,
-                            origin=tag_origin,
-                            added_at=now,
-                        )
-                        for t in to_add
-                    ]
-                )
-                await session.flush()
-
-        # metadata["filename"] hack
-        if out["asset_info_id"] is not None:
-            primary_path = pick_best_live_path(await list_cache_states_by_asset_id(session, asset_id=asset.id))
-            computed_filename = compute_relative_filename(primary_path) if primary_path else None
-
-            current_meta = existing_info.user_metadata or {}
-            new_meta = dict(current_meta)
-            if user_metadata is not None:
-                for k, v in user_metadata.items():
-                    new_meta[k] = v
-            if computed_filename:
-                new_meta["filename"] = computed_filename
-
-            if new_meta != current_meta:
-                await replace_asset_info_metadata_projection(
-                    session,
-                    asset_info_id=out["asset_info_id"],
-                    user_metadata=new_meta,
-                )
-
-    try:
-        await remove_missing_tag_for_asset_id(session, asset_id=asset.id)
-    except Exception:
-        logging.exception("Failed to clear 'missing' tag for asset %s", asset.id)
-    return out
-
-
-async def touch_asset_infos_by_fs_path(
-    session: AsyncSession,
-    *,
-    file_path: str,
-    ts: Optional[datetime] = None,
-    only_if_newer: bool = True,
-) -> None:
-    locator = os.path.abspath(file_path)
-    ts = ts or utcnow()
-    stmt = sa.update(AssetInfo).where(
-        sa.exists(
-            sa.select(sa.literal(1))
-            .select_from(AssetCacheState)
-            .where(
-                AssetCacheState.asset_id == AssetInfo.asset_id,
-                AssetCacheState.file_path == locator,
-            )
-        )
-    )
-    if only_if_newer:
-        stmt = stmt.where(
-            sa.or_(
-                AssetInfo.last_access_time.is_(None),
-                AssetInfo.last_access_time < ts,
-            )
-        )
-    await session.execute(stmt.values(last_access_time=ts))
-
-
-async def list_cache_states_with_asset_under_prefixes(
-    session: AsyncSession,
-    *,
-    prefixes: Sequence[str],
-) -> list[tuple[AssetCacheState, Optional[str], int]]:
-    """Return (AssetCacheState, asset_hash, size_bytes) for rows under any prefix."""
-    if not prefixes:
-        return []
-
-    conds = []
-    for p in prefixes:
-        if not p:
-            continue
-        base = os.path.abspath(p)
-        if not base.endswith(os.sep):
-            base = base + os.sep
-        escaped, esc = escape_like_prefix(base)
-        conds.append(AssetCacheState.file_path.like(escaped + "%", escape=esc))
-
-    if not conds:
-        return []
-
-    rows = (
-        await session.execute(
-            select(AssetCacheState, Asset.hash, Asset.size_bytes)
-            .join(Asset, Asset.id == AssetCacheState.asset_id)
-            .where(sa.or_(*conds))
-            .order_by(AssetCacheState.id.asc())
-        )
-    ).all()
-    return [(r[0], r[1], int(r[2] or 0)) for r in rows]
-
-
-async def _recompute_and_apply_filename_for_asset(session: AsyncSession, *, asset_id: str) -> None:
-    """Compute filename from the first *existing* cache state path and apply it to all AssetInfo (if changed)."""
-    try:
-        primary_path = pick_best_live_path(await list_cache_states_by_asset_id(session, asset_id=asset_id))
-        if not primary_path:
-            return
-        new_filename = compute_relative_filename(primary_path)
-        if not new_filename:
-            return
-        infos = (
-            await session.execute(select(AssetInfo).where(AssetInfo.asset_id == asset_id))
-        ).scalars().all()
-        for info in infos:
-            current_meta = info.user_metadata or {}
-            if current_meta.get("filename") == new_filename:
-                continue
-            updated = dict(current_meta)
-            updated["filename"] = new_filename
-            await replace_asset_info_metadata_projection(session, asset_info_id=info.id, user_metadata=updated)
-    except Exception:
-        logging.exception("Failed to recompute filename metadata for asset %s", asset_id)

app/assets/database/services/info.py

@@ -1,586 +0,0 @@
-from collections import defaultdict
-from datetime import datetime
-from typing import Any, Optional, Sequence
-
-import sqlalchemy as sa
-from sqlalchemy import delete, func, select
-from sqlalchemy.exc import IntegrityError
-from sqlalchemy.ext.asyncio import AsyncSession
-from sqlalchemy.orm import contains_eager, noload
-
-from ..._helpers import compute_relative_filename, normalize_tags
-from ..helpers import (
-    apply_metadata_filter,
-    apply_tag_filters,
-    ensure_tags_exist,
-    escape_like_prefix,
-    project_kv,
-    visible_owner_clause,
-)
-from ..models import Asset, AssetInfo, AssetInfoMeta, AssetInfoTag, Tag
-from ..timeutil import utcnow
-from .queries import (
-    get_asset_by_hash,
-    list_cache_states_by_asset_id,
-    pick_best_live_path,
-)
-
-
-async def list_asset_infos_page(
-    session: AsyncSession,
-    *,
-    owner_id: str = "",
-    include_tags: Optional[Sequence[str]] = None,
-    exclude_tags: Optional[Sequence[str]] = None,
-    name_contains: Optional[str] = None,
-    metadata_filter: Optional[dict] = None,
-    limit: int = 20,
-    offset: int = 0,
-    sort: str = "created_at",
-    order: str = "desc",
-) -> tuple[list[AssetInfo], dict[str, list[str]], int]:
-    base = (
-        select(AssetInfo)
-        .join(Asset, Asset.id == AssetInfo.asset_id)
-        .options(contains_eager(AssetInfo.asset), noload(AssetInfo.tags))
-        .where(visible_owner_clause(owner_id))
-    )
-
-    if name_contains:
-        escaped, esc = escape_like_prefix(name_contains)
-        base = base.where(AssetInfo.name.ilike(f"%{escaped}%", escape=esc))
-
-    base = apply_tag_filters(base, include_tags, exclude_tags)
-    base = apply_metadata_filter(base, metadata_filter)
-
-    sort = (sort or "created_at").lower()
-    order = (order or "desc").lower()
-    sort_map = {
-        "name": AssetInfo.name,
-        "created_at": AssetInfo.created_at,
-        "updated_at": AssetInfo.updated_at,
-        "last_access_time": AssetInfo.last_access_time,
-        "size": Asset.size_bytes,
-    }
-    sort_col = sort_map.get(sort, AssetInfo.created_at)
-    sort_exp = sort_col.desc() if order == "desc" else sort_col.asc()
-
-    base = base.order_by(sort_exp).limit(limit).offset(offset)
-
-    count_stmt = (
-        select(func.count())
-        .select_from(AssetInfo)
-        .join(Asset, Asset.id == AssetInfo.asset_id)
-        .where(visible_owner_clause(owner_id))
-    )
-    if name_contains:
-        escaped, esc = escape_like_prefix(name_contains)
-        count_stmt = count_stmt.where(AssetInfo.name.ilike(f"%{escaped}%", escape=esc))
-    count_stmt = apply_tag_filters(count_stmt, include_tags, exclude_tags)
-    count_stmt = apply_metadata_filter(count_stmt, metadata_filter)
-
-    total = int((await session.execute(count_stmt)).scalar_one() or 0)
-
-    infos = (await session.execute(base)).unique().scalars().all()
-
-    id_list: list[str] = [i.id for i in infos]
-    tag_map: dict[str, list[str]] = defaultdict(list)
-    if id_list:
-        rows = await session.execute(
-            select(AssetInfoTag.asset_info_id, Tag.name)
-            .join(Tag, Tag.name == AssetInfoTag.tag_name)
-            .where(AssetInfoTag.asset_info_id.in_(id_list))
-        )
-        for aid, tag_name in rows.all():
-            tag_map[aid].append(tag_name)
-
-    return infos, tag_map, total
-
-
-async def fetch_asset_info_and_asset(
-    session: AsyncSession,
-    *,
-    asset_info_id: str,
-    owner_id: str = "",
-) -> Optional[tuple[AssetInfo, Asset]]:
-    stmt = (
-        select(AssetInfo, Asset)
-        .join(Asset, Asset.id == AssetInfo.asset_id)
-        .where(
-            AssetInfo.id == asset_info_id,
-            visible_owner_clause(owner_id),
-        )
-        .limit(1)
-        .options(noload(AssetInfo.tags))
-    )
-    row = await session.execute(stmt)
-    pair = row.first()
-    if not pair:
-        return None
-    return pair[0], pair[1]
-
-
-async def fetch_asset_info_asset_and_tags(
-    session: AsyncSession,
-    *,
-    asset_info_id: str,
-    owner_id: str = "",
-) -> Optional[tuple[AssetInfo, Asset, list[str]]]:
-    stmt = (
-        select(AssetInfo, Asset, Tag.name)
-        .join(Asset, Asset.id == AssetInfo.asset_id)
-        .join(AssetInfoTag, AssetInfoTag.asset_info_id == AssetInfo.id, isouter=True)
-        .join(Tag, Tag.name == AssetInfoTag.tag_name, isouter=True)
-        .where(
-            AssetInfo.id == asset_info_id,
-            visible_owner_clause(owner_id),
-        )
-        .options(noload(AssetInfo.tags))
-        .order_by(Tag.name.asc())
-    )
-
-    rows = (await session.execute(stmt)).all()
-    if not rows:
-        return None
-
-    first_info, first_asset, _ = rows[0]
-    tags: list[str] = []
-    seen: set[str] = set()
-    for _info, _asset, tag_name in rows:
-        if tag_name and tag_name not in seen:
-            seen.add(tag_name)
-            tags.append(tag_name)
-    return first_info, first_asset, tags
-
-
-async def create_asset_info_for_existing_asset(
-    session: AsyncSession,
-    *,
-    asset_hash: str,
-    name: str,
-    user_metadata: Optional[dict] = None,
-    tags: Optional[Sequence[str]] = None,
-    tag_origin: str = "manual",
-    owner_id: str = "",
-) -> AssetInfo:
-    """Create or return an existing AssetInfo for an Asset identified by asset_hash."""
-    now = utcnow()
-    asset = await get_asset_by_hash(session, asset_hash=asset_hash)
-    if not asset:
-        raise ValueError(f"Unknown asset hash {asset_hash}")
-
-    info = AssetInfo(
-        owner_id=owner_id,
-        name=name,
-        asset_id=asset.id,
-        preview_id=None,
-        created_at=now,
-        updated_at=now,
-        last_access_time=now,
-    )
-    try:
-        async with session.begin_nested():
-            session.add(info)
-            await session.flush()
-    except IntegrityError:
-        existing = (
-            await session.execute(
-                select(AssetInfo)
-                .options(noload(AssetInfo.tags))
-                .where(
-                    AssetInfo.asset_id == asset.id,
-                    AssetInfo.name == name,
-                    AssetInfo.owner_id == owner_id,
-                )
-                .limit(1)
-            )
-        ).unique().scalars().first()
-        if not existing:
-            raise RuntimeError("AssetInfo upsert failed to find existing row after conflict.")
-        return existing
-
-    # metadata["filename"] hack
-    new_meta = dict(user_metadata or {})
-    computed_filename = None
-    try:
-        p = pick_best_live_path(await list_cache_states_by_asset_id(session, asset_id=asset.id))
-        if p:
-            computed_filename = compute_relative_filename(p)
-    except Exception:
-        computed_filename = None
-    if computed_filename:
-        new_meta["filename"] = computed_filename
-    if new_meta:
-        await replace_asset_info_metadata_projection(
-            session,
-            asset_info_id=info.id,
-            user_metadata=new_meta,
-        )
-
-    if tags is not None:
-        await set_asset_info_tags(
-            session,
-            asset_info_id=info.id,
-            tags=tags,
-            origin=tag_origin,
-        )
-    return info
-
-
-async def set_asset_info_tags(
-    session: AsyncSession,
-    *,
-    asset_info_id: str,
-    tags: Sequence[str],
-    origin: str = "manual",
-) -> dict:
-    desired = normalize_tags(tags)
-
-    current = set(
-        tag_name for (tag_name,) in (
-            await session.execute(select(AssetInfoTag.tag_name).where(AssetInfoTag.asset_info_id == asset_info_id))
-        ).all()
-    )
-
-    to_add = [t for t in desired if t not in current]
-    to_remove = [t for t in current if t not in desired]
-
-    if to_add:
-        await ensure_tags_exist(session, to_add, tag_type="user")
-        session.add_all([
-            AssetInfoTag(asset_info_id=asset_info_id, tag_name=t, origin=origin, added_at=utcnow())
-            for t in to_add
-        ])
-        await session.flush()
-
-    if to_remove:
-        await session.execute(
-            delete(AssetInfoTag)
-            .where(AssetInfoTag.asset_info_id == asset_info_id, AssetInfoTag.tag_name.in_(to_remove))
-        )
-        await session.flush()
-
-    return {"added": to_add, "removed": to_remove, "total": desired}
-
-
-async def update_asset_info_full(
-    session: AsyncSession,
-    *,
-    asset_info_id: str,
-    name: Optional[str] = None,
-    tags: Optional[Sequence[str]] = None,
-    user_metadata: Optional[dict] = None,
-    tag_origin: str = "manual",
-    asset_info_row: Any = None,
-) -> AssetInfo:
-    if not asset_info_row:
-        info = await session.get(AssetInfo, asset_info_id)
-        if not info:
-            raise ValueError(f"AssetInfo {asset_info_id} not found")
-    else:
-        info = asset_info_row
-
-    touched = False
-    if name is not None and name != info.name:
-        info.name = name
-        touched = True
-
-    computed_filename = None
-    try:
-        p = pick_best_live_path(await list_cache_states_by_asset_id(session, asset_id=info.asset_id))
-        if p:
-            computed_filename = compute_relative_filename(p)
-    except Exception:
-        computed_filename = None
-
-    if user_metadata is not None:
-        new_meta = dict(user_metadata)
-        if computed_filename:
-            new_meta["filename"] = computed_filename
-        await replace_asset_info_metadata_projection(
-            session, asset_info_id=asset_info_id, user_metadata=new_meta
-        )
-        touched = True
-    else:
-        if computed_filename:
-            current_meta = info.user_metadata or {}
-            if current_meta.get("filename") != computed_filename:
-                new_meta = dict(current_meta)
-                new_meta["filename"] = computed_filename
-                await replace_asset_info_metadata_projection(
-                    session, asset_info_id=asset_info_id, user_metadata=new_meta
-                )
-                touched = True
-
-    if tags is not None:
-        await set_asset_info_tags(
-            session,
-            asset_info_id=asset_info_id,
-            tags=tags,
-            origin=tag_origin,
-        )
-        touched = True
-
-    if touched and user_metadata is None:
-        info.updated_at = utcnow()
-        await session.flush()
-
-    return info
-
-
-async def replace_asset_info_metadata_projection(
-    session: AsyncSession,
-    *,
-    asset_info_id: str,
-    user_metadata: Optional[dict],
-) -> None:
-    info = await session.get(AssetInfo, asset_info_id)
-    if not info:
-        raise ValueError(f"AssetInfo {asset_info_id} not found")
-
-    info.user_metadata = user_metadata or {}
-    info.updated_at = utcnow()
-    await session.flush()
-
-    await session.execute(delete(AssetInfoMeta).where(AssetInfoMeta.asset_info_id == asset_info_id))
-    await session.flush()
-
-    if not user_metadata:
-        return
-
-    rows: list[AssetInfoMeta] = []
-    for k, v in user_metadata.items():
-        for r in project_kv(k, v):
-            rows.append(
-                AssetInfoMeta(
-                    asset_info_id=asset_info_id,
-                    key=r["key"],
-                    ordinal=int(r["ordinal"]),
-                    val_str=r.get("val_str"),
-                    val_num=r.get("val_num"),
-                    val_bool=r.get("val_bool"),
-                    val_json=r.get("val_json"),
-                )
-            )
-    if rows:
-        session.add_all(rows)
-        await session.flush()
-
-
-async def touch_asset_info_by_id(
-    session: AsyncSession,
-    *,
-    asset_info_id: str,
-    ts: Optional[datetime] = None,
-    only_if_newer: bool = True,
-) -> None:
-    ts = ts or utcnow()
-    stmt = sa.update(AssetInfo).where(AssetInfo.id == asset_info_id)
-    if only_if_newer:
-        stmt = stmt.where(
-            sa.or_(AssetInfo.last_access_time.is_(None), AssetInfo.last_access_time < ts)
-        )
-    await session.execute(stmt.values(last_access_time=ts))
-
-
-async def delete_asset_info_by_id(session: AsyncSession, *, asset_info_id: str, owner_id: str) -> bool:
-    stmt = sa.delete(AssetInfo).where(
-        AssetInfo.id == asset_info_id,
-        visible_owner_clause(owner_id),
-    )
-    return int((await session.execute(stmt)).rowcount or 0) > 0
-
-
-async def add_tags_to_asset_info(
-    session: AsyncSession,
-    *,
-    asset_info_id: str,
-    tags: Sequence[str],
-    origin: str = "manual",
-    create_if_missing: bool = True,
-    asset_info_row: Any = None,
-) -> dict:
-    if not asset_info_row:
-        info = await session.get(AssetInfo, asset_info_id)
-        if not info:
-            raise ValueError(f"AssetInfo {asset_info_id} not found")
-
-    norm = normalize_tags(tags)
-    if not norm:
-        total = await get_asset_tags(session, asset_info_id=asset_info_id)
-        return {"added": [], "already_present": [], "total_tags": total}
-
-    if create_if_missing:
-        await ensure_tags_exist(session, norm, tag_type="user")
-
-    current = {
-        tag_name
-        for (tag_name,) in (
-            await session.execute(
-                sa.select(AssetInfoTag.tag_name).where(AssetInfoTag.asset_info_id == asset_info_id)
-            )
-        ).all()
-    }
-
-    want = set(norm)
-    to_add = sorted(want - current)
-
-    if to_add:
-        async with session.begin_nested() as nested:
-            try:
-                session.add_all(
-                    [
-                        AssetInfoTag(
-                            asset_info_id=asset_info_id,
-                            tag_name=t,
-                            origin=origin,
-                            added_at=utcnow(),
-                        )
-                        for t in to_add
-                    ]
-                )
-                await session.flush()
-            except IntegrityError:
-                await nested.rollback()
-
-    after = set(await get_asset_tags(session, asset_info_id=asset_info_id))
-    return {
-        "added": sorted(((after - current) & want)),
-        "already_present": sorted(want & current),
-        "total_tags": sorted(after),
-    }
-
-
-async def remove_tags_from_asset_info(
-    session: AsyncSession,
-    *,
-    asset_info_id: str,
-    tags: Sequence[str],
-) -> dict:
-    info = await session.get(AssetInfo, asset_info_id)
-    if not info:
-        raise ValueError(f"AssetInfo {asset_info_id} not found")
-
-    norm = normalize_tags(tags)
-    if not norm:
-        total = await get_asset_tags(session, asset_info_id=asset_info_id)
-        return {"removed": [], "not_present": [], "total_tags": total}
-
-    existing = {
-        tag_name
-        for (tag_name,) in (
-            await session.execute(
-                sa.select(AssetInfoTag.tag_name).where(AssetInfoTag.asset_info_id == asset_info_id)
-            )
-        ).all()
-    }
-
-    to_remove = sorted(set(t for t in norm if t in existing))
-    not_present = sorted(set(t for t in norm if t not in existing))
-
-    if to_remove:
-        await session.execute(
-            delete(AssetInfoTag)
-            .where(
-                AssetInfoTag.asset_info_id == asset_info_id,
-                AssetInfoTag.tag_name.in_(to_remove),
-            )
-        )
-        await session.flush()
-
-    total = await get_asset_tags(session, asset_info_id=asset_info_id)
-    return {"removed": to_remove, "not_present": not_present, "total_tags": total}
-
-
-async def list_tags_with_usage(
-    session: AsyncSession,
-    *,
-    prefix: Optional[str] = None,
-    limit: int = 100,
-    offset: int = 0,
-    include_zero: bool = True,
-    order: str = "count_desc",
-    owner_id: str = "",
-) -> tuple[list[tuple[str, str, int]], int]:
-    counts_sq = (
-        select(
-            AssetInfoTag.tag_name.label("tag_name"),
-            func.count(AssetInfoTag.asset_info_id).label("cnt"),
-        )
-        .select_from(AssetInfoTag)
-        .join(AssetInfo, AssetInfo.id == AssetInfoTag.asset_info_id)
-        .where(visible_owner_clause(owner_id))
-        .group_by(AssetInfoTag.tag_name)
-        .subquery()
-    )
-
-    q = (
-        select(
-            Tag.name,
-            Tag.tag_type,
-            func.coalesce(counts_sq.c.cnt, 0).label("count"),
-        )
-        .select_from(Tag)
-        .join(counts_sq, counts_sq.c.tag_name == Tag.name, isouter=True)
-    )
-
-    if prefix:
-        escaped, esc = escape_like_prefix(prefix.strip().lower())
-        q = q.where(Tag.name.like(escaped + "%", escape=esc))
-
-    if not include_zero:
-        q = q.where(func.coalesce(counts_sq.c.cnt, 0) > 0)
-
-    if order == "name_asc":
-        q = q.order_by(Tag.name.asc())
-    else:
-        q = q.order_by(func.coalesce(counts_sq.c.cnt, 0).desc(), Tag.name.asc())
-
-    total_q = select(func.count()).select_from(Tag)
-    if prefix:
-        escaped, esc = escape_like_prefix(prefix.strip().lower())
-        total_q = total_q.where(Tag.name.like(escaped + "%", escape=esc))
-    if not include_zero:
-        total_q = total_q.where(
-            Tag.name.in_(select(AssetInfoTag.tag_name).group_by(AssetInfoTag.tag_name))
-        )
-
-    rows = (await session.execute(q.limit(limit).offset(offset))).all()
-    total = (await session.execute(total_q)).scalar_one()
-
-    rows_norm = [(name, ttype, int(count or 0)) for (name, ttype, count) in rows]
-    return rows_norm, int(total or 0)
-
-
-async def get_asset_tags(session: AsyncSession, *, asset_info_id: str) -> list[str]:
-    return [
-        tag_name
-        for (tag_name,) in (
-            await session.execute(
-                sa.select(AssetInfoTag.tag_name).where(AssetInfoTag.asset_info_id == asset_info_id)
-            )
-        ).all()
-    ]
-
-
-async def set_asset_info_preview(
-    session: AsyncSession,
-    *,
-    asset_info_id: str,
-    preview_asset_id: Optional[str],
-) -> None:
-    """Set or clear preview_id and bump updated_at. Raises on unknown IDs."""
-    info = await session.get(AssetInfo, asset_info_id)
-    if not info:
-        raise ValueError(f"AssetInfo {asset_info_id} not found")
-
-    if preview_asset_id is None:
-        info.preview_id = None
-    else:
-        # validate preview asset exists
-        if not await session.get(Asset, preview_asset_id):
-            raise ValueError(f"Preview Asset {preview_asset_id} not found")
-        info.preview_id = preview_asset_id
-
-    info.updated_at = utcnow()
-    await session.flush()

