ComfyUI v0.3.77

* add check for the format var type in VideoFromComponents.save_to function

* convert "format" to VideoContainer enum

.ci/update_windows/update.py

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

.ci/windows_nvidia_base_files/advanced/run_nvidia_gpu_disable_api_nodes.bat

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

.ci/windows_nvidia_base_files/run_nvidia_gpu.bat

@@ -1,3 +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. 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
+echo If you see this and ComfyUI did not start try updating your Nvidia Drivers to the latest.
 pause

.ci/windows_nvidia_base_files/run_nvidia_gpu_fast_fp16_accumulation.bat

@@ -1,3 +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. 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
+echo If you see this and ComfyUI did not start try updating your Nvidia Drivers to the latest.
 pause

.github/workflows/stable-release.yml

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

.github/workflows/test-build.yml

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

.github/workflows/test-ci.yml

@@ -5,7 +5,6 @@ on:
   push:
     branches:
       - master
-      - release/**
     paths-ignore:
       - 'app/**'
       - 'input/**'

.github/workflows/test-execution.yml

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

.github/workflows/test-launch.yml

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

.github/workflows/test-unit.yml

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

.github/workflows/update-version.yml

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

CODEOWNERS

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

README.md

@@ -81,7 +81,6 @@ See what ComfyUI can do with the [example workflows](https://comfyanonymous.gith
    - [Hunyuan Video](https://comfyanonymous.github.io/ComfyUI_examples/hunyuan_video/)
    - [Wan 2.1](https://comfyanonymous.github.io/ComfyUI_examples/wan/)
    - [Wan 2.2](https://comfyanonymous.github.io/ComfyUI_examples/wan22/)
-   - [Hunyuan Video 1.5](https://docs.comfy.org/tutorials/video/hunyuan/hunyuan-video-1-5)
 - Audio Models
    - [Stable Audio](https://comfyanonymous.github.io/ComfyUI_examples/audio/)
    - [ACE Step](https://comfyanonymous.github.io/ComfyUI_examples/audio/)
@@ -119,9 +118,6 @@ ComfyUI follows a weekly release cycle targeting Monday but this regularly chang
 
 1. **[ComfyUI Core](https://github.com/comfyanonymous/ComfyUI)**
    - Releases a new stable version (e.g., v0.7.0) roughly every week.
-   - Starting from v0.4.0 patch versions will be used for fixes backported onto the current stable release.
-   - Minor versions will be used for releases off the master branch.
-   - Patch versions may still be used for releases on the master branch in cases where a backport would not make sense.
    - Commits outside of the stable release tags may be very unstable and break many custom nodes.
    - Serves as the foundation for the desktop release
 
@@ -183,7 +179,7 @@ Simply download, extract with [7-Zip](https://7-zip.org) or with the windows exp
 
 If you have trouble extracting it, right click the file -> properties -> unblock
 
-The portable above currently comes with python 3.13 and pytorch cuda 13.0. Update your Nvidia drivers if it doesn't start.
+Update your Nvidia drivers if it doesn't start.
 
 #### Alternative Downloads:
 
@@ -212,8 +208,6 @@ Python 3.14 works but you may encounter issues with the torch compile node. The
 
 Python 3.13 is very well supported. If you have trouble with some custom node dependencies on 3.13 you can try 3.12
 
-torch 2.4 and above is supported but some features might only work on newer versions. We generally recommend using the latest major version of pytorch with the latest cuda version unless it is less than 2 weeks old.
-
 ### Instructions:
 
 Git clone this repo.
@@ -325,32 +319,6 @@ For models compatible with Iluvatar Extension for PyTorch. Here's a step-by-step
 1. Install the Iluvatar Corex Toolkit by adhering to the platform-specific instructions on the [Installation](https://support.iluvatar.com/#/DocumentCentre?id=1&nameCenter=2&productId=520117912052801536)
 2. Launch ComfyUI by running `python main.py`
 
