Synchronization

.github/workflows/cla.yml

@@ -22,7 +22,7 @@ jobs:
       - name: Build author-only allowlist
         id: allowlist
         if: >
-          github.event_name == 'pull_request_target' ||
+          (github.event_name == 'pull_request_target' && github.event.action != 'closed') ||
           (github.event_name == 'issue_comment' && github.event.issue.pull_request && (
             github.event.comment.body == 'recheck' ||
             github.event.comment.body == 'I have read and agree to the Contributor License Agreement'
@@ -48,10 +48,10 @@ jobs:
           fi
 
       - name: CLA Assistant
-        # Run on PR events, on "recheck" comment, or when someone posts the signing phrase.
+        # Run on open/update PR events, on "recheck", or when someone posts the signing phrase.
         # IMPORTANT: this phrase must match `custom-pr-sign-comment` below.
         if: >
-          github.event_name == 'pull_request_target' ||
+          (github.event_name == 'pull_request_target' && github.event.action != 'closed') ||
           (github.event_name == 'issue_comment' && github.event.issue.pull_request && (
             github.event.comment.body == 'recheck' ||
             github.event.comment.body == 'I have read and agree to the Contributor License Agreement'
@@ -62,6 +62,8 @@ jobs:
           # PAT required to write to the centralized signatures repo.
           PERSONAL_ACCESS_TOKEN: ${{ secrets.PERSONAL_ACCESS_TOKEN }}
         with:
+          lock-pullrequest-aftermerge: false
+
           # Where the CLA document lives (shown to contributors)
           path-to-document: https://github.com/Comfy-Org/comfy-cla/blob/main/comfyui_icla.md
 
@@ -94,3 +96,25 @@ jobs:
 
           custom-allsigned-prcomment: |
             ✅ All contributors have signed the CLA. Thank you! This PR is ready to be merged.
+
+      - name: Lock PR containing a CLA signature or bot approval
+        if: github.event_name == 'pull_request_target' && github.event.action == 'closed' && github.event.pull_request.merged == true
+        uses: actions/github-script@v7
+        with:
+          retries: 3
+          script: |
+            const pr = context.payload.pull_request;
+            const issue = { ...context.repo, issue_number: pr.number };
+            const comments = await github.paginate(github.rest.issues.listComments, {
+              ...issue,
+              per_page: 100,
+            });
+            const signed = comments.some(comment =>
+              (comment.user?.id === pr.user.id &&
+                comment.body?.trim().toLowerCase() === 'i have read and agree to the contributor license agreement') ||
+              (comment.user?.login === 'github-actions[bot]' &&
+                comment.body?.startsWith('✅ All contributors have signed the CLA. Thank you! This PR is ready to be merged.'))
+            );
+            if (signed) {
+              await github.rest.issues.lock(issue);
+            }