app/assets/database/services/queries.py

@@ -1,76 +0,0 @@
-import os
-from typing import Optional, Sequence, Union
-
-import sqlalchemy as sa
-from sqlalchemy import select
-from sqlalchemy.ext.asyncio import AsyncSession
-
-from ..models import Asset, AssetCacheState, AssetInfo
-
-
-async def asset_exists_by_hash(session: AsyncSession, *, asset_hash: str) -> bool:
-    row = (
-        await session.execute(
-            select(sa.literal(True)).select_from(Asset).where(Asset.hash == asset_hash).limit(1)
-        )
-    ).first()
-    return row is not None
-
-
-async def get_asset_by_hash(session: AsyncSession, *, asset_hash: str) -> Optional[Asset]:
-    return (
-        await session.execute(select(Asset).where(Asset.hash == asset_hash).limit(1))
-    ).scalars().first()
-
-
-async def get_asset_info_by_id(session: AsyncSession, *, asset_info_id: str) -> Optional[AssetInfo]:
-    return await session.get(AssetInfo, asset_info_id)
-
-
-async def asset_info_exists_for_asset_id(session: AsyncSession, *, asset_id: str) -> bool:
-    q = (
-        select(sa.literal(True))
-        .select_from(AssetInfo)
-        .where(AssetInfo.asset_id == asset_id)
-        .limit(1)
-    )
-    return (await session.execute(q)).first() is not None
-
-
-async def get_cache_state_by_asset_id(session: AsyncSession, *, asset_id: str) -> Optional[AssetCacheState]:
-    return (
-        await session.execute(
-            select(AssetCacheState)
-            .where(AssetCacheState.asset_id == asset_id)
-            .order_by(AssetCacheState.id.asc())
-            .limit(1)
-        )
-    ).scalars().first()
-
-
-async def list_cache_states_by_asset_id(
-    session: AsyncSession, *, asset_id: str
-) -> Union[list[AssetCacheState], Sequence[AssetCacheState]]:
-    return (
-        await session.execute(
-            select(AssetCacheState)
-            .where(AssetCacheState.asset_id == asset_id)
-            .order_by(AssetCacheState.id.asc())
-        )
-    ).scalars().all()
-
-
-def pick_best_live_path(states: Union[list[AssetCacheState], Sequence[AssetCacheState]]) -> str:
-    """
-    Return the best on-disk path among cache states:
-      1) Prefer a path that exists with needs_verify == False (already verified).
-      2) Otherwise, pick the first path that exists.
-      3) Otherwise return empty string.
-    """
-    alive = [s for s in states if getattr(s, "file_path", None) and os.path.isfile(s.file_path)]
-    if not alive:
-        return ""
-    for s in alive:
-        if not getattr(s, "needs_verify", False):
-            return s.file_path
-    return alive[0].file_path

app/assets/database/timeutil.py

@@ -1,6 +0,0 @@
-from datetime import datetime, timezone
-
-
-def utcnow() -> datetime:
-    """Naive UTC timestamp (no tzinfo). We always treat DB datetimes as UTC."""
-    return datetime.now(timezone.utc).replace(tzinfo=None)

app/assets/manager.py

@@ -1,556 +0,0 @@
-import contextlib
-import logging
-import mimetypes
-import os
-from typing import Optional, Sequence
-
-from comfy_api.internal import async_to_sync
-
-from ..db import create_session
-from ._helpers import (
-    ensure_within_base,
-    get_name_and_tags_from_asset_path,
-    resolve_destination_from_tags,
-)
-from .api import schemas_in, schemas_out
-from .database.models import Asset
-from .database.services import (
-    add_tags_to_asset_info,
-    asset_exists_by_hash,
-    asset_info_exists_for_asset_id,
-    check_fs_asset_exists_quick,
-    create_asset_info_for_existing_asset,
-    delete_asset_info_by_id,
-    fetch_asset_info_and_asset,
-    fetch_asset_info_asset_and_tags,
-    get_asset_by_hash,
-    get_asset_info_by_id,
-    get_asset_tags,
-    ingest_fs_asset,
-    list_asset_infos_page,
-    list_cache_states_by_asset_id,
-    list_tags_with_usage,
-    pick_best_live_path,
-    remove_tags_from_asset_info,
-    set_asset_info_preview,
-    touch_asset_info_by_id,
-    touch_asset_infos_by_fs_path,
-    update_asset_info_full,
-)
-from .storage import hashing
-
-
-async def asset_exists(*, asset_hash: str) -> bool:
-    async with await create_session() as session:
-        return await asset_exists_by_hash(session, asset_hash=asset_hash)
-
-
-def populate_db_with_asset(file_path: str, tags: Optional[list[str]] = None) -> None:
-    if tags is None:
-        tags = []
-    try:
-        asset_name, path_tags = get_name_and_tags_from_asset_path(file_path)
-        async_to_sync.AsyncToSyncConverter.run_async_in_thread(
-            add_local_asset,
-            tags=list(dict.fromkeys([*path_tags, *tags])),
-            file_name=asset_name,
-            file_path=file_path,
-        )
-    except ValueError as e:
-        logging.warning("Skipping non-asset path %s: %s", file_path, e)
-
-
-async def add_local_asset(tags: list[str], file_name: str, file_path: str) -> None:
-    abs_path = os.path.abspath(file_path)
-    size_bytes, mtime_ns = _get_size_mtime_ns(abs_path)
-    if not size_bytes:
-        return
-
-    async with await create_session() as session:
-        if await check_fs_asset_exists_quick(session, file_path=abs_path, size_bytes=size_bytes, mtime_ns=mtime_ns):
-            await touch_asset_infos_by_fs_path(session, file_path=abs_path)
-            await session.commit()
-            return
-
-    asset_hash = hashing.blake3_hash_sync(abs_path)
-
-    async with await create_session() as session:
-        await ingest_fs_asset(
-            session,
-            asset_hash="blake3:" + asset_hash,
-            abs_path=abs_path,
-            size_bytes=size_bytes,
-            mtime_ns=mtime_ns,
-            mime_type=None,
-            info_name=file_name,
-            tag_origin="automatic",
-            tags=tags,
-        )
-        await session.commit()
-
-
-async def list_assets(
-    *,
-    include_tags: Optional[Sequence[str]] = None,
-    exclude_tags: Optional[Sequence[str]] = None,
-    name_contains: Optional[str] = None,
-    metadata_filter: Optional[dict] = None,
-    limit: int = 20,
-    offset: int = 0,
-    sort: str = "created_at",
-    order: str = "desc",
-    owner_id: str = "",
-) -> schemas_out.AssetsList:
-    sort = _safe_sort_field(sort)
-    order = "desc" if (order or "desc").lower() not in {"asc", "desc"} else order.lower()
-
-    async with await create_session() as session:
-        infos, tag_map, total = await list_asset_infos_page(
-            session,
-            owner_id=owner_id,
-            include_tags=include_tags,
-            exclude_tags=exclude_tags,
-            name_contains=name_contains,
-            metadata_filter=metadata_filter,
-            limit=limit,
-            offset=offset,
-            sort=sort,
-            order=order,
-        )
-
-    summaries: list[schemas_out.AssetSummary] = []
-    for info in infos:
-        asset = info.asset
-        tags = tag_map.get(info.id, [])
-        summaries.append(
-            schemas_out.AssetSummary(
-                id=info.id,
-                name=info.name,
-                asset_hash=asset.hash if asset else None,
-                size=int(asset.size_bytes) if asset else None,
-                mime_type=asset.mime_type if asset else None,
-                tags=tags,
-                preview_url=f"/api/assets/{info.id}/content",
-                created_at=info.created_at,
-                updated_at=info.updated_at,
-                last_access_time=info.last_access_time,
-            )
-        )
-
-    return schemas_out.AssetsList(
-        assets=summaries,
-        total=total,
-        has_more=(offset + len(summaries)) < total,
-    )
-
-
-async def get_asset(*, asset_info_id: str, owner_id: str = "") -> schemas_out.AssetDetail:
-    async with await create_session() as session:
-        res = await fetch_asset_info_asset_and_tags(session, asset_info_id=asset_info_id, owner_id=owner_id)
-        if not res:
-            raise ValueError(f"AssetInfo {asset_info_id} not found")
-        info, asset, tag_names = res
-        preview_id = info.preview_id
-
-    return schemas_out.AssetDetail(
-        id=info.id,
-        name=info.name,
-        asset_hash=asset.hash if asset else None,
-        size=int(asset.size_bytes) if asset and asset.size_bytes is not None else None,
-        mime_type=asset.mime_type if asset else None,
-        tags=tag_names,
-        user_metadata=info.user_metadata or {},
-        preview_id=preview_id,
-        created_at=info.created_at,
-        last_access_time=info.last_access_time,
-    )
-
-
-async def resolve_asset_content_for_download(
-    *,
-    asset_info_id: str,
-    owner_id: str = "",
-) -> tuple[str, str, str]:
-    async with await create_session() as session:
-        pair = await fetch_asset_info_and_asset(session, asset_info_id=asset_info_id, owner_id=owner_id)
-        if not pair:
-            raise ValueError(f"AssetInfo {asset_info_id} not found")
-
-        info, asset = pair
-        states = await list_cache_states_by_asset_id(session, asset_id=asset.id)
-        abs_path = pick_best_live_path(states)
-        if not abs_path:
-            raise FileNotFoundError
-
-        await touch_asset_info_by_id(session, asset_info_id=asset_info_id)
-        await session.commit()
-
-    ctype = asset.mime_type or mimetypes.guess_type(info.name or abs_path)[0] or "application/octet-stream"
-    download_name = info.name or os.path.basename(abs_path)
-    return abs_path, ctype, download_name
-
-
-async def upload_asset_from_temp_path(
-    spec: schemas_in.UploadAssetSpec,
-    *,
-    temp_path: str,
-    client_filename: Optional[str] = None,
-    owner_id: str = "",
-    expected_asset_hash: Optional[str] = None,
-) -> schemas_out.AssetCreated:
-    try:
-        digest = await hashing.blake3_hash(temp_path)
-    except Exception as e:
-        raise RuntimeError(f"failed to hash uploaded file: {e}")
-    asset_hash = "blake3:" + digest
-
-    if expected_asset_hash and asset_hash != expected_asset_hash.strip().lower():
-        raise ValueError("HASH_MISMATCH")
-
-    async with await create_session() as session:
-        existing = await get_asset_by_hash(session, asset_hash=asset_hash)
-        if existing is not None:
-            with contextlib.suppress(Exception):
-                if temp_path and os.path.exists(temp_path):
-                    os.remove(temp_path)
-
-            display_name = _safe_filename(spec.name or (client_filename or ""), fallback=digest)
-            info = await create_asset_info_for_existing_asset(
-                session,
-                asset_hash=asset_hash,
-                name=display_name,
-                user_metadata=spec.user_metadata or {},
-                tags=spec.tags or [],
-                tag_origin="manual",
-                owner_id=owner_id,
-            )
-            tag_names = await get_asset_tags(session, asset_info_id=info.id)
-            await session.commit()
-
-            return schemas_out.AssetCreated(
-                id=info.id,
-                name=info.name,
-                asset_hash=existing.hash,
-                size=int(existing.size_bytes) if existing.size_bytes is not None else None,
-                mime_type=existing.mime_type,
-                tags=tag_names,
-                user_metadata=info.user_metadata or {},
-                preview_id=info.preview_id,
-                created_at=info.created_at,
-                last_access_time=info.last_access_time,
-                created_new=False,
-            )
-
-    base_dir, subdirs = resolve_destination_from_tags(spec.tags)
-    dest_dir = os.path.join(base_dir, *subdirs) if subdirs else base_dir
-    os.makedirs(dest_dir, exist_ok=True)
-
-    src_for_ext = (client_filename or spec.name or "").strip()
-    _ext = os.path.splitext(os.path.basename(src_for_ext))[1] if src_for_ext else ""
-    ext = _ext if 0 < len(_ext) <= 16 else ""
-    hashed_basename = f"{digest}{ext}"
-    dest_abs = os.path.abspath(os.path.join(dest_dir, hashed_basename))
-    ensure_within_base(dest_abs, base_dir)
-
-    content_type = (
-        mimetypes.guess_type(os.path.basename(src_for_ext), strict=False)[0]
-        or mimetypes.guess_type(hashed_basename, strict=False)[0]
-        or "application/octet-stream"
-    )
-
-    try:
-        os.replace(temp_path, dest_abs)
-    except Exception as e:
-        raise RuntimeError(f"failed to move uploaded file into place: {e}")
-
-    try:
-        size_bytes, mtime_ns = _get_size_mtime_ns(dest_abs)
-    except OSError as e:
-        raise RuntimeError(f"failed to stat destination file: {e}")
-
-    async with await create_session() as session:
-        result = await ingest_fs_asset(
-            session,
-            asset_hash=asset_hash,
-            abs_path=dest_abs,
-            size_bytes=size_bytes,
-            mtime_ns=mtime_ns,
-            mime_type=content_type,
-            info_name=_safe_filename(spec.name or (client_filename or ""), fallback=digest),
-            owner_id=owner_id,
-            preview_id=None,
-            user_metadata=spec.user_metadata or {},
-            tags=spec.tags,
-            tag_origin="manual",
-            require_existing_tags=False,
-        )
-        info_id = result["asset_info_id"]
-        if not info_id:
-            raise RuntimeError("failed to create asset metadata")
-
-        pair = await fetch_asset_info_and_asset(session, asset_info_id=info_id, owner_id=owner_id)
-        if not pair:
-            raise RuntimeError("inconsistent DB state after ingest")
-        info, asset = pair
-        tag_names = await get_asset_tags(session, asset_info_id=info.id)
-        await session.commit()
-
-    return schemas_out.AssetCreated(
-        id=info.id,
-        name=info.name,
-        asset_hash=asset.hash,
-        size=int(asset.size_bytes),
-        mime_type=asset.mime_type,
-        tags=tag_names,
-        user_metadata=info.user_metadata or {},
-        preview_id=info.preview_id,
-        created_at=info.created_at,
-        last_access_time=info.last_access_time,
-        created_new=result["asset_created"],
-    )
-
-
-async def update_asset(
-    *,
-    asset_info_id: str,
-    name: Optional[str] = None,
-    tags: Optional[list[str]] = None,
-    user_metadata: Optional[dict] = None,
-    owner_id: str = "",
-) -> schemas_out.AssetUpdated:
-    async with await create_session() as session:
-        info_row = await get_asset_info_by_id(session, asset_info_id=asset_info_id)
-        if not info_row:
-            raise ValueError(f"AssetInfo {asset_info_id} not found")
-        if info_row.owner_id and info_row.owner_id != owner_id:
-            raise PermissionError("not owner")
-
-        info = await update_asset_info_full(
-            session,
-            asset_info_id=asset_info_id,
-            name=name,
-            tags=tags,
-            user_metadata=user_metadata,
-            tag_origin="manual",
-            asset_info_row=info_row,
-        )
-
-        tag_names = await get_asset_tags(session, asset_info_id=asset_info_id)
-        await session.commit()
-
-    return schemas_out.AssetUpdated(
-        id=info.id,
-        name=info.name,
-        asset_hash=info.asset.hash if info.asset else None,
-        tags=tag_names,
-        user_metadata=info.user_metadata or {},
-        updated_at=info.updated_at,
-    )
-
-
-async def set_asset_preview(
-    *,
-    asset_info_id: str,
-    preview_asset_id: Optional[str],
-    owner_id: str = "",
-) -> schemas_out.AssetDetail:
-    async with await create_session() as session:
-        info_row = await get_asset_info_by_id(session, asset_info_id=asset_info_id)
-        if not info_row:
-            raise ValueError(f"AssetInfo {asset_info_id} not found")
-        if info_row.owner_id and info_row.owner_id != owner_id:
-            raise PermissionError("not owner")
-
-        await set_asset_info_preview(
-            session,
-            asset_info_id=asset_info_id,
-            preview_asset_id=preview_asset_id,
-        )
-
-        res = await fetch_asset_info_asset_and_tags(session, asset_info_id=asset_info_id, owner_id=owner_id)
-        if not res:
-            raise RuntimeError("State changed during preview update")
-        info, asset, tags = res
-        await session.commit()
-
-    return schemas_out.AssetDetail(
-        id=info.id,
-        name=info.name,
-        asset_hash=asset.hash if asset else None,
-        size=int(asset.size_bytes) if asset and asset.size_bytes is not None else None,
-        mime_type=asset.mime_type if asset else None,
-        tags=tags,
-        user_metadata=info.user_metadata or {},
-        preview_id=info.preview_id,
-        created_at=info.created_at,
-        last_access_time=info.last_access_time,
-    )
-
-
-async def delete_asset_reference(*, asset_info_id: str, owner_id: str, delete_content_if_orphan: bool = True) -> bool:
-    async with await create_session() as session:
-        info_row = await get_asset_info_by_id(session, asset_info_id=asset_info_id)
-        asset_id = info_row.asset_id if info_row else None
-        deleted = await delete_asset_info_by_id(session, asset_info_id=asset_info_id, owner_id=owner_id)
-        if not deleted:
-            await session.commit()
-            return False
-
-        if not delete_content_if_orphan or not asset_id:
-            await session.commit()
-            return True
-
-        still_exists = await asset_info_exists_for_asset_id(session, asset_id=asset_id)
-        if still_exists:
-            await session.commit()
-            return True
-
-        states = await list_cache_states_by_asset_id(session, asset_id=asset_id)
-        file_paths = [s.file_path for s in (states or []) if getattr(s, "file_path", None)]
-
-        asset_row = await session.get(Asset, asset_id)
-        if asset_row is not None:
-            await session.delete(asset_row)
-
-        await session.commit()
-        for p in file_paths:
-            with contextlib.suppress(Exception):
-                if p and os.path.isfile(p):
-                    os.remove(p)
-    return True
-
-
-async def create_asset_from_hash(
-    *,
-    hash_str: str,
-    name: str,
-    tags: Optional[list[str]] = None,
-    user_metadata: Optional[dict] = None,
-    owner_id: str = "",
-) -> Optional[schemas_out.AssetCreated]:
-    canonical = hash_str.strip().lower()
-    async with await create_session() as session:
-        asset = await get_asset_by_hash(session, asset_hash=canonical)
-        if not asset:
-            return None
-
-        info = await create_asset_info_for_existing_asset(
-            session,
-            asset_hash=canonical,
-            name=_safe_filename(name, fallback=canonical.split(":", 1)[1]),
-            user_metadata=user_metadata or {},
-            tags=tags or [],
-            tag_origin="manual",
-            owner_id=owner_id,
-        )
-        tag_names = await get_asset_tags(session, asset_info_id=info.id)
-        await session.commit()
-
-    return schemas_out.AssetCreated(
-        id=info.id,
-        name=info.name,
-        asset_hash=asset.hash,
-        size=int(asset.size_bytes),
-        mime_type=asset.mime_type,
-        tags=tag_names,
-        user_metadata=info.user_metadata or {},
-        preview_id=info.preview_id,
-        created_at=info.created_at,
-        last_access_time=info.last_access_time,
-        created_new=False,
-    )
-
-
-async def list_tags(
-    *,
-    prefix: Optional[str] = None,
-    limit: int = 100,
-    offset: int = 0,
-    order: str = "count_desc",
-    include_zero: bool = True,
-    owner_id: str = "",
-) -> schemas_out.TagsList:
-    limit = max(1, min(1000, limit))
-    offset = max(0, offset)
-
-    async with await create_session() as session:
-        rows, total = await list_tags_with_usage(
-            session,
-            prefix=prefix,
-            limit=limit,
-            offset=offset,
-            include_zero=include_zero,
-            order=order,
-            owner_id=owner_id,
-        )
-
-    tags = [schemas_out.TagUsage(name=name, count=count, type=tag_type) for (name, tag_type, count) in rows]
-    return schemas_out.TagsList(tags=tags, total=total, has_more=(offset + len(tags)) < total)
-
-
-async def add_tags_to_asset(
-    *,
-    asset_info_id: str,
-    tags: list[str],
-    origin: str = "manual",
-    owner_id: str = "",
-) -> schemas_out.TagsAdd:
-    async with await create_session() as session:
-        info_row = await get_asset_info_by_id(session, asset_info_id=asset_info_id)
-        if not info_row:
-            raise ValueError(f"AssetInfo {asset_info_id} not found")
-        if info_row.owner_id and info_row.owner_id != owner_id:
-            raise PermissionError("not owner")
-        data = await add_tags_to_asset_info(
-            session,
-            asset_info_id=asset_info_id,
-            tags=tags,
-            origin=origin,
-            create_if_missing=True,
-            asset_info_row=info_row,
-        )
-        await session.commit()
-    return schemas_out.TagsAdd(**data)
-
-
-async def remove_tags_from_asset(
-    *,
-    asset_info_id: str,
-    tags: list[str],
-    owner_id: str = "",
-) -> schemas_out.TagsRemove:
-    async with await create_session() as session:
-        info_row = await get_asset_info_by_id(session, asset_info_id=asset_info_id)
-        if not info_row:
-            raise ValueError(f"AssetInfo {asset_info_id} not found")
-        if info_row.owner_id and info_row.owner_id != owner_id:
-            raise PermissionError("not owner")
-
-        data = await remove_tags_from_asset_info(
-            session,
-            asset_info_id=asset_info_id,
-            tags=tags,
-        )
-        await session.commit()
-    return schemas_out.TagsRemove(**data)
-
-
-def _safe_sort_field(requested: Optional[str]) -> str:
-    if not requested:
-        return "created_at"
-    v = requested.lower()
-    if v in {"name", "created_at", "updated_at", "size", "last_access_time"}:
-        return v
-    return "created_at"
-
-
-def _get_size_mtime_ns(path: str) -> tuple[int, int]:
-    st = os.stat(path, follow_symlinks=True)
-    return st.st_size, getattr(st, "st_mtime_ns", int(st.st_mtime * 1_000_000_000))
-
-
-def _safe_filename(name: Optional[str], fallback: str) -> str:
-    n = os.path.basename((name or "").strip() or fallback)
-    if n:
-        return n
-    return fallback