-
-## [ComfyUI-Manager](https://github.com/Comfy-Org/ComfyUI-Manager/tree/manager-v4)
-
-**ComfyUI-Manager** is an extension that allows you to easily install, update, and manage custom nodes for ComfyUI.
-
-### Setup
-
-1. Install the manager dependencies:
-   ```bash
-   pip install -r manager_requirements.txt
-   ```
-
-2. Enable the manager with the `--enable-manager` flag when running ComfyUI:
-   ```bash
-   python main.py --enable-manager
-   ```
-
-### Command Line Options
-
-| Flag | Description |
-|------|-------------|
-| `--enable-manager` | Enable ComfyUI-Manager |
-| `--enable-manager-legacy-ui` | Use the legacy manager UI instead of the new UI (requires `--enable-manager`) |
-| `--disable-manager-ui` | Disable the manager UI and endpoints while keeping background features like security checks and scheduled installation completion (requires `--enable-manager`) |
-
-
 # Running
 
 ```python main.py```

alembic_db/versions/0001_assets.py

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

api_server/routes/internal/internal_routes.py

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

app/assets/api/routes.py

@@ -1,102 +0,0 @@
-import logging
-import uuid
-from aiohttp import web
-
-from pydantic import ValidationError
-
-import app.assets.manager as manager
-from app import user_manager
-from app.assets.api import schemas_in
-from app.assets.helpers import get_query_dict
-
-ROUTES = web.RouteTableDef()
-USER_MANAGER: user_manager.UserManager | None = None
-
-# UUID regex (canonical hyphenated form, case-insensitive)
-UUID_RE = r"[0-9a-fA-F]{8}-[0-9a-fA-F]{4}-[0-9a-fA-F]{4}-[0-9a-fA-F]{4}-[0-9a-fA-F]{12}"
-
-def register_assets_system(app: web.Application, user_manager_instance: user_manager.UserManager) -> None:
-    global USER_MANAGER
-    USER_MANAGER = user_manager_instance
-    app.add_routes(ROUTES)
-
-def _error_response(status: int, code: str, message: str, details: dict | None = None) -> web.Response:
-    return web.json_response({"error": {"code": code, "message": message, "details": details or {}}}, status=status)
-
-
-def _validation_error_response(code: str, ve: ValidationError) -> web.Response:
-    return _error_response(400, code, "Validation failed.", {"errors": ve.json()})
-
-
-@ROUTES.get("/api/assets")
-async def list_assets(request: web.Request) -> web.Response:
-    """
-    GET request to list assets.
-    """
-    query_dict = get_query_dict(request)
-    try:
-        q = schemas_in.ListAssetsQuery.model_validate(query_dict)
-    except ValidationError as ve:
-        return _validation_error_response("INVALID_QUERY", ve)
-
-    payload = manager.list_assets(
-        include_tags=q.include_tags,
-        exclude_tags=q.exclude_tags,
-        name_contains=q.name_contains,
-        metadata_filter=q.metadata_filter,
-        limit=q.limit,
-        offset=q.offset,
-        sort=q.sort,
-        order=q.order,
-        owner_id=USER_MANAGER.get_request_user_id(request),
-    )
-    return web.json_response(payload.model_dump(mode="json"))
-
-
-@ROUTES.get(f"/api/assets/{{id:{UUID_RE}}}")
-async def get_asset(request: web.Request) -> web.Response:
-    """
-    GET request to get an asset's info as JSON.
-    """
-    asset_info_id = str(uuid.UUID(request.match_info["id"]))
-    try:
-        result = manager.get_asset(
-            asset_info_id=asset_info_id,
-            owner_id=USER_MANAGER.get_request_user_id(request),
-        )
-    except ValueError as e:
-        return _error_response(404, "ASSET_NOT_FOUND", str(e), {"id": asset_info_id})
-    except Exception:
-        logging.exception(
-            "get_asset failed for asset_info_id=%s, owner_id=%s",
-            asset_info_id,
-            USER_MANAGER.get_request_user_id(request),
-        )
-        return _error_response(500, "INTERNAL", "Unexpected server error.")
-    return web.json_response(result.model_dump(mode="json"), status=200)
-
-
-@ROUTES.get("/api/tags")
-async def get_tags(request: web.Request) -> web.Response:
-    """
-    GET request to list all tags based on query parameters.
-    """
-    query_map = dict(request.rel_url.query)
-
-    try:
-        query = schemas_in.TagsListQuery.model_validate(query_map)
-    except ValidationError as e:
-        return web.json_response(
-            {"error": {"code": "INVALID_QUERY", "message": "Invalid query parameters", "details": e.errors()}},
-            status=400,
-        )
-
-    result = manager.list_tags(
-        prefix=query.prefix,
-        limit=query.limit,
-        offset=query.offset,
-        order=query.order,
-        include_zero=query.include_zero,
-        owner_id=USER_MANAGER.get_request_user_id(request),
-    )
-    return web.json_response(result.model_dump(mode="json"))

app/assets/api/schemas_in.py

@@ -1,94 +0,0 @@
-import json
-import uuid
-from typing import Any, Literal
-
-from pydantic import (
-    BaseModel,
-    ConfigDict,
-    Field,
-    conint,
-    field_validator,
-)
-
-
-class ListAssetsQuery(BaseModel):
-    include_tags: list[str] = Field(default_factory=list)
-    exclude_tags: list[str] = Field(default_factory=list)
-    name_contains: str | None = None
-
-    # Accept either a JSON string (query param) or a dict
-    metadata_filter: dict[str, Any] | None = None
-
-    limit: conint(ge=1, le=500) = 20
-    offset: conint(ge=0) = 0
-
-    sort: Literal["name", "created_at", "updated_at", "size", "last_access_time"] = "created_at"
-    order: Literal["asc", "desc"] = "desc"
-
-    @field_validator("include_tags", "exclude_tags", mode="before")
-    @classmethod
-    def _split_csv_tags(cls, v):
-        # Accept "a,b,c" or ["a","b"] (we are liberal in what we accept)
-        if v is None:
-            return []
-        if isinstance(v, str):
-            return [t.strip() for t in v.split(",") if t.strip()]
-        if isinstance(v, list):
-            out: list[str] = []
-            for item in v:
-                if isinstance(item, str):
-                    out.extend([t.strip() for t in item.split(",") if t.strip()])
-            return out
-        return v
-
-    @field_validator("metadata_filter", mode="before")
-    @classmethod
-    def _parse_metadata_json(cls, v):
-        if v is None or isinstance(v, dict):
-            return v
-        if isinstance(v, str) and v.strip():
-            try:
-                parsed = json.loads(v)
-            except Exception as e:
-                raise ValueError(f"metadata_filter must be JSON: {e}") from e
-            if not isinstance(parsed, dict):
-                raise ValueError("metadata_filter must be a JSON object")
-            return parsed
-        return None
-
-
-class TagsListQuery(BaseModel):
-    model_config = ConfigDict(extra="ignore", str_strip_whitespace=True)
-
-    prefix: str | None = Field(None, min_length=1, max_length=256)
-    limit: int = Field(100, ge=1, le=1000)
-    offset: int = Field(0, ge=0, le=10_000_000)
-    order: Literal["count_desc", "name_asc"] = "count_desc"
-    include_zero: bool = True
-
-    @field_validator("prefix")
-    @classmethod
-    def normalize_prefix(cls, v: str | None) -> str | None:
-        if v is None:
-            return v
-        v = v.strip()
-        return v.lower() or None
-
-
-class SetPreviewBody(BaseModel):
-    """Set or clear the preview for an AssetInfo. Provide an Asset.id or null."""
-    preview_id: str | None = 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

@@ -1,60 +0,0 @@
-from datetime import datetime
-from typing import Any
-
-from pydantic import BaseModel, ConfigDict, Field, field_serializer
-
-
-class AssetSummary(BaseModel):
-    id: str
-    name: str
-    asset_hash: str | None = None
-    size: int | None = None
-    mime_type: str | None = None
-    tags: list[str] = Field(default_factory=list)
-    preview_url: str | None = None
-    created_at: datetime | None = None
-    updated_at: datetime | None = None
-    last_access_time: datetime | None = None
-
-    model_config = ConfigDict(from_attributes=True)
-
-    @field_serializer("created_at", "updated_at", "last_access_time")
-    def _ser_dt(self, v: datetime | None, _info):
-        return v.isoformat() if v else None
-
-
-class AssetsList(BaseModel):
-    assets: list[AssetSummary]
-    total: int
-    has_more: bool
-
-
-class AssetDetail(BaseModel):
-    id: str
-    name: str
-    asset_hash: str | None = None
-    size: int | None = None
-    mime_type: str | None = None
-    tags: list[str] = Field(default_factory=list)
-    user_metadata: dict[str, Any] = Field(default_factory=dict)
-    preview_id: str | None = None
-    created_at: datetime | None = None
-    last_access_time: datetime | None = None
-
-    model_config = ConfigDict(from_attributes=True)
-
-    @field_serializer("created_at", "last_access_time")
-    def _ser_dt(self, v: datetime | None, _info):
-        return v.isoformat() if v else None
-
-
-class TagUsage(BaseModel):
-    name: str
-    count: int
-    type: str
-
-
-class TagsList(BaseModel):
-    tags: list[TagUsage] = Field(default_factory=list)
-    total: int
-    has_more: bool

app/assets/database/bulk_ops.py

@@ -1,188 +0,0 @@
-import os
-import uuid
-import sqlalchemy
-from typing import Iterable
-from sqlalchemy.orm import Session
-from sqlalchemy.dialects import sqlite
-
-from app.assets.helpers import utcnow
-from app.assets.database.models import Asset, AssetCacheState, AssetInfo, AssetInfoTag, AssetInfoMeta
-
-MAX_BIND_PARAMS = 800
-
-def _chunk_rows(rows: list[dict], cols_per_row: int, max_bind_params: int) -> Iterable[list[dict]]:
-    if not rows:
-        return []
-    rows_per_stmt = max(1, max_bind_params // max(1, cols_per_row))
-    for i in range(0, len(rows), rows_per_stmt):
-        yield rows[i:i + rows_per_stmt]
-
-def _iter_chunks(seq, n: int):
-    for i in range(0, len(seq), n):
-        yield seq[i:i + n]
-
-def _rows_per_stmt(cols: int) -> int:
-    return max(1, MAX_BIND_PARAMS // max(1, cols))
-
-
-def seed_from_paths_batch(
-    session: Session,
-    *,
-    specs: list[dict],
-    owner_id: str = "",
-) -> dict:
-    """Each spec is a dict with keys:
-      - abs_path: str
-      - size_bytes: int
-      - mtime_ns: int
-      - info_name: str
-      - tags: list[str]
-      - fname: Optional[str]
-    """
-    if not specs:
-        return {"inserted_infos": 0, "won_states": 0, "lost_states": 0}
-
-    now = utcnow()
-    asset_rows: list[dict] = []
-    state_rows: list[dict] = []
-    path_to_asset: dict[str, str] = {}
-    asset_to_info: dict[str, dict] = {}  # asset_id -> prepared info row
-    path_list: list[str] = []
-
-    for sp in specs:
-        ap = os.path.abspath(sp["abs_path"])
-        aid = str(uuid.uuid4())
-        iid = str(uuid.uuid4())
-        path_list.append(ap)
-        path_to_asset[ap] = aid
-
-        asset_rows.append(
-            {
-                "id": aid,
-                "hash": None,
-                "size_bytes": sp["size_bytes"],
-                "mime_type": None,
-                "created_at": now,
-            }
-        )
-        state_rows.append(
-            {
-                "asset_id": aid,
-                "file_path": ap,
-                "mtime_ns": sp["mtime_ns"],
-            }
-        )
-        asset_to_info[aid] = {
-            "id": iid,
-            "owner_id": owner_id,
-            "name": sp["info_name"],
-            "asset_id": aid,
-            "preview_id": None,
-            "user_metadata": {"filename": sp["fname"]} if sp["fname"] else None,
-            "created_at": now,
-            "updated_at": now,
-            "last_access_time": now,
-            "_tags": sp["tags"],
-            "_filename": sp["fname"],
-        }
-
-    # insert all seed Assets (hash=NULL)
-    ins_asset = sqlite.insert(Asset)
-    for chunk in _iter_chunks(asset_rows, _rows_per_stmt(5)):
-        session.execute(ins_asset, chunk)
-
-    # try to claim AssetCacheState (file_path)
-    winners_by_path: set[str] = set()
-    ins_state = (
-        sqlite.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((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):
-            session.execute(sqlalchemy.delete(Asset).where(Asset.id.in_(id_chunk)))
-
-    if not winners_by_path:
-        return {"inserted_infos": 0, "won_states": 0, "lost_states": len(losers_by_path)}
-
-    # insert AssetInfo only for winners
-    winner_info_rows = [asset_to_info[path_to_asset[p]] for p in winners_by_path]
-    ins_info = (
-        sqlite.insert(AssetInfo)
-        .on_conflict_do_nothing(index_elements=[AssetInfo.asset_id, AssetInfo.owner_id, AssetInfo.name])
-        .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((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,
-                    }
-                )
-
-    bulk_insert_tags_and_meta(session, tag_rows=tag_rows, meta_rows=meta_rows, max_bind_params=MAX_BIND_PARAMS)
-    return {
-        "inserted_infos": len(inserted_info_ids),
-        "won_states": len(winners_by_path),
-        "lost_states": len(losers_by_path),
-    }
-
-
-def bulk_insert_tags_and_meta(
-    session: Session,
-    *,
-    tag_rows: list[dict],
-    meta_rows: list[dict],
-    max_bind_params: int,
-) -> None:
-    """Batch insert into asset_info_tags and asset_info_meta with ON CONFLICT DO NOTHING.
-    - tag_rows keys: asset_info_id, tag_name, origin, added_at
-    - meta_rows keys: asset_info_id, key, ordinal, val_str, val_num, val_bool, val_json
-    """
-    if tag_rows:
-        ins_links = (
-            sqlite.insert(AssetInfoTag)
-            .on_conflict_do_nothing(index_elements=[AssetInfoTag.asset_info_id, AssetInfoTag.tag_name])
-        )
-        for chunk in _chunk_rows(tag_rows, cols_per_row=4, max_bind_params=max_bind_params):
-            session.execute(ins_links, chunk)
-    if meta_rows:
-        ins_meta = (
-            sqlite.insert(AssetInfoMeta)
-            .on_conflict_do_nothing(
-                index_elements=[AssetInfoMeta.asset_info_id, AssetInfoMeta.key, AssetInfoMeta.ordinal]
-            )
-        )
-        for chunk in _chunk_rows(meta_rows, cols_per_row=7, max_bind_params=max_bind_params):
-            session.execute(ins_meta, chunk)

app/assets/database/models.py

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

app/assets/database/queries.py

@@ -1,267 +0,0 @@
-import sqlalchemy as sa
-from collections import defaultdict
-from sqlalchemy import select, exists, func
-from sqlalchemy.orm import Session, contains_eager, noload
-from app.assets.database.models import Asset, AssetInfo, AssetInfoMeta, AssetInfoTag, Tag
-from app.assets.helpers import escape_like_prefix, normalize_tags
-from typing import Sequence
-
-
-def visible_owner_clause(owner_id: str) -> sa.sql.ClauseElement:
-    """Build owner visibility predicate for reads. Owner-less rows are visible to everyone."""
-    owner_id = (owner_id or "").strip()
-    if owner_id == "":
-        return AssetInfo.owner_id == ""
-    return AssetInfo.owner_id.in_(["", owner_id])
-
-
-def apply_tag_filters(
-    stmt: sa.sql.Select,
-    include_tags: Sequence[str] | None = None,
-    exclude_tags: Sequence[str] | None = None,
-) -> sa.sql.Select:
-    """include_tags: every tag must be present; exclude_tags: none may be present."""
-    include_tags = normalize_tags(include_tags)
-    exclude_tags = normalize_tags(exclude_tags)
-
-    if include_tags:
-        for tag_name in include_tags:
-            stmt = stmt.where(
-                exists().where(
-                    (AssetInfoTag.asset_info_id == AssetInfo.id)
-                    & (AssetInfoTag.tag_name == tag_name)
-                )
-            )
-
-    if exclude_tags:
-        stmt = stmt.where(
-            ~exists().where(
-                (AssetInfoTag.asset_info_id == AssetInfo.id)
-                & (AssetInfoTag.tag_name.in_(exclude_tags))
-            )
-        )
-    return stmt
-
-def apply_metadata_filter(
-    stmt: sa.sql.Select,
-    metadata_filter: dict | None = None,
-) -> sa.sql.Select:
-    """Apply filters using asset_info_meta projection table."""
-    if not metadata_filter:
-        return stmt
-
-    def _exists_for_pred(key: str, *preds) -> sa.sql.ClauseElement:
-        return sa.exists().where(
-            AssetInfoMeta.asset_info_id == AssetInfo.id,
-            AssetInfoMeta.key == key,
-            *preds,
-        )
-
-    def _exists_clause_for_value(key: str, value) -> sa.sql.ClauseElement:
-        if value is None:
-            no_row_for_key = sa.not_(
-                sa.exists().where(
-                    AssetInfoMeta.asset_info_id == AssetInfo.id,
-                    AssetInfoMeta.key == key,
-                )
-            )
-            null_row = _exists_for_pred(
-                key,
-                AssetInfoMeta.val_json.is_(None),
-                AssetInfoMeta.val_str.is_(None),
-                AssetInfoMeta.val_num.is_(None),
-                AssetInfoMeta.val_bool.is_(None),
-            )
-            return sa.or_(no_row_for_key, null_row)
-
-        if isinstance(value, bool):
-            return _exists_for_pred(key, AssetInfoMeta.val_bool == bool(value))
-        if isinstance(value, (int, float)):
-            from decimal import Decimal
-            num = value if isinstance(value, Decimal) else Decimal(str(value))
-            return _exists_for_pred(key, AssetInfoMeta.val_num == num)
-        if isinstance(value, str):
-            return _exists_for_pred(key, AssetInfoMeta.val_str == value)
-        return _exists_for_pred(key, AssetInfoMeta.val_json == value)
-
-    for k, v in metadata_filter.items():
-        if isinstance(v, list):
-            ors = [_exists_clause_for_value(k, elem) for elem in v]
-            if ors:
-                stmt = stmt.where(sa.or_(*ors))
-        else:
-            stmt = stmt.where(_exists_clause_for_value(k, v))
-    return stmt
-
-
-def asset_exists_by_hash(session: Session, asset_hash: str) -> bool:
-    """
-    Check if an asset with a given hash exists in database.
-    """
-    row = (
-        session.execute(
-            select(sa.literal(True)).select_from(Asset).where(Asset.hash == asset_hash).limit(1)
-        )
-    ).first()
-    return row is not None
-
-def get_asset_info_by_id(session: Session, asset_info_id: str) -> AssetInfo | None:
-    return session.get(AssetInfo, asset_info_id)
-
-def list_asset_infos_page(
-    session: Session,
-    owner_id: str = "",
-    include_tags: Sequence[str] | None = None,
-    exclude_tags: Sequence[str] | None = None,
-    name_contains: str | None = None,
-    metadata_filter: dict | None = None,
-    limit: int = 20,
-    offset: int = 0,
-    sort: str = "created_at",
-    order: str = "desc",
-) -> tuple[list[AssetInfo], dict[str, list[str]], int]:
-    base = (
-        select(AssetInfo)
-        .join(Asset, Asset.id == AssetInfo.asset_id)
-        .options(contains_eager(AssetInfo.asset), noload(AssetInfo.tags))
-        .where(visible_owner_clause(owner_id))
-    )
-
-    if name_contains:
-        escaped, esc = escape_like_prefix(name_contains)
-        base = base.where(AssetInfo.name.ilike(f"%{escaped}%", escape=esc))
-
-    base = apply_tag_filters(base, include_tags, exclude_tags)
-    base = apply_metadata_filter(base, metadata_filter)
-
-    sort = (sort or "created_at").lower()
-    order = (order or "desc").lower()
-    sort_map = {
-        "name": AssetInfo.name,
-        "created_at": AssetInfo.created_at,
-        "updated_at": AssetInfo.updated_at,
-        "last_access_time": AssetInfo.last_access_time,
-        "size": Asset.size_bytes,
-    }
-    sort_col = sort_map.get(sort, AssetInfo.created_at)
-    sort_exp = sort_col.desc() if order == "desc" else sort_col.asc()
-
-    base = base.order_by(sort_exp).limit(limit).offset(offset)
-
-    count_stmt = (
-        select(sa.func.count())
-        .select_from(AssetInfo)
-        .join(Asset, Asset.id == AssetInfo.asset_id)
-        .where(visible_owner_clause(owner_id))
-    )
-    if name_contains:
-        escaped, esc = escape_like_prefix(name_contains)
-        count_stmt = count_stmt.where(AssetInfo.name.ilike(f"%{escaped}%", escape=esc))
-    count_stmt = apply_tag_filters(count_stmt, include_tags, exclude_tags)
-    count_stmt = apply_metadata_filter(count_stmt, metadata_filter)
-
-    total = int((session.execute(count_stmt)).scalar_one() or 0)
-
-    infos = (session.execute(base)).unique().scalars().all()
-
-    id_list: list[str] = [i.id for i in infos]
-    tag_map: dict[str, list[str]] = defaultdict(list)
-    if id_list:
-        rows = session.execute(
-            select(AssetInfoTag.asset_info_id, Tag.name)
-            .join(Tag, Tag.name == AssetInfoTag.tag_name)
-            .where(AssetInfoTag.asset_info_id.in_(id_list))
-        )
-        for aid, tag_name in rows.all():
-            tag_map[aid].append(tag_name)
-
-    return infos, tag_map, total
-
-def fetch_asset_info_asset_and_tags(
-    session: Session,
-    asset_info_id: str,
-    owner_id: str = "",
-) -> tuple[AssetInfo, Asset, list[str]] | None:
-    stmt = (
-        select(AssetInfo, Asset, Tag.name)
-        .join(Asset, Asset.id == AssetInfo.asset_id)
-        .join(AssetInfoTag, AssetInfoTag.asset_info_id == AssetInfo.id, isouter=True)
-        .join(Tag, Tag.name == AssetInfoTag.tag_name, isouter=True)
-        .where(
-            AssetInfo.id == asset_info_id,
-            visible_owner_clause(owner_id),
-        )
-        .options(noload(AssetInfo.tags))
-        .order_by(Tag.name.asc())
-    )
-
-    rows = (session.execute(stmt)).all()
-    if not rows:
-        return None
-
-    first_info, first_asset, _ = rows[0]
-    tags: list[str] = []
-    seen: set[str] = set()
-    for _info, _asset, tag_name in rows:
-        if tag_name and tag_name not in seen:
-            seen.add(tag_name)
-            tags.append(tag_name)
-    return first_info, first_asset, tags
-
-def list_tags_with_usage(
-    session: Session,
-    prefix: str | None = None,
-    limit: int = 100,
-    offset: int = 0,
-    include_zero: bool = True,
-    order: str = "count_desc",
-    owner_id: str = "",
-) -> tuple[list[tuple[str, str, int]], int]:
-    counts_sq = (
-        select(
-            AssetInfoTag.tag_name.label("tag_name"),
-            func.count(AssetInfoTag.asset_info_id).label("cnt"),
-        )
-        .select_from(AssetInfoTag)
-        .join(AssetInfo, AssetInfo.id == AssetInfoTag.asset_info_id)
-        .where(visible_owner_clause(owner_id))
-        .group_by(AssetInfoTag.tag_name)
-        .subquery()
-    )
-
-    q = (
-        select(
-            Tag.name,
-            Tag.tag_type,
-            func.coalesce(counts_sq.c.cnt, 0).label("count"),
-        )
-        .select_from(Tag)
-        .join(counts_sq, counts_sq.c.tag_name == Tag.name, isouter=True)
-    )
-
-    if prefix:
-        escaped, esc = escape_like_prefix(prefix.strip().lower())
-        q = q.where(Tag.name.like(escaped + "%", escape=esc))
-
-    if not include_zero:
-        q = q.where(func.coalesce(counts_sq.c.cnt, 0) > 0)
-
-    if order == "name_asc":
-        q = q.order_by(Tag.name.asc())
-    else:
-        q = q.order_by(func.coalesce(counts_sq.c.cnt, 0).desc(), Tag.name.asc())
-
-    total_q = select(func.count()).select_from(Tag)
-    if prefix:
-        escaped, esc = escape_like_prefix(prefix.strip().lower())
-        total_q = total_q.where(Tag.name.like(escaped + "%", escape=esc))
-    if not include_zero:
-        total_q = total_q.where(
-            Tag.name.in_(select(AssetInfoTag.tag_name).group_by(AssetInfoTag.tag_name))
-        )
-
-    rows = (session.execute(q.limit(limit).offset(offset))).all()
-    total = (session.execute(total_q)).scalar_one()
-
-    rows_norm = [(name, ttype, int(count or 0)) for (name, ttype, count) in rows]
-    return rows_norm, int(total or 0)

app/assets/database/tags.py

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

app/assets/hashing.py

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

app/assets/helpers.py

@@ -1,217 +0,0 @@
-import contextlib
-import os
-from aiohttp import web
-from datetime import datetime, timezone
-from pathlib import Path
-from typing import Literal, Any
-
-import folder_paths
-
-
-RootType = Literal["models", "input", "output"]
-ALLOWED_ROOTS: tuple[RootType, ...] = ("models", "input", "output")
-
-def get_query_dict(request: web.Request) -> dict[str, Any]:
-    """
-    Gets a dictionary of query parameters from the request.
-
-    'request.query' is a MultiMapping[str], needs to be converted to a dictionary to be validated by Pydantic.
-    """
-    query_dict = {
-        key: request.query.getall(key) if len(request.query.getall(key)) > 1 else request.query.get(key)
-        for key in request.query.keys()
-    }
-    return query_dict
-
-def list_tree(base_dir: str) -> list[str]:
-    out: list[str] = []
-    base_abs = os.path.abspath(base_dir)
-    if not os.path.isdir(base_abs):
-        return out
-    for dirpath, _subdirs, filenames in os.walk(base_abs, topdown=True, followlinks=False):
-        for name in filenames:
-            out.append(os.path.abspath(os.path.join(dirpath, name)))
-    return out
-
-def prefixes_for_root(root: RootType) -> list[str]:
-    if root == "models":
-        bases: list[str] = []
-        for _bucket, paths in get_comfy_models_folders():
-            bases.extend(paths)
-        return [os.path.abspath(p) for p in bases]
-    if root == "input":
-        return [os.path.abspath(folder_paths.get_input_directory())]
-    if root == "output":
-        return [os.path.abspath(folder_paths.get_output_directory())]
-    return []
-
-def escape_like_prefix(s: str, escape: str = "!") -> tuple[str, str]:
-    """Escapes %, _ and the escape char itself in a LIKE prefix.
-    Returns (escaped_prefix, escape_char). Caller should append '%' and pass escape=escape_char to .like().
-    """
-    s = s.replace(escape, escape + escape)  # escape the escape char first
-    s = s.replace("%", escape + "%").replace("_", escape + "_")  # escape LIKE wildcards
-    return s, escape
-
-def fast_asset_file_check(
-    *,
-    mtime_db: int | None,
-    size_db: int | None,
-    stat_result: os.stat_result,
-) -> bool:
-    if mtime_db is None:
-        return False
-    actual_mtime_ns = getattr(stat_result, "st_mtime_ns", int(stat_result.st_mtime * 1_000_000_000))
-    if int(mtime_db) != int(actual_mtime_ns):
-        return False
-    sz = int(size_db or 0)
-    if sz > 0:
-        return int(stat_result.st_size) == sz
-    return True
-
-def utcnow() -> datetime:
-    """Naive UTC timestamp (no tzinfo). We always treat DB datetimes as UTC."""
-    return datetime.now(timezone.utc).replace(tzinfo=None)
-
-def get_comfy_models_folders() -> list[tuple[str, list[str]]]:
-    """Build a list of (folder_name, base_paths[]) categories that are configured for model locations.
-
-    We trust `folder_paths.folder_names_and_paths` and include a category if
-    *any* of its base paths lies under the Comfy `models_dir`.
-    """
-    targets: list[tuple[str, list[str]]] = []
-    models_root = os.path.abspath(folder_paths.models_dir)
-    for name, values in folder_paths.folder_names_and_paths.items():
-        paths, _exts = values[0], values[1]  # NOTE: this prevents nodepacks that hackily edit folder_... from breaking ComfyUI
-        if any(os.path.abspath(p).startswith(models_root + os.sep) for p in paths):
-            targets.append((name, paths))
-    return targets
-
-def compute_relative_filename(file_path: str) -> str | None:
-    """
-    Return the model's path relative to the last well-known folder (the model category),
-    using forward slashes, eg:
-      /.../models/checkpoints/flux/123/flux.safetensors -> "flux/123/flux.safetensors"
-      /.../models/text_encoders/clip_g.safetensors -> "clip_g.safetensors"
-
-    For non-model paths, returns None.
-    NOTE: this is a temporary helper, used only for initializing metadata["filename"] field.
-    """
-    try:
-        root_category, rel_path = get_relative_to_root_category_path_of_asset(file_path)
-    except ValueError:
-        return None
-
-    p = Path(rel_path)
-    parts = [seg for seg in p.parts if seg not in (".", "..", p.anchor)]
-    if not parts:
-        return None
-
-    if root_category == "models":
-        # parts[0] is the category ("checkpoints", "vae", etc) – drop it
-        inside = parts[1:] if len(parts) > 1 else [parts[0]]
-        return "/".join(inside)
-    return "/".join(parts)  # input/output: keep all parts
-
-
-def get_relative_to_root_category_path_of_asset(file_path: str) -> tuple[Literal["input", "output", "models"], str]:
-    """Given an absolute or relative file path, determine which root category the path belongs to:
-      - 'input' if the file resides under `folder_paths.get_input_directory()`
-      - 'output' if the file resides under `folder_paths.get_output_directory()`
-      - 'models' if the file resides under any base path of categories returned by `get_comfy_models_folders()`
-
-    Returns:
-        (root_category, relative_path_inside_that_root)
-        For 'models', the relative path is prefixed with the category name:
-            e.g. ('models', 'vae/test/sub/ae.safetensors')
-
-    Raises:
-        ValueError: if the path does not belong to input, output, or configured model bases.
-    """
-    fp_abs = os.path.abspath(file_path)
-
-    def _is_within(child: str, parent: str) -> bool:
-        try:
-            return os.path.commonpath([child, parent]) == parent
-        except Exception:
-            return False
-
-    def _rel(child: str, parent: str) -> str:
-        return os.path.relpath(os.path.join(os.sep, os.path.relpath(child, parent)), os.sep)
-
-    # 1) input
-    input_base = os.path.abspath(folder_paths.get_input_directory())
-    if _is_within(fp_abs, input_base):
-        return "input", _rel(fp_abs, input_base)
-
-    # 2) output
-    output_base = os.path.abspath(folder_paths.get_output_directory())
-    if _is_within(fp_abs, output_base):
-        return "output", _rel(fp_abs, output_base)
-
-    # 3) models (check deepest matching base to avoid ambiguity)
-    best: tuple[int, str, str] | None = None  # (base_len, bucket, rel_inside_bucket)
-    for bucket, bases in get_comfy_models_folders():
-        for b in bases:
-            base_abs = os.path.abspath(b)
-            if not _is_within(fp_abs, base_abs):
-                continue
-            cand = (len(base_abs), bucket, _rel(fp_abs, base_abs))
-            if best is None or cand[0] > best[0]:
-                best = cand
-
-    if best is not None:
-        _, bucket, rel_inside = best
-        combined = os.path.join(bucket, rel_inside)
-        return "models", os.path.relpath(os.path.join(os.sep, combined), os.sep)
-
-    raise ValueError(f"Path is not within input, output, or configured model bases: {file_path}")
-
-def get_name_and_tags_from_asset_path(file_path: str) -> tuple[str, list[str]]:
-    """Return a tuple (name, tags) derived from a filesystem path.
-
-    Semantics:
-      - Root category is determined by `get_relative_to_root_category_path_of_asset`.
-      - The returned `name` is the base filename with extension from the relative path.
-      - The returned `tags` are:
-            [root_category] + parent folders of the relative path (in order)
-        For 'models', this means:
-            file '/.../ModelsDir/vae/test_tag/ae.safetensors'
-            -> root_category='models', some_path='vae/test_tag/ae.safetensors'
-            -> name='ae.safetensors', tags=['models', 'vae', 'test_tag']
-
-    Raises:
-        ValueError: if the path does not belong to input, output, or configured model bases.
-    """
-    root_category, some_path = get_relative_to_root_category_path_of_asset(file_path)
-    p = Path(some_path)
-    parent_parts = [part for part in p.parent.parts if part not in (".", "..", p.anchor)]
-    return p.name, list(dict.fromkeys(normalize_tags([root_category, *parent_parts])))
-
-def normalize_tags(tags: list[str] | None) -> list[str]:
-    """
-    Normalize a list of tags by:
-      - Stripping whitespace and converting to lowercase.
-      - Removing duplicates.
-    """
-    return [t.strip().lower() for t in (tags or []) if (t or "").strip()]
-
-def collect_models_files() -> list[str]:
-    out: list[str] = []
-    for folder_name, bases in get_comfy_models_folders():
-        rel_files = folder_paths.get_filename_list(folder_name) or []
-        for rel_path in rel_files:
-            abs_path = folder_paths.get_full_path(folder_name, rel_path)
-            if not abs_path:
-                continue
-            abs_path = os.path.abspath(abs_path)
-            allowed = False
-            for b in bases:
-                base_abs = os.path.abspath(b)
-                with contextlib.suppress(Exception):
-                    if os.path.commonpath([abs_path, base_abs]) == base_abs:
-                        allowed = True
-                        break
-            if allowed:
-                out.append(abs_path)
-    return out

app/assets/manager.py

@@ -1,123 +0,0 @@
-from typing import Sequence
-
-from app.database.db import create_session
-from app.assets.api import schemas_out
-from app.assets.database.queries import (
-    asset_exists_by_hash,
-    fetch_asset_info_asset_and_tags,
-    list_asset_infos_page,
-    list_tags_with_usage,
-)
-
-
-def _safe_sort_field(requested: str | None) -> str:
-    if not requested:
-        return "created_at"
-    v = requested.lower()
-    if v in {"name", "created_at", "updated_at", "size", "last_access_time"}:
-        return v
-    return "created_at"
-
-
-def asset_exists(asset_hash: str) -> bool:
-    with create_session() as session:
-        return asset_exists_by_hash(session, asset_hash=asset_hash)
-
-def list_assets(
-    include_tags: Sequence[str] | None = None,
-    exclude_tags: Sequence[str] | None = None,
-    name_contains: str | None = None,
-    metadata_filter: dict | None = None,
-    limit: int = 20,
-    offset: int = 0,
-    sort: str = "created_at",
-    order: str = "desc",
-    owner_id: str = "",
-) -> schemas_out.AssetsList:
-    sort = _safe_sort_field(sort)
-    order = "desc" if (order or "desc").lower() not in {"asc", "desc"} else order.lower()
-
-    with create_session() as session:
-        infos, tag_map, total = list_asset_infos_page(
-            session,
-            owner_id=owner_id,
-            include_tags=include_tags,
-            exclude_tags=exclude_tags,
-            name_contains=name_contains,
-            metadata_filter=metadata_filter,
-            limit=limit,
-            offset=offset,
-            sort=sort,
-            order=order,
-        )
-
-    summaries: list[schemas_out.AssetSummary] = []
-    for info in infos:
-        asset = info.asset
-        tags = tag_map.get(info.id, [])
-        summaries.append(
-            schemas_out.AssetSummary(
-                id=info.id,
-                name=info.name,
-                asset_hash=asset.hash if asset else None,
-                size=int(asset.size_bytes) if asset else None,
-                mime_type=asset.mime_type if asset else None,
-                tags=tags,
-                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,
-    )
-
-def get_asset(asset_info_id: str, owner_id: str = "") -> schemas_out.AssetDetail:
-    with create_session() as session:
-        res = fetch_asset_info_asset_and_tags(session, asset_info_id=asset_info_id, owner_id=owner_id)
-        if not res:
-            raise ValueError(f"AssetInfo {asset_info_id} not found")
-        info, asset, tag_names = res
-        preview_id = info.preview_id
-
-    return schemas_out.AssetDetail(
-        id=info.id,
-        name=info.name,
-        asset_hash=asset.hash if asset else None,
-        size=int(asset.size_bytes) if asset and asset.size_bytes is not None else None,
-        mime_type=asset.mime_type if asset else None,
-        tags=tag_names,
-        user_metadata=info.user_metadata or {},
-        preview_id=preview_id,
-        created_at=info.created_at,
-        last_access_time=info.last_access_time,
-    )
-
-def list_tags(
-    prefix: str | None = None,
-    limit: int = 100,
-    offset: int = 0,
-    order: str = "count_desc",
-    include_zero: bool = True,
-    owner_id: str = "",
-) -> schemas_out.TagsList:
-    limit = max(1, min(1000, limit))
-    offset = max(0, offset)
-
-    with create_session() as session:
-        rows, total = list_tags_with_usage(
-            session,
-            prefix=prefix,
-            limit=limit,
-            offset=offset,
-            include_zero=include_zero,
-            order=order,
-            owner_id=owner_id,
-        )
-
-    tags = [schemas_out.TagUsage(name=name, count=count, type=tag_type) for (name, tag_type, count) in rows]
-    return schemas_out.TagsList(tags=tags, total=total, has_more=(offset + len(tags)) < total)