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

@@ -70,7 +70,7 @@ 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.
@@ -214,21 +214,30 @@ This is the command to install the nightly with ROCm 7.2 which might have some p
 ```pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/rocm7.2```
 
 
-### AMD GPUs (Experimental: Windows and Linux), RDNA 3, 3.5 and 4 only.
+### AMD GPUs (Windows, ROCm 10.0)
 
-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.
+Use AMD's [multi-architecture PyTorch packages](https://rocm.docs.amd.com/projects/ai-ecosystem/en/latest/frameworks/pytorch/install.html). The `device-*` extras install your GPU's kernels and the matching ROCm runtime automatically; a separate HIP SDK installation is not needed.
 
-RDNA 3 (RX 7000 series):
+Use Windows 11, a current [AMD graphics driver](https://www.amd.com/en/support/download/drivers.html), and 64-bit Python 3.13.
 
-```pip install --pre torch torchvision torchaudio --index-url https://rocm.nightlies.amd.com/v2/gfx110X-all/```
+The install command below uses `device-all` to install kernels for all supported GPUs. To reduce download size and disk usage, optionally replace **both** occurrences of `device-all` with the target for your GPU:
 
-RDNA 3.5 (Strix halo/Ryzen AI Max+ 365):
+| GPU | Device extra |
+| --- | --- |
+| RX 9070 / XT, Radeon AI PRO R9700 | `device-gfx1201` |
+| RX 9060 / XT | `device-gfx1200` |
+| RX 7900 XT / XTX | `device-gfx1100` |
+| RX 7700 XT / 7800 XT | `device-gfx1101` |
+| RX 7600 / XT | `device-gfx1102` |
+| Ryzen AI Max / Max+ (Strix Halo) | `device-gfx1151` |
 
-```pip install --pre torch torchvision torchaudio --index-url https://rocm.nightlies.amd.com/v2/gfx1151/```
+**Note:** This table only lists examples. A GPU missing from it may still be supported: supported architectures include RDNA 2, RDNA 3, RDNA 3.5, and RDNA 4. Keep `device-all` to install kernels for all supported targets. For other models, see AMD's [GPU target table](https://github.com/ROCm/TheRock/blob/main/RELEASES.md#gfx-target-lookup-table) and [ROCm compatibility matrix](https://rocm.docs.amd.com/en/docs-10.0.0/compatibility/compatibility-matrix.html).
 
-RDNA 4 (RX 9000 series):
+**ROCm 10.0.0 with PyTorch 2.13:**
 
-```pip install --pre torch torchvision torchaudio --index-url https://rocm.nightlies.amd.com/v2/gfx120X-all/```
+```bat
+pip install --index-url https://stable.repo.amd.com/rocm/whl-next/ "torch[device-all]==2.13.0+rocm10.0.0" "torchvision[device-all]==0.28.0+rocm10.0.0" "torchaudio==2.11.0.2+rocm10.0.0"
+```
 
 ### Intel GPUs (Windows and Linux)
 

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")

alembic_db/versions/0002_merge_to_asset_references.py

@@ -177,7 +177,10 @@ def downgrade() -> None:
 
     NOTE: Data is not recoverable. The upgrade discards all rows from the old
     tables and truncates assets. After downgrade the old schema will be empty.
-    A filesystem rescan will repopulate data once the older code is running.
+    A filesystem rescan can repopulate only data derived from paths. It will
+    not restore user_metadata, manually-applied tags, preview_id nominations,
+    name renames, or job_id, because build_asset_specs derives names and tags
+    from the path alone.
     """
     # Drop new tables (order matters due to FK constraints)
     op.drop_index("ix_asset_reference_meta_key_val_bool", table_name="asset_reference_meta")

alembic_db/versions/0007_record_content_split.py

@@ -0,0 +1,125 @@
+"""
+Record/content split.
+
+This migration intentionally discards the existing asset database. The
+asset_reference_meta, asset_reference_tags, asset_references, and assets
+tables are dropped, and DELETE FROM tags removes all existing tag rows. No
+data migration is performed.
+
+A filesystem rescan after this migration will not restore user_metadata,
+manually-applied tags, preview_id nominations, name renames, or job_id.
+build_asset_specs derives names and tags from the path alone.