app/assets/scanner.py

@@ -1,501 +0,0 @@
-import asyncio
-import contextlib
-import logging
-import os
-import time
-from dataclasses import dataclass, field
-from typing import Literal, Optional
-
-import sqlalchemy as sa
-
-import folder_paths
-
-from ..db import create_session
-from ._helpers import (
-    collect_models_files,
-    compute_relative_filename,
-    get_comfy_models_folders,
-    get_name_and_tags_from_asset_path,
-    list_tree,
-    new_scan_id,
-    prefixes_for_root,
-    ts_to_iso,
-)
-from .api import schemas_in, schemas_out
-from .database.helpers import (
-    add_missing_tag_for_asset_id,
-    ensure_tags_exist,
-    escape_like_prefix,
-    fast_asset_file_check,
-    remove_missing_tag_for_asset_id,
-    seed_from_paths_batch,
-)
-from .database.models import Asset, AssetCacheState, AssetInfo
-from .database.services import (
-    compute_hash_and_dedup_for_cache_state,
-    list_cache_states_by_asset_id,
-    list_cache_states_with_asset_under_prefixes,
-    list_unhashed_candidates_under_prefixes,
-    list_verify_candidates_under_prefixes,
-)
-
-LOGGER = logging.getLogger(__name__)
-
-SLOW_HASH_CONCURRENCY = 1
-
-
-@dataclass
-class ScanProgress:
-    scan_id: str
-    root: schemas_in.RootType
-    status: Literal["scheduled", "running", "completed", "failed", "cancelled"] = "scheduled"
-    scheduled_at: float = field(default_factory=lambda: time.time())
-    started_at: Optional[float] = None
-    finished_at: Optional[float] = None
-    discovered: int = 0
-    processed: int = 0
-    file_errors: list[dict] = field(default_factory=list)
-
-
-@dataclass
-class SlowQueueState:
-    queue: asyncio.Queue
-    workers: list[asyncio.Task] = field(default_factory=list)
-    closed: bool = False
-
-
-RUNNING_TASKS: dict[schemas_in.RootType, asyncio.Task] = {}
-PROGRESS_BY_ROOT: dict[schemas_in.RootType, ScanProgress] = {}
-SLOW_STATE_BY_ROOT: dict[schemas_in.RootType, SlowQueueState] = {}
-
-
-def current_statuses() -> schemas_out.AssetScanStatusResponse:
-    scans = []
-    for root in schemas_in.ALLOWED_ROOTS:
-        prog = PROGRESS_BY_ROOT.get(root)
-        if not prog:
-            continue
-        scans.append(_scan_progress_to_scan_status_model(prog))
-    return schemas_out.AssetScanStatusResponse(scans=scans)
-
-
-async def schedule_scans(roots: list[schemas_in.RootType]) -> schemas_out.AssetScanStatusResponse:
-    results: list[ScanProgress] = []
-    for root in roots:
-        if root in RUNNING_TASKS and not RUNNING_TASKS[root].done():
-            results.append(PROGRESS_BY_ROOT[root])
-            continue
-
-        prog = ScanProgress(scan_id=new_scan_id(root), root=root, status="scheduled")
-        PROGRESS_BY_ROOT[root] = prog
-        state = SlowQueueState(queue=asyncio.Queue())
-        SLOW_STATE_BY_ROOT[root] = state
-        RUNNING_TASKS[root] = asyncio.create_task(
-            _run_hash_verify_pipeline(root, prog, state),
-            name=f"asset-scan:{root}",
-        )
-        results.append(prog)
-    return _status_response_for(results)
-
-
-async def sync_seed_assets(roots: list[schemas_in.RootType]) -> None:
-    t_total = time.perf_counter()
-    created = 0
-    skipped_existing = 0
-    paths: list[str] = []
-    try:
-        existing_paths: set[str] = set()
-        for r in roots:
-            try:
-                survivors = await _fast_db_consistency_pass(r, collect_existing_paths=True, update_missing_tags=True)
-                if survivors:
-                    existing_paths.update(survivors)
-            except Exception as ex:
-                LOGGER.exception("fast DB reconciliation failed for %s: %s", r, ex)
-
-        if "models" in roots:
-            paths.extend(collect_models_files())
-        if "input" in roots:
-            paths.extend(list_tree(folder_paths.get_input_directory()))
-        if "output" in roots:
-            paths.extend(list_tree(folder_paths.get_output_directory()))
-
-        specs: list[dict] = []
-        tag_pool: set[str] = set()
-        for p in paths:
-            ap = os.path.abspath(p)
-            if ap in existing_paths:
-                skipped_existing += 1
-                continue
-            try:
-                st = os.stat(ap, follow_symlinks=True)
-            except OSError:
-                continue
-            if not st.st_size:
-                continue
-            name, tags = get_name_and_tags_from_asset_path(ap)
-            specs.append(
-                {
-                    "abs_path": ap,
-                    "size_bytes": st.st_size,
-                    "mtime_ns": getattr(st, "st_mtime_ns", int(st.st_mtime * 1_000_000_000)),
-                    "info_name": name,
-                    "tags": tags,
-                    "fname": compute_relative_filename(ap),
-                }
-            )
-            for t in tags:
-                tag_pool.add(t)
-
-        if not specs:
-            return
-        async with await create_session() as sess:
-            if tag_pool:
-                await ensure_tags_exist(sess, tag_pool, tag_type="user")
-
-            result = await seed_from_paths_batch(sess, specs=specs, owner_id="")
-            created += result["inserted_infos"]
-            await sess.commit()
-    finally:
-        LOGGER.info(
-            "Assets scan(roots=%s) completed in %.3fs (created=%d, skipped_existing=%d, total_seen=%d)",
-            roots,
-            time.perf_counter() - t_total,
-            created,
-            skipped_existing,
-            len(paths),
-        )
-
-
-def _status_response_for(progresses: list[ScanProgress]) -> schemas_out.AssetScanStatusResponse:
-    return schemas_out.AssetScanStatusResponse(scans=[_scan_progress_to_scan_status_model(p) for p in progresses])
-
-
-def _scan_progress_to_scan_status_model(progress: ScanProgress) -> schemas_out.AssetScanStatus:
-    return schemas_out.AssetScanStatus(
-        scan_id=progress.scan_id,
-        root=progress.root,
-        status=progress.status,
-        scheduled_at=ts_to_iso(progress.scheduled_at),
-        started_at=ts_to_iso(progress.started_at),
-        finished_at=ts_to_iso(progress.finished_at),
-        discovered=progress.discovered,
-        processed=progress.processed,
-        file_errors=[
-            schemas_out.AssetScanError(
-                path=e.get("path", ""),
-                message=e.get("message", ""),
-                at=e.get("at"),
-            )
-            for e in (progress.file_errors or [])
-        ],
-    )
-
-
-async def _run_hash_verify_pipeline(root: schemas_in.RootType, prog: ScanProgress, state: SlowQueueState) -> None:
-    prog.status = "running"
-    prog.started_at = time.time()
-    try:
-        prefixes = prefixes_for_root(root)
-
-        await _fast_db_consistency_pass(root)
-
-        # collect candidates from DB
-        async with await create_session() as sess:
-            verify_ids = await list_verify_candidates_under_prefixes(sess, prefixes=prefixes)
-            unhashed_ids = await list_unhashed_candidates_under_prefixes(sess, prefixes=prefixes)
-        # dedupe: prioritize verification first
-        seen = set()
-        ordered: list[int] = []
-        for lst in (verify_ids, unhashed_ids):
-            for sid in lst:
-                if sid not in seen:
-                    seen.add(sid)
-                    ordered.append(sid)
-
-        prog.discovered = len(ordered)
-
-        # queue up work
-        for sid in ordered:
-            await state.queue.put(sid)
-        state.closed = True
-        _start_state_workers(root, prog, state)
-        await _await_state_workers_then_finish(root, prog, state)
-    except asyncio.CancelledError:
-        prog.status = "cancelled"
-        raise
-    except Exception as exc:
-        _append_error(prog, path="", message=str(exc))
-        prog.status = "failed"
-        prog.finished_at = time.time()
-        LOGGER.exception("Asset scan failed for %s", root)
-    finally:
-        RUNNING_TASKS.pop(root, None)
-
-
-async def _reconcile_missing_tags_for_root(root: schemas_in.RootType, prog: ScanProgress) -> None:
-    """
-    Detect missing files quickly and toggle 'missing' tag per asset_id.
-
-    Rules:
-      - Only hashed assets (assets.hash != NULL) participate in missing tagging.
-      - We consider ALL cache states of the asset (across roots) before tagging.
-    """
-    if root == "models":
-        bases: list[str] = []
-        for _bucket, paths in get_comfy_models_folders():
-            bases.extend(paths)
-    elif root == "input":
-        bases = [folder_paths.get_input_directory()]
-    else:
-        bases = [folder_paths.get_output_directory()]
-
-    try:
-        async with await create_session() as sess:
-            # state + hash + size for the current root
-            rows = await list_cache_states_with_asset_under_prefixes(sess, prefixes=bases)
-
-            # Track fast_ok within the scanned root and whether the asset is hashed
-            by_asset: dict[str, dict[str, bool]] = {}
-            for state, a_hash, size_db in rows:
-                aid = state.asset_id
-                acc = by_asset.get(aid)
-                if acc is None:
-                    acc = {"any_fast_ok_here": False, "hashed": (a_hash is not None), "size_db": int(size_db or 0)}
-                    by_asset[aid] = acc
-                try:
-                    if acc["hashed"]:
-                        st = os.stat(state.file_path, follow_symlinks=True)
-                        if fast_asset_file_check(mtime_db=state.mtime_ns, size_db=acc["size_db"], stat_result=st):
-                            acc["any_fast_ok_here"] = True
-                except FileNotFoundError:
-                    pass
-                except OSError as e:
-                    _append_error(prog, path=state.file_path, message=str(e))
-
-            # Decide per asset, considering ALL its states (not just this root)
-            for aid, acc in by_asset.items():
-                try:
-                    if not acc["hashed"]:
-                        # Never tag seed assets as missing
-                        continue
-
-                    any_fast_ok_global = acc["any_fast_ok_here"]
-                    if not any_fast_ok_global:
-                        # Check other states outside this root
-                        others = await list_cache_states_by_asset_id(sess, asset_id=aid)
-                        for st in others:
-                            try:
-                                any_fast_ok_global = fast_asset_file_check(
-                                    mtime_db=st.mtime_ns,
-                                    size_db=acc["size_db"],
-                                    stat_result=os.stat(st.file_path, follow_symlinks=True),
-                                )
-                            except OSError:
-                                continue
-
-                    if any_fast_ok_global:
-                        await remove_missing_tag_for_asset_id(sess, asset_id=aid)
-                    else:
-                        await add_missing_tag_for_asset_id(sess, asset_id=aid, origin="automatic")
-                except Exception as ex:
-                    _append_error(prog, path="", message=f"reconcile {aid[:8]}: {ex}")
-
-            await sess.commit()
-    except Exception as e:
-        _append_error(prog, path="", message=f"reconcile failed: {e}")
-
-
-def _start_state_workers(root: schemas_in.RootType, prog: ScanProgress, state: SlowQueueState) -> None:
-    if state.workers:
-        return
-
-    async def _worker(_wid: int):
-        while True:
-            sid = await state.queue.get()
-            try:
-                if sid is None:
-                    return
-                try:
-                    async with await create_session() as sess:
-                        # Optional: fetch path for better error messages
-                        st = await sess.get(AssetCacheState, sid)
-                        try:
-                            await compute_hash_and_dedup_for_cache_state(sess, state_id=sid)
-                            await sess.commit()
-                        except Exception as e:
-                            path = st.file_path if st else f"state:{sid}"
-                            _append_error(prog, path=path, message=str(e))
-                            raise
-                except Exception:
-                    pass
-                finally:
-                    prog.processed += 1
-            finally:
-                state.queue.task_done()
-
-    state.workers = [
-        asyncio.create_task(_worker(i), name=f"asset-hash:{root}:{i}")
-        for i in range(SLOW_HASH_CONCURRENCY)
-    ]
-
-    async def _close_when_ready():
-        while not state.closed:
-            await asyncio.sleep(0.05)
-        for _ in range(SLOW_HASH_CONCURRENCY):
-            await state.queue.put(None)
-
-    asyncio.create_task(_close_when_ready())
-
-
-async def _await_state_workers_then_finish(
-    root: schemas_in.RootType, prog: ScanProgress, state: SlowQueueState
-) -> None:
-    if state.workers:
-        await asyncio.gather(*state.workers, return_exceptions=True)
-    await _reconcile_missing_tags_for_root(root, prog)
-    prog.finished_at = time.time()
-    prog.status = "completed"
-
-
-def _append_error(prog: ScanProgress, *, path: str, message: str) -> None:
-    prog.file_errors.append({
-        "path": path,
-        "message": message,
-        "at": ts_to_iso(time.time()),
-    })
-
-
-async def _fast_db_consistency_pass(
-    root: schemas_in.RootType,
-    *,
-    collect_existing_paths: bool = False,
-    update_missing_tags: bool = False,
-) -> Optional[set[str]]:
-    """Fast DB+FS pass for a root:
-      - Toggle needs_verify per state using fast check
-      - For hashed assets with at least one fast-ok state in this root: delete stale missing states
-      - For seed assets with all states missing: delete Asset and its AssetInfos
-      - Optionally add/remove 'missing' tags based on fast-ok in this root
-      - Optionally return surviving absolute paths
-    """
-    prefixes = prefixes_for_root(root)
-    if not prefixes:
-        return set() if collect_existing_paths else None
-
-    conds = []
-    for p in prefixes:
-        base = os.path.abspath(p)
-        if not base.endswith(os.sep):
-            base += os.sep
-        escaped, esc = escape_like_prefix(base)
-        conds.append(AssetCacheState.file_path.like(escaped + "%", escape=esc))
-
-    async with await create_session() as sess:
-        rows = (
-            await sess.execute(
-                sa.select(
-                    AssetCacheState.id,
-                    AssetCacheState.file_path,
-                    AssetCacheState.mtime_ns,
-                    AssetCacheState.needs_verify,
-                    AssetCacheState.asset_id,
-                    Asset.hash,
-                    Asset.size_bytes,
-                )
-                .join(Asset, Asset.id == AssetCacheState.asset_id)
-                .where(sa.or_(*conds))
-                .order_by(AssetCacheState.asset_id.asc(), AssetCacheState.id.asc())
-            )
-        ).all()
-
-        by_asset: dict[str, dict] = {}
-        for sid, fp, mtime_db, needs_verify, aid, a_hash, a_size in rows:
-            acc = by_asset.get(aid)
-            if acc is None:
-                acc = {"hash": a_hash, "size_db": int(a_size or 0), "states": []}
-                by_asset[aid] = acc
-
-            fast_ok = False
-            try:
-                exists = True
-                fast_ok = fast_asset_file_check(
-                    mtime_db=mtime_db,
-                    size_db=acc["size_db"],
-                    stat_result=os.stat(fp, follow_symlinks=True),
-                )
-            except FileNotFoundError:
-                exists = False
-            except OSError:
-                exists = False
-
-            acc["states"].append({
-                "sid": sid,
-                "fp": fp,
-                "exists": exists,
-                "fast_ok": fast_ok,
-                "needs_verify": bool(needs_verify),
-            })
-
-        to_set_verify: list[int] = []
-        to_clear_verify: list[int] = []
-        stale_state_ids: list[int] = []
-        survivors: set[str] = set()
-
-        for aid, acc in by_asset.items():
-            a_hash = acc["hash"]
-            states = acc["states"]
-            any_fast_ok = any(s["fast_ok"] for s in states)
-            all_missing = all(not s["exists"] for s in states)
-
-            for s in states:
-                if not s["exists"]:
-                    continue
-                if s["fast_ok"] and s["needs_verify"]:
-                    to_clear_verify.append(s["sid"])
-                if not s["fast_ok"] and not s["needs_verify"]:
-                    to_set_verify.append(s["sid"])
-
-            if a_hash is None:
-                if states and all_missing:  # remove seed Asset completely, if no valid AssetCache exists
-                    await sess.execute(sa.delete(AssetInfo).where(AssetInfo.asset_id == aid))
-                    asset = await sess.get(Asset, aid)
-                    if asset:
-                        await sess.delete(asset)
-                else:
-                    for s in states:
-                        if s["exists"]:
-                            survivors.add(os.path.abspath(s["fp"]))
-                continue
-
-            if any_fast_ok:  # if Asset has at least one valid AssetCache record, remove any invalid AssetCache records
-                for s in states:
-                    if not s["exists"]:
-                        stale_state_ids.append(s["sid"])
-                if update_missing_tags:
-                    with contextlib.suppress(Exception):
-                        await remove_missing_tag_for_asset_id(sess, asset_id=aid)
-            elif update_missing_tags:
-                with contextlib.suppress(Exception):
-                    await add_missing_tag_for_asset_id(sess, asset_id=aid, origin="automatic")
-
-            for s in states:
-                if s["exists"]:
-                    survivors.add(os.path.abspath(s["fp"]))
-
-        if stale_state_ids:
-            await sess.execute(sa.delete(AssetCacheState).where(AssetCacheState.id.in_(stale_state_ids)))
-        if to_set_verify:
-            await sess.execute(
-                sa.update(AssetCacheState)
-                .where(AssetCacheState.id.in_(to_set_verify))
-                .values(needs_verify=True)
-            )
-        if to_clear_verify:
-            await sess.execute(
-                sa.update(AssetCacheState)
-                .where(AssetCacheState.id.in_(to_clear_verify))
-                .values(needs_verify=False)
-            )
-        await sess.commit()
-        return survivors if collect_existing_paths else None