app/assets/scanner.py

@@ -1,229 +0,0 @@
-import contextlib
-import time
-import logging
-import os
-import sqlalchemy
-
-import folder_paths
-from app.database.db import create_session, dependencies_available
-from app.assets.helpers import (
-    collect_models_files, compute_relative_filename, fast_asset_file_check, get_name_and_tags_from_asset_path,
-    list_tree,prefixes_for_root, escape_like_prefix,
-    RootType
-)
-from app.assets.database.tags import add_missing_tag_for_asset_id, ensure_tags_exist, remove_missing_tag_for_asset_id
-from app.assets.database.bulk_ops import seed_from_paths_batch
-from app.assets.database.models import Asset, AssetCacheState, AssetInfo
-
-
-def seed_assets(roots: tuple[RootType, ...], enable_logging: bool = False) -> None:
-    """
-    Scan the given roots and seed the assets into the database.
-    """
-    if not dependencies_available():
-        if enable_logging:
-            logging.warning("Database dependencies not available, skipping assets scan")
-        return
-    t_start = time.perf_counter()
-    created = 0
-    skipped_existing = 0
-    paths: list[str] = []
-    try:
-        existing_paths: set[str] = set()
-        for r in roots:
-            try:
-                survivors: set[str] = _fast_db_consistency_pass(r, collect_existing_paths=True, update_missing_tags=True)
-                if survivors:
-                    existing_paths.update(survivors)
-            except Exception as e:
-                logging.exception("fast DB scan failed for %s: %s", r, e)
-
-        if "models" in roots:
-            paths.extend(collect_models_files())
-        if "input" in roots:
-            paths.extend(list_tree(folder_paths.get_input_directory()))
-        if "output" in roots:
-            paths.extend(list_tree(folder_paths.get_output_directory()))
-
-        specs: list[dict] = []
-        tag_pool: set[str] = set()
-        for p in paths:
-            abs_p = os.path.abspath(p)
-            if abs_p in existing_paths:
-                skipped_existing += 1
-                continue
-            try:
-                stat_p = os.stat(abs_p, follow_symlinks=False)
-            except OSError:
-                continue
-            # skip empty files
-            if not stat_p.st_size:
-                continue
-            name, tags = get_name_and_tags_from_asset_path(abs_p)
-            specs.append(
-                {
-                    "abs_path": abs_p,
-                    "size_bytes": stat_p.st_size,
-                    "mtime_ns": getattr(stat_p, "st_mtime_ns", int(stat_p.st_mtime * 1_000_000_000)),
-                    "info_name": name,
-                    "tags": tags,
-                    "fname": compute_relative_filename(abs_p),
-                }
-            )
-            for t in tags:
-                tag_pool.add(t)
-        # if no file specs, nothing to do
-        if not specs:
-            return
-        with create_session() as sess:
-            if tag_pool:
-                ensure_tags_exist(sess, tag_pool, tag_type="user")
-
-            result = seed_from_paths_batch(sess, specs=specs, owner_id="")
-            created += result["inserted_infos"]
-            sess.commit()
-    finally:
-        if enable_logging:
-            logging.info(
-                "Assets scan(roots=%s) completed in %.3fs (created=%d, skipped_existing=%d, total_seen=%d)",
-                roots,
-                time.perf_counter() - t_start,
-                created,
-                skipped_existing,
-                len(paths),
-            )
-
-
-def _fast_db_consistency_pass(
-    root: RootType,
-    *,
-    collect_existing_paths: bool = False,
-    update_missing_tags: bool = False,
-) -> set[str] | None:
-    """Fast DB+FS pass for a root:
-      - Toggle needs_verify per state using fast check
-      - For hashed assets with at least one fast-ok state in this root: delete stale missing states
-      - For seed assets with all states missing: delete Asset and its AssetInfos
-      - Optionally add/remove 'missing' tags based on fast-ok in this root
-      - Optionally return surviving absolute paths
-    """
-    prefixes = prefixes_for_root(root)
-    if not prefixes:
-        return set() if collect_existing_paths else None
-
-    conds = []
-    for p in prefixes:
-        base = os.path.abspath(p)
-        if not base.endswith(os.sep):
-            base += os.sep
-        escaped, esc = escape_like_prefix(base)
-        conds.append(AssetCacheState.file_path.like(escaped + "%", escape=esc))
-
-    with create_session() as sess:
-        rows = (
-            sess.execute(
-                sqlalchemy.select(
-                    AssetCacheState.id,
-                    AssetCacheState.file_path,
-                    AssetCacheState.mtime_ns,
-                    AssetCacheState.needs_verify,
-                    AssetCacheState.asset_id,
-                    Asset.hash,
-                    Asset.size_bytes,
-                )
-                .join(Asset, Asset.id == AssetCacheState.asset_id)
-                .where(sqlalchemy.or_(*conds))
-                .order_by(AssetCacheState.asset_id.asc(), AssetCacheState.id.asc())
-            )
-        ).all()
-
-        by_asset: dict[str, dict] = {}
-        for sid, fp, mtime_db, needs_verify, aid, a_hash, a_size in rows:
-            acc = by_asset.get(aid)
-            if acc is None:
-                acc = {"hash": a_hash, "size_db": int(a_size or 0), "states": []}
-                by_asset[aid] = acc
-
-            fast_ok = False
-            try:
-                exists = True
-                fast_ok = fast_asset_file_check(
-                    mtime_db=mtime_db,
-                    size_db=acc["size_db"],
-                    stat_result=os.stat(fp, follow_symlinks=True),
-                )
-            except FileNotFoundError:
-                exists = False
-            except OSError:
-                exists = False
-
-            acc["states"].append({
-                "sid": sid,
-                "fp": fp,
-                "exists": exists,
-                "fast_ok": fast_ok,
-                "needs_verify": bool(needs_verify),
-            })
-
-        to_set_verify: list[int] = []
-        to_clear_verify: list[int] = []
-        stale_state_ids: list[int] = []
-        survivors: set[str] = set()
-
-        for aid, acc in by_asset.items():
-            a_hash = acc["hash"]
-            states = acc["states"]
-            any_fast_ok = any(s["fast_ok"] for s in states)
-            all_missing = all(not s["exists"] for s in states)
-
-            for s in states:
-                if not s["exists"]:
-                    continue
-                if s["fast_ok"] and s["needs_verify"]:
-                    to_clear_verify.append(s["sid"])
-                if not s["fast_ok"] and not s["needs_verify"]:
-                    to_set_verify.append(s["sid"])
-
-            if a_hash is None:
-                if states and all_missing:  # remove seed Asset completely, if no valid AssetCache exists
-                    sess.execute(sqlalchemy.delete(AssetInfo).where(AssetInfo.asset_id == aid))
-                    asset = sess.get(Asset, aid)
-                    if asset:
-                        sess.delete(asset)
-                else:
-                    for s in states:
-                        if s["exists"]:
-                            survivors.add(os.path.abspath(s["fp"]))
-                continue
-
-            if any_fast_ok:  # if Asset has at least one valid AssetCache record, remove any invalid AssetCache records
-                for s in states:
-                    if not s["exists"]:
-                        stale_state_ids.append(s["sid"])
-                if update_missing_tags:
-                    with contextlib.suppress(Exception):
-                        remove_missing_tag_for_asset_id(sess, asset_id=aid)
-            elif update_missing_tags:
-                with contextlib.suppress(Exception):
-                    add_missing_tag_for_asset_id(sess, asset_id=aid, origin="automatic")
-
-            for s in states:
-                if s["exists"]:
-                    survivors.add(os.path.abspath(s["fp"]))
-
-        if stale_state_ids:
-            sess.execute(sqlalchemy.delete(AssetCacheState).where(AssetCacheState.id.in_(stale_state_ids)))
-        if to_set_verify:
-            sess.execute(
-                sqlalchemy.update(AssetCacheState)
-                .where(AssetCacheState.id.in_(to_set_verify))
-                .values(needs_verify=True)
-            )
-        if to_clear_verify:
-            sess.execute(
-                sqlalchemy.update(AssetCacheState)
-                .where(AssetCacheState.id.in_(to_clear_verify))
-                .values(needs_verify=False)
-            )
-        sess.commit()
-        return survivors if collect_existing_paths else None

app/database/models.py

@@ -1,21 +1,14 @@
-from typing import Any
-from datetime import datetime
-from sqlalchemy.orm import DeclarativeBase
+from sqlalchemy.orm import declarative_base
 
-class Base(DeclarativeBase):
-    pass
+Base = declarative_base()
 
-def to_dict(obj: Any, include_none: bool = False) -> dict[str, Any]:
+
+def to_dict(obj):
     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
+    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/model_manager.py

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

comfy/cli_args.py

@@ -97,13 +97,6 @@ class LatentPreviewMethod(enum.Enum):
     Latent2RGB = "latent2rgb"
     TAESD = "taesd"
 
-    @classmethod
-    def from_string(cls, value: str):
-        for member in cls:
-            if member.value == value:
-                return member
-        return None
-
 parser.add_argument("--preview-method", type=LatentPreviewMethod, default=LatentPreviewMethod.NoPreviews, help="Default preview method for sampler nodes.", action=EnumAction)
 
 parser.add_argument("--preview-size", type=int, default=512, help="Sets the maximum preview size for sampler nodes.")
@@ -128,12 +121,6 @@ upcast.add_argument("--force-upcast-attention", action="store_true", help="Force
 upcast.add_argument("--dont-upcast-attention", action="store_true", help="Disable all upcasting of attention. Should be unnecessary except for debugging.")
 
 
-parser.add_argument("--enable-manager", action="store_true", help="Enable the ComfyUI-Manager feature.")
-manager_group = parser.add_mutually_exclusive_group()
-manager_group.add_argument("--disable-manager-ui", action="store_true", help="Disables only the ComfyUI-Manager UI and endpoints. Scheduled installations and similar background tasks will still operate.")
-manager_group.add_argument("--enable-manager-legacy-ui", action="store_true", help="Enables the legacy UI of ComfyUI-Manager")
-
-
 vram_group = parser.add_mutually_exclusive_group()
 vram_group.add_argument("--gpu-only", action="store_true", help="Store and run everything (text encoders/CLIP models, etc... on the GPU).")
 vram_group.add_argument("--highvram", action="store_true", help="By default models will be unloaded to CPU memory after being used. This option keeps them in GPU memory.")
@@ -181,7 +168,6 @@ parser.add_argument("--multi-user", action="store_true", help="Enables per-user
 parser.add_argument("--verbose", default='INFO', const='DEBUG', nargs="?", choices=['DEBUG', 'INFO', 'WARNING', 'ERROR', 'CRITICAL'], help='Set the logging level')
 parser.add_argument("--log-stdout", action="store_true", help="Send normal process output to stdout instead of stderr (default).")
 
-
 # The default built-in provider hosted under web/
 DEFAULT_VERSION_STRING = "comfyanonymous/ComfyUI@latest"
 
@@ -231,7 +217,6 @@ database_default_path = os.path.abspath(
     os.path.join(os.path.dirname(__file__), "..", "user", "comfyui.db")
 )
 parser.add_argument("--database-url", type=str, default=f"sqlite:///{database_default_path}", help="Specify the database URL, e.g. for an in-memory database you can use 'sqlite:///:memory:'.")
-parser.add_argument("--disable-assets-autoscan", action="store_true", help="Disable asset scanning on startup for database synchronization.")
 
 if comfy.options.args_parsing:
     args = parser.parse_args()

comfy/clip_model.py

@@ -2,25 +2,6 @@ import torch
 from comfy.ldm.modules.attention import optimized_attention_for_device
 import comfy.ops
 
-def clip_preprocess(image, size=224, mean=[0.48145466, 0.4578275, 0.40821073], std=[0.26862954, 0.26130258, 0.27577711], crop=True):
-    image = image[:, :, :, :3] if image.shape[3] > 3 else image
-    mean = torch.tensor(mean, device=image.device, dtype=image.dtype)
-    std = torch.tensor(std, device=image.device, dtype=image.dtype)
-    image = image.movedim(-1, 1)
-    if not (image.shape[2] == size and image.shape[3] == size):
-        if crop:
-            scale = (size / min(image.shape[2], image.shape[3]))
-            scale_size = (round(scale * image.shape[2]), round(scale * image.shape[3]))
-        else:
-            scale_size = (size, size)
-
-        image = torch.nn.functional.interpolate(image, size=scale_size, mode="bicubic", antialias=True)
-        h = (image.shape[2] - size)//2
-        w = (image.shape[3] - size)//2
-        image = image[:,:,h:h+size,w:w+size]
-    image = torch.clip((255. * image), 0, 255).round() / 255.0
-    return (image - mean.view([3,1,1])) / std.view([3,1,1])
-
 class CLIPAttention(torch.nn.Module):
     def __init__(self, embed_dim, heads, dtype, device, operations):
         super().__init__()

comfy/clip_vision.py

@@ -1,5 +1,6 @@
 from .utils import load_torch_file, transformers_convert, state_dict_prefix_replace
 import os
+import torch
 import json
 import logging
 
@@ -16,7 +17,24 @@ class Output:
     def __setitem__(self, key, item):
         setattr(self, key, item)
 
-clip_preprocess = comfy.clip_model.clip_preprocess  # Prevent some stuff from breaking, TODO: remove eventually
+def clip_preprocess(image, size=224, mean=[0.48145466, 0.4578275, 0.40821073], std=[0.26862954, 0.26130258, 0.27577711], crop=True):
+    image = image[:, :, :, :3] if image.shape[3] > 3 else image
+    mean = torch.tensor(mean, device=image.device, dtype=image.dtype)
+    std = torch.tensor(std, device=image.device, dtype=image.dtype)
+    image = image.movedim(-1, 1)
+    if not (image.shape[2] == size and image.shape[3] == size):
+        if crop:
+            scale = (size / min(image.shape[2], image.shape[3]))
+            scale_size = (round(scale * image.shape[2]), round(scale * image.shape[3]))
+        else:
+            scale_size = (size, size)
+
+        image = torch.nn.functional.interpolate(image, size=scale_size, mode="bicubic", antialias=True)
+        h = (image.shape[2] - size)//2
+        w = (image.shape[3] - size)//2
+        image = image[:,:,h:h+size,w:w+size]
+    image = torch.clip((255. * image), 0, 255).round() / 255.0
+    return (image - mean.view([3,1,1])) / std.view([3,1,1])
 
 IMAGE_ENCODERS = {
     "clip_vision_model": comfy.clip_model.CLIPVisionModelProjection,
@@ -55,7 +73,7 @@ class ClipVisionModel():
 
     def encode_image(self, image, crop=True):
         comfy.model_management.load_model_gpu(self.patcher)
-        pixel_values = comfy.clip_model.clip_preprocess(image.to(self.load_device), size=self.image_size, mean=self.image_mean, std=self.image_std, crop=crop).float()
+        pixel_values = clip_preprocess(image.to(self.load_device), size=self.image_size, mean=self.image_mean, std=self.image_std, crop=crop).float()
         out = self.model(pixel_values=pixel_values, intermediate_output='all' if self.return_all_hidden_states else -2)
 
         outputs = Output()

comfy/context_windows.py

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

comfy/hooks.py

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

comfy/k_diffusion/sampling.py

@@ -74,9 +74,6 @@ def get_ancestral_step(sigma_from, sigma_to, eta=1.):
 
 def default_noise_sampler(x, seed=None):
     if seed is not None:
-        if x.device == torch.device("cpu"):
-            seed += 1
-
         generator = torch.Generator(device=x.device)
         generator.manual_seed(seed)
     else:
@@ -1560,13 +1557,10 @@ def sample_er_sde(model, x, sigmas, extra_args=None, callback=None, disable=None
 
 
 @torch.no_grad()
-def sample_seeds_2(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None, r=0.5, solver_type="phi_1"):
+def sample_seeds_2(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None, r=0.5):
     """SEEDS-2 - Stochastic Explicit Exponential Derivative-free Solvers (VP Data Prediction) stage 2.
     arXiv: https://arxiv.org/abs/2305.14267 (NeurIPS 2023)
     """
-    if solver_type not in {"phi_1", "phi_2"}:
-        raise ValueError("solver_type must be 'phi_1' or 'phi_2'")
-
     extra_args = {} if extra_args is None else extra_args
     seed = extra_args.get("seed", None)
     noise_sampler = default_noise_sampler(x, seed=seed) if noise_sampler is None else noise_sampler
@@ -1606,14 +1600,8 @@ def sample_seeds_2(model, x, sigmas, extra_args=None, callback=None, disable=Non
         denoised_2 = model(x_2, sigma_s_1 * s_in, **extra_args)
 
         # Step 2
-        if solver_type == "phi_1":
-            denoised_d = torch.lerp(denoised, denoised_2, fac)
-            x = sigmas[i + 1] / sigmas[i] * (-h * eta).exp() * x - alpha_t * ei_h_phi_1(-h_eta) * denoised_d
-        elif solver_type == "phi_2":
-            b2 = ei_h_phi_2(-h_eta) / r
-            b1 = ei_h_phi_1(-h_eta) - b2
-            x = sigmas[i + 1] / sigmas[i] * (-h * eta).exp() * x - alpha_t * (b1 * denoised + b2 * denoised_2)
-
+        denoised_d = torch.lerp(denoised, denoised_2, fac)
+        x = sigmas[i + 1] / sigmas[i] * (-h * eta).exp() * x - alpha_t * ei_h_phi_1(-h_eta) * denoised_d
         if inject_noise:
             segment_factor = (r - 1) * h * eta
             sde_noise = sde_noise * segment_factor.exp()
@@ -1621,17 +1609,6 @@ def sample_seeds_2(model, x, sigmas, extra_args=None, callback=None, disable=Non
             x = x + sde_noise * sigmas[i + 1] * s_noise
     return x
 
-@torch.no_grad()
-def sample_exp_heun_2_x0(model, x, sigmas, extra_args=None, callback=None, disable=None, solver_type="phi_2"):
-    """Deterministic exponential Heun second order method in data prediction (x0) and logSNR time."""
-    return sample_seeds_2(model, x, sigmas, extra_args=extra_args, callback=callback, disable=disable, eta=0.0, s_noise=0.0, noise_sampler=None, r=1.0, solver_type=solver_type)
-
-
-@torch.no_grad()
-def sample_exp_heun_2_x0_sde(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None, solver_type="phi_2"):
-    """Stochastic exponential Heun second order method in data prediction (x0) and logSNR time."""
-    return sample_seeds_2(model, x, sigmas, extra_args=extra_args, callback=callback, disable=disable, eta=eta, s_noise=s_noise, noise_sampler=noise_sampler, r=1.0, solver_type=solver_type)
-
 
 @torch.no_grad()
 def sample_seeds_3(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None, r_1=1./3, r_2=2./3):
@@ -1779,7 +1756,7 @@ def sample_sa_solver(model, x, sigmas, extra_args=None, callback=None, disable=F
         # Predictor
         if sigmas[i + 1] == 0:
             # Denoising step
-            x_pred = denoised
+            x = denoised
         else:
             tau_t = tau_func(sigmas[i + 1])
             curr_lambdas = lambdas[i - predictor_order_used + 1:i + 1]
@@ -1800,7 +1777,7 @@ def sample_sa_solver(model, x, sigmas, extra_args=None, callback=None, disable=F
             if tau_t > 0 and s_noise > 0:
                 noise = noise_sampler(sigmas[i], sigmas[i + 1]) * sigmas[i + 1] * (-2 * tau_t ** 2 * h).expm1().neg().sqrt() * s_noise
                 x_pred = x_pred + noise
-    return x_pred
+    return x
 
 
 @torch.no_grad()

comfy/latent_formats.py

@@ -407,11 +407,6 @@ class LTXV(LatentFormat):
 
         self.latent_rgb_factors_bias = [-0.0571, -0.1657, -0.2512]
 
-class LTXAV(LTXV):
-    def __init__(self):
-        self.latent_rgb_factors = None
-        self.latent_rgb_factors_bias = None
-
 class HunyuanVideo(LatentFormat):
     latent_channels = 16
     latent_dimensions = 3

comfy/ldm/chroma/model.py

@@ -40,8 +40,7 @@ class ChromaParams:
     out_dim: int
     hidden_dim: int
     n_layers: int
-    txt_ids_dims: list
-    vec_in_dim: int
+
 
 
 

comfy/ldm/chroma_radiance/model.py

@@ -37,7 +37,7 @@ class ChromaRadianceParams(ChromaParams):
     nerf_final_head_type: str
     # None means use the same dtype as the model.
     nerf_embedder_dtype: Optional[torch.dtype]
-    use_x0: bool
+
 
 class ChromaRadiance(Chroma):
     """
@@ -159,9 +159,6 @@ class ChromaRadiance(Chroma):
         self.skip_dit = []
         self.lite = False
 
-        if params.use_x0:
-            self.register_buffer("__x0__", torch.tensor([]))
-
     @property
     def _nerf_final_layer(self) -> nn.Module:
         if self.params.nerf_final_head_type == "linear":
@@ -270,7 +267,7 @@ class ChromaRadiance(Chroma):
         bad_keys = tuple(
             k
             for k, v in overrides.items()
-            if not isinstance(v, type(getattr(params, k))) and (v is not None or k not in nullable_keys)
+            if type(v) != type(getattr(params, k)) and (v is not None or k not in nullable_keys)
         )
         if bad_keys:
             e = f"Invalid value(s) in transformer_options chroma_radiance_options: {', '.join(bad_keys)}"
@@ -279,12 +276,6 @@ class ChromaRadiance(Chroma):
         params_dict |= overrides
         return params.__class__(**params_dict)
 
-    def _apply_x0_residual(self, predicted, noisy, timesteps):
-
-        # non zero during training to prevent 0 div
-        eps = 0.0
-        return (noisy - predicted) / (timesteps.view(-1,1,1,1) + eps)
-
     def _forward(
         self,
         x: Tensor,
@@ -325,11 +316,4 @@ class ChromaRadiance(Chroma):
             transformer_options,
             attn_mask=kwargs.get("attention_mask", None),
         )
-
-        out = self.forward_nerf(img, img_out, params)[:, :, :h, :w]
-
-        # If x0 variant → v-pred, just return this instead
-        if hasattr(self, "__x0__"):
-            out = self._apply_x0_residual(out, img, timestep)
-        return out
-
+        return self.forward_nerf(img, img_out, params)[:, :, :h, :w]

comfy/ldm/flux/layers.py

@@ -57,35 +57,6 @@ class MLPEmbedder(nn.Module):
     def forward(self, x: Tensor) -> Tensor:
         return self.out_layer(self.silu(self.in_layer(x)))
 
-class YakMLP(nn.Module):
-    def __init__(self, hidden_size: int, intermediate_size: int, dtype=None, device=None, operations=None):
-        super().__init__()
-        self.hidden_size = hidden_size
-        self.intermediate_size = intermediate_size
-        self.gate_proj = operations.Linear(self.hidden_size, self.intermediate_size, bias=True, dtype=dtype, device=device)
-        self.up_proj = operations.Linear(self.hidden_size, self.intermediate_size, bias=True, dtype=dtype, device=device)
-        self.down_proj = operations.Linear(self.intermediate_size, self.hidden_size, bias=True, dtype=dtype, device=device)
-        self.act_fn = nn.SiLU()
-
-    def forward(self, x: Tensor) -> Tensor:
-        down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
-        return down_proj
-
-def build_mlp(hidden_size, mlp_hidden_dim, mlp_silu_act=False, yak_mlp=False, dtype=None, device=None, operations=None):
-    if yak_mlp:
-        return YakMLP(hidden_size, mlp_hidden_dim, dtype=dtype, device=device, operations=operations)
-    if mlp_silu_act:
-        return nn.Sequential(
-            operations.Linear(hidden_size, mlp_hidden_dim * 2, bias=False, dtype=dtype, device=device),
-            SiLUActivation(),
-            operations.Linear(mlp_hidden_dim, hidden_size, bias=False, dtype=dtype, device=device),
-        )
-    else:
-        return nn.Sequential(
-            operations.Linear(hidden_size, mlp_hidden_dim, bias=True, dtype=dtype, device=device),
-            nn.GELU(approximate="tanh"),
-            operations.Linear(mlp_hidden_dim, hidden_size, bias=True, dtype=dtype, device=device),
-        )
 
 class RMSNorm(torch.nn.Module):
     def __init__(self, dim: int, dtype=None, device=None, operations=None):
@@ -169,7 +140,7 @@ class SiLUActivation(nn.Module):
 
 
 class DoubleStreamBlock(nn.Module):
-    def __init__(self, hidden_size: int, num_heads: int, mlp_ratio: float, qkv_bias: bool = False, flipped_img_txt=False, modulation=True, mlp_silu_act=False, proj_bias=True, yak_mlp=False, dtype=None, device=None, operations=None):
+    def __init__(self, hidden_size: int, num_heads: int, mlp_ratio: float, qkv_bias: bool = False, flipped_img_txt=False, modulation=True, mlp_silu_act=False, proj_bias=True, dtype=None, device=None, operations=None):
         super().__init__()
 
         mlp_hidden_dim = int(hidden_size * mlp_ratio)
@@ -185,7 +156,18 @@ class DoubleStreamBlock(nn.Module):
 
         self.img_norm2 = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
 
-        self.img_mlp = build_mlp(hidden_size, mlp_hidden_dim, mlp_silu_act=mlp_silu_act, yak_mlp=yak_mlp, dtype=dtype, device=device, operations=operations)
+        if mlp_silu_act:
+            self.img_mlp = nn.Sequential(
+                operations.Linear(hidden_size, mlp_hidden_dim * 2, bias=False, dtype=dtype, device=device),
+                SiLUActivation(),
+                operations.Linear(mlp_hidden_dim, hidden_size, bias=False, dtype=dtype, device=device),
+            )
+        else:
+            self.img_mlp = nn.Sequential(
+                operations.Linear(hidden_size, mlp_hidden_dim, bias=True, dtype=dtype, device=device),
+                nn.GELU(approximate="tanh"),
+                operations.Linear(mlp_hidden_dim, hidden_size, bias=True, dtype=dtype, device=device),
+            )
 
         if self.modulation:
             self.txt_mod = Modulation(hidden_size, double=True, dtype=dtype, device=device, operations=operations)
@@ -195,7 +177,18 @@ class DoubleStreamBlock(nn.Module):
 
         self.txt_norm2 = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
 
-        self.txt_mlp = build_mlp(hidden_size, mlp_hidden_dim, mlp_silu_act=mlp_silu_act, yak_mlp=yak_mlp, dtype=dtype, device=device, operations=operations)
+        if mlp_silu_act:
+            self.txt_mlp = nn.Sequential(
+                operations.Linear(hidden_size, mlp_hidden_dim * 2, bias=False, dtype=dtype, device=device),
+                SiLUActivation(),
+                operations.Linear(mlp_hidden_dim, hidden_size, bias=False, dtype=dtype, device=device),
+            )
+        else:
+            self.txt_mlp = nn.Sequential(
+                operations.Linear(hidden_size, mlp_hidden_dim, bias=True, dtype=dtype, device=device),
+                nn.GELU(approximate="tanh"),
+                operations.Linear(mlp_hidden_dim, hidden_size, bias=True, dtype=dtype, device=device),
+            )
 
         self.flipped_img_txt = flipped_img_txt
 
@@ -282,7 +275,6 @@ class SingleStreamBlock(nn.Module):
         modulation=True,
         mlp_silu_act=False,
         bias=True,
-        yak_mlp=False,
         dtype=None,
         device=None,
         operations=None
@@ -296,17 +288,12 @@ class SingleStreamBlock(nn.Module):
         self.mlp_hidden_dim = int(hidden_size * mlp_ratio)
 
         self.mlp_hidden_dim_first = self.mlp_hidden_dim
-        self.yak_mlp = yak_mlp
         if mlp_silu_act:
             self.mlp_hidden_dim_first = int(hidden_size * mlp_ratio * 2)
             self.mlp_act = SiLUActivation()
         else:
             self.mlp_act = nn.GELU(approximate="tanh")
 
-        if self.yak_mlp:
-            self.mlp_hidden_dim_first *= 2
-            self.mlp_act = nn.SiLU()
-
         # qkv and mlp_in
         self.linear1 = operations.Linear(hidden_size, hidden_size * 3 + self.mlp_hidden_dim_first, bias=bias, dtype=dtype, device=device)
         # proj and mlp_out
@@ -338,10 +325,7 @@ class SingleStreamBlock(nn.Module):
         attn = attention(q, k, v, pe=pe, mask=attn_mask, transformer_options=transformer_options)
         del q, k, v
         # compute activation in mlp stream, cat again and run second linear layer
-        if self.yak_mlp:
-            mlp = self.mlp_act(mlp[..., self.mlp_hidden_dim_first // 2:]) * mlp[..., :self.mlp_hidden_dim_first // 2]
-        else:
-            mlp = self.mlp_act(mlp)
+        mlp = self.mlp_act(mlp)
         output = self.linear2(torch.cat((attn, mlp), 2))
         x += apply_mod(output, mod.gate, None, modulation_dims)
         if x.dtype == torch.float16:

comfy/ldm/flux/math.py

@@ -4,7 +4,6 @@ from torch import Tensor
 
 from comfy.ldm.modules.attention import optimized_attention
 import comfy.model_management
-import logging
 
 
 def attention(q: Tensor, k: Tensor, v: Tensor, pe: Tensor, mask=None, transformer_options={}) -> Tensor:
@@ -14,6 +13,7 @@ def attention(q: Tensor, k: Tensor, v: Tensor, pe: Tensor, mask=None, transforme
     x = optimized_attention(q, k, v, heads, skip_reshape=True, mask=mask, transformer_options=transformer_options)
     return x
 
+
 def rope(pos: Tensor, dim: int, theta: int) -> Tensor:
     assert dim % 2 == 0
     if comfy.model_management.is_device_mps(pos.device) or comfy.model_management.is_intel_xpu() or comfy.model_management.is_directml_enabled():
@@ -28,20 +28,13 @@ def rope(pos: Tensor, dim: int, theta: int) -> Tensor:
     out = rearrange(out, "b n d (i j) -> b n d i j", i=2, j=2)
     return out.to(dtype=torch.float32, device=pos.device)
 
+def apply_rope1(x: Tensor, freqs_cis: Tensor):
+    x_ = x.to(dtype=freqs_cis.dtype).reshape(*x.shape[:-1], -1, 1, 2)
 
-try:
-    import comfy.quant_ops
-    apply_rope = comfy.quant_ops.ck.apply_rope
-    apply_rope1 = comfy.quant_ops.ck.apply_rope1
-except:
-    logging.warning("No comfy kitchen, using old apply_rope functions.")
-    def apply_rope1(x: Tensor, freqs_cis: Tensor):
-        x_ = x.to(dtype=freqs_cis.dtype).reshape(*x.shape[:-1], -1, 1, 2)
+    x_out = freqs_cis[..., 0] * x_[..., 0]
+    x_out.addcmul_(freqs_cis[..., 1], x_[..., 1])
 
-        x_out = freqs_cis[..., 0] * x_[..., 0]
-        x_out.addcmul_(freqs_cis[..., 1], x_[..., 1])
+    return x_out.reshape(*x.shape).type_as(x)
 
-        return x_out.reshape(*x.shape).type_as(x)
-
-    def apply_rope(xq: Tensor, xk: Tensor, freqs_cis: Tensor):
-        return apply_rope1(xq, freqs_cis), apply_rope1(xk, freqs_cis)
+def apply_rope(xq: Tensor, xk: Tensor, freqs_cis: Tensor):
+    return apply_rope1(xq, freqs_cis), apply_rope1(xk, freqs_cis)

comfy/ldm/flux/model.py

@@ -15,8 +15,7 @@ from .layers import (
     MLPEmbedder,
     SingleStreamBlock,
     timestep_embedding,
-    Modulation,
-    RMSNorm
+    Modulation
 )
 
 @dataclass
@@ -35,14 +34,11 @@ class FluxParams:
     patch_size: int
     qkv_bias: bool
     guidance_embed: bool
-    txt_ids_dims: list
     global_modulation: bool = False
     mlp_silu_act: bool = False
     ops_bias: bool = True
     default_ref_method: str = "offset"
     ref_index_scale: float = 1.0
-    yak_mlp: bool = False
-    txt_norm: bool = False
 
 
 class Flux(nn.Module):
@@ -80,11 +76,6 @@ class Flux(nn.Module):
         )
         self.txt_in = operations.Linear(params.context_in_dim, self.hidden_size, bias=params.ops_bias, dtype=dtype, device=device)
 
-        if params.txt_norm:
-            self.txt_norm = RMSNorm(params.context_in_dim, dtype=dtype, device=device, operations=operations)
-        else:
-            self.txt_norm = None
-
         self.double_blocks = nn.ModuleList(
             [
                 DoubleStreamBlock(
@@ -95,7 +86,6 @@ class Flux(nn.Module):
                     modulation=params.global_modulation is False,
                     mlp_silu_act=params.mlp_silu_act,
                     proj_bias=params.ops_bias,
-                    yak_mlp=params.yak_mlp,
                     dtype=dtype, device=device, operations=operations
                 )
                 for _ in range(params.depth)
@@ -104,7 +94,7 @@ class Flux(nn.Module):
 
         self.single_blocks = nn.ModuleList(
             [
-                SingleStreamBlock(self.hidden_size, self.num_heads, mlp_ratio=params.mlp_ratio, modulation=params.global_modulation is False, mlp_silu_act=params.mlp_silu_act, bias=params.ops_bias, yak_mlp=params.yak_mlp, dtype=dtype, device=device, operations=operations)
+                SingleStreamBlock(self.hidden_size, self.num_heads, mlp_ratio=params.mlp_ratio, modulation=params.global_modulation is False, mlp_silu_act=params.mlp_silu_act, bias=params.ops_bias, dtype=dtype, device=device, operations=operations)
                 for _ in range(params.depth_single_blocks)
             ]
         )
@@ -160,8 +150,6 @@ class Flux(nn.Module):
                 y = torch.zeros((img.shape[0], self.params.vec_in_dim), device=img.device, dtype=img.dtype)
             vec = vec + self.vector_in(y[:, :self.params.vec_in_dim])
 
-        if self.txt_norm is not None:
-            txt = self.txt_norm(txt)
         txt = self.txt_in(txt)
 
         vec_orig = vec
@@ -344,9 +332,8 @@ class Flux(nn.Module):
 
         txt_ids = torch.zeros((bs, context.shape[1], len(self.params.axes_dim)), device=x.device, dtype=torch.float32)
 
-        if len(self.params.txt_ids_dims) > 0:
-            for i in self.params.txt_ids_dims:
-                txt_ids[:, :, i] = torch.linspace(0, context.shape[1] - 1, steps=context.shape[1], device=x.device, dtype=torch.float32)
+        if len(self.params.axes_dim) == 4: # Flux 2
+            txt_ids[:, :, 3] = torch.linspace(0, context.shape[1] - 1, steps=context.shape[1], device=x.device, dtype=torch.float32)
 
         out = self.forward_orig(img, img_ids, context, txt_ids, timestep, y, guidance, control, transformer_options, attn_mask=kwargs.get("attention_mask", None))
         out = out[:, :img_tokens]

comfy/ldm/hunyuan_video/model.py

@@ -43,7 +43,6 @@ class HunyuanVideoParams:
     meanflow: bool
     use_cond_type_embedding: bool
     vision_in_dim: int
-    meanflow_sum: bool
 
 
 class SelfAttentionRef(nn.Module):
@@ -318,7 +317,7 @@ class HunyuanVideo(nn.Module):
                 timesteps_r = transformer_options['sample_sigmas'][w[0] + 1]
                 timesteps_r = timesteps_r.unsqueeze(0).to(device=timesteps.device, dtype=timesteps.dtype)
                 vec_r = self.time_r_in(timestep_embedding(timesteps_r, 256, time_factor=1000.0).to(img.dtype))
-                vec = (vec + vec_r) if self.params.meanflow_sum else (vec + vec_r) / 2
+                vec = (vec + vec_r) / 2
 
         if ref_latent is not None:
             ref_latent_ids = self.img_ids(ref_latent)

comfy/ldm/hunyuan_video/upsampler.py

@@ -1,10 +1,8 @@
 import torch
 import torch.nn as nn
 import torch.nn.functional as F
-from comfy.ldm.modules.diffusionmodules.model import ResnetBlock, VideoConv3d
-from comfy.ldm.hunyuan_video.vae_refiner import RMS_norm
-import comfy.model_management
-import comfy.model_patcher
+from comfy.ldm.hunyuan_video.vae_refiner import RMS_norm, ResnetBlock, VideoConv3d
+import model_management, model_patcher
 
 class SRResidualCausalBlock3D(nn.Module):
     def __init__(self, channels: int):
@@ -103,13 +101,13 @@ UPSAMPLERS = {
 
 class HunyuanVideo15SRModel():
     def __init__(self, model_type, config):
-        self.load_device = comfy.model_management.vae_device()
-        offload_device = comfy.model_management.vae_offload_device()
-        self.dtype = comfy.model_management.vae_dtype(self.load_device)
+        self.load_device = model_management.vae_device()
+        offload_device = model_management.vae_offload_device()
+        self.dtype = model_management.vae_dtype(self.load_device)
         self.model_class = UPSAMPLERS.get(model_type)
         self.model = self.model_class(**config).eval()
 
-        self.patcher = comfy.model_patcher.ModelPatcher(self.model, load_device=self.load_device, offload_device=offload_device)
+        self.patcher = model_patcher.ModelPatcher(self.model, load_device=self.load_device, offload_device=offload_device)
 
     def load_sd(self, sd):
         return self.model.load_state_dict(sd, strict=True)
@@ -118,5 +116,5 @@ class HunyuanVideo15SRModel():
         return self.model.state_dict()
 
     def resample_latent(self, latent):
-        comfy.model_management.load_model_gpu(self.patcher)
+        model_management.load_model_gpu(self.patcher)
         return self.model(latent.to(self.load_device))

comfy/ldm/hunyuan_video/vae_refiner.py

@@ -1,12 +1,42 @@
 import torch
 import torch.nn as nn
 import torch.nn.functional as F
-from comfy.ldm.modules.diffusionmodules.model import ResnetBlock, AttnBlock, CarriedConv3d, Normalize, conv_carry_causal_3d, torch_cat_if_needed
+from comfy.ldm.modules.diffusionmodules.model import ResnetBlock, AttnBlock, VideoConv3d, Normalize
 import comfy.ops
 import comfy.ldm.models.autoencoder
 import comfy.model_management