+
+Revision ID: 0007_record_content_split
+Revises: 0006_add_loader_path
+Create Date: 2026-08-26
+"""
+
+from alembic import op
+import sqlalchemy as sa
+
+
+revision = "0007_record_content_split"
+down_revision = "0006_add_loader_path"
+branch_labels = None
+depends_on = None
+
+
+def upgrade() -> None:
+    op.drop_table("asset_reference_meta")
+    op.drop_table("asset_reference_tags")
+    op.drop_table("asset_references")
+    op.drop_table("assets")
+    op.execute("DELETE FROM tags")
+    op.create_table(
+        "asset_contents",
+        sa.Column("id", sa.String(36), primary_key=True),
+        sa.Column("hash", sa.String(256)),
+        sa.Column("size_bytes", sa.BigInteger(), nullable=False),
+        sa.Column("path", sa.Text(), nullable=False),
+        sa.Column("mtime_ns", sa.BigInteger()),
+        sa.Column("is_missing", sa.Boolean(), nullable=False, server_default="0"),
+        sa.Column("created_at", sa.DateTime(), nullable=False),
+        sa.CheckConstraint("size_bytes >= 0", name="ck_asset_contents_size_nonneg"),
+        sa.CheckConstraint("mtime_ns >= 0", name="ck_asset_contents_mtime_nonneg"),
+    )
+    op.create_index("ix_asset_contents_hash", "asset_contents", ["hash"])
+    op.create_index(
+        "uq_asset_contents_path_live", "asset_contents", ["path"], unique=True,
+        sqlite_where=sa.text("is_missing = 0"),
+    )
+    op.create_table(
+        "assets",
+        sa.Column("id", sa.String(36), primary_key=True),
+        sa.Column("content_id", sa.String(36), sa.ForeignKey("asset_contents.id", ondelete="RESTRICT"), nullable=False),
+        sa.Column("name", sa.String(512), nullable=False),
+        sa.Column("mime_type", sa.String(255)),
+        sa.Column("system_metadata", sa.JSON()),
+        sa.Column("job_id", sa.String(36)),
+        sa.Column("user_metadata", sa.JSON()),
+        sa.Column("loader_path", sa.Text()),
+        sa.Column("preview_id", sa.String(36), sa.ForeignKey("assets.id", ondelete="SET NULL")),
+        sa.Column("created_at", sa.DateTime(), nullable=False),
+        sa.Column("updated_at", sa.DateTime(), nullable=False),
+        sa.Column("last_access_time", sa.DateTime()),
+    )
+    op.create_index("ix_assets_content_id", "assets", ["content_id"])
+    op.create_index("ix_assets_name", "assets", ["name"])
+    op.create_index("ix_assets_created_at", "assets", ["created_at"])
+    op.create_index("ix_assets_preview_id", "assets", ["preview_id"])
+    op.create_table(
+        "asset_meta",
+        sa.Column("asset_id", sa.String(36), sa.ForeignKey("assets.id", ondelete="CASCADE"), primary_key=True),
+        sa.Column("key", sa.String(256), primary_key=True),
+        sa.Column("ordinal", sa.Integer(), primary_key=True),
+        sa.Column("val_str", sa.String(2048)), sa.Column("val_num", sa.Numeric(38, 10)),
+        sa.Column("val_bool", sa.Boolean()), sa.Column("val_json", sa.JSON()),
+        sa.CheckConstraint("val_str IS NOT NULL OR val_num IS NOT NULL OR val_bool IS NOT NULL OR val_json IS NOT NULL", name="ck_asset_meta_has_value"),
+    )
+    op.create_index("ix_asset_meta_key", "asset_meta", ["key"])
+    op.create_index("ix_asset_meta_key_val_str", "asset_meta", ["key", "val_str"])
+    op.create_index("ix_asset_meta_key_val_num", "asset_meta", ["key", "val_num"])
+    op.create_index("ix_asset_meta_key_val_bool", "asset_meta", ["key", "val_bool"])
+    op.create_table("asset_tags", sa.Column("asset_id", sa.String(36), sa.ForeignKey("assets.id", ondelete="CASCADE"), primary_key=True), sa.Column("tag_name", sa.String(512), sa.ForeignKey("tags.name", ondelete="RESTRICT"), primary_key=True), sa.Column("origin", sa.String(32), nullable=False), sa.Column("added_at", sa.DateTime(), nullable=False))
+    op.create_index("ix_asset_tags_tag_name", "asset_tags", ["tag_name"])
+    op.create_index("ix_asset_tags_asset_id", "asset_tags", ["asset_id"])