app/assets/storage/hashing.py

@@ -1,72 +0,0 @@
-import asyncio
-import os
-from typing import IO, Union
-
-from blake3 import blake3
-
-DEFAULT_CHUNK = 8 * 1024 * 1024  # 8 MiB
-
-
-def _hash_file_obj_sync(file_obj: IO[bytes], chunk_size: int) -> str:
-    """Hash an already-open binary file object by streaming in chunks.
-    - Seeks to the beginning before reading (if supported).
-    - Restores the original position afterward (if tell/seek are supported).
-    """
-    if chunk_size <= 0:
-        chunk_size = DEFAULT_CHUNK
-
-    orig_pos = None
-    if hasattr(file_obj, "tell"):
-        orig_pos = file_obj.tell()
-
-    try:
-        if hasattr(file_obj, "seek"):
-            file_obj.seek(0)
-
-        h = blake3()
-        while True:
-            chunk = file_obj.read(chunk_size)
-            if not chunk:
-                break
-            h.update(chunk)
-        return h.hexdigest()
-    finally:
-        if hasattr(file_obj, "seek") and orig_pos is not None:
-            file_obj.seek(orig_pos)
-
-
-def blake3_hash_sync(
-    fp: Union[str, bytes, os.PathLike[str], os.PathLike[bytes], IO[bytes]],
-    chunk_size: int = DEFAULT_CHUNK,
-) -> str:
-    """Returns a BLAKE3 hex digest for ``fp``, which may be:
-      - a filename (str/bytes) or PathLike
-      - an open binary file object
-
-    If ``fp`` is a file object, it must be opened in **binary** mode and support
-    ``read``, ``seek``, and ``tell``. The function will seek to the start before
-    reading and will attempt to restore the original position afterward.
-    """
-    if hasattr(fp, "read"):
-        return _hash_file_obj_sync(fp, chunk_size)
-
-    with open(os.fspath(fp), "rb") as f:
-        return _hash_file_obj_sync(f, chunk_size)
-
-
-async def blake3_hash(
-    fp: Union[str, bytes, os.PathLike[str], os.PathLike[bytes], IO[bytes]],
-    chunk_size: int = DEFAULT_CHUNK,
-) -> str:
-    """Async wrapper for ``blake3_hash_sync``.
-    Uses a worker thread so the event loop remains responsive.
-    """
-    # If it is a path, open inside the worker thread to keep I/O off the loop.
-    if hasattr(fp, "read"):
-        return await asyncio.to_thread(blake3_hash_sync, fp, chunk_size)
-
-    def _worker() -> str:
-        with open(os.fspath(fp), "rb") as f:
-            return _hash_file_obj_sync(f, chunk_size)
-
-    return await asyncio.to_thread(_worker)

app/database/db.py

@@ -0,0 +1,112 @@
+import logging
+import os
+import shutil
+from app.logger import log_startup_warning
+from utils.install_util import get_missing_requirements_message
+from comfy.cli_args import args
+
+_DB_AVAILABLE = False
+Session = None
+
+
+try:
+    from alembic import command
+    from alembic.config import Config
+    from alembic.runtime.migration import MigrationContext
+    from alembic.script import ScriptDirectory
+    from sqlalchemy import create_engine
+    from sqlalchemy.orm import sessionmaker
+
+    _DB_AVAILABLE = True
+except ImportError as e:
+    log_startup_warning(
+        f"""
+------------------------------------------------------------------------
+Error importing dependencies: {e}
+{get_missing_requirements_message()}
+This error is happening because ComfyUI now uses a local sqlite database.
+------------------------------------------------------------------------
+""".strip()
+    )
+
+
+def dependencies_available():
+    """
+    Temporary function to check if the dependencies are available
+    """
+    return _DB_AVAILABLE
+
+
+def can_create_session():
+    """
+    Temporary function to check if the database is available to create a session
+    During initial release there may be environmental issues (or missing dependencies) that prevent the database from being created
+    """
+    return dependencies_available() and Session is not None
+
+
+def get_alembic_config():
+    root_path = os.path.join(os.path.dirname(__file__), "../..")
+    config_path = os.path.abspath(os.path.join(root_path, "alembic.ini"))
+    scripts_path = os.path.abspath(os.path.join(root_path, "alembic_db"))
+
+    config = Config(config_path)
+    config.set_main_option("script_location", scripts_path)
+    config.set_main_option("sqlalchemy.url", args.database_url)
+
+    return config
+
+
+def get_db_path():
+    url = args.database_url
+    if url.startswith("sqlite:///"):
+        return url.split("///")[1]
+    else:
+        raise ValueError(f"Unsupported database URL '{url}'.")
+
+
+def init_db():
+    db_url = args.database_url
+    logging.debug(f"Database URL: {db_url}")
+    db_path = get_db_path()
+    db_exists = os.path.exists(db_path)
+
+    config = get_alembic_config()
+
+    # Check if we need to upgrade
+    engine = create_engine(db_url)
+    conn = engine.connect()
+
+    context = MigrationContext.configure(conn)
+    current_rev = context.get_current_revision()
+
+    script = ScriptDirectory.from_config(config)
+    target_rev = script.get_current_head()
+
+    if target_rev is None:
+        logging.warning("No target revision found.")
+    elif current_rev != target_rev:
+        # Backup the database pre upgrade
+        backup_path = db_path + ".bkp"
+        if db_exists:
+            shutil.copy(db_path, backup_path)
+        else:
+            backup_path = None
+
+        try:
+            command.upgrade(config, target_rev)
+            logging.info(f"Database upgraded from {current_rev} to {target_rev}")
+        except Exception as e:
+            if backup_path:
+                # Restore the database from backup if upgrade fails
+                shutil.copy(backup_path, db_path)
+                os.remove(backup_path)
+            logging.exception("Error upgrading database: ")
+            raise e
+
+    global Session
+    Session = sessionmaker(bind=engine)
+
+
+def create_session():
+    return Session()

app/database/models.py

@@ -0,0 +1,14 @@
+from sqlalchemy.orm import declarative_base
+
+Base = declarative_base()
+
+
+def to_dict(obj):
+    fields = obj.__table__.columns.keys()
+    return {
+        field: (val.to_dict() if hasattr(val, "to_dict") else val)
+        for field in fields
+        if (val := getattr(obj, field))
+    }
+
+# TODO: Define models here