 ops = comfy.ops.disable_weight_init
 
+class NoPadConv3d(nn.Module):
+    def __init__(self, n_channels, out_channels, kernel_size, stride=1, dilation=1, padding=0, **kwargs):
+        super().__init__()
+        self.conv = ops.Conv3d(n_channels, out_channels, kernel_size, stride=stride, dilation=dilation, **kwargs)
+
+    def forward(self, x):
+        return self.conv(x)
+
+
+def conv_carry_causal_3d(xl, op, conv_carry_in=None, conv_carry_out=None):
+
+    x = xl[0]
+    xl.clear()
+
+    if conv_carry_out is not None:
+        to_push = x[:, :, -2:, :, :].clone()
+        conv_carry_out.append(to_push)
+
+    if isinstance(op, NoPadConv3d):
+        if conv_carry_in is None:
+            x = torch.nn.functional.pad(x, (1, 1, 1, 1, 2, 0), mode = 'replicate')
+        else:
+            carry_len = conv_carry_in[0].shape[2]
+            x = torch.cat([conv_carry_in.pop(0), x], dim=2)
+            x = torch.nn.functional.pad(x, (1, 1, 1, 1, 2 - carry_len, 0), mode = 'replicate')
+
+    out = op(x)
+
+    return out
+
 
 class RMS_norm(nn.Module):
     def __init__(self, dim):
@@ -19,7 +49,7 @@ class RMS_norm(nn.Module):
         return F.normalize(x, dim=1) * self.scale * comfy.model_management.cast_to(self.gamma, dtype=x.dtype, device=x.device)
 
 class DnSmpl(nn.Module):
-    def __init__(self, ic, oc, tds, refiner_vae, op):
+    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
@@ -79,7 +109,7 @@ class DnSmpl(nn.Module):
 
 
 class UpSmpl(nn.Module):
-    def __init__(self, ic, oc, tus, refiner_vae, op):
+    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 = op(ic, oc * fct, kernel_size=3, stride=1, padding=1)
@@ -133,6 +163,23 @@ class UpSmpl(nn.Module):
 
         return h + x
 
+class HunyuanRefinerResnetBlock(ResnetBlock):
+    def __init__(self, in_channels, out_channels, conv_op=NoPadConv3d, norm_op=RMS_norm):
+        super().__init__(in_channels=in_channels, out_channels=out_channels, temb_channels=0, conv_op=conv_op, norm_op=norm_op)
+
+    def forward(self, x, conv_carry_in=None, conv_carry_out=None):
+        h = x
+        h = [ self.swish(self.norm1(x)) ]
+        h = conv_carry_causal_3d(h, self.conv1, conv_carry_in=conv_carry_in, conv_carry_out=conv_carry_out)
+
+        h = [ self.dropout(self.swish(self.norm2(h))) ]
+        h = conv_carry_causal_3d(h, self.conv2, conv_carry_in=conv_carry_in, conv_carry_out=conv_carry_out)
+
+        if self.in_channels != self.out_channels:
+            x = self.nin_shortcut(x)
+
+        return x+h
+
 class Encoder(nn.Module):
     def __init__(self, in_channels, z_channels, block_out_channels, num_res_blocks,
                  ffactor_spatial, ffactor_temporal, downsample_match_channel=True, refiner_vae=True, **_):
@@ -144,7 +191,7 @@ class Encoder(nn.Module):
 
         self.refiner_vae = refiner_vae
         if self.refiner_vae:
-            conv_op = CarriedConv3d
+            conv_op = NoPadConv3d
             norm_op = RMS_norm
         else:
             conv_op = ops.Conv3d
@@ -159,10 +206,9 @@ class Encoder(nn.Module):
 
         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=conv_op, norm_op=norm_op)
+            stage.block = nn.ModuleList([HunyuanRefinerResnetBlock(in_channels=ch if j == 0 else tgt,
+                                                                   out_channels=tgt,
+                                                                   conv_op=conv_op, norm_op=norm_op)
                                         for j in range(num_res_blocks)])
             ch = tgt
             if i < depth:
@@ -172,9 +218,9 @@ class Encoder(nn.Module):
             self.down.append(stage)
 
         self.mid = nn.Module()
-        self.mid.block_1 = ResnetBlock(in_channels=ch, out_channels=ch, conv_op=conv_op, norm_op=norm_op)
+        self.mid.block_1 = HunyuanRefinerResnetBlock(in_channels=ch, out_channels=ch, conv_op=conv_op, norm_op=norm_op)
         self.mid.attn_1 = AttnBlock(ch, conv_op=ops.Conv3d, norm_op=norm_op)
-        self.mid.block_2 = ResnetBlock(in_channels=ch, out_channels=ch, conv_op=conv_op, norm_op=norm_op)
+        self.mid.block_2 = HunyuanRefinerResnetBlock(in_channels=ch, out_channels=ch, conv_op=conv_op, norm_op=norm_op)
 
         self.norm_out = norm_op(ch)
         self.conv_out = conv_op(ch, z_channels << 1, 3, 1, 1)
@@ -200,20 +246,22 @@ class Encoder(nn.Module):
             conv_carry_out = []
             if i == len(x) - 1:
                 conv_carry_out = None
-
             x1 = [ x1 ]
             x1 = conv_carry_causal_3d(x1, self.conv_in, conv_carry_in, conv_carry_out)
 
             for stage in self.down:
                 for blk in stage.block:
-                    x1 = blk(x1, None, conv_carry_in, conv_carry_out)
+                    x1 = blk(x1, conv_carry_in, conv_carry_out)
                 if hasattr(stage, 'downsample'):
                     x1 = stage.downsample(x1, conv_carry_in, conv_carry_out)
 
             out.append(x1)
             conv_carry_in = conv_carry_out
 
-        out = torch_cat_if_needed(out, dim=2)
+        if len(out) > 1:
+            out = torch.cat(out, dim=2)
+        else:
+            out = out[0]
 
         x = self.mid.block_2(self.mid.attn_1(self.mid.block_1(out)))
         del out
@@ -240,7 +288,7 @@ class Decoder(nn.Module):
 
         self.refiner_vae = refiner_vae
         if self.refiner_vae:
-            conv_op = CarriedConv3d
+            conv_op = NoPadConv3d
             norm_op = RMS_norm
         else:
             conv_op = ops.Conv3d
@@ -250,9 +298,9 @@ class Decoder(nn.Module):
         self.conv_in = conv_op(z_channels, ch, kernel_size=3, stride=1, padding=1)
 
         self.mid = nn.Module()
-        self.mid.block_1 = ResnetBlock(in_channels=ch, out_channels=ch, conv_op=conv_op, norm_op=norm_op)
+        self.mid.block_1 = HunyuanRefinerResnetBlock(in_channels=ch, out_channels=ch, conv_op=conv_op, norm_op=norm_op)
         self.mid.attn_1 = AttnBlock(ch, conv_op=ops.Conv3d, norm_op=norm_op)
-        self.mid.block_2 = ResnetBlock(in_channels=ch, out_channels=ch,  conv_op=conv_op, norm_op=norm_op)
+        self.mid.block_2 = HunyuanRefinerResnetBlock(in_channels=ch, out_channels=ch,  conv_op=conv_op, norm_op=norm_op)
 
         self.up = nn.ModuleList()
         depth = (ffactor_spatial >> 1).bit_length()
@@ -260,10 +308,9 @@ class Decoder(nn.Module):
 
         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=conv_op, norm_op=norm_op)
+            stage.block = nn.ModuleList([HunyuanRefinerResnetBlock(in_channels=ch if j == 0 else tgt,
+                                                                   out_channels=tgt,
+                                                                   conv_op=conv_op, norm_op=norm_op)
                                         for j in range(num_res_blocks + 1)])
             ch = tgt
             if i < depth:
@@ -293,7 +340,7 @@ class Decoder(nn.Module):
                 conv_carry_out = None
             for stage in self.up:
                 for blk in stage.block:
-                    x1 = blk(x1, None, conv_carry_in, conv_carry_out)
+                    x1 = blk(x1, conv_carry_in, conv_carry_out)
                 if hasattr(stage, 'upsample'):
                     x1 = stage.upsample(x1, conv_carry_in, conv_carry_out)
 
@@ -303,7 +350,10 @@ class Decoder(nn.Module):
             conv_carry_in = conv_carry_out
         del x
 
-        out = torch_cat_if_needed(out, dim=2)
+        if len(out) > 1:
+            out = torch.cat(out, dim=2)
+        else:
+            out = out[0]
 
         if not self.refiner_vae:
             if z.shape[-3] == 1:

comfy/ldm/kandinsky5/model.py

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

comfy/ldm/lightricks/av_model.py

@@ -1,837 +0,0 @@
-from typing import Tuple
-import torch
-import torch.nn as nn
-from comfy.ldm.lightricks.model import (
-    CrossAttention,
-    FeedForward,
-    AdaLayerNormSingle,
-    PixArtAlphaTextProjection,
-    LTXVModel,
-)
-from comfy.ldm.lightricks.symmetric_patchifier import AudioPatchifier
-import comfy.ldm.common_dit
-
-class BasicAVTransformerBlock(nn.Module):
-    def __init__(
-        self,
-        v_dim,
-        a_dim,
-        v_heads,
-        a_heads,
-        vd_head,
-        ad_head,
-        v_context_dim=None,
-        a_context_dim=None,
-        attn_precision=None,
-        dtype=None,
-        device=None,
-        operations=None,
-    ):
-        super().__init__()
-
-        self.attn_precision = attn_precision
-
-        self.attn1 = CrossAttention(
-            query_dim=v_dim,
-            heads=v_heads,
-            dim_head=vd_head,
-            context_dim=None,
-            attn_precision=self.attn_precision,
-            dtype=dtype,
-            device=device,
-            operations=operations,
-        )
-        self.audio_attn1 = CrossAttention(
-            query_dim=a_dim,
-            heads=a_heads,
-            dim_head=ad_head,
-            context_dim=None,
-            attn_precision=self.attn_precision,
-            dtype=dtype,
-            device=device,
-            operations=operations,
-        )
-
-        self.attn2 = CrossAttention(
-            query_dim=v_dim,
-            context_dim=v_context_dim,
-            heads=v_heads,
-            dim_head=vd_head,
-            attn_precision=self.attn_precision,
-            dtype=dtype,
-            device=device,
-            operations=operations,
-        )
-        self.audio_attn2 = CrossAttention(
-            query_dim=a_dim,
-            context_dim=a_context_dim,
-            heads=a_heads,
-            dim_head=ad_head,
-            attn_precision=self.attn_precision,
-            dtype=dtype,
-            device=device,
-            operations=operations,
-        )
-
-        # Q: Video, K,V: Audio
-        self.audio_to_video_attn = CrossAttention(
-            query_dim=v_dim,
-            context_dim=a_dim,
-            heads=a_heads,
-            dim_head=ad_head,
-            attn_precision=self.attn_precision,
-            dtype=dtype,
-            device=device,
-            operations=operations,
-        )
-
-        # Q: Audio, K,V: Video
-        self.video_to_audio_attn = CrossAttention(
-            query_dim=a_dim,
-            context_dim=v_dim,
-            heads=a_heads,
-            dim_head=ad_head,
-            attn_precision=self.attn_precision,
-            dtype=dtype,
-            device=device,
-            operations=operations,
-        )
-
-        self.ff = FeedForward(
-            v_dim, dim_out=v_dim, glu=True, dtype=dtype, device=device, operations=operations
-        )
-        self.audio_ff = FeedForward(
-            a_dim, dim_out=a_dim, glu=True, dtype=dtype, device=device, operations=operations
-        )
-
-        self.scale_shift_table = nn.Parameter(torch.empty(6, v_dim, device=device, dtype=dtype))
-        self.audio_scale_shift_table = nn.Parameter(
-            torch.empty(6, a_dim, device=device, dtype=dtype)
-        )
-
-        self.scale_shift_table_a2v_ca_audio = nn.Parameter(
-            torch.empty(5, a_dim, device=device, dtype=dtype)
-        )
-        self.scale_shift_table_a2v_ca_video = nn.Parameter(
-            torch.empty(5, v_dim, device=device, dtype=dtype)
-        )
-
-    def get_ada_values(
-        self, scale_shift_table: torch.Tensor, batch_size: int, timestep: torch.Tensor, indices: slice = slice(None, None)
-    ):
-        num_ada_params = scale_shift_table.shape[0]
-
-        ada_values = (
-            scale_shift_table[indices].unsqueeze(0).unsqueeze(0).to(device=timestep.device, dtype=timestep.dtype)
-            + timestep.reshape(batch_size, timestep.shape[1], num_ada_params, -1)[:, :, indices, :]
-        ).unbind(dim=2)
-        return ada_values
-
-    def get_av_ca_ada_values(
-        self,
-        scale_shift_table: torch.Tensor,
-        batch_size: int,
-        scale_shift_timestep: torch.Tensor,
-        gate_timestep: torch.Tensor,
-        num_scale_shift_values: int = 4,
-    ):
-        scale_shift_ada_values = self.get_ada_values(
-            scale_shift_table[:num_scale_shift_values, :],
-            batch_size,
-            scale_shift_timestep,
-        )
-        gate_ada_values = self.get_ada_values(
-            scale_shift_table[num_scale_shift_values:, :],
-            batch_size,
-            gate_timestep,
-        )
-
-        scale_shift_chunks = [t.squeeze(2) for t in scale_shift_ada_values]
-        gate_ada_values = [t.squeeze(2) for t in gate_ada_values]
-
-        return (*scale_shift_chunks, *gate_ada_values)
-
-    def forward(
-        self,
-        x: Tuple[torch.Tensor, torch.Tensor],
-        v_context=None,
-        a_context=None,
-        attention_mask=None,
-        v_timestep=None,
-        a_timestep=None,
-        v_pe=None,
-        a_pe=None,
-        v_cross_pe=None,
-        a_cross_pe=None,
-        v_cross_scale_shift_timestep=None,
-        a_cross_scale_shift_timestep=None,
-        v_cross_gate_timestep=None,
-        a_cross_gate_timestep=None,
-        transformer_options=None,
-    ) -> Tuple[torch.Tensor, torch.Tensor]:
-        run_vx = transformer_options.get("run_vx", True)
-        run_ax = transformer_options.get("run_ax", True)
-
-        vx, ax = x
-        run_ax = run_ax and ax.numel() > 0
-        run_a2v = run_vx and transformer_options.get("a2v_cross_attn", True) and ax.numel() > 0
-        run_v2a = run_ax and transformer_options.get("v2a_cross_attn", True)
-
-        if run_vx:
-            vshift_msa, vscale_msa, vgate_msa = (
-                self.get_ada_values(self.scale_shift_table, vx.shape[0], v_timestep, slice(0, 3))
-            )
-
-            norm_vx = comfy.ldm.common_dit.rms_norm(vx) * (1 + vscale_msa) + vshift_msa
-            vx += self.attn1(norm_vx, pe=v_pe, transformer_options=transformer_options) * vgate_msa
-            vx += self.attn2(
-                comfy.ldm.common_dit.rms_norm(vx),
-                context=v_context,
-                mask=attention_mask,
-                transformer_options=transformer_options,
-            )
-
-            del vshift_msa, vscale_msa, vgate_msa
-
-        if run_ax:
-            ashift_msa, ascale_msa, agate_msa = (
-                self.get_ada_values(self.audio_scale_shift_table, ax.shape[0], a_timestep, slice(0, 3))
-            )
-
-            norm_ax = comfy.ldm.common_dit.rms_norm(ax) * (1 + ascale_msa) + ashift_msa
-            ax += (
-                self.audio_attn1(norm_ax, pe=a_pe, transformer_options=transformer_options)
-                * agate_msa
-            )
-            ax += self.audio_attn2(
-                comfy.ldm.common_dit.rms_norm(ax),
-                context=a_context,
-                mask=attention_mask,
-                transformer_options=transformer_options,
-            )
-
-            del ashift_msa, ascale_msa, agate_msa
-
-        # Audio - Video cross attention.
-        if run_a2v or run_v2a:
-            # norm3
-            vx_norm3 = comfy.ldm.common_dit.rms_norm(vx)
-            ax_norm3 = comfy.ldm.common_dit.rms_norm(ax)
-
-            (
-                scale_ca_audio_hidden_states_a2v,
-                shift_ca_audio_hidden_states_a2v,
-                scale_ca_audio_hidden_states_v2a,
-                shift_ca_audio_hidden_states_v2a,
-                gate_out_v2a,
-            ) = self.get_av_ca_ada_values(
-                self.scale_shift_table_a2v_ca_audio,
-                ax.shape[0],
-                a_cross_scale_shift_timestep,
-                a_cross_gate_timestep,
-            )
-
-            (
-                scale_ca_video_hidden_states_a2v,
-                shift_ca_video_hidden_states_a2v,
-                scale_ca_video_hidden_states_v2a,
-                shift_ca_video_hidden_states_v2a,
-                gate_out_a2v,
-            ) = self.get_av_ca_ada_values(
-                self.scale_shift_table_a2v_ca_video,
-                vx.shape[0],
-                v_cross_scale_shift_timestep,
-                v_cross_gate_timestep,
-            )
-
-            if run_a2v:
-                vx_scaled = (
-                    vx_norm3 * (1 + scale_ca_video_hidden_states_a2v)
-                    + shift_ca_video_hidden_states_a2v
-                )
-                ax_scaled = (
-                    ax_norm3 * (1 + scale_ca_audio_hidden_states_a2v)
-                    + shift_ca_audio_hidden_states_a2v
-                )
-                vx += (
-                    self.audio_to_video_attn(
-                        vx_scaled,
-                        context=ax_scaled,
-                        pe=v_cross_pe,
-                        k_pe=a_cross_pe,
-                        transformer_options=transformer_options,
-                    )
-                    * gate_out_a2v
-                )
-
-                del gate_out_a2v
-                del scale_ca_video_hidden_states_a2v,\
-                    shift_ca_video_hidden_states_a2v,\
-                    scale_ca_audio_hidden_states_a2v,\
-                    shift_ca_audio_hidden_states_a2v,\
-
-            if run_v2a:
-                ax_scaled = (
-                    ax_norm3 * (1 + scale_ca_audio_hidden_states_v2a)
-                    + shift_ca_audio_hidden_states_v2a
-                )
-                vx_scaled = (
-                    vx_norm3 * (1 + scale_ca_video_hidden_states_v2a)
-                    + shift_ca_video_hidden_states_v2a
-                )
-                ax += (
-                    self.video_to_audio_attn(
-                        ax_scaled,
-                        context=vx_scaled,
-                        pe=a_cross_pe,
-                        k_pe=v_cross_pe,
-                        transformer_options=transformer_options,
-                    )
-                    * gate_out_v2a
-                )
-
-                del gate_out_v2a
-                del scale_ca_video_hidden_states_v2a,\
-                    shift_ca_video_hidden_states_v2a,\
-                    scale_ca_audio_hidden_states_v2a,\
-                    shift_ca_audio_hidden_states_v2a
-
-        if run_vx:
-            vshift_mlp, vscale_mlp, vgate_mlp = (
-                self.get_ada_values(self.scale_shift_table, vx.shape[0], v_timestep, slice(3, None))
-            )
-
-            vx_scaled = comfy.ldm.common_dit.rms_norm(vx) * (1 + vscale_mlp) + vshift_mlp
-            vx += self.ff(vx_scaled) * vgate_mlp
-            del vshift_mlp, vscale_mlp, vgate_mlp
-
-        if run_ax:
-            ashift_mlp, ascale_mlp, agate_mlp = (
-                self.get_ada_values(self.audio_scale_shift_table, ax.shape[0], a_timestep, slice(3, None))
-            )
-
-            ax_scaled = comfy.ldm.common_dit.rms_norm(ax) * (1 + ascale_mlp) + ashift_mlp
-            ax += self.audio_ff(ax_scaled) * agate_mlp
-
-            del ashift_mlp, ascale_mlp, agate_mlp
-
-
-        return vx, ax
-
-
-class LTXAVModel(LTXVModel):
-    """LTXAV model for audio-video generation."""
-
-    def __init__(
-        self,
-        in_channels=128,
-        audio_in_channels=128,
-        cross_attention_dim=4096,
-        audio_cross_attention_dim=2048,
-        attention_head_dim=128,
-        audio_attention_head_dim=64,
-        num_attention_heads=32,
-        audio_num_attention_heads=32,
-        caption_channels=3840,
-        num_layers=48,
-        positional_embedding_theta=10000.0,
-        positional_embedding_max_pos=[20, 2048, 2048],
-        audio_positional_embedding_max_pos=[20],
-        causal_temporal_positioning=False,
-        vae_scale_factors=(8, 32, 32),
-        use_middle_indices_grid=False,
-        timestep_scale_multiplier=1000.0,
-        av_ca_timestep_scale_multiplier=1.0,
-        dtype=None,
-        device=None,
-        operations=None,
-        **kwargs,
-    ):
-        # Store audio-specific parameters
-        self.audio_in_channels = audio_in_channels
-        self.audio_cross_attention_dim = audio_cross_attention_dim
-        self.audio_attention_head_dim = audio_attention_head_dim
-        self.audio_num_attention_heads = audio_num_attention_heads
-        self.audio_positional_embedding_max_pos = audio_positional_embedding_max_pos
-
-        # Calculate audio dimensions
-        self.audio_inner_dim = audio_num_attention_heads * audio_attention_head_dim
-        self.audio_out_channels = audio_in_channels
-
-        # Audio-specific constants
-        self.num_audio_channels = 8
-        self.audio_frequency_bins = 16
-
-        self.av_ca_timestep_scale_multiplier = av_ca_timestep_scale_multiplier
-
-        super().__init__(
-            in_channels=in_channels,
-            cross_attention_dim=cross_attention_dim,
-            attention_head_dim=attention_head_dim,
-            num_attention_heads=num_attention_heads,
-            caption_channels=caption_channels,
-            num_layers=num_layers,
-            positional_embedding_theta=positional_embedding_theta,
-            positional_embedding_max_pos=positional_embedding_max_pos,
-            causal_temporal_positioning=causal_temporal_positioning,
-            vae_scale_factors=vae_scale_factors,
-            use_middle_indices_grid=use_middle_indices_grid,
-            timestep_scale_multiplier=timestep_scale_multiplier,
-            dtype=dtype,
-            device=device,
-            operations=operations,
-            **kwargs,
-        )
-
-    def _init_model_components(self, device, dtype, **kwargs):
-        """Initialize LTXAV-specific components."""
-        # Audio-specific projections
-        self.audio_patchify_proj = self.operations.Linear(
-            self.audio_in_channels, self.audio_inner_dim, bias=True, dtype=dtype, device=device
-        )
-
-        # Audio-specific AdaLN
-        self.audio_adaln_single = AdaLayerNormSingle(
-            self.audio_inner_dim,
-            use_additional_conditions=False,
-            dtype=dtype,
-            device=device,
-            operations=self.operations,
-        )
-
-        num_scale_shift_values = 4
-        self.av_ca_video_scale_shift_adaln_single = AdaLayerNormSingle(
-            self.inner_dim,
-            use_additional_conditions=False,
-            embedding_coefficient=num_scale_shift_values,
-            dtype=dtype,
-            device=device,
-            operations=self.operations,
-        )
-        self.av_ca_a2v_gate_adaln_single = AdaLayerNormSingle(
-            self.inner_dim,
-            use_additional_conditions=False,
-            embedding_coefficient=1,
-            dtype=dtype,
-            device=device,
-            operations=self.operations,
-        )
-        self.av_ca_audio_scale_shift_adaln_single = AdaLayerNormSingle(
-            self.audio_inner_dim,
-            use_additional_conditions=False,
-            embedding_coefficient=num_scale_shift_values,
-            dtype=dtype,
-            device=device,
-            operations=self.operations,
-        )
-        self.av_ca_v2a_gate_adaln_single = AdaLayerNormSingle(
-            self.audio_inner_dim,
-            use_additional_conditions=False,
-            embedding_coefficient=1,
-            dtype=dtype,
-            device=device,
-            operations=self.operations,
-        )
-
-        # Audio caption projection
-        self.audio_caption_projection = PixArtAlphaTextProjection(
-            in_features=self.caption_channels,
-            hidden_size=self.audio_inner_dim,
-            dtype=dtype,
-            device=device,
-            operations=self.operations,
-        )
-
-    def _init_transformer_blocks(self, device, dtype, **kwargs):
-        """Initialize transformer blocks for LTXAV."""
-        self.transformer_blocks = nn.ModuleList(
-            [
-                BasicAVTransformerBlock(
-                    v_dim=self.inner_dim,
-                    a_dim=self.audio_inner_dim,
-                    v_heads=self.num_attention_heads,
-                    a_heads=self.audio_num_attention_heads,
-                    vd_head=self.attention_head_dim,
-                    ad_head=self.audio_attention_head_dim,
-                    v_context_dim=self.cross_attention_dim,
-                    a_context_dim=self.audio_cross_attention_dim,
-                    dtype=dtype,
-                    device=device,
-                    operations=self.operations,
-                )
-                for _ in range(self.num_layers)
-            ]
-        )
-
-    def _init_output_components(self, device, dtype):
-        """Initialize output components for LTXAV."""
-        # Video output components
-        super()._init_output_components(device, dtype)
-        # Audio output components
-        self.audio_scale_shift_table = nn.Parameter(
-            torch.empty(2, self.audio_inner_dim, dtype=dtype, device=device)
-        )
-        self.audio_norm_out = self.operations.LayerNorm(
-            self.audio_inner_dim, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device
-        )
-        self.audio_proj_out = self.operations.Linear(
-            self.audio_inner_dim, self.audio_out_channels, dtype=dtype, device=device
-        )
-        self.a_patchifier = AudioPatchifier(1, start_end=True)
-
-    def separate_audio_and_video_latents(self, x, audio_length):
-        """Separate audio and video latents from combined input."""
-        # vx = x[:, : self.in_channels]
-        # ax = x[:, self.in_channels :]
-        #
-        # ax = ax.reshape(ax.shape[0], -1)
-        # ax = ax[:, : audio_length * self.num_audio_channels * self.audio_frequency_bins]
-        #
-        # ax = ax.reshape(
-        #     ax.shape[0], self.num_audio_channels, audio_length, self.audio_frequency_bins
-        # )
-
-        vx = x[0]
-        ax = x[1] if len(x) > 1 else torch.zeros(
-            (vx.shape[0], self.num_audio_channels, 0, self.audio_frequency_bins),
-            device=vx.device, dtype=vx.dtype
-        )
-        return vx, ax
-
-    def recombine_audio_and_video_latents(self, vx, ax, target_shape=None):
-        if ax.numel() == 0:
-            return vx
-        else:
-            return [vx, ax]
-        """Recombine audio and video latents for output."""
-        # if ax.device != vx.device or ax.dtype != vx.dtype:
-        #     logging.warning("Audio and video latents are on different devices or dtypes.")
-        #     ax = ax.to(device=vx.device, dtype=vx.dtype)
-        #     logging.warning(f"Audio audio latent moved to device: {ax.device}, dtype: {ax.dtype}")
-        #
-        # ax = ax.reshape(ax.shape[0], -1)
-        # # pad to f x h x w of the video latents
-        # divisor = vx.shape[-1] * vx.shape[-2] * vx.shape[-3]
-        # if target_shape is None:
-        #     repetitions = math.ceil(ax.shape[-1] / divisor)
-        # else:
-        #     repetitions = target_shape[1] - vx.shape[1]
-        # padded_len = repetitions * divisor
-        # ax = F.pad(ax, (0, padded_len - ax.shape[-1]))
-        # ax = ax.reshape(ax.shape[0], -1, vx.shape[-3], vx.shape[-2], vx.shape[-1])
-        # return torch.cat([vx, ax], dim=1)
-
-    def _process_input(self, x, keyframe_idxs, denoise_mask, **kwargs):
-        """Process input for LTXAV - separate audio and video, then patchify."""
-        audio_length = kwargs.get("audio_length", 0)
-        # Separate audio and video latents
-        vx, ax = self.separate_audio_and_video_latents(x, audio_length)
-        [vx, v_pixel_coords, additional_args] = super()._process_input(
-            vx, keyframe_idxs, denoise_mask, **kwargs
-        )
-
-        ax, a_latent_coords = self.a_patchifier.patchify(ax)
-        ax = self.audio_patchify_proj(ax)
-
-        # additional_args.update({"av_orig_shape": list(x.shape)})
-        return [vx, ax], [v_pixel_coords, a_latent_coords], additional_args
-
-    def _prepare_timestep(self, timestep, batch_size, hidden_dtype, **kwargs):
-        """Prepare timestep embeddings."""
-        # TODO: some code reuse is needed here.
-        grid_mask = kwargs.get("grid_mask", None)
-        if grid_mask is not None:
-            timestep = timestep[:, grid_mask]
-
-        timestep = timestep * self.timestep_scale_multiplier
-        v_timestep, v_embedded_timestep = self.adaln_single(
-            timestep.flatten(),
-            {"resolution": None, "aspect_ratio": None},
-            batch_size=batch_size,
-            hidden_dtype=hidden_dtype,
-        )
-
-        # Second dimension is 1 or number of tokens (if timestep_per_token)
-        v_timestep = v_timestep.view(batch_size, -1, v_timestep.shape[-1])
-        v_embedded_timestep = v_embedded_timestep.view(
-            batch_size, -1, v_embedded_timestep.shape[-1]
-        )
-
-        # Prepare audio timestep
-        a_timestep = kwargs.get("a_timestep")
-        if a_timestep is not None:
-            a_timestep = a_timestep * self.timestep_scale_multiplier
-            av_ca_factor = self.av_ca_timestep_scale_multiplier / self.timestep_scale_multiplier
-
-            av_ca_audio_scale_shift_timestep, _ = self.av_ca_audio_scale_shift_adaln_single(
-                a_timestep.flatten(),
-                {"resolution": None, "aspect_ratio": None},
-                batch_size=batch_size,
-                hidden_dtype=hidden_dtype,
-            )
-            av_ca_video_scale_shift_timestep, _ = self.av_ca_video_scale_shift_adaln_single(
-                timestep.flatten(),
-                {"resolution": None, "aspect_ratio": None},
-                batch_size=batch_size,
-                hidden_dtype=hidden_dtype,
-            )
-            av_ca_a2v_gate_noise_timestep, _ = self.av_ca_a2v_gate_adaln_single(
-                timestep.flatten() * av_ca_factor,
-                {"resolution": None, "aspect_ratio": None},
-                batch_size=batch_size,
-                hidden_dtype=hidden_dtype,
-            )
-            av_ca_v2a_gate_noise_timestep, _ = self.av_ca_v2a_gate_adaln_single(
-                a_timestep.flatten() * av_ca_factor,
-                {"resolution": None, "aspect_ratio": None},
-                batch_size=batch_size,
-                hidden_dtype=hidden_dtype,
-            )
-
-            a_timestep, a_embedded_timestep = self.audio_adaln_single(
-                a_timestep.flatten(),
-                {"resolution": None, "aspect_ratio": None},
-                batch_size=batch_size,
-                hidden_dtype=hidden_dtype,
-            )
-            a_timestep = a_timestep.view(batch_size, -1, a_timestep.shape[-1])
-            a_embedded_timestep = a_embedded_timestep.view(
-                batch_size, -1, a_embedded_timestep.shape[-1]
-            )
-            cross_av_timestep_ss = [
-                av_ca_audio_scale_shift_timestep,
-                av_ca_video_scale_shift_timestep,
-                av_ca_a2v_gate_noise_timestep,
-                av_ca_v2a_gate_noise_timestep,
-            ]
-            cross_av_timestep_ss = list(
-                [t.view(batch_size, -1, t.shape[-1]) for t in cross_av_timestep_ss]
-            )
-        else:
-            a_timestep = timestep
-            a_embedded_timestep = kwargs.get("embedded_timestep")
-            cross_av_timestep_ss = []
-
-        return [v_timestep, a_timestep, cross_av_timestep_ss], [
-            v_embedded_timestep,
-            a_embedded_timestep,
-        ]
-
-    def _prepare_context(self, context, batch_size, x, attention_mask=None):
-        vx = x[0]
-        ax = x[1]
-        v_context, a_context = torch.split(
-            context, int(context.shape[-1] / 2), len(context.shape) - 1
-        )
-
-        v_context, attention_mask = super()._prepare_context(
-            v_context, batch_size, vx, attention_mask
-        )
-        if self.audio_caption_projection is not None:
-            a_context = self.audio_caption_projection(a_context)
-            a_context = a_context.view(batch_size, -1, ax.shape[-1])
-
-        return [v_context, a_context], attention_mask
-
-    def _prepare_positional_embeddings(self, pixel_coords, frame_rate, x_dtype):
-        v_pixel_coords = pixel_coords[0]
-        v_pe = super()._prepare_positional_embeddings(v_pixel_coords, frame_rate, x_dtype)
-
-        a_latent_coords = pixel_coords[1]
-        a_pe = self._precompute_freqs_cis(
-            a_latent_coords,
-            dim=self.audio_inner_dim,
-            out_dtype=x_dtype,
-            max_pos=self.audio_positional_embedding_max_pos,
-            use_middle_indices_grid=self.use_middle_indices_grid,
-            num_attention_heads=self.audio_num_attention_heads,
-        )
-
-        # calculate positional embeddings for the middle of the token duration, to use in av cross attention layers.
-        max_pos = max(
-            self.positional_embedding_max_pos[0], self.audio_positional_embedding_max_pos[0]
-        )
-        v_pixel_coords = v_pixel_coords.to(torch.float32)
-        v_pixel_coords[:, 0] = v_pixel_coords[:, 0] * (1.0 / frame_rate)
-        av_cross_video_freq_cis = self._precompute_freqs_cis(
-            v_pixel_coords[:, 0:1, :],
-            dim=self.audio_cross_attention_dim,
-            out_dtype=x_dtype,
-            max_pos=[max_pos],
-            use_middle_indices_grid=True,
-            num_attention_heads=self.audio_num_attention_heads,
-        )
-        av_cross_audio_freq_cis = self._precompute_freqs_cis(
-            a_latent_coords[:, 0:1, :],
-            dim=self.audio_cross_attention_dim,
-            out_dtype=x_dtype,
-            max_pos=[max_pos],
-            use_middle_indices_grid=True,
-            num_attention_heads=self.audio_num_attention_heads,
-        )
-
-        return [(v_pe, av_cross_video_freq_cis), (a_pe, av_cross_audio_freq_cis)]
-
-    def _process_transformer_blocks(
-        self, x, context, attention_mask, timestep, pe, transformer_options={}, **kwargs
-    ):
-        vx = x[0]
-        ax = x[1]
-        v_context = context[0]
-        a_context = context[1]
-        v_timestep = timestep[0]
-        a_timestep = timestep[1]
-        v_pe, av_cross_video_freq_cis = pe[0]
-        a_pe, av_cross_audio_freq_cis = pe[1]
-
-        (
-            av_ca_audio_scale_shift_timestep,
-            av_ca_video_scale_shift_timestep,
-            av_ca_a2v_gate_noise_timestep,
-            av_ca_v2a_gate_noise_timestep,
-        ) = timestep[2]
-
-        """Process transformer blocks for LTXAV."""
-        patches_replace = transformer_options.get("patches_replace", {})
-        blocks_replace = patches_replace.get("dit", {})
-
-        # Process transformer blocks
-        for i, block in enumerate(self.transformer_blocks):
-            if ("double_block", i) in blocks_replace:
-
-                def block_wrap(args):
-                    out = {}
-                    out["img"] = block(
-                        args["img"],
-                        v_context=args["v_context"],
-                        a_context=args["a_context"],
-                        attention_mask=args["attention_mask"],
-                        v_timestep=args["v_timestep"],
-                        a_timestep=args["a_timestep"],
-                        v_pe=args["v_pe"],
-                        a_pe=args["a_pe"],
-                        v_cross_pe=args["v_cross_pe"],
-                        a_cross_pe=args["a_cross_pe"],
-                        v_cross_scale_shift_timestep=args["v_cross_scale_shift_timestep"],
-                        a_cross_scale_shift_timestep=args["a_cross_scale_shift_timestep"],
-                        v_cross_gate_timestep=args["v_cross_gate_timestep"],
-                        a_cross_gate_timestep=args["a_cross_gate_timestep"],
-                        transformer_options=args["transformer_options"],
-                    )
-                    return out
-
-                out = blocks_replace[("double_block", i)](
-                    {
-                        "img": (vx, ax),
-                        "v_context": v_context,
-                        "a_context": a_context,
-                        "attention_mask": attention_mask,
-                        "v_timestep": v_timestep,
-                        "a_timestep": a_timestep,
-                        "v_pe": v_pe,
-                        "a_pe": a_pe,
-                        "v_cross_pe": av_cross_video_freq_cis,
-                        "a_cross_pe": av_cross_audio_freq_cis,
-                        "v_cross_scale_shift_timestep": av_ca_video_scale_shift_timestep,
-                        "a_cross_scale_shift_timestep": av_ca_audio_scale_shift_timestep,
-                        "v_cross_gate_timestep": av_ca_a2v_gate_noise_timestep,
-                        "a_cross_gate_timestep": av_ca_v2a_gate_noise_timestep,
-                        "transformer_options": transformer_options,
-                    },
-                    {"original_block": block_wrap},
-                )
-                vx, ax = out["img"]
-            else:
-                vx, ax = block(
-                    (vx, ax),
