Merge private DynamicGroup parent and update LoRA inputs

Author: jaeone94

.ci/windows_nvidia_base_files/advanced/run_nvidia_gpu_disable_api_nodes.bat

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

.github/workflows/ruff.yml

@@ -38,7 +38,7 @@ jobs:
     - 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 torch torchvision --index-url https://download.pytorch.org/whl/cpu
         pip install -r requirements.txt
 
     - name: Install Pylint

.github/workflows/test-execution.yml

@@ -22,7 +22,7 @@ jobs:
     - 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 torch torchvision --index-url https://download.pytorch.org/whl/cpu
         pip install -r requirements.txt
         pip install -r tests-unit/requirements.txt
     - name: Run Execution Tests

.github/workflows/test-launch.yml

@@ -21,7 +21,7 @@ jobs:
     - 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 torch torchvision --index-url https://download.pytorch.org/whl/cpu
         pip install -r requirements.txt
         pip install wait-for-it
       working-directory: ComfyUI

.github/workflows/test-unit.yml

@@ -22,7 +22,7 @@ jobs:
     - 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 torch torchvision --index-url https://download.pytorch.org/whl/cpu
         pip install -r requirements.txt
     - name: Run Unit Tests
       run: |

AGENTS.md

@@ -186,6 +186,10 @@
   handwritten implementations of the same operation. Remove duplicate local
   kernels and adapt inputs to the shared operation's documented layout while
   preserving the model's original math and output contract.
+- Model implementations must use `AttentionTensorContainer` and per-module
+  `self.comfy_attention = ComfyAttention()`, passed to attention as `preferred_attention`.
+- Integrate model block loops with the memory compiler and prefetch helpers in
+  `comfy.model_prefetch`, following existing model patterns.
 - All models should use the optimized attention function selected by ComfyUI.
   Treat optimized backend functions, dispatch helpers, and capability-selected
   callables as opaque. Higher-level code must not inspect function identity,

CODEOWNERS

@@ -1,4 +1,4 @@
-* @comfyanonymous @kosinkadink @guill @alexisrolland @rattus128 @kijai
+* @comfyanonymous @guill @alexisrolland @rattus128 @kijai
 
 /CODEOWNERS @comfyanonymous
 /AGENTS.md @comfyanonymous

QUANTIZATION.md

@@ -152,6 +152,27 @@ Example:
 
 To create compatible checkpoints, use any quantization tool provided the output follows the checkpoint format described above and uses a layout defined in `QUANT_ALGOS`.
 
+### Diffusion attention preferences
+
+A diffusion attention module can have a `<module path>.comfy_attention.config` entry whose
+uint8 tensor contains UTF-8 JSON:
+
+```json
+{"attention": "comfy_kitchen_int8"}
+```
+
+Use the module that performs attention, such as `transformer_blocks.0.attn` for
+Qwen Image 2.1 or `blocks.0.attn` for MiniMax H3.
+
+Only `comfy_kitchen_int8` is supported. Invalid targets and other method names
+are ignored with a warning during loading, leaving normal attention selection.
+Kitchen INT8 support is checked when each preference is loaded
+for the primary device; unsupported devices keep normal attention selection.
+Explicit attention overrides retain priority.
+The `ComfyAttention` child module loads and saves its own metadata through normal
+state-dict loading and saving. Preferences do not enable weight
+quantization. Text encoder and VAE loaders do not apply these preferences.
+
 ### Weight Quantization
 
 Weight quantization is straightforward - compute the scaling factor directly from the weight tensor using the absolute maximum method described earlier. Each layer's weights are quantized independently and stored with their corresponding `weight_scale` parameter.

README.md

@@ -37,6 +37,8 @@
 
 ComfyUI is the AI creation engine for visual professionals who demand control over every model, every parameter, and every output. Its powerful and modular node graph interface empowers creatives to generate images, videos, 3D models, audio, and more...
 - ComfyUI natively supports the latest open-source state of the art models.
+- It is by far the most optimized inference engine for diffusion models in the world with a focus on local consumer hardware.
+- Can run even the biggest open source models on as low as 4GB vram + 8GB ram relatively quickly (saturating your GPU compute) using our state of the art asynchronous weight streaming technology.
 - [Partner nodes](https://docs.comfy.org/tutorials/partner-nodes/overview#partner-nodes) provide access to the best closed source models such as Nano Banana, Seedance, Hunyuan3D, etc.
 - It is available on Windows, Linux, and macOS, locally with our [desktop application](https://www.comfy.org/download), our [portable install](#installing) or on our [cloud](https://www.comfy.org/cloud).
 - The most sophisticated workflows can be exposed through a simple UI thanks to App Mode.
@@ -70,13 +72,13 @@ See what ComfyUI can do with the [newer template workflows](https://comfy.org/wo
   - [Image editing](https://comfy.org/workflows/tag/image-edit/): Flux Kontext, Flux.2 Klein, Qwen Image Edit, HiDream E1.1 and O1, OmniGen2, Boogu, JoyImage Edit, MageFlow Edit, and LongCat Image Edit.
   - [Video generation](https://comfy.org/workflows/tag/video-generation/): Wan 2.1 and 2.2, LTX-Video 2 and 2.3, HunyuanVideo 1.5, Kandinsky 5 Video, CogVideoX, Cosmos Predict2, Bernini-R, SCAIL 2, and Mochi.
   - [Audio and video generation](https://comfy.org/workflows/): MiniMax H3 and LTX-AV.
-  - [Audio generation](https://comfy.org/workflows/tag/text-to-audio/): ACE-Step 1.5, Stable Audio 3 and MiniMax Music 3
+  - [Audio generation](https://comfy.org/workflows/tag/text-to-audio/): ACE-Step 1.5, Stable Audio 3, MiniMax Music 3 and Yue 2.
   - [3D and vision](https://comfy.org/workflows/): Hunyuan3D 2.1, TripoSplat, SeedVR2, SUPIR, Depth Anything 3, MoGe, SAM 3 and 3.1, RT-DETRv4, and BiRefNet.
   - [Text generation](https://comfy.org/workflows/tag/text-generation/): Gemma 3 and 4, Qwen3, Qwen3.5, and Qwen3-VL, including multimodal inputs.
 - Load complete checkpoints or separate diffusion models, VAEs, text encoders, LoRAs, ControlNets, adapters, and upscalers from supported model formats.
 - Built-in tools for inpainting, outpainting, reference conditioning, masks and compositing, model merging, upscaling, frame interpolation, segmentation, depth estimation, and media processing.
 - Save and load workflows as JSON, or recover complete workflows and seeds from supported generated media.
-- Runs fully offline: core does not download anything unless you request it. Use `--disable-api-nodes` to disable the optional paid [Comfy API nodes](https://docs.comfy.org/tutorials/api-nodes/overview) and force all built-in functionality to stay offline.
+- Runs fully offline: core does not download anything unless you request it. Use `--offline` to disable the optional paid [Comfy API nodes](https://docs.comfy.org/tutorials/api-nodes/overview) and force all built-in functionality to stay offline, or `--disable-partner-nodes` to only disable the API nodes.
 - Extend ComfyUI with custom nodes
 - Configure additional model locations with [`extra_model_paths.yaml`](extra_model_paths.yaml.example).
 - Support for saving and loading high bit depth images and videos: 16 bit PNG images, 32 bit EXR, 10 bit AVIF are supported and more.

alembic_db/versions/0001_assets.py

@@ -150,7 +150,6 @@ def downgrade() -> None:
 
     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")
@@ -160,7 +159,6 @@ def downgrade() -> None:
     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")