+    op.create_table("asset_system_state", sa.Column("key", sa.String(256), primary_key=True), sa.Column("value", sa.Text(), nullable=False))
+
+
+def downgrade() -> None:
+    op.drop_table("asset_system_state")
+    op.drop_table("asset_tags")
+    op.drop_table("asset_meta")
+    op.drop_table("assets")
+    op.drop_table("asset_contents")
+    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"])
+    op.create_table("asset_references", sa.Column("id", sa.String(36), primary_key=True), sa.Column("asset_id", sa.String(36), sa.ForeignKey("assets.id", ondelete="CASCADE"), nullable=False), sa.Column("file_path", sa.Text()), sa.Column("loader_path", sa.Text()), sa.Column("mtime_ns", sa.BigInteger()), sa.Column("needs_verify", sa.Boolean(), nullable=False), sa.Column("is_missing", sa.Boolean(), nullable=False), sa.Column("enrichment_level", sa.Integer(), nullable=False), sa.Column("owner_id", sa.String(128), nullable=False), sa.Column("name", sa.String(512), nullable=False), sa.Column("preview_id", sa.String(36), sa.ForeignKey("asset_references.id", ondelete="SET NULL")), sa.Column("user_metadata", sa.JSON()), sa.Column("system_metadata", sa.JSON()), sa.Column("job_id", sa.String(36)), sa.Column("created_at", sa.DateTime(), nullable=False), sa.Column("updated_at", sa.DateTime(), nullable=False), sa.Column("last_access_time", sa.DateTime(), nullable=False), sa.Column("deleted_at", sa.DateTime()))
+    op.create_index("uq_asset_references_file_path", "asset_references", ["file_path"], unique=True)
+    op.create_index("ix_asset_references_asset_id", "asset_references", ["asset_id"])
+    op.create_index("ix_asset_references_owner_id", "asset_references", ["owner_id"])
+    op.create_index("ix_asset_references_name", "asset_references", ["name"])
+    op.create_index("ix_asset_references_is_missing", "asset_references", ["is_missing"])
+    op.create_index("ix_asset_references_enrichment_level", "asset_references", ["enrichment_level"])
+    op.create_index("ix_asset_references_created_at", "asset_references", ["created_at"])
+    op.create_index("ix_asset_references_last_access_time", "asset_references", ["last_access_time"])
+    op.create_index("ix_asset_references_owner_name", "asset_references", ["owner_id", "name"])
+    op.create_index("ix_asset_references_deleted_at", "asset_references", ["deleted_at"])
+    op.create_index("ix_asset_references_preview_id", "asset_references", ["preview_id"])
+    op.create_table("asset_reference_meta", sa.Column("asset_reference_id", sa.String(36), sa.ForeignKey("asset_references.id", ondelete="CASCADE"), primary_key=True), sa.Column("key", sa.String(256), primary_key=True), sa.Column("ordinal", sa.Integer(), primary_key=True), sa.Column("val_str", sa.String(2048)), sa.Column("val_num", sa.Numeric(38, 10)), sa.Column("val_bool", sa.Boolean()), sa.Column("val_json", sa.JSON()), sa.CheckConstraint("val_str IS NOT NULL OR val_num IS NOT NULL OR val_bool IS NOT NULL OR val_json IS NOT NULL", name="ck_asset_reference_meta_has_value"))
+    op.create_index("ix_asset_reference_meta_key", "asset_reference_meta", ["key"])
+    op.create_index("ix_asset_reference_meta_key_val_str", "asset_reference_meta", ["key", "val_str"])
+    op.create_index("ix_asset_reference_meta_key_val_num", "asset_reference_meta", ["key", "val_num"])
+    op.create_index("ix_asset_reference_meta_key_val_bool", "asset_reference_meta", ["key", "val_bool"])
+    op.create_table("asset_reference_tags", sa.Column("asset_reference_id", sa.String(36), sa.ForeignKey("asset_references.id", ondelete="CASCADE"), primary_key=True), sa.Column("tag_name", sa.String(512), sa.ForeignKey("tags.name", ondelete="RESTRICT"), primary_key=True), sa.Column("origin", sa.String(32), nullable=False), sa.Column("added_at", sa.DateTime(), nullable=False))
+    op.create_index("ix_asset_reference_tags_tag_name", "asset_reference_tags", ["tag_name"])
+    op.create_index("ix_asset_reference_tags_asset_reference_id", "asset_reference_tags", ["asset_reference_id"])