-                    v_context=v_context,
-                    a_context=a_context,
-                    attention_mask=attention_mask,
-                    v_timestep=v_timestep,
-                    a_timestep=a_timestep,
-                    v_pe=v_pe,
-                    a_pe=a_pe,
-                    v_cross_pe=av_cross_video_freq_cis,
-                    a_cross_pe=av_cross_audio_freq_cis,
-                    v_cross_scale_shift_timestep=av_ca_video_scale_shift_timestep,
-                    a_cross_scale_shift_timestep=av_ca_audio_scale_shift_timestep,
-                    v_cross_gate_timestep=av_ca_a2v_gate_noise_timestep,
-                    a_cross_gate_timestep=av_ca_v2a_gate_noise_timestep,
-                    transformer_options=transformer_options,
-                )
-
-        return [vx, ax]
-
-    def _process_output(self, x, embedded_timestep, keyframe_idxs, **kwargs):
-        vx = x[0]
-        ax = x[1]
-        v_embedded_timestep = embedded_timestep[0]
-        a_embedded_timestep = embedded_timestep[1]
-        vx = super()._process_output(vx, v_embedded_timestep, keyframe_idxs, **kwargs)
-
-        # Process audio output
-        a_scale_shift_values = (
-            self.audio_scale_shift_table[None, None].to(device=a_embedded_timestep.device, dtype=a_embedded_timestep.dtype)
-            + a_embedded_timestep[:, :, None]
-        )
-        a_shift, a_scale = a_scale_shift_values[:, :, 0], a_scale_shift_values[:, :, 1]
-
-        ax = self.audio_norm_out(ax)
-        ax = ax * (1 + a_scale) + a_shift
-        ax = self.audio_proj_out(ax)
-
-        # Unpatchify audio
-        ax = self.a_patchifier.unpatchify(
-            ax, channels=self.num_audio_channels, freq=self.audio_frequency_bins
-        )
-
-        # Recombine audio and video
-        original_shape = kwargs.get("av_orig_shape")
-        return self.recombine_audio_and_video_latents(vx, ax, original_shape)
-
-    def forward(
-        self,
-        x,
-        timestep,
-        context,
-        attention_mask=None,
-        frame_rate=25,
-        transformer_options={},
-        keyframe_idxs=None,
-        **kwargs,
-    ):
-        """
-        Forward pass for LTXAV model.
-
-        Args:
-            x: Combined audio-video input tensor
-            timestep: Tuple of (video_timestep, audio_timestep) or single timestep
-            context: Context tensor (e.g., text embeddings)
-            attention_mask: Attention mask tensor
-            frame_rate: Frame rate for temporal processing
-            transformer_options: Additional options for transformer blocks
-            keyframe_idxs: Keyframe indices for temporal processing
-            **kwargs: Additional keyword arguments including audio_length
-
-        Returns:
-            Combined audio-video output tensor
-        """
-        # Handle timestep format
-        if isinstance(timestep, (tuple, list)) and len(timestep) == 2:
-            v_timestep, a_timestep = timestep
-            kwargs["a_timestep"] = a_timestep
-            timestep = v_timestep
-        else:
-            kwargs["a_timestep"] = timestep
-
-        # Call parent forward method
-        return super().forward(
-            x,
-            timestep,
-            context,
-            attention_mask,
-            frame_rate,
-            transformer_options,
-            keyframe_idxs,
-            **kwargs,
-        )

comfy/ldm/lightricks/embeddings_connector.py

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

comfy/ldm/lightricks/latent_upsampler.py

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

comfy/ldm/lightricks/model.py

@@ -1,47 +1,13 @@
-from abc import ABC, abstractmethod
-from enum import Enum
-import functools
-import math
-from typing import Dict, Optional, Tuple
-
-from einops import rearrange
-import numpy as np
 import torch
 from torch import nn
 import comfy.patcher_extension
 import comfy.ldm.modules.attention
 import comfy.ldm.common_dit
+import math
+from typing import Dict, Optional, Tuple
 
 from .symmetric_patchifier import SymmetricPatchifier, latent_to_pixel_coords
-
-def _log_base(x, base):
-    return np.log(x) / np.log(base)
-
-class LTXRopeType(str, Enum):
-    INTERLEAVED = "interleaved"
-    SPLIT = "split"
-
-    KEY = "rope_type"
-
-    @classmethod
-    def from_dict(cls, kwargs, default=None):
-        if default is None:
-            default = cls.INTERLEAVED
-        return cls(kwargs.get(cls.KEY, default))
-
-
-class LTXFrequenciesPrecision(str, Enum):
-    FLOAT32 = "float32"
-    FLOAT64 = "float64"
-
-    KEY = "frequencies_precision"
-
-    @classmethod
-    def from_dict(cls, kwargs, default=None):
-        if default is None:
-            default = cls.FLOAT32
-        return cls(kwargs.get(cls.KEY, default))
-
+from comfy.ldm.flux.math import apply_rope1
 
 def get_timestep_embedding(
     timesteps: torch.Tensor,
@@ -73,7 +39,9 @@ def get_timestep_embedding(
     assert len(timesteps.shape) == 1, "Timesteps should be a 1d-array"
 
     half_dim = embedding_dim // 2
-    exponent = -math.log(max_period) * torch.arange(start=0, end=half_dim, dtype=torch.float32, device=timesteps.device)
+    exponent = -math.log(max_period) * torch.arange(
+        start=0, end=half_dim, dtype=torch.float32, device=timesteps.device
+    )
     exponent = exponent / (half_dim - downscale_freq_shift)
 
     emb = torch.exp(exponent)
@@ -105,9 +73,7 @@ class TimestepEmbedding(nn.Module):
         post_act_fn: Optional[str] = None,
         cond_proj_dim=None,
         sample_proj_bias=True,
-        dtype=None,
-        device=None,
-        operations=None,
+        dtype=None, device=None, operations=None,
     ):
         super().__init__()
 
@@ -124,9 +90,7 @@ class TimestepEmbedding(nn.Module):
             time_embed_dim_out = out_dim
         else:
             time_embed_dim_out = time_embed_dim
-        self.linear_2 = operations.Linear(
-            time_embed_dim, time_embed_dim_out, sample_proj_bias, dtype=dtype, device=device
-        )
+        self.linear_2 = operations.Linear(time_embed_dim, time_embed_dim_out, sample_proj_bias, dtype=dtype, device=device)
 
         if post_act_fn is None:
             self.post_act = None
@@ -175,22 +139,12 @@ class PixArtAlphaCombinedTimestepSizeEmbeddings(nn.Module):
     https://github.com/PixArt-alpha/PixArt-alpha/blob/0f55e922376d8b797edd44d25d0e7464b260dcab/diffusion/model/nets/PixArtMS.py#L164C9-L168C29
     """
 
-    def __init__(
-        self,
-        embedding_dim,
-        size_emb_dim,
-        use_additional_conditions: bool = False,
-        dtype=None,
-        device=None,
-        operations=None,
-    ):
+    def __init__(self, embedding_dim, size_emb_dim, use_additional_conditions: bool = False, dtype=None, device=None, operations=None):
         super().__init__()
 
         self.outdim = size_emb_dim
         self.time_proj = Timesteps(num_channels=256, flip_sin_to_cos=True, downscale_freq_shift=0)
-        self.timestep_embedder = TimestepEmbedding(
-            in_channels=256, time_embed_dim=embedding_dim, dtype=dtype, device=device, operations=operations
-        )
+        self.timestep_embedder = TimestepEmbedding(in_channels=256, time_embed_dim=embedding_dim, dtype=dtype, device=device, operations=operations)
 
     def forward(self, timestep, resolution, aspect_ratio, batch_size, hidden_dtype):
         timesteps_proj = self.time_proj(timestep)
@@ -209,22 +163,15 @@ class AdaLayerNormSingle(nn.Module):
         use_additional_conditions (`bool`): To use additional conditions for normalization or not.
     """
 
-    def __init__(
-        self, embedding_dim: int, embedding_coefficient: int = 6, use_additional_conditions: bool = False, dtype=None, device=None, operations=None
-    ):
+    def __init__(self, embedding_dim: int, use_additional_conditions: bool = False, dtype=None, device=None, operations=None):
         super().__init__()
 
         self.emb = PixArtAlphaCombinedTimestepSizeEmbeddings(
-            embedding_dim,
-            size_emb_dim=embedding_dim // 3,
-            use_additional_conditions=use_additional_conditions,
-            dtype=dtype,
-            device=device,
-            operations=operations,
+            embedding_dim, size_emb_dim=embedding_dim // 3, use_additional_conditions=use_additional_conditions, dtype=dtype, device=device, operations=operations
         )
 
         self.silu = nn.SiLU()
-        self.linear = operations.Linear(embedding_dim, embedding_coefficient * embedding_dim, bias=True, dtype=dtype, device=device)
+        self.linear = operations.Linear(embedding_dim, 6 * embedding_dim, bias=True, dtype=dtype, device=device)
 
     def forward(
         self,
@@ -238,7 +185,6 @@ class AdaLayerNormSingle(nn.Module):
         embedded_timestep = self.emb(timestep, **added_cond_kwargs, batch_size=batch_size, hidden_dtype=hidden_dtype)
         return self.linear(self.silu(embedded_timestep)), embedded_timestep
 
-
 class PixArtAlphaTextProjection(nn.Module):
     """
     Projects caption embeddings. Also handles dropout for classifier-free guidance.
@@ -246,24 +192,18 @@ class PixArtAlphaTextProjection(nn.Module):
     Adapted from https://github.com/PixArt-alpha/PixArt-alpha/blob/master/diffusion/model/nets/PixArt_blocks.py
     """
 
-    def __init__(
-        self, in_features, hidden_size, out_features=None, act_fn="gelu_tanh", dtype=None, device=None, operations=None
-    ):
+    def __init__(self, in_features, hidden_size, out_features=None, act_fn="gelu_tanh", dtype=None, device=None, operations=None):
         super().__init__()
         if out_features is None:
             out_features = hidden_size
-        self.linear_1 = operations.Linear(
-            in_features=in_features, out_features=hidden_size, bias=True, dtype=dtype, device=device
-        )
+        self.linear_1 = operations.Linear(in_features=in_features, out_features=hidden_size, bias=True, dtype=dtype, device=device)
         if act_fn == "gelu_tanh":
             self.act_1 = nn.GELU(approximate="tanh")
         elif act_fn == "silu":
             self.act_1 = nn.SiLU()
         else:
             raise ValueError(f"Unknown activation function: {act_fn}")
-        self.linear_2 = operations.Linear(
-            in_features=hidden_size, out_features=out_features, bias=True, dtype=dtype, device=device
-        )
+        self.linear_2 = operations.Linear(in_features=hidden_size, out_features=out_features, bias=True, dtype=dtype, device=device)
 
     def forward(self, caption):
         hidden_states = self.linear_1(caption)
@@ -282,68 +222,23 @@ class GELU_approx(nn.Module):
 
 
 class FeedForward(nn.Module):
-    def __init__(self, dim, dim_out, mult=4, glu=False, dropout=0.0, dtype=None, device=None, operations=None):
+    def __init__(self, dim, dim_out, mult=4, glu=False, dropout=0., dtype=None, device=None, operations=None):
         super().__init__()
         inner_dim = int(dim * mult)
         project_in = GELU_approx(dim, inner_dim, dtype=dtype, device=device, operations=operations)
 
         self.net = nn.Sequential(
-            project_in, nn.Dropout(dropout), operations.Linear(inner_dim, dim_out, dtype=dtype, device=device)
+            project_in,
+            nn.Dropout(dropout),
+            operations.Linear(inner_dim, dim_out, dtype=dtype, device=device)
         )
 
     def forward(self, x):
         return self.net(x)
 
-def apply_rotary_emb(input_tensor, freqs_cis):
-    cos_freqs, sin_freqs = freqs_cis[0], freqs_cis[1]
-    split_pe = freqs_cis[2] if len(freqs_cis) > 2 else False
-    return (
-        apply_split_rotary_emb(input_tensor, cos_freqs, sin_freqs)
-        if split_pe else
-        apply_interleaved_rotary_emb(input_tensor, cos_freqs, sin_freqs)
-    )
-
-def apply_interleaved_rotary_emb(input_tensor, cos_freqs, sin_freqs):  # TODO: remove duplicate funcs and pick the best/fastest one
-    t_dup = rearrange(input_tensor, "... (d r) -> ... d r", r=2)
-    t1, t2 = t_dup.unbind(dim=-1)
-    t_dup = torch.stack((-t2, t1), dim=-1)
-    input_tensor_rot = rearrange(t_dup, "... d r -> ... (d r)")
-
-    out = input_tensor * cos_freqs + input_tensor_rot * sin_freqs
-
-    return out
-
-def apply_split_rotary_emb(input_tensor, cos, sin):
-    needs_reshape = False
-    if input_tensor.ndim != 4 and cos.ndim == 4:
-        B, H, T, _ = cos.shape
-        input_tensor = input_tensor.reshape(B, T, H, -1).swapaxes(1, 2)
-        needs_reshape = True
-    split_input = rearrange(input_tensor, "... (d r) -> ... d r", d=2)
-    first_half_input = split_input[..., :1, :]
-    second_half_input = split_input[..., 1:, :]
-    output = split_input * cos.unsqueeze(-2)
-    first_half_output = output[..., :1, :]
-    second_half_output = output[..., 1:, :]
-    first_half_output.addcmul_(-sin.unsqueeze(-2), second_half_input)
-    second_half_output.addcmul_(sin.unsqueeze(-2), first_half_input)
-    output = rearrange(output, "... d r -> ... (d r)")
-    return output.swapaxes(1, 2).reshape(B, T, -1) if needs_reshape else output
-
 
 class CrossAttention(nn.Module):
-    def __init__(
-        self,
-        query_dim,
-        context_dim=None,
-        heads=8,
-        dim_head=64,
-        dropout=0.0,
-        attn_precision=None,
-        dtype=None,
-        device=None,
-        operations=None,
-    ):
+    def __init__(self, query_dim, context_dim=None, heads=8, dim_head=64, dropout=0., attn_precision=None, dtype=None, device=None, operations=None):
         super().__init__()
         inner_dim = dim_head * heads
         context_dim = query_dim if context_dim is None else context_dim
@@ -359,11 +254,9 @@ class CrossAttention(nn.Module):
         self.to_k = operations.Linear(context_dim, inner_dim, bias=True, dtype=dtype, device=device)
         self.to_v = operations.Linear(context_dim, inner_dim, bias=True, dtype=dtype, device=device)
 
-        self.to_out = nn.Sequential(
-            operations.Linear(inner_dim, query_dim, dtype=dtype, device=device), nn.Dropout(dropout)
-        )
+        self.to_out = nn.Sequential(operations.Linear(inner_dim, query_dim, dtype=dtype, device=device), nn.Dropout(dropout))
 
-    def forward(self, x, context=None, mask=None, pe=None, k_pe=None, transformer_options={}):
+    def forward(self, x, context=None, mask=None, pe=None, transformer_options={}):
         q = self.to_q(x)
         context = x if context is None else context
         k = self.to_k(context)
@@ -373,8 +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 if k_pe is None else 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)
@@ -384,34 +277,14 @@ class CrossAttention(nn.Module):
 
 
 class BasicTransformerBlock(nn.Module):
-    def __init__(
-        self, dim, n_heads, d_head, context_dim=None, attn_precision=None, dtype=None, device=None, operations=None
-    ):
+    def __init__(self, dim, n_heads, d_head, context_dim=None, attn_precision=None, dtype=None, device=None, operations=None):
         super().__init__()
 
         self.attn_precision = attn_precision
-        self.attn1 = CrossAttention(
-            query_dim=dim,
-            heads=n_heads,
-            dim_head=d_head,
-            context_dim=None,
-            attn_precision=self.attn_precision,
-            dtype=dtype,
-            device=device,
-            operations=operations,
-        )
+        self.attn1 = CrossAttention(query_dim=dim, heads=n_heads, dim_head=d_head, context_dim=None, attn_precision=self.attn_precision, dtype=dtype, device=device, operations=operations)
         self.ff = FeedForward(dim, dim_out=dim, glu=True, dtype=dtype, device=device, operations=operations)
 
-        self.attn2 = CrossAttention(
-            query_dim=dim,
-            context_dim=context_dim,
-            heads=n_heads,
-            dim_head=d_head,
-            attn_precision=self.attn_precision,
-            dtype=dtype,
-            device=device,
-            operations=operations,
-        )
+        self.attn2 = CrossAttention(query_dim=dim, context_dim=context_dim, heads=n_heads, dim_head=d_head, attn_precision=self.attn_precision, dtype=dtype, device=device, operations=operations)
 
         self.scale_shift_table = nn.Parameter(torch.empty(6, dim, device=device, dtype=dtype))
 
@@ -433,446 +306,116 @@ class BasicTransformerBlock(nn.Module):
         return x
 
 def get_fractional_positions(indices_grid, max_pos):
-    n_pos_dims = indices_grid.shape[1]
-    assert n_pos_dims == len(max_pos), f'Number of position dimensions ({n_pos_dims}) must match max_pos length ({len(max_pos)})'
     fractional_positions = torch.stack(
-        [indices_grid[:, i] / max_pos[i] for i in range(n_pos_dims)],
-        axis=-1,
+        [
+            indices_grid[:, i] / max_pos[i]
+            for i in range(3)
+        ],
+        dim=-1,
     )
     return fractional_positions
 
 
-@functools.lru_cache(maxsize=5)
-def generate_freq_grid_np(positional_embedding_theta, positional_embedding_max_pos_count, inner_dim, _ = None):
-    theta = positional_embedding_theta
-    start = 1
-    end = theta
-
-    n_elem = 2 * positional_embedding_max_pos_count
-    pow_indices = np.power(
-        theta,
-        np.linspace(
-            _log_base(start, theta),
-            _log_base(end, theta),
-            inner_dim // n_elem,
-            dtype=np.float64,
-        ),
-    )
-    return torch.tensor(pow_indices * math.pi / 2, dtype=torch.float32)
-
-def generate_freq_grid_pytorch(positional_embedding_theta, positional_embedding_max_pos_count, inner_dim, device):
-    theta = positional_embedding_theta
-    start = 1
-    end = theta
-    n_elem = 2 * positional_embedding_max_pos_count
-
-    indices = theta ** (
-        torch.linspace(
-            math.log(start, theta),
-            math.log(end, theta),
-            inner_dim // n_elem,
-            device=device,
-            dtype=torch.float32,
-        )
-    )
-    indices = indices.to(dtype=torch.float32)
-
-    indices = indices * math.pi / 2
-
-    return indices
-
-def generate_freqs(indices, indices_grid, max_pos, use_middle_indices_grid):
-    if use_middle_indices_grid:
-        assert(len(indices_grid.shape) == 4 and indices_grid.shape[-1] ==2)
-        indices_grid_start, indices_grid_end = indices_grid[..., 0], indices_grid[..., 1]
-        indices_grid = (indices_grid_start + indices_grid_end) / 2.0
-    elif len(indices_grid.shape) == 4:
-        indices_grid = indices_grid[..., 0]
+def precompute_freqs_cis(indices_grid, dim, out_dtype, theta=10000.0, max_pos=[20, 2048, 2048]):
+    dtype = torch.float32
+    device = indices_grid.device
 
     # Get fractional positions and compute frequency indices
     fractional_positions = get_fractional_positions(indices_grid, max_pos)
-    indices = indices.to(device=fractional_positions.device)
+    indices = theta ** torch.linspace(0, 1, dim // 6, device=device, dtype=dtype) * math.pi / 2
 
-    freqs = (
-        (indices * (fractional_positions.unsqueeze(-1) * 2 - 1))
-        .transpose(-1, -2)
-        .flatten(2)
-    )
-    return freqs
+    # 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)
 
-def interleaved_freqs_cis(freqs, pad_size):
-    cos_freq = freqs.cos().repeat_interleave(2, dim=-1)
-    sin_freq = freqs.sin().repeat_interleave(2, dim=-1)
-    if pad_size != 0:
-        cos_padding = torch.ones_like(cos_freq[:, :, : pad_size])
-        sin_padding = torch.zeros_like(cos_freq[:, :, : pad_size])
-        cos_freq = torch.cat([cos_padding, cos_freq], dim=-1)
-        sin_freq = torch.cat([sin_padding, sin_freq], dim=-1)
-    return cos_freq, sin_freq
+    # Pad if dim is not divisible by 6
+    if dim % 6 != 0:
+        padding_size = dim % 6
+        cos_vals = torch.cat([torch.ones_like(cos_vals[:, :, :padding_size]), cos_vals], dim=-1)
+        sin_vals = torch.cat([torch.zeros_like(sin_vals[:, :, :padding_size]), sin_vals], dim=-1)
 
-def split_freqs_cis(freqs, pad_size, num_attention_heads):
-    cos_freq = freqs.cos()
-    sin_freq = freqs.sin()
+    # 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]
 
-    if pad_size != 0:
-        cos_padding = torch.ones_like(cos_freq[:, :, :pad_size])
-        sin_padding = torch.zeros_like(sin_freq[:, :, :pad_size])
+    # 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]
 
-        cos_freq = torch.concatenate([cos_padding, cos_freq], axis=-1)
-        sin_freq = torch.concatenate([sin_padding, sin_freq], axis=-1)
+    return freqs_cis
 
-    # Reshape freqs to be compatible with multi-head attention
-    B , T, half_HD = cos_freq.shape
 
-    cos_freq = cos_freq.reshape(B, T, num_attention_heads, half_HD // num_attention_heads)
-    sin_freq = sin_freq.reshape(B, T, num_attention_heads, half_HD // num_attention_heads)
+class LTXVModel(torch.nn.Module):
+    def __init__(self,
+                 in_channels=128,
+                 cross_attention_dim=2048,
+                 attention_head_dim=64,
+                 num_attention_heads=32,
 
-    cos_freq = torch.swapaxes(cos_freq, 1, 2)  # (B,H,T,D//2)
-    sin_freq = torch.swapaxes(sin_freq, 1, 2)  # (B,H,T,D//2)
-    return cos_freq, sin_freq
+                 caption_channels=4096,
+                 num_layers=28,
 
-class LTXBaseModel(torch.nn.Module, ABC):
-    """
-    Abstract base class for LTX models (Lightricks Transformer models).
 
-    This class defines the common interface and shared functionality for all LTX models,
-    including LTXV (video) and LTXAV (audio-video) variants.
-    """
-
-    def __init__(
-        self,
-        in_channels: int,
-        cross_attention_dim: int,
-        attention_head_dim: int,
-        num_attention_heads: int,
-        caption_channels: int,
-        num_layers: int,
-        positional_embedding_theta: float = 10000.0,
-        positional_embedding_max_pos: list = [20, 2048, 2048],
-        causal_temporal_positioning: bool = False,
-        vae_scale_factors: tuple = (8, 32, 32),
-        use_middle_indices_grid=False,
-        timestep_scale_multiplier = 1000.0,
-        dtype=None,
-        device=None,
-        operations=None,
-        **kwargs,
-    ):
+                 positional_embedding_theta=10000.0,
+                 positional_embedding_max_pos=[20, 2048, 2048],
+                 causal_temporal_positioning=False,
+                 vae_scale_factors=(8, 32, 32),
+                 dtype=None, device=None, operations=None, **kwargs):
         super().__init__()
         self.generator = None
         self.vae_scale_factors = vae_scale_factors
-        self.use_middle_indices_grid = use_middle_indices_grid
         self.dtype = dtype
-        self.in_channels = in_channels
-        self.cross_attention_dim = cross_attention_dim
-        self.attention_head_dim = attention_head_dim
-        self.num_attention_heads = num_attention_heads
-        self.caption_channels = caption_channels
-        self.num_layers = num_layers
-        self.positional_embedding_theta = positional_embedding_theta
-        self.positional_embedding_max_pos = positional_embedding_max_pos
-        self.split_positional_embedding = LTXRopeType.from_dict(kwargs)
-        self.freq_grid_generator = (
-            generate_freq_grid_np if LTXFrequenciesPrecision.from_dict(kwargs) == LTXFrequenciesPrecision.FLOAT64
-            else generate_freq_grid_pytorch
-        )
-        self.causal_temporal_positioning = causal_temporal_positioning
-        self.operations = operations
-        self.timestep_scale_multiplier = timestep_scale_multiplier
-
-        # Common dimensions
-        self.inner_dim = num_attention_heads * attention_head_dim
         self.out_channels = in_channels
+        self.inner_dim = num_attention_heads * attention_head_dim
+        self.causal_temporal_positioning = causal_temporal_positioning
 
-        # Initialize common components
-        self._init_common_components(device, dtype)
-
-        # Initialize model-specific components
-        self._init_model_components(device, dtype, **kwargs)
-
-        # Initialize transformer blocks
-        self._init_transformer_blocks(device, dtype, **kwargs)
-
-        # Initialize output components
-        self._init_output_components(device, dtype)
-
-    def _init_common_components(self, device, dtype):
-        """Initialize components common to all LTX models
-        - patchify_proj: Linear projection for patchifying input
-        - adaln_single: AdaLN layer for timestep embedding
-        - caption_projection: Linear projection for caption embedding
-        """
-        self.patchify_proj = self.operations.Linear(
-            self.in_channels, self.inner_dim, bias=True, dtype=dtype, device=device
-        )
+        self.patchify_proj = operations.Linear(in_channels, self.inner_dim, bias=True, dtype=dtype, device=device)
 
         self.adaln_single = AdaLayerNormSingle(
-            self.inner_dim, use_additional_conditions=False, dtype=dtype, device=device, operations=self.operations
+            self.inner_dim, use_additional_conditions=False, dtype=dtype, device=device, operations=operations
         )
 
+        # self.adaln_single.linear = operations.Linear(self.inner_dim, 4 * self.inner_dim, bias=True, dtype=dtype, device=device)
+
         self.caption_projection = PixArtAlphaTextProjection(
-            in_features=self.caption_channels,
-            hidden_size=self.inner_dim,
-            dtype=dtype,
-            device=device,
-            operations=self.operations,
+            in_features=caption_channels, hidden_size=self.inner_dim, dtype=dtype, device=device, operations=operations
         )
 
-    @abstractmethod
-    def _init_model_components(self, device, dtype, **kwargs):
-        """Initialize model-specific components. Must be implemented by subclasses."""
-        pass
-
-    @abstractmethod
-    def _init_transformer_blocks(self, device, dtype, **kwargs):
-        """Initialize transformer blocks. Must be implemented by subclasses."""
-        pass
-
-    @abstractmethod
-    def _init_output_components(self, device, dtype):
-        """Initialize output components. Must be implemented by subclasses."""
-        pass
-
-    @abstractmethod
-    def _process_input(self, x, keyframe_idxs, denoise_mask, **kwargs):
-        """Process input data. Must be implemented by subclasses."""
-        pass
-
-    @abstractmethod
-    def _process_transformer_blocks(self, x, context, attention_mask, timestep, pe, **kwargs):
-        """Process transformer blocks. Must be implemented by subclasses."""
-        pass
-
-    @abstractmethod
-    def _process_output(self, x, embedded_timestep, keyframe_idxs, **kwargs):
-        """Process output data. Must be implemented by subclasses."""
-        pass
-
-    def _prepare_timestep(self, timestep, batch_size, hidden_dtype, **kwargs):
-        """Prepare timestep embeddings."""
-        grid_mask = kwargs.get("grid_mask", None)
-        if grid_mask is not None:
-            timestep = timestep[:, grid_mask]
-
-        timestep = timestep * self.timestep_scale_multiplier
-        timestep, embedded_timestep = self.adaln_single(
-            timestep.flatten(),
-            {"resolution": None, "aspect_ratio": None},
-            batch_size=batch_size,
-            hidden_dtype=hidden_dtype,
-        )
-
-        # Second dimension is 1 or number of tokens (if timestep_per_token)
-        timestep = timestep.view(batch_size, -1, timestep.shape[-1])
-        embedded_timestep = embedded_timestep.view(batch_size, -1, embedded_timestep.shape[-1])
-
-        return timestep, embedded_timestep
-
-    def _prepare_context(self, context, batch_size, x, attention_mask=None):
-        """Prepare context for transformer blocks."""
-        if self.caption_projection is not None:
-            context = self.caption_projection(context)
-            context = context.view(batch_size, -1, x.shape[-1])
-
-        return context, attention_mask
-
-    def _precompute_freqs_cis(
-        self,
-        indices_grid,
-        dim,
-        out_dtype,
-        theta=10000.0,
-        max_pos=[20, 2048, 2048],
-        use_middle_indices_grid=False,
-        num_attention_heads=32,
-    ):
-        split_mode = self.split_positional_embedding == LTXRopeType.SPLIT
-        indices = self.freq_grid_generator(theta, indices_grid.shape[1], dim, indices_grid.device)
-        freqs = generate_freqs(indices, indices_grid, max_pos, use_middle_indices_grid)
-
-        if split_mode:
-            expected_freqs = dim // 2
-            current_freqs = freqs.shape[-1]
-            pad_size = expected_freqs - current_freqs
-            cos_freq, sin_freq = split_freqs_cis(freqs, pad_size, num_attention_heads)
-        else:
-            # 2 because of cos and sin by 3 for (t, x, y), 1 for temporal only
-            n_elem = 2 * indices_grid.shape[1]
-            cos_freq, sin_freq = interleaved_freqs_cis(freqs, dim % n_elem)
-        return cos_freq.to(out_dtype), sin_freq.to(out_dtype), split_mode
-
-    def _prepare_positional_embeddings(self, pixel_coords, frame_rate, x_dtype):
-        """Prepare positional embeddings."""
-        fractional_coords = pixel_coords.to(torch.float32)
-        fractional_coords[:, 0] = fractional_coords[:, 0] * (1.0 / frame_rate)
-        pe = self._precompute_freqs_cis(
-            fractional_coords,
-            dim=self.inner_dim,
-            out_dtype=x_dtype,
-            max_pos=self.positional_embedding_max_pos,
-            use_middle_indices_grid=self.use_middle_indices_grid,
-            num_attention_heads=self.num_attention_heads,
-        )
-        return pe
-
-    def _prepare_attention_mask(self, attention_mask, x_dtype):
-        """Prepare attention mask."""
-        if attention_mask is not None and not torch.is_floating_point(attention_mask):
-            attention_mask = (attention_mask - 1).to(x_dtype).reshape(
-                (attention_mask.shape[0], 1, -1, attention_mask.shape[-1])
-            ) * torch.finfo(x_dtype).max
-        return attention_mask
-
-    def forward(
-        self, x, timestep, context, attention_mask, frame_rate=25, transformer_options={}, keyframe_idxs=None, denoise_mask=None, **kwargs
-    ):
-        """
-        Forward pass for LTX models.
-
-        Args:
-            x: Input tensor
-            timestep: Timestep tensor
-            context: Context tensor (e.g., text embeddings)
-            attention_mask: Attention mask tensor
-            frame_rate: Frame rate for temporal processing
-            transformer_options: Additional options for transformer blocks
-            keyframe_idxs: Keyframe indices for temporal processing
-            **kwargs: Additional keyword arguments
-
-        Returns:
-            Processed output tensor
-        """
-        return comfy.patcher_extension.WrapperExecutor.new_class_executor(
-            self._forward,
-            self,
-            comfy.patcher_extension.get_all_wrappers(
-                comfy.patcher_extension.WrappersMP.DIFFUSION_MODEL, transformer_options
-            ),
-        ).execute(x, timestep, context, attention_mask, frame_rate, transformer_options, keyframe_idxs, denoise_mask=denoise_mask, **kwargs)
-
-    def _forward(
-        self, x, timestep, context, attention_mask, frame_rate=25, transformer_options={}, keyframe_idxs=None, denoise_mask=None, **kwargs
-    ):
-        """
-        Internal forward pass for LTX models.
-
-        Args:
-            x: Input tensor
-            timestep: Timestep tensor
-            context: Context tensor (e.g., text embeddings)
-            attention_mask: Attention mask tensor
-            frame_rate: Frame rate for temporal processing
-            transformer_options: Additional options for transformer blocks
-            keyframe_idxs: Keyframe indices for temporal processing
-            **kwargs: Additional keyword arguments
-
-        Returns:
-            Processed output tensor
-        """
-        if isinstance(x, list):
-            input_dtype = x[0].dtype
-            batch_size = x[0].shape[0]
-        else:
-            input_dtype = x.dtype
-            batch_size = x.shape[0]
-        # Process input
-        merged_args = {**transformer_options, **kwargs}
-        x, pixel_coords, additional_args = self._process_input(x, keyframe_idxs, denoise_mask, **merged_args)
-        merged_args.update(additional_args)
-
-        # Prepare timestep and context
-        timestep, embedded_timestep = self._prepare_timestep(timestep, batch_size, input_dtype, **merged_args)
-        context, attention_mask = self._prepare_context(context, batch_size, x, attention_mask)
-
-        # Prepare attention mask and positional embeddings
-        attention_mask = self._prepare_attention_mask(attention_mask, input_dtype)
-        pe = self._prepare_positional_embeddings(pixel_coords, frame_rate, input_dtype)
-
-        # Process transformer blocks
-        x = self._process_transformer_blocks(
-            x, context, attention_mask, timestep, pe, transformer_options=transformer_options, **merged_args
-        )
-
-        # Process output
-        x = self._process_output(x, embedded_timestep, keyframe_idxs, **merged_args)
-        return x
-
-
-class LTXVModel(LTXBaseModel):
-    """LTXV model for video generation."""
-
-    def __init__(
-        self,
-        in_channels=128,
-        cross_attention_dim=2048,
-        attention_head_dim=64,
-        num_attention_heads=32,
-        caption_channels=4096,
-        num_layers=28,
-        positional_embedding_theta=10000.0,
-        positional_embedding_max_pos=[20, 2048, 2048],
-        causal_temporal_positioning=False,
-        vae_scale_factors=(8, 32, 32),
-        use_middle_indices_grid=False,
-        timestep_scale_multiplier = 1000.0,
-        dtype=None,
-        device=None,
-        operations=None,
-        **kwargs,
-    ):
-        super().__init__(
-            in_channels=in_channels,
-            cross_attention_dim=cross_attention_dim,
-            attention_head_dim=attention_head_dim,
-            num_attention_heads=num_attention_heads,
-            caption_channels=caption_channels,
-            num_layers=num_layers,
-            positional_embedding_theta=positional_embedding_theta,
-            positional_embedding_max_pos=positional_embedding_max_pos,
-            causal_temporal_positioning=causal_temporal_positioning,
-            vae_scale_factors=vae_scale_factors,
-            use_middle_indices_grid=use_middle_indices_grid,
-            timestep_scale_multiplier=timestep_scale_multiplier,
-            dtype=dtype,
-            device=device,
-            operations=operations,
-            **kwargs,
-        )
-
-    def _init_model_components(self, device, dtype, **kwargs):
-        """Initialize LTXV-specific components."""
-        # No additional components needed for LTXV beyond base class
-        pass
-
-    def _init_transformer_blocks(self, device, dtype, **kwargs):
-        """Initialize transformer blocks for LTXV."""
         self.transformer_blocks = nn.ModuleList(
             [
                 BasicTransformerBlock(
                     self.inner_dim,
-                    self.num_attention_heads,
-                    self.attention_head_dim,
-                    context_dim=self.cross_attention_dim,
-                    dtype=dtype,
-                    device=device,
-                    operations=self.operations,
+                    num_attention_heads,
+                    attention_head_dim,
+                    context_dim=cross_attention_dim,
+                    # attn_precision=attn_precision,
+                    dtype=dtype, device=device, operations=operations
                 )
-                for _ in range(self.num_layers)
+                for d in range(num_layers)
             ]
         )
 
-    def _init_output_components(self, device, dtype):
-        """Initialize output components for LTXV."""
         self.scale_shift_table = nn.Parameter(torch.empty(2, self.inner_dim, dtype=dtype, device=device))
-        self.norm_out = self.operations.LayerNorm(
-            self.inner_dim, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device
-        )
-        self.proj_out = self.operations.Linear(self.inner_dim, self.out_channels, dtype=dtype, device=device)
-        self.patchifier = SymmetricPatchifier(1, start_end=True)
+        self.norm_out = operations.LayerNorm(self.inner_dim, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
+        self.proj_out = operations.Linear(self.inner_dim, self.out_channels, dtype=dtype, device=device)
+
+        self.patchifier = SymmetricPatchifier(1)
+
+    def forward(self, x, timestep, context, attention_mask, frame_rate=25, transformer_options={}, keyframe_idxs=None, **kwargs):
+        return comfy.patcher_extension.WrapperExecutor.new_class_executor(
+            self._forward,
+            self,
+            comfy.patcher_extension.get_all_wrappers(comfy.patcher_extension.WrappersMP.DIFFUSION_MODEL, transformer_options)