app/db.py

@@ -1,255 +0,0 @@
-import logging
-import os
-import shutil
-from contextlib import asynccontextmanager
-from typing import Optional
-
-from alembic import command
-from alembic.config import Config
-from alembic.runtime.migration import MigrationContext
-from alembic.script import ScriptDirectory
-from sqlalchemy import create_engine, text
-from sqlalchemy.engine import make_url
-from sqlalchemy.ext.asyncio import (
-    AsyncEngine,
-    AsyncSession,
-    async_sessionmaker,
-    create_async_engine,
-)
-
-from comfy.cli_args import args
-
-LOGGER = logging.getLogger(__name__)
-ENGINE: Optional[AsyncEngine] = None
-SESSION: Optional[async_sessionmaker] = None
-
-
-def _root_paths():
-    """Resolve alembic.ini and migrations script folder."""
-    root_path = os.path.abspath(os.path.dirname(__file__))
-    config_path = os.path.abspath(os.path.join(root_path, "../alembic.ini"))
-    scripts_path = os.path.abspath(os.path.join(root_path, "alembic_db"))
-    return config_path, scripts_path
-
-
-def _absolutize_sqlite_url(db_url: str) -> str:
-    """Make SQLite database path absolute. No-op for non-SQLite URLs."""
-    try:
-        u = make_url(db_url)
-    except Exception:
-        return db_url
-
-    if not u.drivername.startswith("sqlite"):
-        return db_url
-
-    db_path: str = u.database or ""
-    if isinstance(db_path, str) and db_path.startswith("file:"):
-        return str(u)  # Do not touch SQLite URI databases like: "file:xxx?mode=memory&cache=shared"
-    if not os.path.isabs(db_path):
-        db_path = os.path.abspath(os.path.join(os.getcwd(), db_path))
-        u = u.set(database=db_path)
-    return str(u)
-
-
-def _normalize_sqlite_memory_url(db_url: str) -> tuple[str, bool]:
-    """
-    If db_url points at an in-memory SQLite DB (":memory:" or file:... mode=memory),
-    rewrite it to a *named* shared in-memory URI and ensure 'uri=true' is present.
-    Returns: (normalized_url, is_memory)
-    """
-    try:
-        u = make_url(db_url)
-    except Exception:
-        return db_url, False
-    if not u.drivername.startswith("sqlite"):
-        return db_url, False
-
-    db = u.database or ""
-    if db == ":memory:":
-        u = u.set(database=f"file:comfyui_db_{os.getpid()}?mode=memory&cache=shared&uri=true")
-        return str(u), True
-    if isinstance(db, str) and db.startswith("file:") and "mode=memory" in db:
-        if "uri=true" not in db:
-            u = u.set(database=(db + ("&" if "?" in db else "?") + "uri=true"))
-        return str(u), True
-    return str(u), False
-
-
-def _get_sqlite_file_path(sync_url: str) -> Optional[str]:
-    """Return the on-disk path for a SQLite URL, else None."""
-    try:
-        u = make_url(sync_url)
-    except Exception:
-        return None
-
-    if not u.drivername.startswith("sqlite"):
-        return None
-    db_path = u.database
-    if isinstance(db_path, str) and db_path.startswith("file:"):
-        return None  # Not a real file if it is a URI like "file:...?"
-    return db_path
-
-
-def _get_alembic_config(sync_url: str) -> Config:
-    """Prepare Alembic Config with script location and DB URL."""
-    config_path, scripts_path = _root_paths()
-    cfg = Config(config_path)
-    cfg.set_main_option("script_location", scripts_path)
-    cfg.set_main_option("sqlalchemy.url", sync_url)
-    return cfg
-
-
-async def init_db_engine() -> None:
-    """Initialize async engine + sessionmaker and run migrations to head.
-
-    This must be called once on application startup before any DB usage.
-    """
-    global ENGINE, SESSION
-
-    if ENGINE is not None:
-        return
-
-    raw_url = args.database_url
-    if not raw_url:
-        raise RuntimeError("Database URL is not configured.")
-
-    db_url, is_mem = _normalize_sqlite_memory_url(raw_url)
-    db_url = _absolutize_sqlite_url(db_url)
-
-    # Prepare async engine
-    connect_args = {}
-    if db_url.startswith("sqlite"):
-        connect_args = {
-            "check_same_thread": False,
-            "timeout": 12,
-        }
-        if is_mem:
-            connect_args["uri"] = True
-
-    ENGINE = create_async_engine(
-        db_url,
-        connect_args=connect_args,
-        pool_pre_ping=True,
-        future=True,
-    )
-
-    # Enforce SQLite pragmas on the async engine
-    if db_url.startswith("sqlite"):
-        async with ENGINE.begin() as conn:
-            if not is_mem:
-                # WAL for concurrency and durability, Foreign Keys for referential integrity
-                current_mode = (await conn.execute(text("PRAGMA journal_mode;"))).scalar()
-                if str(current_mode).lower() != "wal":
-                    new_mode = (await conn.execute(text("PRAGMA journal_mode=WAL;"))).scalar()
-                    if str(new_mode).lower() != "wal":
-                        raise RuntimeError("Failed to set SQLite journal mode to WAL.")
-                    LOGGER.info("SQLite journal mode set to WAL.")
-
-            await conn.execute(text("PRAGMA foreign_keys = ON;"))
-            await conn.execute(text("PRAGMA synchronous = NORMAL;"))
-
-    await _run_migrations(database_url=db_url, connect_args=connect_args)
-
-    SESSION = async_sessionmaker(
-        bind=ENGINE,
-        class_=AsyncSession,
-        expire_on_commit=False,
-        autoflush=False,
-        autocommit=False,
-    )
-
-
-async def _run_migrations(database_url: str, connect_args: dict) -> None:
-    if database_url.find("postgresql+psycopg") == -1:
-        """SQLite: Convert an async SQLAlchemy URL to a sync URL for Alembic."""
-        u = make_url(database_url)
-        driver = u.drivername
-        if not driver.startswith("sqlite+aiosqlite"):
-            raise ValueError(f"Unsupported DB driver: {driver}")
-        database_url, is_mem = _normalize_sqlite_memory_url(str(u.set(drivername="sqlite")))
-        database_url = _absolutize_sqlite_url(database_url)
-
-    cfg = _get_alembic_config(database_url)
-    engine = create_engine(database_url, future=True, connect_args=connect_args)
-    with engine.connect() as conn:
-        context = MigrationContext.configure(conn)
-        current_rev = context.get_current_revision()
-
-    script = ScriptDirectory.from_config(cfg)
-    target_rev = script.get_current_head()
-
-    if target_rev is None:
-        LOGGER.warning("Alembic: no target revision found.")
-        return
-
-    if current_rev == target_rev:
-        LOGGER.debug("Alembic: database already at head %s", target_rev)
-        return
-
-    LOGGER.info("Alembic: upgrading database from %s to %s", current_rev, target_rev)
-
-    # Optional backup for SQLite file DBs
-    backup_path = None
-    sqlite_path = _get_sqlite_file_path(database_url)
-    if sqlite_path and os.path.exists(sqlite_path):
-        backup_path = sqlite_path + ".bkp"
-        try:
-            shutil.copy(sqlite_path, backup_path)
-        except Exception as exc:
-            LOGGER.warning("Failed to create SQLite backup before migration: %s", exc)
-
-    try:
-        command.upgrade(cfg, target_rev)
-    except Exception:
-        if backup_path and os.path.exists(backup_path):
-            LOGGER.exception("Error upgrading database, attempting restore from backup.")
-            try:
-                shutil.copy(backup_path, sqlite_path)  # restore
-                os.remove(backup_path)
-            except Exception as re:
-                LOGGER.error("Failed to restore SQLite backup: %s", re)
-        else:
-            LOGGER.exception("Error upgrading database, backup is not available.")
-        raise
-
-
-def get_engine():
-    """Return the global async engine (initialized after init_db_engine())."""
-    if ENGINE is None:
-        raise RuntimeError("Engine is not initialized. Call init_db_engine() first.")
-    return ENGINE
-
-
-def get_session_maker():
-    """Return the global async_sessionmaker (initialized after init_db_engine())."""
-    if SESSION is None:
-        raise RuntimeError("Session maker is not initialized. Call init_db_engine() first.")
-    return SESSION
-
-
-@asynccontextmanager
-async def session_scope():
-    """Async context manager for a unit of work:
-
-    async with session_scope() as sess:
-        ... use sess ...
-    """
-    maker = get_session_maker()
-    async with maker() as sess:
-        try:
-            yield sess
-            await sess.commit()
-        except Exception:
-            await sess.rollback()
-            raise
-
-
-async def create_session():
-    """Convenience helper to acquire a single AsyncSession instance.
-
-    Typical usage:
-        async with (await create_session()) as sess:
-            ...
-    """
-    maker = get_session_maker()
-    return maker()

app/frontend_management.py

@@ -10,7 +10,8 @@ import importlib
 from dataclasses import dataclass
 from functools import cached_property
 from pathlib import Path
-from typing import TypedDict, Optional
+from typing import Dict, TypedDict, Optional
+from aiohttp import web
 from importlib.metadata import version
 
 import requests
@@ -42,6 +43,7 @@ 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:
@@ -63,6 +65,7 @@ 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."""
 
@@ -196,17 +199,6 @@ def download_release_asset_zip(release: Release, destination_path: str) -> None:
 
 
 class FrontendManager:
-    """
-    A class to manage ComfyUI frontend versions and installations.
-
-    This class handles the initialization and management of different frontend versions,
-    including the default frontend from the pip package and custom frontend versions
-    from GitHub repositories.
-
-    Attributes:
-        CUSTOM_FRONTENDS_ROOT (str): The root directory where custom frontend versions are stored.
-    """
-
     CUSTOM_FRONTENDS_ROOT = str(Path(__file__).parents[1] / "web_custom_versions")
 
     @classmethod
@@ -214,17 +206,39 @@ 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:
-        """
-        Get the path to the default frontend installation from the pip package.
-
-        Returns:
-            str: The path to the default frontend static files.
-
-        Raises:
-            SystemExit: If the comfyui-frontend-package is not installed.
-        """
         try:
             import comfyui_frontend_package
 
@@ -244,16 +258,54 @@ comfyui-frontend-package is not installed.
             sys.exit(-1)
 
     @classmethod
-    def templates_path(cls) -> str:
-        """
-        Get the path to the workflow templates.
+    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 ***********
 
-        Returns:
-            str: The path to the workflow templates directory.
+comfyui-workflow-templates is not installed.
 
-        Raises:
-            SystemExit: If the comfyui-workflow-templates package is not installed.
-        """
+{frontend_install_warning_message()}
+
+********** ERROR ***********
+""".strip()
+            )
+            return None
+
+        try:
+            template_entries = list(iter_templates())
+        except Exception as exc:
+            logging.error(f"Failed to enumerate workflow templates: {exc}")
+            return None
+
+        asset_map: Dict[str, str] = {}
+        try:
+            for entry in template_entries:
+                for asset in entry.assets:
+                    asset_map[asset.filename] = get_asset_path(
+                        entry.template_id, asset.filename
+                    )
+        except Exception as exc:
+            logging.error(f"Failed to resolve template asset paths: {exc}")
+            return None
+
+        if not asset_map:
+            logging.error("No workflow template assets found. Did the packages install correctly?")
+            return None
+
+        return asset_map
+
+
+    @classmethod
+    def legacy_templates_path(cls) -> Optional[str]:
+        """Return the legacy templates directory shipped inside the meta package."""
         try:
             import comfyui_workflow_templates
 
@@ -272,6 +324,7 @@ comfyui-workflow-templates is not installed.
 ********** ERROR ***********
 """.strip()
             )
+            return None
 
     @classmethod
     def embedded_docs_path(cls) -> str:
@@ -289,16 +342,11 @@ comfyui-workflow-templates is not installed.
     @classmethod
     def parse_version_string(cls, value: str) -> tuple[str, str, str]:
         """
-        Parse a version string into its components.
-
-        The version string should be in the format: 'owner/repo@version'
-        where version can be either a semantic version (v1.2.3) or 'latest'.
-
         Args:
             value (str): The version string to parse.
 
         Returns:
-            tuple[str, str, str]: A tuple containing (owner, repo, version).
+            tuple[str, str]: A tuple containing provider name and version.
 
         Raises:
             argparse.ArgumentTypeError: If the version string is invalid.
@@ -315,22 +363,18 @@ comfyui-workflow-templates is not installed.
         cls, version_string: str, provider: Optional[FrontEndProvider] = None
     ) -> str:
         """
-        Initialize a frontend version without error handling.
-
-        This method attempts to initialize a specific frontend version, either from
-        the default pip package or from a custom GitHub repository. It will download
-        and extract the frontend files if necessary.
+        Initializes the frontend for the specified version.
 
         Args:
-            version_string (str): The version string specifying which frontend to use.
-            provider (FrontEndProvider, optional): The provider to use for custom frontends.
+            version_string (str): The version string.
+            provider (FrontEndProvider, optional): The provider to use. Defaults to None.
 
         Returns:
             str: The path to the initialized frontend.
 
         Raises:
-            Exception: If there is an error during initialization (e.g., network timeout,
-                      invalid URL, or missing assets).
+            Exception: If there is an error during the initialization process.
+            main error source might be request timeout or invalid URL.
         """
         if version_string == DEFAULT_VERSION_STRING:
             check_frontend_version()
@@ -382,17 +426,13 @@ comfyui-workflow-templates is not installed.
     @classmethod
     def init_frontend(cls, version_string: str) -> str:
         """
-        Initialize a frontend version with error handling.
-
-        This is the main method to initialize a frontend version. It wraps init_frontend_unsafe
-        with error handling, falling back to the default frontend if initialization fails.
+        Initializes the frontend with the specified version string.
 
         Args:
-            version_string (str): The version string specifying which frontend to use.
+            version_string (str): The version string to initialize the frontend with.
 
         Returns:
-            str: The path to the initialized frontend. If initialization fails,
-                 returns the path to the default frontend.
+            str: The path of the initialized frontend.
         """
         try:
             return cls.init_frontend_unsafe(version_string)
@@ -401,3 +441,17 @@ comfyui-workflow-templates is not installed.
             logging.info("Falling back to the default frontend.")
             check_frontend_version()
             return cls.default_frontend_path()
+    @classmethod
+    def template_asset_handler(cls):
+        assets = cls.template_asset_map()
+        if not assets:
+            return None
+
+        async def serve_template(request: web.Request) -> web.StreamResponse:
+            rel_path = request.match_info.get("path", "")
+            target = assets.get(rel_path)
+            if target is None:
+                raise web.HTTPNotFound()
+            return web.FileResponse(target)
+
+        return serve_template

app/subgraph_manager.py

@@ -0,0 +1,112 @@
+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

@@ -105,6 +105,7 @@ 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.")
@@ -145,7 +146,9 @@ 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. --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("--fast", nargs="*", type=PerformanceFeature, help="Enable some untested and potentially quality deteriorating optimizations. This is used to test new features so using it might crash your comfyui. --fast with no arguments enables everything. You can pass a list specific optimizations if you only want to enable specific ones. Current valid optimizations: {}".format(" ".join(map(lambda c: c.value, PerformanceFeature))))
+
+parser.add_argument("--disable-pinned-memory", action="store_true", help="Disable pinned memory use.")
 
 parser.add_argument("--mmap-torch-files", action="store_true", help="Use mmap when loading ckpt/pt files.")
 parser.add_argument("--disable-mmap", action="store_true", help="Don't use mmap when loading safetensors.")
@@ -212,8 +215,7 @@ parser.add_argument(
 database_default_path = os.path.abspath(
     os.path.join(os.path.dirname(__file__), "..", "user", "comfyui.db")
 )
-parser.add_argument("--database-url", type=str, default=f"sqlite+aiosqlite:///{database_default_path}", help="Specify the database URL, e.g. for an in-memory database you can use 'sqlite+aiosqlite:///:memory:'.")
-parser.add_argument("--disable-assets-autoscan", action="store_true", help="Disable asset scanning on startup for database synchronization.")
+parser.add_argument("--database-url", type=str, default=f"sqlite:///{database_default_path}", help="Specify the database URL, e.g. for an in-memory database you can use 'sqlite:///:memory:'.")
 
 if comfy.options.args_parsing:
     args = parser.parse_args()

comfy/controlnet.py

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

comfy/ldm/ace/vae/music_dcae_pipeline.py

@@ -23,8 +23,6 @@ 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),
         ])
@@ -37,10 +35,6 @@ 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)):
@@ -73,10 +67,8 @@ 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):
@@ -91,9 +83,7 @@ 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)
@@ -101,7 +91,6 @@ 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_radiance/model.py

@@ -189,15 +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(img_out, nerf_pixels, B, C, num_patches, patch_size, params)
+            img_dct = self.forward_tiled_nerf(nerf_hidden, 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)
 
@@ -240,17 +240,8 @@ class ChromaRadiance(Chroma):
             end = min(i + tile_size, num_patches)
 
             # Slice the current tile from the input tensors
-            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)
+            nerf_hidden_tile = nerf_hidden[i * batch:end * batch]
+            nerf_pixels_tile = nerf_pixels[i * batch:end * batch]
 
             # get DCT-encoded pixel embeddings [pixel-dct]
             img_dct_tile = self.nerf_image_embedder(nerf_pixels_tile)

comfy/ldm/flux/layers.py

@@ -195,8 +195,8 @@ class DoubleStreamBlock(nn.Module):
             txt_attn, img_attn = attn[:, : txt.shape[1]], attn[:, txt.shape[1]:]
 
         # calculate the img bloks
-        img = img + apply_mod(self.img_attn.proj(img_attn), img_mod1.gate, None, modulation_dims_img)
-        img = img + apply_mod(self.img_mlp(apply_mod(self.img_norm2(img), (1 + img_mod2.scale), img_mod2.shift, modulation_dims_img)), img_mod2.gate, None, modulation_dims_img)
+        img += apply_mod(self.img_attn.proj(img_attn), img_mod1.gate, None, modulation_dims_img)
+        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)

comfy/ldm/flux/math.py

@@ -7,15 +7,7 @@ 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 = q.to(dtype=pe.dtype).reshape(*q.shape[:-1], -1, 1, 2)
-        k = k.to(dtype=pe.dtype).reshape(*k.shape[:-1], -1, 1, 2)
-        q = (pe[..., 0] * q[..., 0] + pe[..., 1] * q[..., 1]).reshape(*q_shape).type_as(v)
-        k = (pe[..., 0] * k[..., 0] + pe[..., 1] * k[..., 1]).reshape(*k_shape).type_as(v)
-
+    q, k = apply_rope(q, k, pe)
     heads = q.shape[1]
     x = optimized_attention(q, k, v, heads, skip_reshape=True, mask=mask, transformer_options=transformer_options)
     return x
@@ -37,7 +29,10 @@ 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] + freqs_cis[..., 1] * x_[..., 1]
+
+    x_out = freqs_cis[..., 0] * x_[..., 0]
+    x_out.addcmul_(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/hunyuan_video/vae_refiner.py

@@ -1,7 +1,7 @@
 import torch
 import torch.nn as nn
 import torch.nn.functional as F
-from comfy.ldm.modules.diffusionmodules.model import ResnetBlock, AttnBlock, VideoConv3d
+from comfy.ldm.modules.diffusionmodules.model import ResnetBlock, AttnBlock, VideoConv3d, Normalize
 import comfy.ops
 import comfy.ldm.models.autoencoder
 ops = comfy.ops.disable_weight_init
@@ -17,11 +17,12 @@ class RMS_norm(nn.Module):
         return F.normalize(x, dim=1) * self.scale * self.gamma
 
 class DnSmpl(nn.Module):
-    def __init__(self, ic, oc, tds=True):
+    def __init__(self, ic, oc, tds=True, refiner_vae=True, op=VideoConv3d):
         super().__init__()
         fct = 2 * 2 * 2 if tds else 1 * 2 * 2
         assert oc % fct == 0
-        self.conv = VideoConv3d(ic, oc // fct, kernel_size=3)
+        self.conv = op(ic, oc // fct, kernel_size=3, stride=1, padding=1)
+        self.refiner_vae = refiner_vae
 
         self.tds = tds
         self.gs = fct * ic // oc
@@ -30,7 +31,7 @@ class DnSmpl(nn.Module):
         r1 = 2 if self.tds else 1
         h = self.conv(x)
 
-        if self.tds:
+        if self.tds and self.refiner_vae:
             hf = h[:, :, :1, :, :]
             b, c, f, ht, wd = hf.shape
             hf = hf.reshape(b, c, f, ht // 2, 2, wd // 2, 2)
@@ -66,6 +67,7 @@ class DnSmpl(nn.Module):
             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)
@@ -83,10 +85,11 @@ class DnSmpl(nn.Module):
 
 
 class UpSmpl(nn.Module):
-    def __init__(self, ic, oc, tus=True):
+    def __init__(self, ic, oc, tus=True, refiner_vae=True, op=VideoConv3d):
         super().__init__()
         fct = 2 * 2 * 2 if tus else 1 * 2 * 2
-        self.conv = VideoConv3d(ic, oc * fct, kernel_size=3)
+        self.conv = op(ic, oc * fct, kernel_size=3, stride=1, padding=1)
+        self.refiner_vae = refiner_vae
 
         self.tus = tus
         self.rp = fct * oc // ic
@@ -95,7 +98,7 @@ class UpSmpl(nn.Module):
         r1 = 2 if self.tus else 1
         h = self.conv(x)
 