+        ).execute(x, timestep, context, attention_mask, frame_rate, transformer_options, keyframe_idxs, **kwargs)
+
+    def _forward(self, x, timestep, context, attention_mask, frame_rate=25, transformer_options={}, keyframe_idxs=None, **kwargs):
+        patches_replace = transformer_options.get("patches_replace", {})
+
+        orig_shape = list(x.shape)
 
-    def _process_input(self, x, keyframe_idxs, denoise_mask, **kwargs):
-        """Process input for LTXV."""
-        additional_args = {"orig_shape": list(x.shape)}
         x, latent_coords = self.patchifier.patchify(x)
         pixel_coords = latent_to_pixel_coords(
             latent_coords=latent_coords,
@@ -880,30 +423,44 @@ class LTXVModel(LTXBaseModel):
             causal_fix=self.causal_temporal_positioning,
         )
 
-        grid_mask = None
         if keyframe_idxs is not None:
-            additional_args.update({ "orig_patchified_shape": list(x.shape)})
-            denoise_mask = self.patchifier.patchify(denoise_mask)[0]
-            grid_mask = ~torch.any(denoise_mask < 0, dim=-1)[0]
-            additional_args.update({"grid_mask": grid_mask})
-            x = x[:, grid_mask, :]
-            pixel_coords = pixel_coords[:, :, grid_mask, ...]
+            pixel_coords[:, :, -keyframe_idxs.shape[2]:] = keyframe_idxs
 
-            kf_grid_mask = grid_mask[-keyframe_idxs.shape[2]:]
-            keyframe_idxs = keyframe_idxs[..., kf_grid_mask, :]
-            pixel_coords[:, :, -keyframe_idxs.shape[2]:, :] = keyframe_idxs
+        fractional_coords = pixel_coords.to(torch.float32)
+        fractional_coords[:, 0] = fractional_coords[:, 0] * (1.0 / frame_rate)
 
         x = self.patchify_proj(x)
-        return x, pixel_coords, additional_args
+        timestep = timestep * 1000.0
+
+        if attention_mask is not None and not torch.is_floating_point(attention_mask):
+            attention_mask = (attention_mask - 1).to(x.dtype).reshape((attention_mask.shape[0], 1, -1, attention_mask.shape[-1])) * torch.finfo(x.dtype).max
+
+        pe = precompute_freqs_cis(fractional_coords, dim=self.inner_dim, out_dtype=x.dtype)
+
+        batch_size = x.shape[0]
+        timestep, embedded_timestep = self.adaln_single(
+            timestep.flatten(),
+            {"resolution": None, "aspect_ratio": None},
+            batch_size=batch_size,
+            hidden_dtype=x.dtype,
+        )
+        # Second dimension is 1 or number of tokens (if timestep_per_token)
+        timestep = timestep.view(batch_size, -1, timestep.shape[-1])
+        embedded_timestep = embedded_timestep.view(
+            batch_size, -1, embedded_timestep.shape[-1]
+        )
+
+        # 2. Blocks
+        if self.caption_projection is not None:
+            batch_size = x.shape[0]
+            context = self.caption_projection(context)
+            context = context.view(
+                batch_size, -1, x.shape[-1]
+            )
 
-    def _process_transformer_blocks(self, x, context, attention_mask, timestep, pe, transformer_options={}, **kwargs):
-        """Process transformer blocks for LTXV."""
-        patches_replace = transformer_options.get("patches_replace", {})
         blocks_replace = patches_replace.get("dit", {})
-
         for i, block in enumerate(self.transformer_blocks):
             if ("double_block", i) in blocks_replace:
-
                 def block_wrap(args):
                     out = {}
                     out["img"] = block(args["img"], context=args["txt"], attention_mask=args["attention_mask"], timestep=args["vec"], pe=args["pe"], transformer_options=args["transformer_options"])
@@ -921,28 +478,16 @@ class LTXVModel(LTXBaseModel):
                     transformer_options=transformer_options,
                 )
 
-        return x
-
-    def _process_output(self, x, embedded_timestep, keyframe_idxs, **kwargs):
-        """Process output for LTXV."""
-        # Apply scale-shift modulation
+        # 3. Output
         scale_shift_values = (
             self.scale_shift_table[None, None].to(device=x.device, dtype=x.dtype) + embedded_timestep[:, :, None]
         )
         shift, scale = scale_shift_values[:, :, 0], scale_shift_values[:, :, 1]
-
         x = self.norm_out(x)
-        x = x * (1 + scale) + shift
+        # Modulation
+        x = torch.addcmul(x, x, scale).add_(shift)
         x = self.proj_out(x)
 
-        if keyframe_idxs is not None:
-            grid_mask = kwargs["grid_mask"]
-            orig_patchified_shape = kwargs["orig_patchified_shape"]
-            full_x = torch.zeros(orig_patchified_shape, dtype=x.dtype, device=x.device)
-            full_x[:, grid_mask, :] = x
-            x = full_x
-        # Unpatchify to restore original dimensions
-        orig_shape = kwargs["orig_shape"]
         x = self.patchifier.unpatchify(
             latents=x,
             output_height=orig_shape[3],

comfy/ldm/lightricks/symmetric_patchifier.py

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

comfy/ldm/lightricks/vae/audio_vae.py

@@ -1,286 +0,0 @@
-import json
-from dataclasses import dataclass
-import math
-import torch
-import torchaudio
-
-import comfy.model_management
-import comfy.model_patcher
-import comfy.utils as utils
-from comfy.ldm.mmaudio.vae.distributions import DiagonalGaussianDistribution
-from comfy.ldm.lightricks.symmetric_patchifier import AudioPatchifier
-from comfy.ldm.lightricks.vae.causal_audio_autoencoder import (
-    CausalityAxis,
-    CausalAudioAutoencoder,
-)
-from comfy.ldm.lightricks.vocoders.vocoder import Vocoder
-
-LATENT_DOWNSAMPLE_FACTOR = 4
-
-
-@dataclass(frozen=True)
-class AudioVAEComponentConfig:
-    """Container for model component configuration extracted from metadata."""
-
-    autoencoder: dict
-    vocoder: dict
-
-    @classmethod
-    def from_metadata(cls, metadata: dict) -> "AudioVAEComponentConfig":
-        assert metadata is not None and "config" in metadata, "Metadata is required for audio VAE"
-
-        raw_config = metadata["config"]
-        if isinstance(raw_config, str):
-            parsed_config = json.loads(raw_config)
-        else:
-            parsed_config = raw_config
-
-        audio_config = parsed_config.get("audio_vae")
-        vocoder_config = parsed_config.get("vocoder")
-
-        assert audio_config is not None, "Audio VAE config is required for audio VAE"
-        assert vocoder_config is not None, "Vocoder config is required for audio VAE"
-
-        return cls(autoencoder=audio_config, vocoder=vocoder_config)
-
-
-class ModelDeviceManager:
-    """Manages device placement and GPU residency for the composed model."""
-
-    def __init__(self, module: torch.nn.Module):
-        load_device = comfy.model_management.get_torch_device()
-        offload_device = comfy.model_management.vae_offload_device()
-        self.patcher = comfy.model_patcher.ModelPatcher(module, load_device, offload_device)
-
-    def ensure_model_loaded(self) -> None:
-        comfy.model_management.free_memory(
-            self.patcher.model_size(),
-            self.patcher.load_device,
-        )
-        comfy.model_management.load_model_gpu(self.patcher)
-
-    def move_to_load_device(self, tensor: torch.Tensor) -> torch.Tensor:
-        return tensor.to(self.patcher.load_device)
-
-    @property
-    def load_device(self):
-        return self.patcher.load_device
-
-
-class AudioLatentNormalizer:
-    """Applies per-channel statistics in patch space and restores original layout."""
-
-    def __init__(self, patchfier: AudioPatchifier, statistics_processor: torch.nn.Module):
-        self.patchifier = patchfier
-        self.statistics = statistics_processor
-
-    def normalize(self, latents: torch.Tensor) -> torch.Tensor:
-        channels = latents.shape[1]
-        freq = latents.shape[3]
-        patched, _ = self.patchifier.patchify(latents)
-        normalized = self.statistics.normalize(patched)
-        return self.patchifier.unpatchify(normalized, channels=channels, freq=freq)
-
-    def denormalize(self, latents: torch.Tensor) -> torch.Tensor:
-        channels = latents.shape[1]
-        freq = latents.shape[3]
-        patched, _ = self.patchifier.patchify(latents)
-        denormalized = self.statistics.un_normalize(patched)
-        return self.patchifier.unpatchify(denormalized, channels=channels, freq=freq)
-
-
-class AudioPreprocessor:
-    """Prepares raw waveforms for the autoencoder by matching training conditions."""
-
-    def __init__(self, target_sample_rate: int, mel_bins: int, mel_hop_length: int, n_fft: int):
-        self.target_sample_rate = target_sample_rate
-        self.mel_bins = mel_bins
-        self.mel_hop_length = mel_hop_length
-        self.n_fft = n_fft
-
-    def resample(self, waveform: torch.Tensor, source_rate: int) -> torch.Tensor:
-        if source_rate == self.target_sample_rate:
-            return waveform
-        return torchaudio.functional.resample(waveform, source_rate, self.target_sample_rate)
-
-    @staticmethod
-    def normalize_amplitude(
-        waveform: torch.Tensor, max_amplitude: float = 0.5, eps: float = 1e-5
-    ) -> torch.Tensor:
-        waveform = waveform - waveform.mean(dim=2, keepdim=True)
-        peak = torch.max(torch.abs(waveform)) + eps
-        scale = peak.clamp(max=max_amplitude) / peak
-        return waveform * scale
-
-    def waveform_to_mel(
-        self, waveform: torch.Tensor, waveform_sample_rate: int, device
-    ) -> torch.Tensor:
-        waveform = self.resample(waveform, waveform_sample_rate)
-        waveform = self.normalize_amplitude(waveform)
-
-        mel_transform = torchaudio.transforms.MelSpectrogram(
-            sample_rate=self.target_sample_rate,
-            n_fft=self.n_fft,
-            win_length=self.n_fft,
-            hop_length=self.mel_hop_length,
-            f_min=0.0,
-            f_max=self.target_sample_rate / 2.0,
-            n_mels=self.mel_bins,
-            window_fn=torch.hann_window,
-            center=True,
-            pad_mode="reflect",
-            power=1.0,
-            mel_scale="slaney",
-            norm="slaney",
-        ).to(device)
-
-        mel = mel_transform(waveform)
-        mel = torch.log(torch.clamp(mel, min=1e-5))
-        return mel.permute(0, 1, 3, 2).contiguous()
-
-
-class AudioVAE(torch.nn.Module):
-    """High-level Audio VAE wrapper exposing encode and decode entry points."""
-
-    def __init__(self, state_dict: dict, metadata: dict):
-        super().__init__()
-
-        component_config = AudioVAEComponentConfig.from_metadata(metadata)
-
-        vae_sd = utils.state_dict_prefix_replace(state_dict, {"audio_vae.": ""}, filter_keys=True)
-        vocoder_sd = utils.state_dict_prefix_replace(state_dict, {"vocoder.": ""}, filter_keys=True)
-
-        self.autoencoder = CausalAudioAutoencoder(config=component_config.autoencoder)
-        self.vocoder = Vocoder(config=component_config.vocoder)
-
-        self.autoencoder.load_state_dict(vae_sd, strict=False)
-        self.vocoder.load_state_dict(vocoder_sd, strict=False)
-
-        autoencoder_config = self.autoencoder.get_config()
-        self.normalizer = AudioLatentNormalizer(
-            AudioPatchifier(
-                patch_size=1,
-                audio_latent_downsample_factor=LATENT_DOWNSAMPLE_FACTOR,
-                sample_rate=autoencoder_config["sampling_rate"],
-                hop_length=autoencoder_config["mel_hop_length"],
-                is_causal=autoencoder_config["is_causal"],
-            ),
-            self.autoencoder.per_channel_statistics,
-        )
-
-        self.preprocessor = AudioPreprocessor(
-            target_sample_rate=autoencoder_config["sampling_rate"],
-            mel_bins=autoencoder_config["mel_bins"],
-            mel_hop_length=autoencoder_config["mel_hop_length"],
-            n_fft=autoencoder_config["n_fft"],
-        )
-
-        self.device_manager = ModelDeviceManager(self)
-
-    def encode(self, audio: dict) -> torch.Tensor:
-        """Encode a waveform dictionary into normalized latent tensors."""
-
-        waveform = audio["waveform"]
-        waveform_sample_rate = audio["sample_rate"]
-        input_device = waveform.device
-        # Ensure that Audio VAE is loaded on the correct device.
-        self.device_manager.ensure_model_loaded()
-
-        waveform = self.device_manager.move_to_load_device(waveform)
-        expected_channels = self.autoencoder.encoder.in_channels
-        if waveform.shape[1] != expected_channels:
-            raise ValueError(
-                f"Input audio must have {expected_channels} channels, got {waveform.shape[1]}"
-            )
-
-        mel_spec = self.preprocessor.waveform_to_mel(
-            waveform, waveform_sample_rate, device=self.device_manager.load_device
-        )
-
-        latents = self.autoencoder.encode(mel_spec)
-        posterior = DiagonalGaussianDistribution(latents)
-        latent_mode = posterior.mode()
-
-        normalized = self.normalizer.normalize(latent_mode)
-        return normalized.to(input_device)
-
-    def decode(self, latents: torch.Tensor) -> torch.Tensor:
-        """Decode normalized latent tensors into an audio waveform."""
-        original_shape = latents.shape
-
-        # Ensure that Audio VAE is loaded on the correct device.
-        self.device_manager.ensure_model_loaded()
-
-        latents = self.device_manager.move_to_load_device(latents)
-        latents = self.normalizer.denormalize(latents)
-
-        target_shape = self.target_shape_from_latents(original_shape)
-        mel_spec = self.autoencoder.decode(latents, target_shape=target_shape)
-
-        waveform = self.run_vocoder(mel_spec)
-        return self.device_manager.move_to_load_device(waveform)
-
-    def target_shape_from_latents(self, latents_shape):
-        batch, _, time, _ = latents_shape
-        target_length = time * LATENT_DOWNSAMPLE_FACTOR
-        if self.autoencoder.causality_axis != CausalityAxis.NONE:
-            target_length -= LATENT_DOWNSAMPLE_FACTOR - 1
-        return (
-            batch,
-            self.autoencoder.decoder.out_ch,
-            target_length,
-            self.autoencoder.mel_bins,
-        )
-
-    def num_of_latents_from_frames(self, frames_number: int, frame_rate: int) -> int:
-        return math.ceil((float(frames_number) / frame_rate) * self.latents_per_second)
-
-    def run_vocoder(self, mel_spec: torch.Tensor) -> torch.Tensor:
-        audio_channels = self.autoencoder.decoder.out_ch
-        vocoder_input = mel_spec.transpose(2, 3)
-
-        if audio_channels == 1:
-            vocoder_input = vocoder_input.squeeze(1)
-        elif audio_channels != 2:
-            raise ValueError(f"Unsupported audio_channels: {audio_channels}")
-
-        return self.vocoder(vocoder_input)
-
-    @property
-    def sample_rate(self) -> int:
-        return int(self.autoencoder.sampling_rate)
-
-    @property
-    def mel_hop_length(self) -> int:
-        return int(self.autoencoder.mel_hop_length)
-
-    @property
-    def mel_bins(self) -> int:
-        return int(self.autoencoder.mel_bins)
-
-    @property
-    def latent_channels(self) -> int:
-        return int(self.autoencoder.decoder.z_channels)
-
-    @property
-    def latent_frequency_bins(self) -> int:
-        return int(self.mel_bins // LATENT_DOWNSAMPLE_FACTOR)
-
-    @property
-    def latents_per_second(self) -> float:
-        return self.sample_rate / self.mel_hop_length / LATENT_DOWNSAMPLE_FACTOR
-
-    @property
-    def output_sample_rate(self) -> int:
-        output_rate = getattr(self.vocoder, "output_sample_rate", None)
-        if output_rate is not None:
-            return int(output_rate)
-        upsample_factor = getattr(self.vocoder, "upsample_factor", None)
-        if upsample_factor is None:
-            raise AttributeError(
-                "Vocoder is missing upsample_factor; cannot infer output sample rate"
-            )
-        return int(self.sample_rate * upsample_factor / self.mel_hop_length)
-
-    def memory_required(self, input_shape):
-        return self.device_manager.patcher.model_size()

comfy/ldm/lightricks/vae/causal_audio_autoencoder.py

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

comfy/ldm/lightricks/vocoders/vocoder.py

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

comfy/ldm/lumina/controlnet.py

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

comfy/ldm/lumina/model.py

@@ -22,10 +22,6 @@ def modulate(x, scale):
 #                               Core NextDiT Model                              #
 #############################################################################
 
-def clamp_fp16(x):
-    if x.dtype == torch.float16:
-        return torch.nan_to_num(x, nan=0.0, posinf=65504, neginf=-65504)
-    return x
 
 class JointAttention(nn.Module):
     """Multi-head attention module."""
@@ -173,7 +169,7 @@ class FeedForward(nn.Module):
 
     # @torch.compile
     def _forward_silu_gating(self, x1, x3):
-        return clamp_fp16(F.silu(x1) * x3)
+        return F.silu(x1) * x3
 
     def forward(self, x):
         return self.w2(self._forward_silu_gating(self.w1(x), self.w3(x)))
@@ -277,27 +273,27 @@ class JointTransformerBlock(nn.Module):
             scale_msa, gate_msa, scale_mlp, gate_mlp = self.adaLN_modulation(adaln_input).chunk(4, dim=1)
 
             x = x + gate_msa.unsqueeze(1).tanh() * self.attention_norm2(
-                clamp_fp16(self.attention(
+                self.attention(
                     modulate(self.attention_norm1(x), scale_msa),
                     x_mask,
                     freqs_cis,
                     transformer_options=transformer_options,
-                ))
+                )
             )
             x = x + gate_mlp.unsqueeze(1).tanh() * self.ffn_norm2(
-                clamp_fp16(self.feed_forward(
+                self.feed_forward(
                     modulate(self.ffn_norm1(x), scale_mlp),
-                ))
+                )
             )
         else:
             assert adaln_input is None
             x = x + self.attention_norm2(
-                clamp_fp16(self.attention(
+                self.attention(
                     self.attention_norm1(x),
                     x_mask,
                     freqs_cis,
                     transformer_options=transformer_options,
-                ))
+                )
             )
             x = x + self.ffn_norm2(
                 self.feed_forward(
@@ -377,7 +373,6 @@ class NextDiT(nn.Module):
         z_image_modulation=False,
         time_scale=1.0,
         pad_tokens_multiple=None,
-        clip_text_dim=None,
         image_model=None,
         device=None,
         dtype=None,
@@ -448,31 +443,6 @@ class NextDiT(nn.Module):
             ),
         )
 
-        self.clip_text_pooled_proj = None
-
-        if clip_text_dim is not None:
-            self.clip_text_dim = clip_text_dim
-            self.clip_text_pooled_proj = nn.Sequential(
-                operation_settings.get("operations").RMSNorm(clip_text_dim, eps=norm_eps, elementwise_affine=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")),
-                operation_settings.get("operations").Linear(
-                    clip_text_dim,
-                    clip_text_dim,
-                    bias=True,
-                    device=operation_settings.get("device"),
-                    dtype=operation_settings.get("dtype"),
-                ),
-            )
-            self.time_text_embed = nn.Sequential(
-                nn.SiLU(),
-                operation_settings.get("operations").Linear(
-                    min(dim, 1024) + clip_text_dim,
-                    min(dim, 1024),
-                    bias=True,
-                    device=operation_settings.get("device"),
-                    dtype=operation_settings.get("dtype"),
-                ),
-            )
-
         self.layers = nn.ModuleList(
             [
                 JointTransformerBlock(
@@ -491,8 +461,7 @@ class NextDiT(nn.Module):
                 for layer_id in range(n_layers)
             ]
         )
-        # This norm final is in the lumina 2.0 code but isn't actually used for anything.
-        # self.norm_final = operation_settings.get("operations").RMSNorm(dim, eps=norm_eps, elementwise_affine=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.norm_final = operation_settings.get("operations").RMSNorm(dim, eps=norm_eps, elementwise_affine=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
         self.final_layer = FinalLayer(dim, patch_size, self.out_channels, z_image_modulation=z_image_modulation, operation_settings=operation_settings)
 
         if self.pad_tokens_multiple is not None:
@@ -537,7 +506,6 @@ class NextDiT(nn.Module):
         bsz = len(x)
         pH = pW = self.patch_size
         device = x[0].device
-        orig_x = x
 
         if self.pad_tokens_multiple is not None:
             pad_extra = (-cap_feats.shape[1]) % self.pad_tokens_multiple
@@ -574,21 +542,13 @@ class NextDiT(nn.Module):
 
         freqs_cis = self.rope_embedder(torch.cat((cap_pos_ids, x_pos_ids), dim=1)).movedim(1, 2)
 
-        patches = transformer_options.get("patches", {})
-
         # refine context
         for layer in self.context_refiner:
             cap_feats = layer(cap_feats, cap_mask, freqs_cis[:, :cap_pos_ids.shape[1]], transformer_options=transformer_options)
 
         padded_img_mask = None
-        x_input = x
-        for i, layer in enumerate(self.noise_refiner):
+        for layer in self.noise_refiner:
             x = layer(x, padded_img_mask, freqs_cis[:, cap_pos_ids.shape[1]:], t, transformer_options=transformer_options)
-            if "noise_refiner" in patches:
-                for p in patches["noise_refiner"]:
-                    out = p({"img": x, "img_input": x_input, "txt": cap_feats, "pe": freqs_cis[:, cap_pos_ids.shape[1]:], "vec": t, "x": orig_x, "block_index": i, "transformer_options": transformer_options, "block_type": "noise_refiner"})
-                    if "img" in out:
-                        x = out["img"]
 
         padded_full_embed = torch.cat((cap_feats, x), dim=1)
         mask = None
@@ -604,7 +564,7 @@ class NextDiT(nn.Module):
         ).execute(x, timesteps, context, num_tokens, attention_mask, **kwargs)
 
     # def forward(self, x, t, cap_feats, cap_mask):
-    def _forward(self, x, timesteps, context, num_tokens, attention_mask=None, transformer_options={}, **kwargs):
+    def _forward(self, x, timesteps, context, num_tokens, attention_mask=None, **kwargs):
         t = 1.0 - timesteps
         cap_feats = context
         cap_mask = attention_mask
@@ -621,36 +581,16 @@ class NextDiT(nn.Module):
 
         cap_feats = self.cap_embedder(cap_feats)  # (N, L, D)  # todo check if able to batchify w.o. redundant compute
 
-        if self.clip_text_pooled_proj is not None:
-            pooled = kwargs.get("clip_text_pooled", None)
-            if pooled is not None:
-                pooled = self.clip_text_pooled_proj(pooled)
-            else:
-                pooled = torch.zeros((x.shape[0], self.clip_text_dim), device=x.device, dtype=x.dtype)
-
-            adaln_input = self.time_text_embed(torch.cat((t, pooled), dim=-1))
-
-        patches = transformer_options.get("patches", {})
+        transformer_options = kwargs.get("transformer_options", {})
         x_is_tensor = isinstance(x, torch.Tensor)
-        img, mask, img_size, cap_size, freqs_cis = self.patchify_and_embed(x, cap_feats, cap_mask, adaln_input, num_tokens, transformer_options=transformer_options)
-        freqs_cis = freqs_cis.to(img.device)
+        x, mask, img_size, cap_size, freqs_cis = self.patchify_and_embed(x, cap_feats, cap_mask, t, num_tokens, transformer_options=transformer_options)
+        freqs_cis = freqs_cis.to(x.device)
 
-        transformer_options["total_blocks"] = len(self.layers)
-        transformer_options["block_type"] = "double"
-        img_input = img
-        for i, layer in enumerate(self.layers):
-            transformer_options["block_index"] = i
-            img = layer(img, mask, freqs_cis, adaln_input, transformer_options=transformer_options)
-            if "double_block" in patches:
-                for p in patches["double_block"]:
-                    out = p({"img": img[:, cap_size[0]:], "img_input": img_input[:, cap_size[0]:], "txt": img[:, :cap_size[0]], "pe": freqs_cis[:, cap_size[0]:], "vec": adaln_input, "x": x, "block_index": i, "transformer_options": transformer_options})
-                    if "img" in out:
-                        img[:, cap_size[0]:] = out["img"]
-                    if "txt" in out:
-                        img[:, :cap_size[0]] = out["txt"]
+        for layer in self.layers:
+            x = layer(x, mask, freqs_cis, adaln_input, transformer_options=transformer_options)
 
-        img = self.final_layer(img, adaln_input)
-        img = self.unpatchify(img, img_size, cap_size, return_tensor=x_is_tensor)[:, :, :h, :w]
+        x = self.final_layer(x, adaln_input)
+        x = self.unpatchify(x, img_size, cap_size, return_tensor=x_is_tensor)[:,:,:h,:w]
 
-        return -img
+        return -x
 

comfy/ldm/modules/attention.py

@@ -30,13 +30,6 @@ except ImportError as e:
             raise e
         exit(-1)
 
-SAGE_ATTENTION3_IS_AVAILABLE = False
-try:
-    from sageattn3 import sageattn3_blackwell
-    SAGE_ATTENTION3_IS_AVAILABLE = True
-except ImportError:
-    pass
-
 FLASH_ATTENTION_IS_AVAILABLE = False
 try:
     from flash_attn import flash_attn_func
@@ -524,7 +517,6 @@ def attention_pytorch(q, k, v, heads, mask=None, attn_precision=None, skip_resha
 
 @wrap_attn
 def attention_sage(q, k, v, heads, mask=None, attn_precision=None, skip_reshape=False, skip_output_reshape=False, **kwargs):
-    exception_fallback = False
     if skip_reshape:
         b, _, _, dim_head = q.shape
         tensor_layout = "HND"
@@ -549,8 +541,6 @@ def attention_sage(q, k, v, heads, mask=None, attn_precision=None, skip_reshape=
         out = sageattn(q, k, v, attn_mask=mask, is_causal=False, tensor_layout=tensor_layout)
     except Exception as e:
         logging.error("Error running sage attention: {}, using pytorch attention instead.".format(e))
-        exception_fallback = True
-    if exception_fallback:
         if tensor_layout == "NHD":
             q, k, v = map(
                 lambda t: t.transpose(1, 2),
@@ -570,93 +560,6 @@ def attention_sage(q, k, v, heads, mask=None, attn_precision=None, skip_reshape=
             out = out.reshape(b, -1, heads * dim_head)
     return out
 
-@wrap_attn
-def attention3_sage(q, k, v, heads, mask=None, attn_precision=None, skip_reshape=False, skip_output_reshape=False, **kwargs):
-    exception_fallback = False
-    if (q.device.type != "cuda" or
-        q.dtype not in (torch.float16, torch.bfloat16) or
-        mask is not None):
-        return attention_pytorch(
-            q, k, v, heads,
-            mask=mask,
-            attn_precision=attn_precision,
-            skip_reshape=skip_reshape,
-            skip_output_reshape=skip_output_reshape,
-            **kwargs
-        )
-
-    if skip_reshape:
-        B, H, L, D = q.shape
-        if H != heads:
-            return attention_pytorch(
-                q, k, v, heads,
-                mask=mask,
-                attn_precision=attn_precision,
-                skip_reshape=True,
-                skip_output_reshape=skip_output_reshape,
-                **kwargs
-            )
-        q_s, k_s, v_s = q, k, v
-        N = q.shape[2]
-        dim_head = D
-    else:
-        B, N, inner_dim = q.shape
-        if inner_dim % heads != 0:
-            return attention_pytorch(
-                q, k, v, heads,
-                mask=mask,
-                attn_precision=attn_precision,
-                skip_reshape=False,
-                skip_output_reshape=skip_output_reshape,
-                **kwargs
-            )
-        dim_head = inner_dim // heads
-
-    if dim_head >= 256 or N <= 1024:
-        return attention_pytorch(
-                q, k, v, heads,
-                mask=mask,
-                attn_precision=attn_precision,
-                skip_reshape=skip_reshape,
-                skip_output_reshape=skip_output_reshape,
-                **kwargs
-            )
-
-    if not skip_reshape:
-        q_s, k_s, v_s = map(
-            lambda t: t.view(B, -1, heads, dim_head).permute(0, 2, 1, 3).contiguous(),
-            (q, k, v),
-        )
-        B, H, L, D = q_s.shape
-
-    try:
-        out = sageattn3_blackwell(q_s, k_s, v_s, is_causal=False)
-    except Exception as e:
-        exception_fallback = True
-        logging.error("Error running SageAttention3: %s, falling back to pytorch attention.", e)
-
-    if exception_fallback:
-        if not skip_reshape:
-            del q_s, k_s, v_s
-        return attention_pytorch(
-                q, k, v, heads,
-                mask=mask,
-                attn_precision=attn_precision,
-                skip_reshape=False,
-                skip_output_reshape=skip_output_reshape,
-                **kwargs
-            )
-
-    if skip_reshape:
-        if not skip_output_reshape:
-            out = out.permute(0, 2, 1, 3).reshape(B, L, H * D)
-    else:
-        if skip_output_reshape:
-            pass
-        else:
-            out = out.permute(0, 2, 1, 3).reshape(B, L, H * D)
-
-    return out
 
 try:
     @torch.library.custom_op("flash_attention::flash_attn", mutates_args=())
@@ -744,8 +647,6 @@ optimized_attention_masked = optimized_attention
 # register core-supported attention functions
 if SAGE_ATTENTION_IS_AVAILABLE:
     register_attention_function("sage", attention_sage)
-if SAGE_ATTENTION3_IS_AVAILABLE:
-    register_attention_function("sage3", attention3_sage)
 if FLASH_ATTENTION_IS_AVAILABLE:
     register_attention_function("flash", attention_flash)
 if model_management.xformers_enabled():

comfy/ldm/modules/diffusionmodules/model.py

@@ -13,12 +13,6 @@ if model_management.xformers_enabled_vae():
     import xformers
     import xformers.ops
 
-def torch_cat_if_needed(xl, dim):
-    if len(xl) > 1:
-        return torch.cat(xl, dim)
-    else:
-        return xl[0]
-
 def get_timestep_embedding(timesteps, embedding_dim):
     """
     This matches the implementation in Denoising Diffusion Probabilistic Models:
@@ -49,37 +43,6 @@ def Normalize(in_channels, num_groups=32):
     return ops.GroupNorm(num_groups=num_groups, num_channels=in_channels, eps=1e-6, affine=True)
 
 
-class CarriedConv3d(nn.Module):
-    def __init__(self, n_channels, out_channels, kernel_size, stride=1, dilation=1, padding=0, **kwargs):
-        super().__init__()
-        self.conv = ops.Conv3d(n_channels, out_channels, kernel_size, stride=stride, dilation=dilation, **kwargs)
-
-    def forward(self, x):
-        return self.conv(x)
-
-
-def conv_carry_causal_3d(xl, op, conv_carry_in=None, conv_carry_out=None):
-
-    x = xl[0]
-    xl.clear()
-
-    if isinstance(op, CarriedConv3d):
-        if conv_carry_in is None:
-            x = torch.nn.functional.pad(x, (1, 1, 1, 1, 2, 0), mode = 'replicate')
-        else:
-            carry_len = conv_carry_in[0].shape[2]
-            x = torch.nn.functional.pad(x, (1, 1, 1, 1, 2 - carry_len, 0), mode = 'replicate')
-            x = torch.cat([conv_carry_in.pop(0), x], dim=2)
-
-        if conv_carry_out is not None:
-            to_push = x[:, :, -2:, :, :].clone()
-            conv_carry_out.append(to_push)
-
-    out = op(x)
-
-    return out
-
-
 class VideoConv3d(nn.Module):
     def __init__(self, n_channels, out_channels, kernel_size, stride=1, dilation=1, padding_mode='replicate', padding=1, **kwargs):
         super().__init__()
@@ -126,24 +89,29 @@ class Upsample(nn.Module):
                                         stride=1,
                                         padding=1)
 
-    def forward(self, x, conv_carry_in=None, conv_carry_out=None):
+    def forward(self, x):
         scale_factor = self.scale_factor
         if isinstance(scale_factor, (int, float)):
             scale_factor = (scale_factor,) * (x.ndim - 2)
 
         if x.ndim == 5 and scale_factor[0] > 1.0:
-            results = []
-            if conv_carry_in is None:
-                first = x[:, :, :1, :, :]
-                results.append(interpolate_up(first.squeeze(2), scale_factor=scale_factor[1:]).unsqueeze(2))
-                x = x[:, :, 1:, :, :]
-            if x.shape[2] > 0:
-                results.append(interpolate_up(x, scale_factor))
-            x = torch_cat_if_needed(results, dim=2)
+            t = x.shape[2]
+            if t > 1:
+                a, b = x.split((1, t - 1), dim=2)
+                del x
+                b = interpolate_up(b, scale_factor)
+            else:
+                a = x
+
+            a = interpolate_up(a.squeeze(2), scale_factor=scale_factor[1:]).unsqueeze(2)
+            if t > 1:
+                x = torch.cat((a, b), dim=2)
+            else:
+                x = a
         else:
             x = interpolate_up(x, scale_factor)
         if self.with_conv:
-            x = conv_carry_causal_3d([x], self.conv, conv_carry_in, conv_carry_out)
+            x = self.conv(x)
         return x
 
 
@@ -159,20 +127,17 @@ class Downsample(nn.Module):
                                         stride=stride,
                                         padding=0)
 
-    def forward(self, x, conv_carry_in=None, conv_carry_out=None):
+    def forward(self, x):
         if self.with_conv:
-            if isinstance(self.conv, CarriedConv3d):
-                x = conv_carry_causal_3d([x], self.conv, conv_carry_in, conv_carry_out)
-            elif x.ndim == 4:
+            if x.ndim == 4:
                 pad = (0, 1, 0, 1)
                 mode = "constant"
                 x = torch.nn.functional.pad(x, pad, mode=mode, value=0)
-                x = self.conv(x)
             elif x.ndim == 5:
                 pad = (1, 1, 1, 1, 2, 0)
                 mode = "replicate"
                 x = torch.nn.functional.pad(x, pad, mode=mode)
-                x = self.conv(x)
+            x = self.conv(x)
         else:
             x = torch.nn.functional.avg_pool2d(x, kernel_size=2, stride=2)
         return x
@@ -218,23 +183,23 @@ class ResnetBlock(nn.Module):
                                                     stride=1,
                                                     padding=0)
 
-    def forward(self, x, temb=None, conv_carry_in=None, conv_carry_out=None):
+    def forward(self, x, temb=None):
         h = x
         h = self.norm1(h)
-        h = [ self.swish(h) ]
-        h = conv_carry_causal_3d(h, self.conv1, conv_carry_in=conv_carry_in, conv_carry_out=conv_carry_out)
+        h = self.swish(h)
+        h = self.conv1(h)
 
         if temb is not None:
             h = h + self.temb_proj(self.swish(temb))[:,:,None,None]
 
         h = self.norm2(h)
         h = self.swish(h)
-        h = [ self.dropout(h) ]
-        h = conv_carry_causal_3d(h, self.conv2, conv_carry_in=conv_carry_in, conv_carry_out=conv_carry_out)
+        h = self.dropout(h)
+        h = self.conv2(h)
 
         if self.in_channels != self.out_channels:
             if self.use_conv_shortcut:
-                x = conv_carry_causal_3d([x], self.conv_shortcut, conv_carry_in=conv_carry_in, conv_carry_out=conv_carry_out)
+                x = self.conv_shortcut(x)
             else:
                 x = self.nin_shortcut(x)
 
@@ -314,7 +279,6 @@ def pytorch_attention(q, k, v):
     orig_shape = q.shape
     B = orig_shape[0]
     C = orig_shape[1]
-    oom_fallback = False
     q, k, v = map(
         lambda t: t.view(B, 1, C, -1).transpose(2, 3).contiguous(),
         (q, k, v),
@@ -325,8 +289,6 @@ def pytorch_attention(q, k, v):
         out = out.transpose(2, 3).reshape(orig_shape)
     except model_management.OOM_EXCEPTION:
         logging.warning("scaled_dot_product_attention OOMed: switched to slice attention")
-        oom_fallback = True
-    if oom_fallback:
         out = slice_attention(q.view(B, -1, C), k.view(B, -1, C).transpose(1, 2), v.view(B, -1, C).transpose(1, 2)).reshape(orig_shape)
     return out
 
@@ -394,8 +356,7 @@ class Model(nn.Module):
                  attn_resolutions, dropout=0.0, resamp_with_conv=True, in_channels,
                  resolution, use_timestep=True, use_linear_attn=False, attn_type="vanilla"):
         super().__init__()
-        if use_linear_attn:
-            attn_type = "linear"
+        if use_linear_attn: attn_type = "linear"
         self.ch = ch
         self.temb_ch = self.ch*4
         self.num_resolutions = len(ch_mult)
@@ -549,22 +510,16 @@ class Encoder(nn.Module):
                  conv3d=False, time_compress=None,
                  **ignore_kwargs):
         super().__init__()
-        if use_linear_attn:
-            attn_type = "linear"
+        if use_linear_attn: attn_type = "linear"
         self.ch = ch
         self.temb_ch = 0
         self.num_resolutions = len(ch_mult)
         self.num_res_blocks = num_res_blocks
         self.resolution = resolution
         self.in_channels = in_channels
-        self.carried = False
 
         if conv3d:
-            if not attn_resolutions:
-                conv_op = CarriedConv3d
-                self.carried = True
-            else:
-                conv_op = VideoConv3d
+            conv_op = VideoConv3d
             mid_attn_conv_op = ops.Conv3d
         else:
             conv_op = ops.Conv2d
@@ -577,7 +532,6 @@ class Encoder(nn.Module):
                                        stride=1,
                                        padding=1)
 
-        self.time_compress = 1
         curr_res = resolution
         in_ch_mult = (1,)+tuple(ch_mult)
         self.in_ch_mult = in_ch_mult
@@ -604,15 +558,10 @@ class Encoder(nn.Module):
                 if time_compress is not None:
                     if (self.num_resolutions - 1 - i_level) > math.log2(time_compress):
                         stride = (1, 2, 2)
-                else:
-                    self.time_compress *= 2
                 down.downsample = Downsample(block_in, resamp_with_conv, stride=stride, conv_op=conv_op)
                 curr_res = curr_res // 2
             self.down.append(down)
 
-        if time_compress is not None:
-            self.time_compress = time_compress
-
         # middle
         self.mid = nn.Module()
         self.mid.block_1 = ResnetBlock(in_channels=block_in,
@@ -638,42 +587,15 @@ class Encoder(nn.Module):
     def forward(self, x):
         # timestep embedding
         temb = None
-
-        if self.carried:
-            xl = [x[:, :, :1, :, :]]
-            if x.shape[2] > self.time_compress:
-                tc = self.time_compress
-                xl += torch.split(x[:, :, 1: 1 + ((x.shape[2] - 1) // tc) * tc, :, :], tc * 2, dim = 2)
-            x = xl
-        else:
-            x = [x]
-        out = []
-
-        conv_carry_in = None
-
-        for i, x1 in enumerate(x):
-            conv_carry_out = []
-            if i == len(x) - 1:
-                conv_carry_out = None
-
-            # downsampling
-            x1 = [ x1 ]
-            h1 = conv_carry_causal_3d(x1, self.conv_in, conv_carry_in, conv_carry_out)
-
-            for i_level in range(self.num_resolutions):
-                for i_block in range(self.num_res_blocks):
-                    h1 = self.down[i_level].block[i_block](h1, temb, conv_carry_in, conv_carry_out)
-                    if len(self.down[i_level].attn) > 0:
-                        assert i == 0 #carried should not happen if attn exists
-                        h1 = self.down[i_level].attn[i_block](h1)
-                if i_level != self.num_resolutions-1:
-                    h1 = self.down[i_level].downsample(h1, conv_carry_in, conv_carry_out)
-
-            out.append(h1)
-            conv_carry_in = conv_carry_out
-
-        h = torch_cat_if_needed(out, dim=2)
-        del out
+        # downsampling
+        h = self.conv_in(x)
+        for i_level in range(self.num_resolutions):
+            for i_block in range(self.num_res_blocks):
+                h = self.down[i_level].block[i_block](h, temb)
+                if len(self.down[i_level].attn) > 0:
+                    h = self.down[i_level].attn[i_block](h)
+            if i_level != self.num_resolutions-1:
+                h = self.down[i_level].downsample(h)
 
         # middle
         h = self.mid.block_1(h, temb)
@@ -682,15 +604,15 @@ class Encoder(nn.Module):
 
         # end
         h = self.norm_out(h)
-        h = [ nonlinearity(h) ]
-        h = conv_carry_causal_3d(h, self.conv_out)
+        h = nonlinearity(h)
+        h = self.conv_out(h)
         return h
 
 
 class Decoder(nn.Module):
     def __init__(self, *, ch, out_ch, ch_mult=(1,2,4,8), num_res_blocks,
                  attn_resolutions, dropout=0.0, resamp_with_conv=True, in_channels,
-                 resolution, z_channels, tanh_out=False, use_linear_attn=False,
+                 resolution, z_channels, give_pre_end=False, tanh_out=False, use_linear_attn=False,
                  conv_out_op=ops.Conv2d,
                  resnet_op=ResnetBlock,
                  attn_op=AttnBlock,
@@ -704,18 +626,12 @@ class Decoder(nn.Module):
         self.num_res_blocks = num_res_blocks
         self.resolution = resolution
         self.in_channels = in_channels
+        self.give_pre_end = give_pre_end
         self.tanh_out = tanh_out
-        self.carried = False
 
         if conv3d:
-            if not attn_resolutions and resnet_op == ResnetBlock:
-                conv_op = CarriedConv3d
-                conv_out_op = CarriedConv3d
-                self.carried = True
-            else:
-                conv_op = VideoConv3d
-                conv_out_op = VideoConv3d
-
+            conv_op = VideoConv3d
+            conv_out_op = VideoConv3d
             mid_attn_conv_op = ops.Conv3d
         else:
             conv_op = ops.Conv2d
@@ -790,43 +706,29 @@ class Decoder(nn.Module):
         temb = None
 
         # z to block_in
-        h = conv_carry_causal_3d([z], self.conv_in)
+        h = self.conv_in(z)
 
         # middle
         h = self.mid.block_1(h, temb, **kwargs)
         h = self.mid.attn_1(h, **kwargs)
         h = self.mid.block_2(h, temb, **kwargs)
 
-        if self.carried:
-            h = torch.split(h, 2, dim=2)
-        else:
-            h = [ h ]
-        out = []
-
-        conv_carry_in = None
-
         # upsampling
-        for i, h1 in enumerate(h):
-            conv_carry_out = []
-            if i == len(h) - 1:
-                conv_carry_out = None
-            for i_level in reversed(range(self.num_resolutions)):
-                for i_block in range(self.num_res_blocks+1):
-                    h1 = self.up[i_level].block[i_block](h1, temb, conv_carry_in, conv_carry_out, **kwargs)
-                    if len(self.up[i_level].attn) > 0:
-                        assert i == 0 #carried should not happen if attn exists
-                        h1 = self.up[i_level].attn[i_block](h1, **kwargs)
-                if i_level != 0:
-                    h1 = self.up[i_level].upsample(h1, conv_carry_in, conv_carry_out)
+        for i_level in reversed(range(self.num_resolutions)):
+            for i_block in range(self.num_res_blocks+1):
+                h = self.up[i_level].block[i_block](h, temb, **kwargs)
+                if len(self.up[i_level].attn) > 0:
+                    h = self.up[i_level].attn[i_block](h, **kwargs)
+            if i_level != 0:
+                h = self.up[i_level].upsample(h)
 
-            h1 = self.norm_out(h1)
-            h1 = [ nonlinearity(h1) ]
-            h1 = conv_carry_causal_3d(h1, self.conv_out, conv_carry_in, conv_carry_out)
-            if self.tanh_out:
-                h1 = torch.tanh(h1)
-            out.append(h1)
-            conv_carry_in = conv_carry_out
+        # end
+        if self.give_pre_end:
+            return h
 
-        out = torch_cat_if_needed(out, dim=2)
-
-        return out
+        h = self.norm_out(h)
+        h = nonlinearity(h)
+        h = self.conv_out(h, **kwargs)
+        if self.tanh_out:
+            h = torch.tanh(h)
+        return h

comfy/ldm/modules/ema.py

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

comfy/ldm/qwen_image/model.py

@@ -61,7 +61,7 @@ def apply_rotary_emb(x, freqs_cis):
 
 
 class QwenTimestepProjEmbeddings(nn.Module):
-    def __init__(self, embedding_dim, pooled_projection_dim, use_additional_t_cond=False, dtype=None, device=None, operations=None):
+    def __init__(self, embedding_dim, pooled_projection_dim, dtype=None, device=None, operations=None):
         super().__init__()
         self.time_proj = Timesteps(num_channels=256, flip_sin_to_cos=True, downscale_freq_shift=0, scale=1000)
         self.timestep_embedder = TimestepEmbedding(
@@ -72,19 +72,9 @@ class QwenTimestepProjEmbeddings(nn.Module):
             operations=operations
         )
 
-        self.use_additional_t_cond = use_additional_t_cond
-        if self.use_additional_t_cond:
-            self.addition_t_embedding = operations.Embedding(2, embedding_dim, device=device, dtype=dtype)
-
-    def forward(self, timestep, hidden_states, addition_t_cond=None):
+    def forward(self, timestep, hidden_states):
         timesteps_proj = self.time_proj(timestep)
         timesteps_emb = self.timestep_embedder(timesteps_proj.to(dtype=hidden_states.dtype))
-
-        if self.use_additional_t_cond:
-            if addition_t_cond is None:
-                addition_t_cond = torch.zeros((timesteps_emb.shape[0]), device=timesteps_emb.device, dtype=torch.long)
-            timesteps_emb += self.addition_t_embedding(addition_t_cond, out_dtype=timesteps_emb.dtype)
-
         return timesteps_emb
 
 
@@ -228,24 +218,9 @@ class QwenImageTransformerBlock(nn.Module):
             operations=operations,
         )
 
-    def _apply_gate(self, x, y, gate, timestep_zero_index=None):
-        if timestep_zero_index is not None:
-            return y + torch.cat((x[:, :timestep_zero_index] * gate[0], x[:, timestep_zero_index:] * gate[1]), dim=1)
-        else:
-            return torch.addcmul(y, gate, x)
-
-    def _modulate(self, x: torch.Tensor, mod_params: torch.Tensor, timestep_zero_index=None) -> Tuple[torch.Tensor, torch.Tensor]:
+    def _modulate(self, x: torch.Tensor, mod_params: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
         shift, scale, gate = torch.chunk(mod_params, 3, dim=-1)
-        if timestep_zero_index is not None:
-            actual_batch = shift.size(0) // 2
-            shift, shift_0 = shift[:actual_batch], shift[actual_batch:]
-            scale, scale_0 = scale[:actual_batch], scale[actual_batch:]
-            gate, gate_0 = gate[:actual_batch], gate[actual_batch:]
-            reg = torch.addcmul(shift.unsqueeze(1), x[:, :timestep_zero_index], 1 + scale.unsqueeze(1))
-            zero = torch.addcmul(shift_0.unsqueeze(1), x[:, timestep_zero_index:], 1 + scale_0.unsqueeze(1))
-            return torch.cat((reg, zero), dim=1), (gate.unsqueeze(1), gate_0.unsqueeze(1))
-        else:
-            return torch.addcmul(shift.unsqueeze(1), x, 1 + scale.unsqueeze(1)), gate.unsqueeze(1)
+        return torch.addcmul(shift.unsqueeze(1), x, 1 + scale.unsqueeze(1)), gate.unsqueeze(1)
 
     def forward(
         self,
@@ -254,19 +229,14 @@ class QwenImageTransformerBlock(nn.Module):
         encoder_hidden_states_mask: torch.Tensor,
         temb: torch.Tensor,
         image_rotary_emb: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
-        timestep_zero_index=None,
         transformer_options={},
     ) -> Tuple[torch.Tensor, torch.Tensor]:
         img_mod_params = self.img_mod(temb)
-
-        if timestep_zero_index is not None:
-            temb = temb.chunk(2, dim=0)[0]
-
         txt_mod_params = self.txt_mod(temb)
         img_mod1, img_mod2 = img_mod_params.chunk(2, dim=-1)
         txt_mod1, txt_mod2 = txt_mod_params.chunk(2, dim=-1)
 
-        img_modulated, img_gate1 = self._modulate(self.img_norm1(hidden_states), img_mod1, timestep_zero_index)
+        img_modulated, img_gate1 = self._modulate(self.img_norm1(hidden_states), img_mod1)
         del img_mod1
         txt_modulated, txt_gate1 = self._modulate(self.txt_norm1(encoder_hidden_states), txt_mod1)
         del txt_mod1
@@ -281,15 +251,15 @@ class QwenImageTransformerBlock(nn.Module):
         del img_modulated
         del txt_modulated
 
-        hidden_states = self._apply_gate(img_attn_output, hidden_states, img_gate1, timestep_zero_index)
+        hidden_states = hidden_states + img_gate1 * img_attn_output
         encoder_hidden_states = encoder_hidden_states + txt_gate1 * txt_attn_output
         del img_attn_output
         del txt_attn_output
         del img_gate1
         del txt_gate1
 
-        img_modulated2, img_gate2 = self._modulate(self.img_norm2(hidden_states), img_mod2, timestep_zero_index)
-        hidden_states = self._apply_gate(self.img_mlp(img_modulated2), hidden_states, img_gate2, timestep_zero_index)
+        img_modulated2, img_gate2 = self._modulate(self.img_norm2(hidden_states), img_mod2)
+        hidden_states = torch.addcmul(hidden_states, img_gate2, self.img_mlp(img_modulated2))
 
         txt_modulated2, txt_gate2 = self._modulate(self.txt_norm2(encoder_hidden_states), txt_mod2)
         encoder_hidden_states = torch.addcmul(encoder_hidden_states, txt_gate2, self.txt_mlp(txt_modulated2))
@@ -330,11 +300,10 @@ class QwenImageTransformer2DModel(nn.Module):
         num_attention_heads: int = 24,
         joint_attention_dim: int = 3584,
         pooled_projection_dim: int = 768,
+        guidance_embeds: bool = False,
         axes_dims_rope: Tuple[int, int, int] = (16, 56, 56),
-        default_ref_method="index",
         image_model=None,
         final_layer=True,
-        use_additional_t_cond=False,
         dtype=None,
         device=None,
         operations=None,
@@ -345,14 +314,12 @@ class QwenImageTransformer2DModel(nn.Module):
         self.in_channels = in_channels
         self.out_channels = out_channels or in_channels
         self.inner_dim = num_attention_heads * attention_head_dim
-        self.default_ref_method = default_ref_method
 
         self.pe_embedder = EmbedND(dim=attention_head_dim, theta=10000, axes_dim=list(axes_dims_rope))
 
         self.time_text_embed = QwenTimestepProjEmbeddings(
             embedding_dim=self.inner_dim,
             pooled_projection_dim=pooled_projection_dim,
-            use_additional_t_cond=use_additional_t_cond,
             dtype=dtype,
             device=device,
             operations=operations
@@ -374,9 +341,6 @@ class QwenImageTransformer2DModel(nn.Module):
             for _ in range(num_layers)
         ])
 
-        if self.default_ref_method == "index_timestep_zero":
-            self.register_buffer("__index_timestep_zero__", torch.tensor([]))
-
         if final_layer:
             self.norm_out = LastLayer(self.inner_dim, self.inner_dim, dtype=dtype, device=device, operations=operations)
             self.proj_out = operations.Linear(self.inner_dim, patch_size * patch_size * self.out_channels, bias=True, dtype=dtype, device=device)
@@ -386,33 +350,27 @@ class QwenImageTransformer2DModel(nn.Module):
         patch_size = self.patch_size
         hidden_states = comfy.ldm.common_dit.pad_to_patch_size(x, (1, self.patch_size, self.patch_size))
         orig_shape = hidden_states.shape
-        hidden_states = hidden_states.view(orig_shape[0], orig_shape[1], orig_shape[-3], orig_shape[-2] // 2, 2, orig_shape[-1] // 2, 2)
-        hidden_states = hidden_states.permute(0, 2, 3, 5, 1, 4, 6)
-        hidden_states = hidden_states.reshape(orig_shape[0], orig_shape[-3] * (orig_shape[-2] // 2) * (orig_shape[-1] // 2), orig_shape[1] * 4)
-        t_len = t
+        hidden_states = hidden_states.view(orig_shape[0], orig_shape[1], orig_shape[-2] // 2, 2, orig_shape[-1] // 2, 2)
+        hidden_states = hidden_states.permute(0, 2, 4, 1, 3, 5)
+        hidden_states = hidden_states.reshape(orig_shape[0], (orig_shape[-2] // 2) * (orig_shape[-1] // 2), orig_shape[1] * 4)
         h_len = ((h + (patch_size // 2)) // patch_size)
         w_len = ((w + (patch_size // 2)) // patch_size)
 
         h_offset = ((h_offset + (patch_size // 2)) // patch_size)
         w_offset = ((w_offset + (patch_size // 2)) // patch_size)
 
-        img_ids = torch.zeros((t_len, h_len, w_len, 3), device=x.device)
+        img_ids = torch.zeros((h_len, w_len, 3), device=x.device)
+        img_ids[:, :, 0] = img_ids[:, :, 1] + index
+        img_ids[:, :, 1] = img_ids[:, :, 1] + torch.linspace(h_offset, h_len - 1 + h_offset, steps=h_len, device=x.device, dtype=x.dtype).unsqueeze(1) - (h_len // 2)
+        img_ids[:, :, 2] = img_ids[:, :, 2] + torch.linspace(w_offset, w_len - 1 + w_offset, steps=w_len, device=x.device, dtype=x.dtype).unsqueeze(0) - (w_len // 2)
+        return hidden_states, repeat(img_ids, "h w c -> b (h w) c", b=bs), orig_shape
 
-        if t_len > 1:
-            img_ids[:, :, :, 0] = img_ids[:, :, :, 0] + torch.linspace(0, t_len - 1, steps=t_len, device=x.device, dtype=x.dtype).unsqueeze(1).unsqueeze(1)
-        else:
-            img_ids[:, :, :, 0] = img_ids[:, :, :, 0] + index
-
-        img_ids[:, :, :, 1] = img_ids[:, :, :, 1] + torch.linspace(h_offset, h_len - 1 + h_offset, steps=h_len, device=x.device, dtype=x.dtype).unsqueeze(1).unsqueeze(0) - (h_len // 2)
-        img_ids[:, :, :, 2] = img_ids[:, :, :, 2] + torch.linspace(w_offset, w_len - 1 + w_offset, steps=w_len, device=x.device, dtype=x.dtype).unsqueeze(0).unsqueeze(0) - (w_len // 2)
-        return hidden_states, repeat(img_ids, "t h w c -> b (t h w) c", b=bs), orig_shape
-
-    def forward(self, x, timestep, context, attention_mask=None, ref_latents=None, additional_t_cond=None, transformer_options={}, **kwargs):
+    def forward(self, x, timestep, context, attention_mask=None, guidance=None, ref_latents=None, transformer_options={}, **kwargs):
         return comfy.patcher_extension.WrapperExecutor.new_class_executor(
             self._forward,
             self,
             comfy.patcher_extension.get_all_wrappers(comfy.patcher_extension.WrappersMP.DIFFUSION_MODEL, transformer_options)
-        ).execute(x, timestep, context, attention_mask, ref_latents, additional_t_cond, transformer_options, **kwargs)
+        ).execute(x, timestep, context, attention_mask, guidance, ref_latents, transformer_options, **kwargs)
 
     def _forward(
         self,
@@ -420,8 +378,8 @@ class QwenImageTransformer2DModel(nn.Module):
         timesteps,
         context,
         attention_mask=None,
+        guidance: torch.Tensor = None,
         ref_latents=None,
-        additional_t_cond=None,
         transformer_options={},
         control=None,
         **kwargs
@@ -433,24 +391,16 @@ class QwenImageTransformer2DModel(nn.Module):
         hidden_states, img_ids, orig_shape = self.process_img(x)
         num_embeds = hidden_states.shape[1]
 
-        timestep_zero_index = None
         if ref_latents is not None:
             h = 0
             w = 0
             index = 0
-            ref_method = kwargs.get("ref_latents_method", self.default_ref_method)
-            index_ref_method = (ref_method == "index") or (ref_method == "index_timestep_zero")
-            negative_ref_method = ref_method == "negative_index"
-            timestep_zero = ref_method == "index_timestep_zero"
+            index_ref_method = kwargs.get("ref_latents_method", "index") == "index"
             for ref in ref_latents:
                 if index_ref_method:
                     index += 1
                     h_offset = 0
                     w_offset = 0
-                elif negative_ref_method:
-                    index -= 1
-                    h_offset = 0
-                    w_offset = 0
                 else:
                     index = 1
                     h_offset = 0
@@ -465,10 +415,6 @@ class QwenImageTransformer2DModel(nn.Module):
                 kontext, kontext_ids, _ = self.process_img(ref, index=index, h_offset=h_offset, w_offset=w_offset)
                 hidden_states = torch.cat([hidden_states, kontext], dim=1)
                 img_ids = torch.cat([img_ids, kontext_ids], dim=1)
-            if timestep_zero:
-                if index > 0:
-                    timestep = torch.cat([timestep, timestep * 0], dim=0)
-                    timestep_zero_index = num_embeds
 
         txt_start = round(max(((x.shape[-1] + (self.patch_size // 2)) // self.patch_size) // 2, ((x.shape[-2] + (self.patch_size // 2)) // self.patch_size) // 2))
         txt_ids = torch.arange(txt_start, txt_start + context.shape[1], device=x.device).reshape(1, -1, 1).repeat(x.shape[0], 1, 3)
@@ -480,7 +426,14 @@ class QwenImageTransformer2DModel(nn.Module):
         encoder_hidden_states = self.txt_norm(encoder_hidden_states)
         encoder_hidden_states = self.txt_in(encoder_hidden_states)
 
-        temb = self.time_text_embed(timestep, hidden_states, additional_t_cond)
+        if guidance is not None:
+            guidance = guidance * 1000
+
+        temb = (
+            self.time_text_embed(timestep, hidden_states)
+            if guidance is None
+            else self.time_text_embed(timestep, guidance, hidden_states)
+        )
 
         patches_replace = transformer_options.get("patches_replace", {})
         patches = transformer_options.get("patches", {})
@@ -493,7 +446,7 @@ class QwenImageTransformer2DModel(nn.Module):
             if ("double_block", i) in blocks_replace:
                 def block_wrap(args):
                     out = {}
-                    out["txt"], out["img"] = block(hidden_states=args["img"], encoder_hidden_states=args["txt"], encoder_hidden_states_mask=encoder_hidden_states_mask, temb=args["vec"], image_rotary_emb=args["pe"], timestep_zero_index=timestep_zero_index, transformer_options=args["transformer_options"])
+                    out["txt"], out["img"] = block(hidden_states=args["img"], encoder_hidden_states=args["txt"], encoder_hidden_states_mask=encoder_hidden_states_mask, temb=args["vec"], image_rotary_emb=args["pe"], transformer_options=args["transformer_options"])
                     return out
                 out = blocks_replace[("double_block", i)]({"img": hidden_states, "txt": encoder_hidden_states, "vec": temb, "pe": image_rotary_emb, "transformer_options": transformer_options}, {"original_block": block_wrap})
                 hidden_states = out["img"]
@@ -505,7 +458,6 @@ class QwenImageTransformer2DModel(nn.Module):
                     encoder_hidden_states_mask=encoder_hidden_states_mask,
                     temb=temb,
                     image_rotary_emb=image_rotary_emb,
-                    timestep_zero_index=timestep_zero_index,
                     transformer_options=transformer_options,
                 )
 
@@ -522,12 +474,9 @@ class QwenImageTransformer2DModel(nn.Module):
                     if add is not None:
                         hidden_states[:, :add.shape[1]] += add
 
-        if timestep_zero_index is not None:
-            temb = temb.chunk(2, dim=0)[0]
-
         hidden_states = self.norm_out(hidden_states, temb)
         hidden_states = self.proj_out(hidden_states)
 
-        hidden_states = hidden_states[:, :num_embeds].view(orig_shape[0], orig_shape[-3], orig_shape[-2] // 2, orig_shape[-1] // 2, orig_shape[1], 2, 2)
-        hidden_states = hidden_states.permute(0, 4, 1, 2, 5, 3, 6)
+        hidden_states = hidden_states[:, :num_embeds].view(orig_shape[0], orig_shape[-2] // 2, orig_shape[-1] // 2, orig_shape[1], 2, 2)
+        hidden_states = hidden_states.permute(0, 3, 1, 4, 2, 5)
         return hidden_states.reshape(orig_shape)[:, :, :, :x.shape[-2], :x.shape[-1]]

comfy/ldm/util.py

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

comfy/ldm/wan/model.py

@@ -568,10 +568,7 @@ class WanModel(torch.nn.Module):
 
         patches_replace = transformer_options.get("patches_replace", {})
         blocks_replace = patches_replace.get("dit", {})
-        transformer_options["total_blocks"] = len(self.blocks)
-        transformer_options["block_type"] = "double"
         for i, block in enumerate(self.blocks):
-            transformer_options["block_index"] = i
             if ("double_block", i) in blocks_replace:
                 def block_wrap(args):
                     out = {}
@@ -766,10 +763,7 @@ class VaceWanModel(WanModel):
 
         patches_replace = transformer_options.get("patches_replace", {})
         blocks_replace = patches_replace.get("dit", {})
-        transformer_options["total_blocks"] = len(self.blocks)
-        transformer_options["block_type"] = "double"
         for i, block in enumerate(self.blocks):
-            transformer_options["block_index"] = i
             if ("double_block", i) in blocks_replace:
                 def block_wrap(args):
                     out = {}
@@ -868,10 +862,7 @@ class CameraWanModel(WanModel):
 
         patches_replace = transformer_options.get("patches_replace", {})
         blocks_replace = patches_replace.get("dit", {})
-        transformer_options["total_blocks"] = len(self.blocks)
-        transformer_options["block_type"] = "double"
         for i, block in enumerate(self.blocks):
-            transformer_options["block_index"] = i
             if ("double_block", i) in blocks_replace:
                 def block_wrap(args):
                     out = {}
@@ -1335,19 +1326,16 @@ class WanModel_S2V(WanModel):
 
         patches_replace = transformer_options.get("patches_replace", {})
         blocks_replace = patches_replace.get("dit", {})
-        transformer_options["total_blocks"] = len(self.blocks)
-        transformer_options["block_type"] = "double"
         for i, block in enumerate(self.blocks):
-            transformer_options["block_index"] = i
             if ("double_block", i) in blocks_replace:
                 def block_wrap(args):
                     out = {}
-                    out["img"] = block(args["img"], context=args["txt"], e=args["vec"], freqs=args["pe"], transformer_options=args["transformer_options"])
+                    out["img"] = block(args["img"], context=args["txt"], e=args["vec"], freqs=args["pe"])
                     return out
-                out = blocks_replace[("double_block", i)]({"img": x, "txt": context, "vec": e0, "pe": freqs, "transformer_options": transformer_options}, {"original_block": block_wrap})
+                out = blocks_replace[("double_block", i)]({"img": x, "txt": context, "vec": e0, "pe": freqs}, {"original_block": block_wrap})
                 x = out["img"]
             else:
-                x = block(x, e=e0, freqs=freqs, context=context, transformer_options=transformer_options)
+                x = block(x, e=e0, freqs=freqs, context=context)
             if audio_emb is not None:
                 x = self.audio_injector(x, i, audio_emb, audio_emb_global, seq_len)
         # head
@@ -1586,10 +1574,7 @@ class HumoWanModel(WanModel):
 
         patches_replace = transformer_options.get("patches_replace", {})
         blocks_replace = patches_replace.get("dit", {})
-        transformer_options["total_blocks"] = len(self.blocks)
-        transformer_options["block_type"] = "double"
         for i, block in enumerate(self.blocks):
-            transformer_options["block_index"] = i
             if ("double_block", i) in blocks_replace:
                 def block_wrap(args):
                     out = {}

comfy/ldm/wan/model_animate.py

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

comfy/ldm/wan/vae.py

@@ -227,7 +227,6 @@ class Encoder3d(nn.Module):
     def __init__(self,
                  dim=128,
                  z_dim=4,
-                 input_channels=3,
                  dim_mult=[1, 2, 4, 4],
                  num_res_blocks=2,
                  attn_scales=[],
@@ -246,7 +245,7 @@ class Encoder3d(nn.Module):
         scale = 1.0
 
         # init block
-        self.conv1 = CausalConv3d(input_channels, dims[0], 3, padding=1)
+        self.conv1 = CausalConv3d(3, dims[0], 3, padding=1)
 
         # downsample blocks
         downsamples = []
@@ -332,7 +331,6 @@ class Decoder3d(nn.Module):
     def __init__(self,
                  dim=128,
                  z_dim=4,
-                 output_channels=3,
                  dim_mult=[1, 2, 4, 4],
                  num_res_blocks=2,
                  attn_scales=[],
@@ -380,7 +378,7 @@ class Decoder3d(nn.Module):
         # output blocks
         self.head = nn.Sequential(
             RMS_norm(out_dim, images=False), nn.SiLU(),
-            CausalConv3d(out_dim, output_channels, 3, padding=1))
+            CausalConv3d(out_dim, 3, 3, padding=1))
 
     def forward(self, x, feat_cache=None, feat_idx=[0]):
         ## conv1
@@ -451,7 +449,6 @@ class WanVAE(nn.Module):
                  num_res_blocks=2,
                  attn_scales=[],
                  temperal_downsample=[True, True, False],
-                 image_channels=3,
                  dropout=0.0):
         super().__init__()
         self.dim = dim
@@ -463,11 +460,11 @@ class WanVAE(nn.Module):
         self.temperal_upsample = temperal_downsample[::-1]
 
         # modules
-        self.encoder = Encoder3d(dim, z_dim * 2, image_channels, dim_mult, num_res_blocks,
+        self.encoder = Encoder3d(dim, z_dim * 2, dim_mult, num_res_blocks,
                                  attn_scales, self.temperal_downsample, dropout)
         self.conv1 = CausalConv3d(z_dim * 2, z_dim * 2, 1)
         self.conv2 = CausalConv3d(z_dim, z_dim, 1)
-        self.decoder = Decoder3d(dim, z_dim, image_channels, dim_mult, num_res_blocks,
+        self.decoder = Decoder3d(dim, z_dim, dim_mult, num_res_blocks,
                                  attn_scales, self.temperal_upsample, dropout)
 
     def encode(self, x):

comfy/lora.py

@@ -320,16 +320,8 @@ def model_lora_keys_unet(model, key_map={}):
                 to = diffusers_keys[k]
                 key_lora = k[:-len(".weight")]
                 key_map["diffusion_model.{}".format(key_lora)] = to
-                key_map["transformer.{}".format(key_lora)] = to
                 key_map["lycoris_{}".format(key_lora.replace(".", "_"))] = to
 
-    if isinstance(model, comfy.model_base.Kandinsky5):
-        for k in sdk:
-            if k.startswith("diffusion_model.") and k.endswith(".weight"):
-                key_lora = k[len("diffusion_model."):-len(".weight")]
-                key_map["{}".format(key_lora)] = k
-                key_map["transformer.{}".format(key_lora)] = k
-
     return key_map
 
 

comfy/model_base.py

@@ -20,7 +20,6 @@ import comfy.ldm.hunyuan3dv2_1
 import comfy.ldm.hunyuan3dv2_1.hunyuandit
 import torch
 import logging
-import comfy.ldm.lightricks.av_model
 from comfy.ldm.modules.diffusionmodules.openaimodel import UNetModel, Timestep
 from comfy.ldm.cascade.stage_c import StageC
 from comfy.ldm.cascade.stage_b import StageB
@@ -48,7 +47,6 @@ import comfy.ldm.chroma_radiance.model
 import comfy.ldm.ace.model
 import comfy.ldm.omnigen.omnigen2
 import comfy.ldm.qwen_image.model
-import comfy.ldm.kandinsky5.model
 
 import comfy.model_management
 import comfy.patcher_extension
@@ -136,7 +134,7 @@ class BaseModel(torch.nn.Module):
         if not unet_config.get("disable_unet_model_creation", False):
             if model_config.custom_operations is None:
                 fp8 = model_config.optimizations.get("fp8", False)
-                operations = comfy.ops.pick_operations(unet_config.get("dtype", None), self.manual_cast_dtype, fp8_optimizations=fp8, model_config=model_config)
+                operations = comfy.ops.pick_operations(unet_config.get("dtype", None), self.manual_cast_dtype, fp8_optimizations=fp8, scaled_fp8=model_config.scaled_fp8, model_config=model_config)
             else:
                 operations = model_config.custom_operations
             self.diffusion_model = unet_model(**unet_config, device=device, operations=operations)
@@ -331,6 +329,18 @@ class BaseModel(torch.nn.Module):
             extra_sds.append(self.model_config.process_clip_vision_state_dict_for_saving(clip_vision_state_dict))
 
         unet_state_dict = self.diffusion_model.state_dict()
+
+        if self.model_config.scaled_fp8 is not None:
+            unet_state_dict["scaled_fp8"] = torch.tensor([], dtype=self.model_config.scaled_fp8)
+
+        # Save mixed precision metadata
+        if hasattr(self.model_config, 'layer_quant_config') and self.model_config.layer_quant_config:
+            metadata = {
+                "format_version": "1.0",
+                "layers": self.model_config.layer_quant_config
+            }
+            unet_state_dict["_quantization_metadata"] = metadata
+
         unet_state_dict = self.model_config.process_unet_state_dict_for_saving(unet_state_dict)
 
         if self.model_type == ModelType.V_PREDICTION:
@@ -947,7 +957,7 @@ class GenmoMochi(BaseModel):
 
 class LTXV(BaseModel):
     def __init__(self, model_config, model_type=ModelType.FLUX, device=None):
-        super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.lightricks.model.LTXVModel)
+        super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.lightricks.model.LTXVModel) #TODO
 
     def extra_conds(self, **kwargs):
         out = super().extra_conds(**kwargs)
@@ -978,60 +988,6 @@ class LTXV(BaseModel):
     def scale_latent_inpaint(self, sigma, noise, latent_image, **kwargs):
         return latent_image
 
-class LTXAV(BaseModel):
-    def __init__(self, model_config, model_type=ModelType.FLUX, device=None):
-        super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.lightricks.av_model.LTXAVModel) #TODO
-
-    def extra_conds(self, **kwargs):
-        out = super().extra_conds(**kwargs)
-        attention_mask = kwargs.get("attention_mask", None)
-        if attention_mask is not None:
-            out['attention_mask'] = comfy.conds.CONDRegular(attention_mask)
-        cross_attn = kwargs.get("cross_attn", None)
-        if cross_attn is not None:
-            out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn)
-
-        out['frame_rate'] = comfy.conds.CONDConstant(kwargs.get("frame_rate", 25))
-
-        denoise_mask = kwargs.get("concat_mask", kwargs.get("denoise_mask", None))
-
-        audio_denoise_mask = None
-        if denoise_mask is not None and "latent_shapes" in kwargs:
-            denoise_mask = utils.unpack_latents(denoise_mask, kwargs["latent_shapes"])
-            if len(denoise_mask) > 1:
-                audio_denoise_mask = denoise_mask[1]
-            denoise_mask = denoise_mask[0]
-
-        if denoise_mask is not None:
-            out["denoise_mask"] = comfy.conds.CONDRegular(denoise_mask)
-
-        if audio_denoise_mask is not None:
-            out["audio_denoise_mask"] = comfy.conds.CONDRegular(audio_denoise_mask)
-
-        keyframe_idxs = kwargs.get("keyframe_idxs", None)
-        if keyframe_idxs is not None:
-            out['keyframe_idxs'] = comfy.conds.CONDRegular(keyframe_idxs)
-
-        latent_shapes = kwargs.get("latent_shapes", None)
-        if latent_shapes is not None:
-            out['latent_shapes'] = comfy.conds.CONDConstant(latent_shapes)
-
-        return out
-
-    def process_timestep(self, timestep, x, denoise_mask=None, audio_denoise_mask=None, **kwargs):
-        v_timestep = timestep
-        a_timestep = timestep
-
-        if denoise_mask is not None:
-            v_timestep = self.diffusion_model.patchifier.patchify(((denoise_mask) * timestep.view([timestep.shape[0]] + [1] * (denoise_mask.ndim - 1)))[:, :1])[0]
-        if audio_denoise_mask is not None:
-            a_timestep = self.diffusion_model.a_patchifier.patchify(((audio_denoise_mask) * timestep.view([timestep.shape[0]] + [1] * (audio_denoise_mask.ndim - 1)))[:, :1, :, :1])[0]
-
-        return v_timestep, a_timestep
-
-    def scale_latent_inpaint(self, sigma, noise, latent_image, **kwargs):
-        return latent_image
-
 class HunyuanVideo(BaseModel):
     def __init__(self, model_config, model_type=ModelType.FLOW, device=None):
         super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.hunyuan_video.model.HunyuanVideo)
@@ -1165,10 +1121,6 @@ class Lumina2(BaseModel):
             if 'num_tokens' not in out:
                 out['num_tokens'] = comfy.conds.CONDConstant(cross_attn.shape[1])
 
-        clip_text_pooled = kwargs.get("pooled_output", None)  # NewBie
-        if clip_text_pooled is not None:
-            out['clip_text_pooled'] = comfy.conds.CONDRegular(clip_text_pooled)
-
         return out
 
 class WAN21(BaseModel):
@@ -1690,49 +1642,3 @@ class HunyuanVideo15_SR_Distilled(HunyuanVideo15):
         out = super().extra_conds(**kwargs)
         out['disable_time_r'] = comfy.conds.CONDConstant(False)
         return out
-
-class Kandinsky5(BaseModel):
-    def __init__(self, model_config, model_type=ModelType.FLOW, device=None):
-        super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.kandinsky5.model.Kandinsky5)
-
-    def encode_adm(self, **kwargs):
-        return kwargs["pooled_output"]
-
-    def concat_cond(self, **kwargs):
-        noise = kwargs.get("noise", None)
-        device = kwargs["device"]
-        image = torch.zeros_like(noise)
-
-        mask = kwargs.get("concat_mask", kwargs.get("denoise_mask", None))
-        if mask is None:
-            mask = torch.zeros_like(noise)[:, :1]
-        else:
-            mask = 1.0 - mask
-            mask = utils.common_upscale(mask.to(device), noise.shape[-1], noise.shape[-2], "bilinear", "center")
-            if mask.shape[-3] < noise.shape[-3]:
-                mask = torch.nn.functional.pad(mask, (0, 0, 0, 0, 0, noise.shape[-3] - mask.shape[-3]), mode='constant', value=0)
-            mask = utils.resize_to_batch_size(mask, noise.shape[0])
-