-        if self.tus:
+        if self.tus and self.refiner_vae:
             hf = h[:, :, :1, :, :]
             b, c, f, ht, wd = hf.shape
             nc = c // (2 * 2)
@@ -148,43 +151,56 @@ class UpSmpl(nn.Module):
 
 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, **_):
+                 ffactor_spatial, ffactor_temporal, downsample_match_channel=True, refiner_vae=True, **_):
         super().__init__()
         self.z_channels = z_channels
         self.block_out_channels = block_out_channels
         self.num_res_blocks = num_res_blocks
-        self.conv_in = VideoConv3d(in_channels, block_out_channels[0], 3, 1, 1)
+        self.ffactor_temporal = ffactor_temporal
+
+        self.refiner_vae = refiner_vae
+        if self.refiner_vae:
+            conv_op = VideoConv3d
+            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.down = nn.ModuleList()
         ch = block_out_channels[0]
         depth = (ffactor_spatial >> 1).bit_length()
-        depth_temporal = ((ffactor_spatial // ffactor_temporal) >> 1).bit_length()
+        depth_temporal = ((ffactor_spatial // self.ffactor_temporal) >> 1).bit_length()
 
         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_op=VideoConv3d, norm_op=RMS_norm)
+                                                     conv_op=conv_op, norm_op=norm_op)
                                         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)
+                stage.downsample = DnSmpl(ch, nxt, tds=i >= depth_temporal, refiner_vae=self.refiner_vae, op=conv_op)
                 ch = nxt
             self.down.append(stage)
 
         self.mid = nn.Module()
-        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.mid.block_1 = ResnetBlock(in_channels=ch, out_channels=ch, temb_channels=0, conv_op=conv_op, norm_op=norm_op)
+        self.mid.attn_1 = AttnBlock(ch, conv_op=ops.Conv3d, norm_op=norm_op)
+        self.mid.block_2 = ResnetBlock(in_channels=ch, out_channels=ch, temb_channels=0, conv_op=conv_op, norm_op=norm_op)
 
-        self.norm_out = RMS_norm(ch)
-        self.conv_out = VideoConv3d(ch, z_channels << 1, 3, 1, 1)
+        self.norm_out = norm_op(ch)
+        self.conv_out = conv_op(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)
 
         for stage in self.down:
@@ -200,31 +216,42 @@ class Encoder(nn.Module):
         skip = x.view(b, c // grp, grp, t, h, w).mean(2)
 
         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()
+        if self.refiner_vae:
+            out = self.regul(out)[0]
+
+            out = torch.cat((out[:, :, :1], out), dim=2)
+            out = out.permute(0, 2, 1, 3, 4)
+            b, f_times_2, c, h, w = out.shape
+            out = out.reshape(b, f_times_2 // 2, 2 * c, h, w)
+            out = out.permute(0, 2, 1, 3, 4).contiguous()
+
         return out
 
 class Decoder(nn.Module):
     def __init__(self, z_channels, out_channels, block_out_channels, num_res_blocks,
-                 ffactor_spatial, ffactor_temporal, upsample_match_channel=True, **_):
+                 ffactor_spatial, ffactor_temporal, upsample_match_channel=True, refiner_vae=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 = VideoConv3d
+            norm_op = RMS_norm
+        else:
+            conv_op = ops.Conv3d
+            norm_op = Normalize
+
         ch = block_out_channels[0]
-        self.conv_in = VideoConv3d(z_channels, ch, 3)
+        self.conv_in = conv_op(z_channels, ch, kernel_size=3, stride=1, padding=1)
 
         self.mid = nn.Module()
-        self.mid.block_1 = ResnetBlock(in_channels=ch, out_channels=ch, temb_channels=0, conv_op=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.mid.block_1 = ResnetBlock(in_channels=ch, out_channels=ch, temb_channels=0, conv_op=conv_op, norm_op=norm_op)
+        self.mid.attn_1 = AttnBlock(ch, conv_op=ops.Conv3d, norm_op=norm_op)
+        self.mid.block_2 = ResnetBlock(in_channels=ch, out_channels=ch, temb_channels=0, conv_op=conv_op, norm_op=norm_op)
 
         self.up = nn.ModuleList()
         depth = (ffactor_spatial >> 1).bit_length()
@@ -235,25 +262,26 @@ class Decoder(nn.Module):
             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)
+                                                     conv_op=conv_op, norm_op=norm_op)
                                         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)
+                stage.upsample = UpSmpl(ch, nxt, tus=i < depth_temporal, refiner_vae=self.refiner_vae, op=conv_op)
                 ch = nxt
             self.up.append(stage)
 
-        self.norm_out = RMS_norm(ch)
-        self.conv_out = VideoConv3d(ch, out_channels, 3)
+        self.norm_out = norm_op(ch)
+        self.conv_out = conv_op(ch, out_channels, 3, stride=1, padding=1)
 
     def forward(self, z):
-        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:]
+        if self.refiner_vae:
+            z = z.permute(0, 2, 1, 3, 4)
+            b, f, c, h, w = z.shape
+            z = z.reshape(b, f, 2, c // 2, h, w)
+            z = z.permute(0, 1, 2, 3, 4, 5).reshape(b, f * 2, c // 2, h, w)
+            z = z.permute(0, 2, 1, 3, 4)
+            z = z[:, :, 1:]
 
         x = self.conv_in(z) + z.repeat_interleave(self.block_out_channels[0] // self.z_channels, 1)
         x = self.mid.block_2(self.mid.attn_1(self.mid.block_1(x)))
@@ -264,4 +292,10 @@ class Decoder(nn.Module):
             if hasattr(stage, 'upsample'):
                 x = stage.upsample(x)
 
-        return self.conv_out(F.silu(self.norm_out(x)))
+        out = self.conv_out(F.silu(self.norm_out(x)))
+
+        if not self.refiner_vae:
+            if z.shape[-3] == 1:
+                out = out[:, :, -1:]
+
+        return out

comfy/ldm/lightricks/model.py

@@ -3,12 +3,11 @@ 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,
@@ -238,20 +237,6 @@ 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__()
@@ -281,8 +266,8 @@ class CrossAttention(nn.Module):
         k = self.k_norm(k)
 
         if pe is not None:
-            q = apply_rotary_emb(q, pe)
-            k = apply_rotary_emb(k, pe)
+            q = apply_rope1(q.unsqueeze(1), pe).squeeze(1)
+            k = apply_rope1(k.unsqueeze(1), pe).squeeze(1)
 
         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)
@@ -306,12 +291,17 @@ 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)
 
-        x += self.attn1(comfy.ldm.common_dit.rms_norm(x) * (1 + scale_msa) + shift_msa, pe=pe, transformer_options=transformer_options) * gate_msa
+        attn1_input = comfy.ldm.common_dit.rms_norm(x)
+        attn1_input = torch.addcmul(attn1_input, attn1_input, scale_msa).add_(shift_msa)
+        attn1_input = self.attn1(attn1_input, pe=pe, transformer_options=transformer_options)
+        x.addcmul_(attn1_input, gate_msa)
+        del attn1_input
 
         x += self.attn2(x, context=context, mask=attention_mask, transformer_options=transformer_options)
 
-        y = comfy.ldm.common_dit.rms_norm(x) * (1 + scale_mlp) + shift_mlp
-        x += self.ff(y) * gate_mlp
+        y = comfy.ldm.common_dit.rms_norm(x)
+        y = torch.addcmul(y, y, scale_mlp).add_(shift_mlp)
+        x.addcmul_(self.ff(y), gate_mlp)
 
         return x
 
@@ -327,41 +317,35 @@ 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 #self.dtype
+    dtype = torch.float32
+    device = indices_grid.device
 
+    # 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
 
-    start = 1
-    end = theta
-    device = fractional_positions.device
+    # 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)
 
-    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)
+    # Pad if dim is not divisible by 6
     if dim % 6 != 0:
-        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)
+        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
 
 
 class LTXVModel(torch.nn.Module):
@@ -501,7 +485,7 @@ class LTXVModel(torch.nn.Module):
         shift, scale = scale_shift_values[:, :, 0], scale_shift_values[:, :, 1]
         x = self.norm_out(x)
         # Modulation
-        x = x * (1 + scale) + shift
+        x = torch.addcmul(x, x, scale).add_(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.int32, device=device)
+        position_ids = torch.zeros(bsz, max_seq_len, 3, dtype=torch.float32, device=device)
 
         for i in range(bsz):
             cap_len = l_effective_cap_len[i]
@@ -531,10 +531,22 @@ class NextDiT(nn.Module):
             H_tokens, W_tokens = H // pH, W // pW
             assert H_tokens * W_tokens == img_len
 
-            position_ids[i, :cap_len, 0] = torch.arange(cap_len, dtype=torch.int32, device=device)
+            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:cap_len+img_len, 0] = cap_len
-            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()
+            row_ids = (torch.arange(H_tokens, dtype=torch.float32, device=device) * h_scale + h_start).view(-1, 1).repeat(1, W_tokens).flatten()
+            col_ids = (torch.arange(W_tokens, dtype=torch.float32, device=device) * w_scale + w_start).view(1, -1).repeat(H_tokens, 1).flatten()
             position_ids[i, cap_len:cap_len+img_len, 1] = row_ids
             position_ids[i, cap_len:cap_len+img_len, 2] = col_ids
 

comfy/ldm/mmaudio/vae/activations.py

@@ -0,0 +1,120 @@
+# 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

@@ -0,0 +1,157 @@
+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

@@ -0,0 +1,156 @@
+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

@@ -0,0 +1,219 @@
+# 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

@@ -0,0 +1,92 @@
+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

@@ -0,0 +1,358 @@
+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

@@ -0,0 +1,121 @@
+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).squeeze(1).unsqueeze(2).to(x.dtype)
+        image_rotary_emb = self.pe_embedder(ids).to(x.dtype).contiguous()
         del ids, txt_ids, img_ids
 
         hidden_states = self.img_in(hidden_states) + self.controlnet_x_embedder(hint)

comfy/ldm/qwen_image/model.py

@@ -10,6 +10,7 @@ from comfy.ldm.modules.attention import optimized_attention_masked
 from comfy.ldm.flux.layers import EmbedND
 import comfy.ldm.common_dit
 import comfy.patcher_extension
+from comfy.ldm.flux.math import apply_rope1
 
 class GELU(nn.Module):
     def __init__(self, dim_in: int, dim_out: int, approximate: str = "none", bias: bool = True, dtype=None, device=None, operations=None):
@@ -134,33 +135,34 @@ class Attention(nn.Module):
         image_rotary_emb: Optional[torch.Tensor] = None,
         transformer_options={},
     ) -> Tuple[torch.Tensor, torch.Tensor]:
+        batch_size = hidden_states.shape[0]
+        seq_img = hidden_states.shape[1]
         seq_txt = encoder_hidden_states.shape[1]
 
-        img_query = self.to_q(hidden_states).unflatten(-1, (self.heads, -1))
-        img_key = self.to_k(hidden_states).unflatten(-1, (self.heads, -1))
-        img_value = self.to_v(hidden_states).unflatten(-1, (self.heads, -1))
+        # Project and reshape to BHND format (batch, heads, seq, dim)
+        img_query = self.to_q(hidden_states).view(batch_size, seq_img, self.heads, -1).transpose(1, 2).contiguous()
+        img_key = self.to_k(hidden_states).view(batch_size, seq_img, self.heads, -1).transpose(1, 2).contiguous()
+        img_value = self.to_v(hidden_states).view(batch_size, seq_img, self.heads, -1).transpose(1, 2)
 
-        txt_query = self.add_q_proj(encoder_hidden_states).unflatten(-1, (self.heads, -1))
-        txt_key = self.add_k_proj(encoder_hidden_states).unflatten(-1, (self.heads, -1))
-        txt_value = self.add_v_proj(encoder_hidden_states).unflatten(-1, (self.heads, -1))
+        txt_query = self.add_q_proj(encoder_hidden_states).view(batch_size, seq_txt, self.heads, -1).transpose(1, 2).contiguous()
+        txt_key = self.add_k_proj(encoder_hidden_states).view(batch_size, seq_txt, self.heads, -1).transpose(1, 2).contiguous()
+        txt_value = self.add_v_proj(encoder_hidden_states).view(batch_size, seq_txt, self.heads, -1).transpose(1, 2)
 
         img_query = self.norm_q(img_query)
         img_key = self.norm_k(img_key)
         txt_query = self.norm_added_q(txt_query)
         txt_key = self.norm_added_k(txt_key)
 
-        joint_query = torch.cat([txt_query, img_query], dim=1)
-        joint_key = torch.cat([txt_key, img_key], dim=1)
-        joint_value = torch.cat([txt_value, img_value], dim=1)
+        joint_query = torch.cat([txt_query, img_query], dim=2)
+        joint_key = torch.cat([txt_key, img_key], dim=2)
+        joint_value = torch.cat([txt_value, img_value], dim=2)
 
-        joint_query = apply_rotary_emb(joint_query, image_rotary_emb)
-        joint_key = apply_rotary_emb(joint_key, image_rotary_emb)
+        joint_query = apply_rope1(joint_query, image_rotary_emb)
+        joint_key = apply_rope1(joint_key, image_rotary_emb)
 
-        joint_query = joint_query.flatten(start_dim=2)
-        joint_key = joint_key.flatten(start_dim=2)
-        joint_value = joint_value.flatten(start_dim=2)
-
-        joint_hidden_states = optimized_attention_masked(joint_query, joint_key, joint_value, self.heads, attention_mask, transformer_options=transformer_options)
+        joint_hidden_states = optimized_attention_masked(joint_query, joint_key, joint_value, self.heads,
+                                                         attention_mask, transformer_options=transformer_options,
+                                                         skip_reshape=True)
 
         txt_attn_output = joint_hidden_states[:, :seq_txt, :]
         img_attn_output = joint_hidden_states[:, seq_txt:, :]
@@ -413,7 +415,7 @@ 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).squeeze(1).unsqueeze(2).to(x.dtype)
+        image_rotary_emb = self.pe_embedder(ids).to(x.dtype).contiguous()
         del ids, txt_ids, img_ids
 
         hidden_states = self.img_in(hidden_states)

comfy/ldm/wan/model.py

@@ -232,11 +232,13 @@ 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)
@@ -587,7 +589,7 @@ 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):
+    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={}):
         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])
@@ -600,10 +602,22 @@ 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(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[:, :, :, 1] = img_ids[:, :, :, 1] + torch.linspace(h_start, h_start + (h_len - 1), steps=steps_h, device=device, dtype=dtype).reshape(1, -1, 1)
+        img_ids[:, :, :, 2] = img_ids[:, :, :, 2] + torch.linspace(w_start, w_start + (w_len - 1), steps=steps_w, device=device, dtype=dtype).reshape(1, 1, -1)
         img_ids = img_ids.reshape(1, -1, img_ids.shape[-1])
 
         freqs = self.rope_embedder(img_ids).movedim(1, 2)
@@ -629,7 +643,7 @@ 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)
+        freqs = self.rope_encode(t_len, h, w, device=x.device, dtype=x.dtype, transformer_options=transformer_options)
         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):
@@ -902,7 +916,7 @@ 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,

comfy/ldm/wan/vae.py

@@ -468,55 +468,46 @@ class WanVAE(nn.Module):
                                  attn_scales, self.temperal_upsample, dropout)
 
     def encode(self, x):
-        self.clear_cache()
+        conv_idx = [0]
+        feat_map = [None] * count_conv3d(self.decoder)
         ## 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_):
-            self._enc_conv_idx = [0]
+            conv_idx = [0]
             if i == 0:
                 out = self.encoder(
                     x[:, :, :1, :, :],
-                    feat_cache=self._enc_feat_map,
-                    feat_idx=self._enc_conv_idx)
+                    feat_cache=feat_map,
+                    feat_idx=conv_idx)
             else:
                 out_ = self.encoder(
                     x[:, :, 1 + 4 * (i - 1):1 + 4 * i, :, :],
-                    feat_cache=self._enc_feat_map,
-                    feat_idx=self._enc_conv_idx)
+                    feat_cache=feat_map,
+                    feat_idx=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):
-        self.clear_cache()
+        conv_idx = [0]
+        feat_map = [None] * count_conv3d(self.decoder)
         # z: [b,c,t,h,w]
 
         iter_ = z.shape[2]
         x = self.conv2(z)
         for i in range(iter_):
-            self._conv_idx = [0]
+            conv_idx = [0]
             if i == 0:
                 out = self.decoder(
                     x[:, :, i:i + 1, :, :],
-                    feat_cache=self._feat_map,
-                    feat_idx=self._conv_idx)
+                    feat_cache=feat_map,
+                    feat_idx=conv_idx)
             else:
                 out_ = self.decoder(
                     x[:, :, i:i + 1, :, :],
-                    feat_cache=self._feat_map,
-                    feat_idx=self._conv_idx)
+                    feat_cache=feat_map,
+                    feat_idx=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):
-        self.clear_cache()
+        conv_idx = [0]
+        feat_map = [None] * count_conv3d(self.encoder)
         x = patchify(x, patch_size=2)
         t = x.shape[2]
         iter_ = 1 + (t - 1) // 4
         for i in range(iter_):
-            self._enc_conv_idx = [0]
+            conv_idx = [0]
             if i == 0:
                 out = self.encoder(
                     x[:, :, :1, :, :],
-                    feat_cache=self._enc_feat_map,
-                    feat_idx=self._enc_conv_idx,
+                    feat_cache=feat_map,
+                    feat_idx=conv_idx,
                 )
             else:
                 out_ = self.encoder(
                     x[:, :, 1 + 4 * (i - 1):1 + 4 * i, :, :],
-                    feat_cache=self._enc_feat_map,
-                    feat_idx=self._enc_conv_idx,
+                    feat_cache=feat_map,
+                    feat_idx=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):
-        self.clear_cache()
+        conv_idx = [0]
+        feat_map = [None] * count_conv3d(self.decoder)
         iter_ = z.shape[2]
         x = self.conv2(z)
         for i in range(iter_):
-            self._conv_idx = [0]
+            conv_idx = [0]
             if i == 0:
                 out = self.decoder(
                     x[:, :, i:i + 1, :, :],
-                    feat_cache=self._feat_map,
-                    feat_idx=self._conv_idx,
+                    feat_cache=feat_map,
+                    feat_idx=conv_idx,
                     first_chunk=True,
                 )
             else:
                 out_ = self.decoder(
                     x[:, :, i:i + 1, :, :],
-                    feat_cache=self._feat_map,
-                    feat_idx=self._conv_idx,
+                    feat_cache=feat_map,
+                    feat_idx=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):
@@ -715,12 +715,3 @@ 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

@@ -134,10 +134,11 @@ 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)
+                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)
             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")
@@ -196,8 +197,14 @@ class BaseModel(torch.nn.Module):
             extra_conds[o] = extra
 
         t = self.process_timestep(t, x=x, **extra_conds)
-        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)
+        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)
 
     def process_timestep(self, timestep, **kwargs):
         return timestep
@@ -326,6 +333,14 @@ 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:
@@ -669,7 +684,6 @@ 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 = {}
@@ -698,7 +712,6 @@ 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 = {}

comfy/model_detection.py

@@ -6,6 +6,20 @@ 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:
@@ -213,7 +227,7 @@ 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"] = 32
+                dit_config["nerf_tile_size"] = 512
                 dit_config["nerf_final_head_type"] = "conv" if f"{key_prefix}nerf_final_layer_conv.norm.scale" in state_dict_keys else "linear"
                 dit_config["nerf_embedder_dtype"] = torch.float32
         else:
@@ -365,8 +379,8 @@ 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"] = 2304
-        dit_config["n_layers"] = 26
+        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["n_heads"] = 24
         dit_config["n_kv_heads"] = 8
         dit_config["qk_norm"] = True
@@ -701,6 +715,12 @@ 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

@@ -89,6 +89,7 @@ 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
@@ -330,13 +331,21 @@ 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:
@@ -344,11 +353,11 @@ 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 torch_version_numeric >= (2, 8):
-#                    if any((a in arch) for a in ["gfx1201"]):
-#                        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, 7) and rocm_version >= (6, 4):
-            if any((a in arch) for a in ["gfx1200", "gfx1201", "gfx942", "gfx950"]):  # TODO: more arches
+            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
                 SUPPORT_FP8_OPS = True
 
 except:
@@ -370,6 +379,9 @@ 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)
@@ -925,11 +937,7 @@ def vae_dtype(device=None, allowed_dtypes=[]):
         if d == torch.float16 and should_use_fp16(device):
             return d
 
-        # 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):
+        if d == torch.bfloat16 and should_use_bf16(device):
             return d
 
     return torch.float32
@@ -991,12 +999,6 @@ 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
@@ -1011,6 +1013,16 @@ 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)
@@ -1019,21 +1031,17 @@ def get_offload_stream(device):
 
     if device in STREAMS:
         ss = STREAMS[device]
-        s = ss[stream_counter]
+        #Sync the oldest stream in the queue with the current
+        ss[stream_counter].wait_stream(current_stream(device))
         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 s
+        return ss[stream_counter]
     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):
@@ -1042,18 +1050,14 @@ 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:
+    if stream is None or current_stream(device) is None:
         return
-    if is_device_cuda(device):
-        torch.cuda.current_stream().wait_stream(stream)
-    elif is_device_xpu(device):
-        torch.xpu.current_stream().wait_stream(stream)
+    current_stream(device).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:
@@ -1078,6 +1082,73 @@ 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 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
+
+    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
 
@@ -1330,7 +1401,7 @@ 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 ["gfx1030", "gfx1031", "gfx1010", "gfx1011", "gfx1012", "gfx906", "gfx900", "gfx803"]):  # RDNA2 and older don't support bf16
+        if any((a in arch) for a in AMD_RDNA2_AND_OLDER_ARCH):  # RDNA2 and older don't support bf16
             if manual_cast:
                 return True
             return False

comfy/model_patcher.py

@@ -123,16 +123,30 @@ def move_weight_functions(m, device):
     return memory
 
 class LowVramPatch:
-    def __init__(self, key, patches):
+    def __init__(self, key, patches, convert_func=None, set_func=None):
         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
-            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.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.lora.calculate_weight(self.patches[self.key], weight, self.key, intermediate_dtype=intermediate_dtype)
+        out = comfy.lora.calculate_weight(self.patches[self.key], weight, self.key, intermediate_dtype=intermediate_dtype)
+        if self.set_func is not None:
+            return self.set_func(out, seed=string_to_seed(self.key), return_weight=True).to(dtype=intermediate_dtype)
+        else:
+            return out
 
 def get_key_weight(model, key):
     set_func = None
@@ -224,6 +238,7 @@ 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]] = {}
@@ -261,6 +276,9 @@ 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
 
@@ -280,6 +298,7 @@ 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
 
@@ -436,6 +455,19 @@ 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
 
@@ -604,6 +636,21 @@ 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():
@@ -625,9 +672,11 @@ 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]
@@ -644,6 +693,7 @@ 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
 
@@ -657,16 +707,19 @@ class ModelPatcher:
                         if force_patch_weights:
                             self.patch_weight_to_device(weight_key)
                         else:
-                            m.weight_function = [LowVramPatch(weight_key, self.patches)]
+                            _, set_func, convert_func = get_key_weight(self.model, weight_key)
+                            m.weight_function = [LowVramPatch(weight_key, self.patches, convert_func, set_func)]
                             patch_counter += 1
                     if bias_key in self.patches:
                         if force_patch_weights:
                             self.patch_weight_to_device(bias_key)
                         else:
-                            m.bias_function = [LowVramPatch(bias_key, self.patches)]
+                            _, set_func, convert_func = get_key_weight(self.model, bias_key)
+                            m.bias_function = [LowVramPatch(bias_key, self.patches, convert_func, set_func)]
                             patch_counter += 1
 
                     cast_weight = True
+                    offloaded.append((module_mem, n, m, params))
                 else:
                     if hasattr(m, "comfy_cast_weights"):
                         wipe_lowvram_weight(m)
@@ -697,7 +750,9 @@ class ModelPatcher:
                         continue
 
                 for param in params:
-                    self.patch_weight_to_device("{}.{}".format(n, param), device_to=device_to)
+                    key = "{}.{}".format(n, param)
+                    self.unpin_weight(key)
+                    self.patch_weight_to_device(key, device_to=device_to)
 
                 logging.debug("lowvram: loaded module regularly {} {}".format(n, m))
                 m.comfy_patched_weights = True
@@ -705,11 +760,17 @@ 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 {} {} {}".format(lowvram_model_memory / (1024 * 1024), mem_counter / (1024 * 1024), patch_counter))
+                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))
                 self.model.model_lowvram = True
             else:
-                logging.info("loaded completely {} {} {}".format(lowvram_model_memory / (1024 * 1024), mem_counter / (1024 * 1024), full_load))
+                logging.info("loaded completely; {:.2f} MB usable, {:.2f} MB loaded, full load: {}".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)
@@ -746,6 +807,7 @@ 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)
@@ -825,10 +887,12 @@ class ModelPatcher:
                         module_mem += move_weight_functions(m, device_to)
                         if lowvram_possible:
                             if weight_key in self.patches:
-                                m.weight_function.append(LowVramPatch(weight_key, self.patches))
+                                _, set_func, convert_func = get_key_weight(self.model, weight_key)
+                                m.weight_function.append(LowVramPatch(weight_key, self.patches, convert_func, set_func))
                                 patch_counter += 1
                             if bias_key in self.patches:
-                                m.bias_function.append(LowVramPatch(bias_key, self.patches))
+                                _, set_func, convert_func = get_key_weight(self.model, bias_key)
+                                m.bias_function.append(LowVramPatch(bias_key, self.patches, convert_func, set_func))
                                 patch_counter += 1
                             cast_weight = True
 
@@ -839,6 +903,9 @@ 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
@@ -1241,5 +1308,6 @@ class ModelPatcher:
         self.clear_cached_hook_weights()
 
     def __del__(self):
+        self.unpin_all_weights()
         self.detach(unpatch_all=False)
 

comfy/model_sampling.py

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

comfy/nested_tensor.py

@@ -0,0 +1,91 @@
+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

@@ -24,13 +24,18 @@ 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():
+    if torch.cuda.is_available() and comfy.model_management.WINDOWS:
         from torch.nn.attention import SDPBackend, sdpa_kernel
         import inspect
         if "set_priority" in inspect.signature(sdpa_kernel).parameters:
@@ -50,15 +55,26 @@ try:
 except (ModuleNotFoundError, TypeError):
     logging.warning("Could not set sdpa backend priority.")
 
-cast_to = comfy.model_management.cast_to #TODO: remove once no more references
+NVIDIA_MEMORY_CONV_BUG_WORKAROUND = False
+try:
+    if comfy.model_management.is_nvidia():
+        if torch.backends.cudnn.version() >= 91002 and comfy.model_management.torch_version_numeric >= (2, 9) and comfy.model_management.torch_version_numeric <= (2, 10):
+            #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
 
-if torch.cuda.is_available() and torch.backends.cudnn.is_available() and PerformanceFeature.AutoTune in args.fast:
-    torch.backends.cudnn.benchmark = True
+cast_to = comfy.model_management.cast_to #TODO: remove once no more references
 
 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):
+
+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.
     if input is not None:
         if dtype is None:
             dtype = input.dtype
@@ -67,32 +83,58 @@ def cast_bias_weight(s, input=None, dtype=None, device=None, bias_dtype=None):
         if device is None:
             device = input.device
 
-    offload_stream = comfy.model_management.get_offload_stream(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
+
     if offload_stream is not None:
         wf_context = offload_stream
     else:
         wf_context = contextlib.nullcontext()
 
-    bias = None
     non_blocking = comfy.model_management.device_supports_non_blocking(device)
-    if s.bias is not None:
-        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 has_function:
+    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
+    if s.bias is not None:
+        bias = comfy.model_management.cast_to(s.bias, bias_dtype, device, non_blocking=non_blocking, copy=bias_has_function, stream=offload_stream)
+
+        if bias_has_function:
             with wf_context:
                 for f in s.bias_function:
                     bias = f(bias)
 
-    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:
+    weight = weight.to(dtype=dtype)
+    if weight_has_function:
         with wf_context:
             for f in s.weight_function:
                 weight = f(weight)
 
     comfy.model_management.sync_stream(device, offload_stream)
-    return weight, bias
+    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))
+
 
 class CastWeightBiasOp:
     comfy_cast_weights = False
@@ -105,10 +147,13 @@ class disable_weight_init:
             return None
 
         def forward_comfy_cast_weights(self, input):
-            weight, bias = cast_bias_weight(self, input)
-            return torch.nn.functional.linear(input, weight, bias)
+            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
 
         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:
@@ -119,10 +164,13 @@ class disable_weight_init:
             return None
 
         def forward_comfy_cast_weights(self, input):
-            weight, bias = cast_bias_weight(self, input)
-            return self._conv_forward(input, weight, bias)
+            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
 
         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:
@@ -133,10 +181,13 @@ class disable_weight_init:
             return None
 
         def forward_comfy_cast_weights(self, input):
-            weight, bias = cast_bias_weight(self, input)
-            return self._conv_forward(input, weight, bias)
+            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
 
         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:
@@ -146,11 +197,23 @@ 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 = cast_bias_weight(self, input)
-            return self._conv_forward(input, weight, bias)
+            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
 
         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:
@@ -161,10 +224,13 @@ class disable_weight_init:
             return None
 
         def forward_comfy_cast_weights(self, input):
-            weight, bias = cast_bias_weight(self, input)
-            return torch.nn.functional.group_norm(input, self.num_groups, weight, bias, self.eps)
+            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
 
         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:
@@ -176,13 +242,17 @@ class disable_weight_init:
 
         def forward_comfy_cast_weights(self, input):
             if self.weight is not None:
-                weight, bias = cast_bias_weight(self, input)
+                weight, bias, offload_stream = cast_bias_weight(self, input, offloadable=True)
             else:
                 weight = None
                 bias = None
-            return torch.nn.functional.layer_norm(input, self.normalized_shape, weight, bias, self.eps)
+                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
 
         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:
@@ -195,13 +265,18 @@ class disable_weight_init:
 
         def forward_comfy_cast_weights(self, input):
             if self.weight is not None:
-                weight, bias = cast_bias_weight(self, input)
+                weight, bias, offload_stream = cast_bias_weight(self, input, offloadable=True)
             else:
                 weight = None
-            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)
+                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
 
         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:
@@ -217,12 +292,15 @@ class disable_weight_init:
                 input, output_size, self.stride, self.padding, self.kernel_size,
                 num_spatial_dims, self.dilation)
 
-            weight, bias = cast_bias_weight(self, input)
-            return torch.nn.functional.conv_transpose2d(
+            weight, bias, offload_stream = cast_bias_weight(self, input, offloadable=True)
+            x = 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:
@@ -238,12 +316,15 @@ class disable_weight_init:
                 input, output_size, self.stride, self.padding, self.kernel_size,
                 num_spatial_dims, self.dilation)
 
-            weight, bias = cast_bias_weight(self, input)
-            return torch.nn.functional.conv_transpose1d(
+            weight, bias, offload_stream = cast_bias_weight(self, input, offloadable=True)
+            x = 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:
@@ -258,10 +339,14 @@ 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 = 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)
+            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
+
 
         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:
@@ -312,20 +397,18 @@ 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
 
-    tensor_2d = False
-    if len(input.shape) == 2:
-        tensor_2d = True
-        input = input.unsqueeze(1)
-
-    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()
+
+    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)
 
         scale_weight = self.scale_weight
         scale_input = self.scale_input
@@ -337,23 +420,20 @@ 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)
-            input = input.reshape(-1, input_shape[2]).to(dtype).contiguous()
+            layout_params_weight = {'scale': scale_input, 'orig_dtype': input_dtype}
+            quantized_input = QuantizedTensor(input.to(dtype).contiguous(), "TensorCoreFP8Layout", layout_params_weight)
         else:
             scale_input = scale_input.to(input.device)
-            input = (input * (1.0 / scale_input).to(input_dtype)).reshape(-1, input_shape[2]).to(dtype).contiguous()
+            quantized_input = QuantizedTensor.from_float(input, "TensorCoreFP8Layout", scale=scale_input, dtype=dtype)
 
-        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)
+        # 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 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]))
+        uncast_bias_weight(self, w, bias, offload_stream)
+        return o
 
     return None
 
@@ -373,8 +453,10 @@ class fp8_ops(manual_cast):
                 except Exception as e:
                     logging.info("Exception during fp8 op: {}".format(e))
 
-            weight, bias = cast_bias_weight(self, input)
-            return torch.nn.functional.linear(input, weight, bias)
+            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
 
 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))
@@ -402,12 +484,14 @@ def scaled_fp8_ops(fp8_matrix_mult=False, scale_input=False, override_dtype=None
                     if out is not None:
                         return out
 
-                weight, bias = cast_bias_weight(self, input)
+                weight, bias, offload_stream = cast_bias_weight(self, input, offloadable=True)
 
                 if weight.numel() < input.numel(): #TODO: optimize
-                    return torch.nn.functional.linear(input, weight * self.scale_weight.to(device=weight.device, dtype=weight.dtype), bias)
+                    x = torch.nn.functional.linear(input, weight * self.scale_weight.to(device=weight.device, dtype=weight.dtype), bias)
                 else:
-                    return torch.nn.functional.linear(input * self.scale_weight.to(device=weight.device, dtype=weight.dtype), weight, bias)
+                    x = torch.nn.functional.linear(input * self.scale_weight.to(device=weight.device, dtype=weight.dtype), weight, bias)
+                uncast_bias_weight(self, weight, bias, offload_stream)
+                return x
 
             def convert_weight(self, weight, inplace=False, **kwargs):
                 if inplace:
@@ -416,8 +500,10 @@ 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, **kwargs):
+            def set_weight(self, weight, inplace_update=False, seed=None, return_weight=False, **kwargs):
                 weight = comfy.float.stochastic_rounding(weight / self.scale_weight.to(device=weight.device, dtype=weight.dtype), self.weight.dtype, seed=seed)
+                if return_weight:
+                    return weight
                 if inplace_update:
                     self.weight.data.copy_(weight)
                 else:
@@ -444,7 +530,130 @@ if CUBLAS_IS_AVAILABLE:
             def forward(self, *args, **kwargs):
                 return super().forward(*args, **kwargs)
 
-def pick_operations(weight_dtype, compute_dtype, load_device=None, disable_fast_fp8=False, fp8_optimizations=False, scaled_fp8=None):
+
+# ==============================================================================
+# Mixed Precision Operations
+# ==============================================================================
+from .quant_ops import QuantizedTensor
+
+QUANT_FORMAT_MIXINS = {
+    "float8_e4m3fn": {
+        "dtype": torch.float8_e4m3fn,
+        "layout_type": "TensorCoreFP8Layout",
+        "parameters": {
+            "weight_scale": torch.nn.Parameter(torch.zeros((), dtype=torch.float32), requires_grad=False),
+            "input_scale": torch.nn.Parameter(torch.zeros((), dtype=torch.float32), requires_grad=False),
+        }
+    }
+}
+
+class MixedPrecisionOps(disable_weight_init):
+    _layer_quant_config = {}
+    _compute_dtype = torch.bfloat16
+
+    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}")
+
+                mixin = QUANT_FORMAT_MIXINS[quant_format]
+                self.layout_type = mixin["layout_type"]
+
+                scale_key = f"{prefix}weight_scale"
+                layout_params = {
+                    'scale': state_dict.pop(scale_key, None),
+                    'orig_dtype': MixedPrecisionOps._compute_dtype
+                }
+                if layout_params['scale'] is not None:
+                    manually_loaded_keys.append(scale_key)
+
+                self.weight = torch.nn.Parameter(
+                    QuantizedTensor(weight.to(device=device, dtype=mixin["dtype"]), self.layout_type, layout_params),
+                    requires_grad=False
+                )
+
+                for param_name, param_value in mixin["parameters"].items():
+                    param_key = f"{prefix}{param_name}"
+                    _v = state_dict.pop(param_key, None)
+                    if _v is None:
+                        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, fp8_dtype=self.weight.dtype)
+            return self._forward(input, self.weight, self.bias)
+
+
+def pick_operations(weight_dtype, compute_dtype, load_device=None, disable_fast_fp8=False, fp8_optimizations=False, scaled_fp8=None, model_config=None):
+    if model_config and hasattr(model_config, 'layer_quant_config') and model_config.layer_quant_config:
+        MixedPrecisionOps._layer_quant_config = model_config.layer_quant_config
+        MixedPrecisionOps._compute_dtype = compute_dtype
+        logging.info(f"Using mixed precision operations: {len(model_config.layer_quant_config)} quantized layers")
+        return MixedPrecisionOps
+
     fp8_compute = comfy.model_management.supports_fp8_compute(load_device)
     if scaled_fp8 is not None:
         return scaled_fp8_ops(fp8_matrix_mult=fp8_compute and fp8_optimizations, scale_input=fp8_optimizations, override_dtype=scaled_fp8)

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(value, curr_value)
+            merged_dict[key] = merge_nested_dicts(curr_value, value)
         elif isinstance(value, list):
             merged_dict.setdefault(key, []).extend(value)
         else:

comfy/quant_ops.py

@@ -0,0 +1,512 @@
+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
+
+
+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]
+
+    if isinstance(qt_dest, QuantizedTensor):
+        if isinstance(src, QuantizedTensor):
+            # Copy from another quantized tensor
+            qt_dest._qdata.copy_(src._qdata)
+            qt_dest._layout_type = src._layout_type
+            qt_dest._layout_params = _copy_layout_params(src._layout_params)
+        else:
+            # Copy from regular tensor - just copy raw data
+            qt_dest._qdata.copy_(src)
+        return qt_dest
+    return func(*args, **kwargs)
+
+
+@register_generic_util(torch.ops.aten._has_compatible_shallow_copy_type.default)
+def generic_has_compatible_shallow_copy_type(func, args, kwargs):
+    return True
+
+# ==============================================================================
+# FP8 Layout + Operation Handlers
+# ==============================================================================
+class TensorCoreFP8Layout(QuantizedLayout):
+    """
+    Storage format:
+    - qdata: FP8 tensor (torch.float8_e4m3fn or torch.float8_e5m2)
+    - scale: Scalar tensor (float32) for dequantization
+    - orig_dtype: Original dtype before quantization (for casting back)
+    """
+    @classmethod
+    def quantize(cls, tensor, scale=None, dtype=torch.float8_e4m3fn):
+        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']
+
+
+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

@@ -4,13 +4,9 @@ import comfy.samplers
 import comfy.utils
 import numpy as np
 import logging
+import comfy.nested_tensor
 
-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)
+def prepare_noise_inner(latent_image, generator, noise_inds=None):
     if noise_inds is None:
         return torch.randn(latent_image.size(), dtype=latent_image.dtype, layout=latent_image.layout, generator=generator, device="cpu")
 
@@ -21,10 +17,29 @@ def prepare_noise(latent_image, seed, noise_inds=None):
         if i in unique_inds:
             noises.append(noise)
     noises = [noises[i] for i in inverse]
-    noises = torch.cat(noises, axis=0)
+    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)
+
     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

@@ -306,17 +306,10 @@ def _calc_cond_batch(model: BaseModel, conds: list[list[dict]], x_in: torch.Tens
                                                                                  copy_dict1=False)
 
             if patches is not None:
-                # 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["patches"] = comfy.patcher_extension.merge_nested_dicts(
+                    transformer_options.get("patches", {}),
+                    patches
+                )
 
             transformer_options["cond_or_uncond"] = cond_or_uncond[:]
             transformer_options["uuids"] = uuids[:]
@@ -789,7 +782,7 @@ 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):
+def process_conds(model, noise, conds, device, latent_image=None, denoise_mask=None, seed=None, latent_shapes=None):
     for k in conds:
         conds[k] = conds[k][:]
         resolve_areas_and_cond_masks_multidim(conds[k], noise.shape[2:], device)
@@ -799,7 +792,7 @@ 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)
+            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)
 
     #make sure each cond area has an opposite one with the same area
     for k in conds:
@@ -969,11 +962,11 @@ 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):
+    def inner_sample(self, noise, latent_image, device, sampler, sigmas, denoise_mask, callback, disable_pbar, seed, latent_shapes=None):
         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)
+        self.conds = process_conds(self.inner_model, noise, self.conds, device, latent_image, denoise_mask, seed, latent_shapes=latent_shapes)
 
         extra_model_options = comfy.model_patcher.create_model_options_clone(self.model_options)
         extra_model_options.setdefault("transformer_options", {})["sample_sigmas"] = sigmas
@@ -987,7 +980,7 @@ 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):
+    def outer_sample(self, noise, latent_image, sampler, sigmas, denoise_mask=None, callback=None, disable_pbar=False, seed=None, latent_shapes=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
 
@@ -1001,7 +994,7 @@ class CFGGuider:
 
         try:
             self.model_patcher.pre_run()
-            output = self.inner_sample(noise, latent_image, device, sampler, sigmas, denoise_mask, callback, disable_pbar, seed)
+            output = self.inner_sample(noise, latent_image, device, sampler, sigmas, denoise_mask, callback, disable_pbar, seed, latent_shapes=latent_shapes)
         finally:
             self.model_patcher.cleanup()
 
@@ -1014,6 +1007,12 @@ 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]))
@@ -1033,7 +1032,7 @@ 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)
+            output = executor.execute(noise, latent_image, sampler, sigmas, denoise_mask, callback, disable_pbar, seed, latent_shapes=latent_shapes)
         finally:
             cast_to_load_options(self.model_options, device=self.model_patcher.offload_device)
             self.model_options = orig_model_options
@@ -1041,6 +1040,9 @@ 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

@@ -18,6 +18,7 @@ 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
@@ -142,6 +143,9 @@ 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)
 
@@ -275,8 +279,13 @@ class VAE:
         if 'decoder.up_blocks.0.resnets.0.norm1.weight' in sd.keys(): #diffusers format
             sd = diffusers_convert.convert_vae_state_dict(sd)
 
-        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)
+        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.downscale_ratio = 8
         self.upscale_ratio = 8
         self.latent_channels = 4
@@ -287,10 +296,12 @@ 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:
@@ -332,35 +343,51 @@ 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:
-                #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:
+                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.modules.diffusionmodules.model.Encoder", 'params': ddconfig},
-                                                                decoder_config={'target': "comfy.ldm.modules.diffusionmodules.model.Decoder", 'params': ddconfig})
+                                                                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 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})
+
+                    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)
+                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})
             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)
@@ -526,6 +553,25 @@ 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
@@ -553,12 +599,25 @@ 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]
@@ -636,6 +695,7 @@ 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)
@@ -651,6 +711,13 @@ 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)
@@ -697,6 +764,7 @@ 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)
@@ -718,6 +786,13 @@ 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
@@ -858,6 +933,7 @@ 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:
@@ -880,6 +956,8 @@ 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']
@@ -984,6 +1062,10 @@ 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)
@@ -1194,7 +1276,7 @@ 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={}):
+def load_diffusion_model_state_dict(sd, model_options={}, metadata=None):
     """
     Loads a UNet diffusion model from a state dictionary, supporting both diffusers and regular formats.
 
@@ -1228,7 +1310,7 @@ def load_diffusion_model_state_dict(sd, model_options={}):
     weight_dtype = comfy.utils.weight_dtype(sd)
 
     load_device = model_management.get_torch_device()
-    model_config = model_detection.model_config_from_unet(sd, "")
+    model_config = model_detection.model_config_from_unet(sd, "", metadata=metadata)
 
     if model_config is not None:
         new_sd = sd
@@ -1262,7 +1344,10 @@ def load_diffusion_model_state_dict(sd, model_options={}):
     else:
         unet_dtype = dtype
 
-    manual_cast_dtype = model_management.unet_manual_cast(unet_dtype, load_device, model_config.supported_inference_dtypes)
+    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)
     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):
@@ -1278,8 +1363,8 @@ def load_diffusion_model_state_dict(sd, model_options={}):
 
 
 def load_diffusion_model(unet_path, model_options={}):
-    sd = comfy.utils.load_torch_file(unet_path)
-    model = load_diffusion_model_state_dict(sd, model_options=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)
     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/supported_models_base.py

@@ -50,6 +50,7 @@ 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/llama.py

@@ -3,6 +3,7 @@ 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
@@ -28,6 +29,9 @@ class Llama2Config:
     mlp_activation = "silu"
     qkv_bias = False
     rope_dims = None
+    q_norm = None
+    k_norm = None
+    rope_scale = None
 
 @dataclass
 class Qwen25_3BConfig:
@@ -46,6 +50,9 @@ class Qwen25_3BConfig:
     mlp_activation = "silu"
     qkv_bias = True
     rope_dims = None
+    q_norm = None
+    k_norm = None
+    rope_scale = None
 
 @dataclass
 class Qwen25_7BVLI_Config:
@@ -64,6 +71,9 @@ class Qwen25_7BVLI_Config:
     mlp_activation = "silu"
     qkv_bias = True
     rope_dims = [16, 24, 24]
+    q_norm = None
+    k_norm = None
+    rope_scale = None
 
 @dataclass
 class Gemma2_2B_Config:
@@ -82,6 +92,32 @@ 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
+
+@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]
 
 class RMSNorm(nn.Module):
     def __init__(self, dim: int, eps: float = 1e-5, add=False, device=None, dtype=None):
@@ -106,25 +142,40 @@ def rotate_half(x):
     return torch.cat((-x2, x1), dim=-1)
 
 
-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))
+def precompute_freqs_cis(head_dim, position_ids, theta, rope_scale=None, rope_dims=None, device=None):
+    if not isinstance(theta, list):
+        theta = [theta]
 
-    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 = []
+    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))
 
-    return (cos, sin)
+        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
 
 
 def apply_rope(xq, xk, freqs_cis):
@@ -152,6 +203,14 @@ 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,
@@ -168,6 +227,11 @@ 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)
@@ -192,7 +256,7 @@ 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, device=None, dtype=None, ops: Any = None):
+    def __init__(self, config: Llama2Config, index, 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)
@@ -226,7 +290,7 @@ class TransformerBlock(nn.Module):
         return x
 
 class TransformerBlockGemma2(nn.Module):
-    def __init__(self, config: Llama2Config, device=None, dtype=None, ops: Any = None):
+    def __init__(self, config: Llama2Config, index, 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)
@@ -235,6 +299,13 @@ 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,
@@ -242,6 +313,14 @@ 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)
@@ -276,7 +355,7 @@ class Llama2_(nn.Module):
             device=device,
             dtype=dtype
         )
-        if self.config.transformer_type == "gemma2":
+        if self.config.transformer_type == "gemma2" or self.config.transformer_type == "gemma3":
             transformer = TransformerBlockGemma2
             self.normalize_in = True
         else:
@@ -284,8 +363,8 @@ class Llama2_(nn.Module):
             self.normalize_in = False
 
         self.layers = nn.ModuleList([
-            transformer(config, device=device, dtype=dtype, ops=ops)
-            for _ in range(config.num_hidden_layers)
+            transformer(config, index=i, device=device, dtype=dtype, ops=ops)
+            for i in range(config.num_hidden_layers)
         ])
         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)
@@ -305,6 +384,7 @@ 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)
 
@@ -433,3 +513,12 @@ 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

@@ -11,23 +11,41 @@ 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={}):
-        super().__init__(device=device, dtype=dtype, name="gemma2_2b", clip_model=Gemma2_2BModel, model_options=model_options)
+    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 te(dtype_llama=None, llama_scaled_fp8=None):
+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
+
     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:
@@ -35,5 +53,5 @@ def te(dtype_llama=None, llama_scaled_fp8=None):
                 model_options["scaled_fp8"] = llama_scaled_fp8
             if dtype_llama is not None:
                 dtype = dtype_llama
-            super().__init__(device=device, dtype=dtype, model_options=model_options)
+            super().__init__(device=device, dtype=dtype, name=model_type, model_options=model_options, clip_model=model)
     return LuminaTEModel_

comfy/utils.py

@@ -39,7 +39,11 @@ if hasattr(torch.serialization, "add_safe_globals"):  # TODO: this was added in
         pass
     ModelCheckpoint.__module__ = "pytorch_lightning.callbacks.model_checkpoint"
 
-    from numpy.core.multiarray import scalar
+    def scalar(*args, **kwargs):
+        from numpy.core.multiarray import scalar as sc
+        return sc(*args, **kwargs)
+    scalar.__module__ = "numpy.core.multiarray"
+
     from numpy import dtype
     from numpy.dtypes import Float64DType
     from _codecs import encode
@@ -50,16 +54,10 @@ if hasattr(torch.serialization, "add_safe_globals"):  # TODO: this was added in
 else:
     logging.info("Warning, you are using an old pytorch version and some ckpt/pt files might be loaded unsafely. Upgrading to 2.4 or above is recommended.")
 
-def is_html_file(file_path):
-    with open(file_path, "rb") as f:
-        content = f.read(100)
-        return b"<!DOCTYPE html>" in content or b"<html" in content
-
 def load_torch_file(ckpt, safe_load=False, device=None, return_metadata=False):
     if device is None:
         device = torch.device("cpu")
     metadata = None
-
     if ckpt.lower().endswith(".safetensors") or ckpt.lower().endswith(".sft"):
         try:
             with safetensors.safe_open(ckpt, framework="pt", device=device.type) as f:
@@ -72,8 +70,6 @@ def load_torch_file(ckpt, safe_load=False, device=None, return_metadata=False):
                 if return_metadata:
                     metadata = f.metadata()
         except Exception as e:
-            if is_html_file(ckpt):
-                raise ValueError("{}\n\nFile path: {}\n\nThe requested file is an HTML document not a safetensors file. Please re-download the file, not the web page.".format(e, ckpt))
             if len(e.args) > 0:
                 message = e.args[0]
                 if "HeaderTooLarge" in message:
@@ -101,8 +97,6 @@ def load_torch_file(ckpt, safe_load=False, device=None, return_metadata=False):
                     sd = pl_sd
             else:
                 sd = pl_sd
-
-    # populate_db_with_asset(ckpt) # surprise tool that can help us later - performs hashing on model file
     return (sd, metadata) if return_metadata else sd
 
 def save_torch_file(sd, ckpt, metadata=None):
@@ -1112,3 +1106,25 @@ 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

@@ -8,8 +8,8 @@ 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
-from comfy_api.latest._io import _IO as io  #noqa: F401
-from comfy_api.latest._ui import _UI as ui  #noqa: F401
+from . import _io as io
+from . import _ui as ui
 # 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
@@ -114,6 +114,10 @@ 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",
@@ -121,4 +125,8 @@ __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
+from typing import Optional, Union, IO
 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: str,
+        path: Union[str, IO[bytes]],
         format: VideoContainer = VideoContainer.AUTO,
         codec: VideoCodec = VideoCodec.AUTO,
         metadata: Optional[dict] = None

comfy_api/latest/_io.py

@@ -336,11 +336,25 @@ class Combo(ComfyTypeIO):
     class Input(WidgetInput):
         """Combo input (dropdown)."""
         Type = str
-        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):
+        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
             super().__init__(id, display_name, optional, tooltip, lazy, default, socketless)
             self.multiselect = False
             self.options = options
@@ -392,20 +406,6 @@ class MultiCombo(ComfyTypeI):
             })
             return to_return
 
-@comfytype(io_type="ASSET")
-class Asset(ComfyTypeI):
-    class Input(WidgetInput):
-        def __init__(self, id: str, query_tags: list[str], display_name: str=None, optional=False, tooltip: str=None, lazy: bool=None,
-                    default: str=None, socketless: bool=None):
-            super().__init__(id, display_name, optional, tooltip, lazy, default, socketless)
-            self.query_tags = query_tags
-
-        def as_dict(self):
-            to_return = super().as_dict() | prune_dict({
-                "query_tags": self.query_tags
-            })
-            return to_return
-
 @comfytype(io_type="IMAGE")
 class Image(ComfyTypeIO):
     Type = torch.Tensor
@@ -1582,77 +1582,78 @@ class _UIOutput(ABC):
         ...
 
 
-class _IO:
-    FolderType = FolderType
-    UploadType = UploadType
-    RemoteOptions = RemoteOptions
-    NumberDisplay = NumberDisplay
+__all__ = [
+    "FolderType",
+    "UploadType",
+    "RemoteOptions",
+    "NumberDisplay",
 
-    comfytype = staticmethod(comfytype)
-    Custom = staticmethod(Custom)
-    Input = Input
-    WidgetInput = WidgetInput
-    Output = Output
-    ComfyTypeI = ComfyTypeI
-    ComfyTypeIO = ComfyTypeIO
-    #---------------------------------
+    "comfytype",
+    "Custom",
+    "Input",
+    "WidgetInput",
+    "Output",
+    "ComfyTypeI",
+    "ComfyTypeIO",
     # Supported Types
-    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)
+    "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",
+]

comfy_api/latest/_ui.py

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

comfy_api_nodes/apinode_utils.py

@@ -1,691 +0,0 @@
-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

@@ -1,17 +0,0 @@
-# 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

@@ -1,57 +0,0 @@
-# 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

@@ -2,6 +2,7 @@
 #   filename:  filtered-openapi.yaml
 #   timestamp: 2025-07-30T08:54:00+00:00
 
+# pylint: disable
 from __future__ import annotations
 
 from datetime import date, datetime
@@ -1320,6 +1321,7 @@ 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):
@@ -1354,6 +1356,7 @@ 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

@@ -50,44 +50,6 @@ 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(
@@ -160,15 +122,8 @@ class BFLStatus(str, Enum):
     error = "Error"
 
 
-class BFLFluxProStatusResponse(BaseModel):
+class BFLFluxStatusResponse(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: 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)."
-    )
+    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)

comfy_api_nodes/apis/client.py

@@ -1,957 +0,0 @@
-"""
-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."
-        )