-        return torch.cat((image, mask), dim=1)
-
-    def extra_conds(self, **kwargs):
-        out = super().extra_conds(**kwargs)
-        attention_mask = kwargs.get("attention_mask", None)
-        if attention_mask is not None:
-            out['attention_mask'] = comfy.conds.CONDRegular(attention_mask)
-        cross_attn = kwargs.get("cross_attn", None)
-        if cross_attn is not None:
-            out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn)
-
-        time_dim_replace = kwargs.get("time_dim_replace", None)
-        if time_dim_replace is not None:
-            out['time_dim_replace'] = comfy.conds.CONDRegular(self.process_latent_in(time_dim_replace))
-
-        return out
-
-class Kandinsky5Image(Kandinsky5):
-    def __init__(self, model_config, model_type=ModelType.FLOW, device=None):
-        super().__init__(model_config, model_type, device=device)
-
-    def concat_cond(self, **kwargs):
-        return None

comfy/model_detection.py

@@ -6,6 +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:
@@ -180,10 +194,8 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
             dit_config["use_cond_type_embedding"] = False
         if '{}vision_in.proj.0.weight'.format(key_prefix) in state_dict_keys:
             dit_config["vision_in_dim"] = state_dict['{}vision_in.proj.0.weight'.format(key_prefix)].shape[0]
-            dit_config["meanflow_sum"] = True
         else:
             dit_config["vision_in_dim"] = None
-            dit_config["meanflow_sum"] = False
         return dit_config
 
     if '{}double_blocks.0.img_attn.norm.key_norm.scale'.format(key_prefix) in state_dict_keys and ('{}img_in.weight'.format(key_prefix) in state_dict_keys or f"{key_prefix}distilled_guidance_layer.norms.0.scale" in state_dict_keys): #Flux, Chroma or Chroma Radiance (has no img_in.weight)
@@ -196,12 +208,12 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
             dit_config["theta"] = 2000
             dit_config["out_channels"] = 128
             dit_config["global_modulation"] = True
+            dit_config["vec_in_dim"] = None
             dit_config["mlp_silu_act"] = True
             dit_config["qkv_bias"] = False
             dit_config["ops_bias"] = False
             dit_config["default_ref_method"] = "index"
             dit_config["ref_index_scale"] = 10.0
-            dit_config["txt_ids_dims"] = [3]
             patch_size = 1
         else:
             dit_config["image_model"] = "flux"
@@ -211,7 +223,6 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
             dit_config["theta"] = 10000
             dit_config["out_channels"] = 16
             dit_config["qkv_bias"] = True
-            dit_config["txt_ids_dims"] = []
             patch_size = 2
 
         dit_config["in_channels"] = 16
@@ -234,10 +245,6 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
         vec_in_key = '{}vector_in.in_layer.weight'.format(key_prefix)
         if vec_in_key in state_dict_keys:
             dit_config["vec_in_dim"] = state_dict[vec_in_key].shape[1]
-        else:
-            dit_config["vec_in_dim"] = None
-
-        dit_config["num_heads"] = dit_config["hidden_size"] // sum(dit_config["axes_dim"])
 
         dit_config["depth"] = count_blocks(state_dict_keys, '{}double_blocks.'.format(key_prefix) + '{}.')
         dit_config["depth_single_blocks"] = count_blocks(state_dict_keys, '{}single_blocks.'.format(key_prefix) + '{}.')
@@ -261,17 +268,8 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
                 dit_config["nerf_tile_size"] = 512
                 dit_config["nerf_final_head_type"] = "conv" if f"{key_prefix}nerf_final_layer_conv.norm.scale" in state_dict_keys else "linear"
                 dit_config["nerf_embedder_dtype"] = torch.float32
-                if "{}__x0__".format(key_prefix) in state_dict_keys: # x0 pred
-                    dit_config["use_x0"] = True
-                else:
-                    dit_config["use_x0"] = False
         else:
             dit_config["guidance_embed"] = "{}guidance_in.in_layer.weight".format(key_prefix) in state_dict_keys
-            dit_config["yak_mlp"] = '{}double_blocks.0.img_mlp.gate_proj.weight'.format(key_prefix) in state_dict_keys
-            dit_config["txt_norm"] = "{}txt_norm.scale".format(key_prefix) in state_dict_keys
-            if dit_config["yak_mlp"] and dit_config["txt_norm"]:  # Ovis model
-                dit_config["txt_ids_dims"] = [1, 2]
-
         return dit_config
 
     if '{}t5_yproj.weight'.format(key_prefix) in state_dict_keys: #Genmo mochi preview
@@ -307,7 +305,7 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
 
     if '{}adaln_single.emb.timestep_embedder.linear_1.bias'.format(key_prefix) in state_dict_keys: #Lightricks ltxv
         dit_config = {}
-        dit_config["image_model"] = "ltxav" if f'{key_prefix}audio_adaln_single.linear.weight' in state_dict_keys else "ltxv"
+        dit_config["image_model"] = "ltxv"
         dit_config["num_layers"] = count_blocks(state_dict_keys, '{}transformer_blocks.'.format(key_prefix) + '{}.')
         shape = state_dict['{}transformer_blocks.0.attn2.to_k.weight'.format(key_prefix)].shape
         dit_config["attention_head_dim"] = shape[0] // 32
@@ -431,10 +429,6 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
             dit_config["axes_lens"] = [300, 512, 512]
             dit_config["rope_theta"] = 10000.0
             dit_config["ffn_dim_multiplier"] = 4.0
-            ctd_weight = state_dict.get('{}clip_text_pooled_proj.0.weight'.format(key_prefix), None)
-            if ctd_weight is not None:  # NewBie
-                dit_config["clip_text_dim"] = ctd_weight.shape[0]
-                # NewBie also sets axes_lens = [1024, 512, 512] but it's not used in ComfyUI
         elif dit_config["dim"] == 3840:  # Z image
             dit_config["n_heads"] = 30
             dit_config["n_kv_heads"] = 30
@@ -621,29 +615,6 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
         dit_config["image_model"] = "qwen_image"
         dit_config["in_channels"] = state_dict['{}img_in.weight'.format(key_prefix)].shape[1]
         dit_config["num_layers"] = count_blocks(state_dict_keys, '{}transformer_blocks.'.format(key_prefix) + '{}.')
-        if "{}__index_timestep_zero__".format(key_prefix) in state_dict_keys:  # 2511
-            dit_config["default_ref_method"] = "index_timestep_zero"
-        if "{}time_text_embed.addition_t_embedding.weight".format(key_prefix) in state_dict_keys:  # Layered
-            dit_config["use_additional_t_cond"] = True
-            dit_config["default_ref_method"] = "negative_index"
-        return dit_config
-
-    if '{}visual_transformer_blocks.0.cross_attention.key_norm.weight'.format(key_prefix) in state_dict_keys: # Kandinsky 5
-        dit_config = {}
-        model_dim = state_dict['{}visual_embeddings.in_layer.bias'.format(key_prefix)].shape[0]
-        dit_config["model_dim"] = model_dim
-        if model_dim in [4096, 2560]: # pro video and lite image
-            dit_config["axes_dims"] = (32, 48, 48)
-            if model_dim == 2560: # lite image
-                dit_config["rope_scale_factor"] = (1.0, 1.0, 1.0)
-        elif model_dim == 1792: # lite video
-            dit_config["axes_dims"] = (16, 24, 24)
-        dit_config["time_dim"] = state_dict['{}time_embeddings.in_layer.bias'.format(key_prefix)].shape[0]
-        dit_config["image_model"] = "kandinsky5"
-        dit_config["ff_dim"] = state_dict['{}visual_transformer_blocks.0.feed_forward.in_layer.weight'.format(key_prefix)].shape[0]
-        dit_config["visual_embed_dim"] = state_dict['{}visual_embeddings.in_layer.weight'.format(key_prefix)].shape[1]
-        dit_config["num_text_blocks"] = count_blocks(state_dict_keys, '{}text_transformer_blocks.'.format(key_prefix) + '{}.')
-        dit_config["num_visual_blocks"] = count_blocks(state_dict_keys, '{}visual_transformer_blocks.'.format(key_prefix) + '{}.')
         return dit_config
 
     if '{}input_blocks.0.0.weight'.format(key_prefix) not in state_dict_keys:
@@ -788,11 +759,22 @@ def model_config_from_unet(state_dict, unet_key_prefix, use_base_if_no_match=Fal
     if model_config is None and use_base_if_no_match:
         model_config = comfy.supported_models_base.BASE(unet_config)
 
+    scaled_fp8_key = "{}scaled_fp8".format(unet_key_prefix)
+    if scaled_fp8_key in state_dict:
+        scaled_fp8_weight = state_dict.pop(scaled_fp8_key)
+        model_config.scaled_fp8 = scaled_fp8_weight.dtype
+        if model_config.scaled_fp8 == torch.float32:
+            model_config.scaled_fp8 = torch.float8_e4m3fn
+        if scaled_fp8_weight.nelement() == 2:
+            model_config.optimizations["fp8"] = False
+        else:
+            model_config.optimizations["fp8"] = True
+
     # Detect per-layer quantization (mixed precision)
-    quant_config = comfy.utils.detect_layer_quantization(state_dict, unet_key_prefix)
-    if quant_config:
-        model_config.quant_config = quant_config
-        logging.info("Detected mixed precision quantization")
+    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
 

comfy/model_management.py

@@ -22,10 +22,10 @@ from enum import Enum
 from comfy.cli_args import args, PerformanceFeature
 import torch
 import sys
+import importlib
 import platform
 import weakref
 import gc
-import os
 
 class VRAMState(Enum):
     DISABLED = 0    #No vram present: no need to move models to vram
@@ -333,42 +333,28 @@ except:
 SUPPORT_FP8_OPS = args.supports_fp8_compute
 
 AMD_RDNA2_AND_OLDER_ARCH = ["gfx1030", "gfx1031", "gfx1010", "gfx1011", "gfx1012", "gfx906", "gfx900", "gfx803"]
-AMD_ENABLE_MIOPEN_ENV = 'COMFYUI_ENABLE_MIOPEN'
 
 try:
     if is_amd():
         arch = torch.cuda.get_device_properties(get_torch_device()).gcnArchName
         if not (any((a in arch) for a in AMD_RDNA2_AND_OLDER_ARCH)):
-            if os.getenv(AMD_ENABLE_MIOPEN_ENV) != '1':
-                torch.backends.cudnn.enabled = False  # Seems to improve things a lot on AMD
-                logging.info("Set: torch.backends.cudnn.enabled = False for better AMD performance.")
+            torch.backends.cudnn.enabled = False  # Seems to improve things a lot on AMD
+            logging.info("Set: torch.backends.cudnn.enabled = False for better AMD performance.")
 
         try:
             rocm_version = tuple(map(int, str(torch.version.hip).split(".")[:2]))
         except:
             rocm_version = (6, -1)
 
-        def aotriton_supported(gpu_arch):
-            path = torch.__path__[0]
-            path = os.path.join(os.path.join(path, "lib"), "aotriton.images")
-            gfx = set(map(lambda a: a[4:], filter(lambda a: a.startswith("amd-gfx"), os.listdir(path))))
-            if gpu_arch in gfx:
-                return True
-            if "{}x".format(gpu_arch[:-1]) in gfx:
-                return True
-            if "{}xx".format(gpu_arch[:-2]) in gfx:
-                return True
-            return False
-
         logging.info("AMD arch: {}".format(arch))
         logging.info("ROCm version: {}".format(rocm_version))
         if args.use_split_cross_attention == False and args.use_quad_cross_attention == False:
-            if aotriton_supported(arch):  # AMD efficient attention implementation depends on aotriton.
+            if importlib.util.find_spec('triton') is not None:  # AMD efficient attention implementation depends on triton. TODO: better way of detecting if it's compiled in or not.
                 if torch_version_numeric >= (2, 7):  # works on 2.6 but doesn't actually seem to improve much
                     if any((a in arch) for a in ["gfx90a", "gfx942", "gfx1100", "gfx1101", "gfx1151"]):  # TODO: more arches, TODO: gfx950
                         ENABLE_PYTORCH_ATTENTION = True
                 if rocm_version >= (7, 0):
-                   if any((a in arch) for a in ["gfx1200", "gfx1201"]):
+                   if any((a in arch) for a in ["gfx1201"]):
                        ENABLE_PYTORCH_ATTENTION = True
         if torch_version_numeric >= (2, 7) and rocm_version >= (6, 4):
             if any((a in arch) for a in ["gfx1200", "gfx1201", "gfx950"]):  # TODO: more arches, "gfx942" gives error on pytorch nightly 2.10 1013 rocm7.0
@@ -467,7 +453,7 @@ def module_size(module):
     sd = module.state_dict()
     for k in sd:
         t = sd[k]
-        module_mem += t.nbytes
+        module_mem += t.nelement() * t.element_size()
     return module_mem
 
 class LoadedModel:
@@ -1030,8 +1016,8 @@ NUM_STREAMS = 0
 if args.async_offload is not None:
     NUM_STREAMS = args.async_offload
 else:
-    #  Enable by default on Nvidia and AMD
-    if is_nvidia() or is_amd():
+    #  Enable by default on Nvidia
+    if is_nvidia():
         NUM_STREAMS = 2
 
 if args.disable_async_offload:
@@ -1137,16 +1123,6 @@ if not args.disable_pinned_memory:
 
 PINNING_ALLOWED_TYPES = set(["Parameter", "QuantizedTensor"])
 
-def discard_cuda_async_error():
-    try:
-        a = torch.tensor([1], dtype=torch.uint8, device=get_torch_device())
-        b = torch.tensor([1], dtype=torch.uint8, device=get_torch_device())
-        _ = a + b
-        torch.cuda.synchronize()
-    except torch.AcceleratorError:
-        #Dump it! We already know about it from the synchronous return
-        pass
-
 def pin_memory(tensor):
     global TOTAL_PINNED_MEMORY
     if MAX_PINNED_MEMORY <= 0:
@@ -1167,7 +1143,7 @@ def pin_memory(tensor):
     if not tensor.is_contiguous():
         return False
 
-    size = tensor.nbytes
+    size = tensor.numel() * tensor.element_size()
     if (TOTAL_PINNED_MEMORY + size) > MAX_PINNED_MEMORY:
         return False
 
@@ -1179,9 +1155,6 @@ def pin_memory(tensor):
         PINNED_MEMORY[ptr] = size
         TOTAL_PINNED_MEMORY += size
         return True
-    else:
-        logging.warning("Pin error.")
-        discard_cuda_async_error()
 
     return False
 
@@ -1194,7 +1167,7 @@ def unpin_memory(tensor):
         return False
 
     ptr = tensor.data_ptr()
-    size = tensor.nbytes
+    size = tensor.numel() * tensor.element_size()
 
     size_stored = PINNED_MEMORY.get(ptr, None)
     if size_stored is None:
@@ -1210,9 +1183,6 @@ def unpin_memory(tensor):
         if len(PINNED_MEMORY) == 0:
             TOTAL_PINNED_MEMORY = 0
         return True
-    else:
-        logging.warning("Unpin error.")
-        discard_cuda_async_error()
 
     return False
 
@@ -1515,16 +1485,6 @@ def supports_fp8_compute(device=None):
 
     return True
 
-def supports_nvfp4_compute(device=None):
-    if not is_nvidia():
-        return False
-
-    props = torch.cuda.get_device_properties(device)
-    if props.major < 10:
-        return False
-
-    return True
-
 def extended_fp16_support():
     # TODO: check why some models work with fp16 on newer torch versions but not on older
     if torch_version_numeric < (2, 7):
@@ -1532,20 +1492,6 @@ def extended_fp16_support():
 
     return True
 
-LORA_COMPUTE_DTYPES = {}
-def lora_compute_dtype(device):
-    dtype = LORA_COMPUTE_DTYPES.get(device, None)
-    if dtype is not None:
-        return dtype
-
-    if should_use_fp16(device):
-        dtype = torch.float16
-    else:
-        dtype = torch.float32
-
-    LORA_COMPUTE_DTYPES[device] = dtype
-    return dtype
-
 def soft_empty_cache(force=False):
     global cpu_state
     if cpu_state == CPUState.MPS:
@@ -1563,10 +1509,6 @@ def soft_empty_cache(force=False):
 def unload_all_models():
     free_memory(1e30, get_torch_device())
 
-def debug_memory_summary():
-    if is_amd() or is_nvidia():
-        return torch.cuda.memory.memory_summary()
-    return ""
 
 #TODO: might be cleaner to put this somewhere else
 import threading

comfy/model_patcher.py

@@ -35,7 +35,6 @@ import comfy.model_management
 import comfy.patcher_extension
 import comfy.utils
 from comfy.comfy_types import UnetWrapperFunction
-from comfy.quant_ops import QuantizedTensor
 from comfy.patcher_extension import CallbacksMP, PatcherInjection, WrappersMP
 
 
@@ -127,23 +126,36 @@ class LowVramPatch:
     def __init__(self, key, patches, convert_func=None, set_func=None):
         self.key = key
         self.patches = patches
-        self.convert_func = convert_func # TODO: remove
+        self.convert_func = convert_func
         self.set_func = set_func
 
     def __call__(self, weight):
-        return comfy.lora.calculate_weight(self.patches[self.key], weight, self.key, intermediate_dtype=weight.dtype)
+        intermediate_dtype = weight.dtype
+        if self.convert_func is not None:
+            weight = self.convert_func(weight, inplace=False)
 
-LOWVRAM_PATCH_ESTIMATE_MATH_FACTOR = 2
+        if intermediate_dtype not in [torch.float32, torch.float16, torch.bfloat16]: #intermediate_dtype has to be one that is supported in math ops
+            intermediate_dtype = torch.float32
+            out = comfy.lora.calculate_weight(self.patches[self.key], weight.to(intermediate_dtype), self.key, intermediate_dtype=intermediate_dtype)
+            if self.set_func is None:
+                return comfy.float.stochastic_rounding(out, weight.dtype, seed=string_to_seed(self.key))
+            else:
+                return self.set_func(out, seed=string_to_seed(self.key), return_weight=True)
+
+        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
+
+#The above patch logic may cast up the weight to fp32, and do math. Go with fp32 x 3
+LOWVRAM_PATCH_ESTIMATE_MATH_FACTOR = 3
 
 def low_vram_patch_estimate_vram(model, key):
     weight, set_func, convert_func = get_key_weight(model, key)
     if weight is None:
         return 0
-    model_dtype = getattr(model, "manual_cast_dtype", torch.float32)
-    if model_dtype is None:
-        model_dtype = weight.dtype
-
-    return weight.numel() * model_dtype.itemsize * LOWVRAM_PATCH_ESTIMATE_MATH_FACTOR
+    return weight.numel() * torch.float32.itemsize * LOWVRAM_PATCH_ESTIMATE_MATH_FACTOR
 
 def get_key_weight(model, key):
     set_func = None
@@ -454,9 +466,6 @@ class ModelPatcher:
     def set_model_post_input_patch(self, patch):
         self.set_model_patch(patch, "post_input")
 
-    def set_model_noise_refiner_patch(self, patch):
-        self.set_model_patch(patch, "noise_refiner")
-
     def set_model_rope_options(self, scale_x, shift_x, scale_y, shift_y, scale_t, shift_t, **kwargs):
         rope_options = self.model_options["transformer_options"].get("rope_options", {})
         rope_options["scale_x"] = scale_x
@@ -621,11 +630,10 @@ class ModelPatcher:
         if key not in self.backup:
             self.backup[key] = collections.namedtuple('Dimension', ['weight', 'inplace_update'])(weight.to(device=self.offload_device, copy=inplace_update), inplace_update)
 
-        temp_dtype = comfy.model_management.lora_compute_dtype(device_to)
         if device_to is not None:
-            temp_weight = comfy.model_management.cast_to_device(weight, device_to, temp_dtype, copy=True)
+            temp_weight = comfy.model_management.cast_to_device(weight, device_to, torch.float32, copy=True)
         else:
-            temp_weight = weight.to(temp_dtype, copy=True)
+            temp_weight = weight.to(torch.float32, copy=True)
         if convert_func is not None:
             temp_weight = convert_func(temp_weight, inplace=True)
 
@@ -669,18 +677,12 @@ class ModelPatcher:
                 module_mem = comfy.model_management.module_size(m)
                 module_offload_mem = module_mem
                 if hasattr(m, "comfy_cast_weights"):
-                    def check_module_offload_mem(key):
-                        if key in self.patches:
-                            return low_vram_patch_estimate_vram(self.model, key)
-                        model_dtype = getattr(self.model, "manual_cast_dtype", None)
-                        weight, _, _ = get_key_weight(self.model, key)
-                        if model_dtype is None or weight is None:
-                            return 0
-                        if (weight.dtype != model_dtype or isinstance(weight, QuantizedTensor)):
-                            return weight.numel() * model_dtype.itemsize
-                        return 0
-                    module_offload_mem += check_module_offload_mem("{}.weight".format(n))
-                    module_offload_mem += check_module_offload_mem("{}.bias".format(n))
+                    weight_key = "{}.weight".format(n)
+                    bias_key = "{}.bias".format(n)
+                    if weight_key in self.patches:
+                        module_offload_mem += low_vram_patch_estimate_vram(self.model, weight_key)
+                    if bias_key in self.patches:
+                        module_offload_mem += low_vram_patch_estimate_vram(self.model, bias_key)
                 loading.append((module_offload_mem, module_mem, n, m, params))
         return loading
 
@@ -697,12 +699,12 @@ class ModelPatcher:
             offloaded = []
             offload_buffer = 0
             loading.sort(reverse=True)
-            for i, x in enumerate(loading):
+            for x in loading:
                 module_offload_mem, module_mem, n, m, params = x
 
                 lowvram_weight = False
 
-                potential_offload = max(offload_buffer, module_offload_mem + sum([ x1[1] for x1 in loading[i+1:i+1+comfy.model_management.NUM_STREAMS]]))
+                potential_offload = max(offload_buffer, module_offload_mem * (comfy.model_management.NUM_STREAMS + 1))
                 lowvram_fits = mem_counter + module_mem + potential_offload < lowvram_model_memory
 
                 weight_key = "{}.weight".format(n)
@@ -718,7 +720,6 @@ class ModelPatcher:
                             continue
 
                 cast_weight = self.force_cast_weights
-                m.comfy_force_cast_weights = self.force_cast_weights
                 if lowvram_weight:
                     if hasattr(m, "comfy_cast_weights"):
                         m.weight_function = []
@@ -776,8 +777,6 @@ class ModelPatcher:
                     key = "{}.{}".format(n, param)
                     self.unpin_weight(key)
                     self.patch_weight_to_device(key, device_to=device_to)
-                if comfy.model_management.is_device_cuda(device_to):
-                    torch.cuda.synchronize()
 
                 logging.debug("lowvram: loaded module regularly {} {}".format(n, m))
                 m.comfy_patched_weights = True
@@ -791,12 +790,11 @@ class ModelPatcher:
                 for param in params:
                     self.pin_weight_to_device("{}.{}".format(n, param))
 
-            usable_stat = "{:.2f} MB usable,".format(lowvram_model_memory / (1024 * 1024)) if lowvram_model_memory < 1e32 else ""
             if lowvram_counter > 0:
-                logging.info("loaded partially; {} {:.2f} MB loaded, {:.2f} MB offloaded, {:.2f} MB buffer reserved, lowvram patches: {}".format(usable_stat, mem_counter / (1024 * 1024), lowvram_mem_counter / (1024 * 1024), offload_buffer / (1024 * 1024), patch_counter))
+                logging.info("loaded partially; {:.2f} MB usable, {:.2f} MB loaded, {:.2f} MB offloaded, {:.2f} MB buffer reserved, lowvram patches: {}".format(lowvram_model_memory / (1024 * 1024), mem_counter / (1024 * 1024), lowvram_mem_counter / (1024 * 1024), offload_buffer / (1024 * 1024), patch_counter))
                 self.model.model_lowvram = True
             else:
-                logging.info("loaded completely; {} {:.2f} MB loaded, full load: {}".format(usable_stat, 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)
@@ -878,18 +876,14 @@ class ModelPatcher:
             patch_counter = 0
             unload_list = self._load_list()
             unload_list.sort()
-
             offload_buffer = self.model.model_offload_buffer_memory
-            if len(unload_list) > 0:
-                NS = comfy.model_management.NUM_STREAMS
-                offload_weight_factor = [ min(offload_buffer / (NS + 1), unload_list[0][1]) ] * NS
 
             for unload in unload_list:
                 if memory_to_free + offload_buffer - self.model.model_offload_buffer_memory < memory_freed:
                     break
                 module_offload_mem, module_mem, n, m, params = unload
 
-                potential_offload = module_offload_mem + sum(offload_weight_factor)
+                potential_offload = (comfy.model_management.NUM_STREAMS + 1) * module_offload_mem
 
                 lowvram_possible = hasattr(m, "comfy_cast_weights")
                 if hasattr(m, "comfy_patched_weights") and m.comfy_patched_weights == True:
@@ -935,14 +929,12 @@ class ModelPatcher:
                                     patch_counter += 1
                             cast_weight = True
 
-                        if cast_weight and hasattr(m, "comfy_cast_weights"):
+                        if cast_weight:
                             m.prev_comfy_cast_weights = m.comfy_cast_weights
                             m.comfy_cast_weights = True
                         m.comfy_patched_weights = False
                         memory_freed += module_mem
                         offload_buffer = max(offload_buffer, potential_offload)
-                        offload_weight_factor.append(module_mem)
-                        offload_weight_factor.pop(0)
                         logging.debug("freed {}".format(n))
 
                         for param in params:

comfy/ops.py

@@ -22,7 +22,7 @@ import comfy.model_management
 from comfy.cli_args import args, PerformanceFeature
 import comfy.float
 import comfy.rmsnorm
-import json
+import contextlib
 
 def run_every_op():
     if torch.compiler.is_compiling():
@@ -79,7 +79,7 @@ def cast_bias_weight(s, input=None, dtype=None, device=None, bias_dtype=None, of
     if input is not None:
         if dtype is None:
             if isinstance(input, QuantizedTensor):
-                dtype = input.params.orig_dtype
+                dtype = input._layout_params["orig_dtype"]
             else:
                 dtype = input.dtype
         if bias_dtype is None:
@@ -93,6 +93,13 @@ def cast_bias_weight(s, input=None, dtype=None, device=None, bias_dtype=None, of
     else:
         offload_stream = None
 
+    if offload_stream is not None:
+        wf_context = offload_stream
+        if hasattr(wf_context, "as_context"):
+            wf_context = wf_context.as_context(offload_stream)
+    else:
+        wf_context = contextlib.nullcontext()
+
     non_blocking = comfy.model_management.device_supports_non_blocking(device)
 
     weight_has_function = len(s.weight_function) > 0
@@ -104,24 +111,22 @@ def cast_bias_weight(s, input=None, dtype=None, device=None, bias_dtype=None, of
     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)
 
-    comfy.model_management.sync_stream(device, offload_stream)
-
-    bias_a = bias
-    weight_a = weight
-
-    if s.bias is not None:
-        for f in s.bias_function:
-            bias = f(bias)
+        if bias_has_function:
+            with wf_context:
+                for f in s.bias_function:
+                    bias = f(bias)
 
     if weight_has_function or weight.dtype != dtype:
-        weight = weight.to(dtype=dtype)
-        if isinstance(weight, QuantizedTensor):
-            weight = weight.dequantize()
-        for f in s.weight_function:
-            weight = f(weight)
+        with wf_context:
+            weight = weight.to(dtype=dtype)
+            if isinstance(weight, QuantizedTensor):
+                weight = weight.dequantize()
+            for f in s.weight_function:
+                weight = f(weight)
 
+    comfy.model_management.sync_stream(device, offload_stream)
     if offloadable:
-        return weight, bias, (offload_stream, weight_a, bias_a)
+        return weight, bias, offload_stream
     else:
         #Legacy function signature
         return weight, bias
@@ -130,16 +135,13 @@ def cast_bias_weight(s, input=None, dtype=None, device=None, bias_dtype=None, of
 def uncast_bias_weight(s, weight, bias, offload_stream):
     if offload_stream is None:
         return
-    os, weight_a, bias_a = offload_stream
-    if os is None:
-        return
-    if weight_a is not None:
-        device = weight_a.device
+    if weight is not None:
+        device = weight.device
     else:
-        if bias_a is None:
+        if bias is None:
             return
-        device = bias_a.device
-    os.wait_stream(comfy.model_management.current_stream(device))
+        device = bias.device
+    offload_stream.wait_stream(comfy.model_management.current_stream(device))
 
 
 class CastWeightBiasOp:
@@ -412,34 +414,36 @@ def fp8_linear(self, input):
         return None
 
     input_dtype = input.dtype
-    input_shape = input.shape
-    tensor_3d = input.ndim == 3
 
-    if tensor_3d:
-        input = input.reshape(-1, input_shape[2])
+    if input.ndim == 3 or input.ndim == 2:
+        w, bias, offload_stream = cast_bias_weight(self, input, dtype=dtype, bias_dtype=input_dtype, offloadable=True)
 
-    if input.ndim != 2:
-        return None
-    w, bias, offload_stream = cast_bias_weight(self, input, dtype=dtype, bias_dtype=input_dtype, offloadable=True)
-    scale_weight = torch.ones((), device=input.device, dtype=torch.float32)
+        scale_weight = self.scale_weight
+        scale_input = self.scale_input
+        if scale_weight is None:
+            scale_weight = torch.ones((), device=input.device, dtype=torch.float32)
+        else:
+            scale_weight = scale_weight.to(input.device)
 
-    scale_input = torch.ones((), device=input.device, dtype=torch.float32)
-    input = torch.clamp(input, min=-448, max=448, out=input)
-    input_fp8 = input.to(dtype).contiguous()
-    layout_params_input = TensorCoreFP8Layout.Params(scale=scale_input, orig_dtype=input_dtype, orig_shape=tuple(input_fp8.shape))
-    quantized_input = QuantizedTensor(input_fp8, "TensorCoreFP8Layout", layout_params_input)
+        if scale_input is None:
+            scale_input = torch.ones((), device=input.device, dtype=torch.float32)
+            input = torch.clamp(input, min=-448, max=448, out=input)
+            layout_params_weight = {'scale': scale_input, 'orig_dtype': input_dtype}
+            quantized_input = QuantizedTensor(input.to(dtype).contiguous(), "TensorCoreFP8Layout", layout_params_weight)
+        else:
+            scale_input = scale_input.to(input.device)
+            quantized_input = QuantizedTensor.from_float(input, "TensorCoreFP8Layout", scale=scale_input, dtype=dtype)
 
-    # Wrap weight in QuantizedTensor - this enables unified dispatch
-    # Call F.linear - __torch_dispatch__ routes to fp8_linear handler in quant_ops.py!
-    layout_params_weight = TensorCoreFP8Layout.Params(scale=scale_weight, orig_dtype=input_dtype, orig_shape=tuple(w.shape))
-    quantized_weight = QuantizedTensor(w, "TensorCoreFP8Layout", layout_params_weight)
-    o = torch.nn.functional.linear(quantized_input, quantized_weight, bias)
+        # 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)
 
-    uncast_bias_weight(self, w, bias, offload_stream)
-    if tensor_3d:
-        o = o.reshape((input_shape[0], input_shape[1], w.shape[0]))
+        uncast_bias_weight(self, w, bias, offload_stream)
+        return o
 
-    return o
+    return None
 
 class fp8_ops(manual_cast):
     class Linear(manual_cast.Linear):
@@ -449,7 +453,7 @@ class fp8_ops(manual_cast):
             return None
 
         def forward_comfy_cast_weights(self, input):
-            if len(self.weight_function) == 0 and len(self.bias_function) == 0:
+            if not self.training:
                 try:
                     out = fp8_linear(self, input)
                     if out is not None:
@@ -462,6 +466,59 @@ class fp8_ops(manual_cast):
             uncast_bias_weight(self, weight, bias, offload_stream)
             return x
 
+def scaled_fp8_ops(fp8_matrix_mult=False, scale_input=False, override_dtype=None):
+    logging.info("Using scaled fp8: fp8 matrix mult: {}, scale input: {}".format(fp8_matrix_mult, scale_input))
+    class scaled_fp8_op(manual_cast):
+        class Linear(manual_cast.Linear):
+            def __init__(self, *args, **kwargs):
+                if override_dtype is not None:
+                    kwargs['dtype'] = override_dtype
+                super().__init__(*args, **kwargs)
+
+            def reset_parameters(self):
+                if not hasattr(self, 'scale_weight'):
+                    self.scale_weight = torch.nn.parameter.Parameter(data=torch.ones((), device=self.weight.device, dtype=torch.float32), requires_grad=False)
+
+                if not scale_input:
+                    self.scale_input = None
+
+                if not hasattr(self, 'scale_input'):
+                    self.scale_input = torch.nn.parameter.Parameter(data=torch.ones((), device=self.weight.device, dtype=torch.float32), requires_grad=False)
+                return None
+
+            def forward_comfy_cast_weights(self, input):
+                if fp8_matrix_mult:
+                    out = fp8_linear(self, input)
+                    if out is not None:
+                        return out
+
+                weight, bias, offload_stream = cast_bias_weight(self, input, offloadable=True)
+
+                if weight.numel() < input.numel(): #TODO: optimize
+                    x = torch.nn.functional.linear(input, weight * self.scale_weight.to(device=weight.device, dtype=weight.dtype), bias)
+                else:
+                    x = torch.nn.functional.linear(input * self.scale_weight.to(device=weight.device, dtype=weight.dtype), weight, bias)
+                uncast_bias_weight(self, weight, bias, offload_stream)
+                return x
+
+            def convert_weight(self, weight, inplace=False, **kwargs):
+                if inplace:
+                    weight *= self.scale_weight.to(device=weight.device, dtype=weight.dtype)
+                    return weight
+                else:
+                    return weight.to(dtype=torch.float32) * self.scale_weight.to(device=weight.device, dtype=torch.float32)
+
+            def set_weight(self, weight, inplace_update=False, seed=None, return_weight=False, **kwargs):
+                weight = comfy.float.stochastic_rounding(weight / self.scale_weight.to(device=weight.device, dtype=weight.dtype), self.weight.dtype, seed=seed)
+                if return_weight:
+                    return weight
+                if inplace_update:
+                    self.weight.data.copy_(weight)
+                else:
+                    self.weight = torch.nn.Parameter(weight, requires_grad=False)
+
+    return scaled_fp8_op
+
 CUBLAS_IS_AVAILABLE = False
 try:
     from cublas_ops import CublasLinear
@@ -485,20 +542,14 @@ if CUBLAS_IS_AVAILABLE:
 # ==============================================================================
 # Mixed Precision Operations
 # ==============================================================================
-from .quant_ops import (
-    QuantizedTensor,
-    QUANT_ALGOS,
-    TensorCoreFP8Layout,
-    get_layout_class,
-)
+from .quant_ops import QuantizedTensor, QUANT_ALGOS
 
 
-def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_precision_mm=False, disabled=[]):
+def mixed_precision_ops(layer_quant_config={}, compute_dtype=torch.bfloat16, full_precision_mm=False):
     class MixedPrecisionOps(manual_cast):
-        _quant_config = quant_config
+        _layer_quant_config = layer_quant_config
         _compute_dtype = compute_dtype
         _full_precision_mm = full_precision_mm
-        _disabled = disabled
 
         class Linear(torch.nn.Module, CastWeightBiasOp):
             def __init__(
@@ -523,21 +574,10 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec
 
                 self.tensor_class = None
                 self._full_precision_mm = MixedPrecisionOps._full_precision_mm
-                self._full_precision_mm_config = False
 
             def reset_parameters(self):
                 return None
 
-            def _load_scale_param(self, state_dict, prefix, param_name, device, manually_loaded_keys, dtype=None):
-                key = f"{prefix}{param_name}"
-                value = state_dict.pop(key, None)
-                if value is not None:
-                    value = value.to(device=device)
-                    if dtype is not None:
-                        value = value.view(dtype=dtype)
-                    manually_loaded_keys.append(key)
-                return value
-
             def _load_from_state_dict(self, state_dict, prefix, local_metadata,
                                     strict, missing_keys, unexpected_keys, error_msgs):
 
@@ -546,76 +586,40 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec
                 weight_key = f"{prefix}weight"
                 weight = state_dict.pop(weight_key, None)
                 if weight is None:
-                    logging.warning(f"Missing weight for layer {layer_name}")
-                    return
+                    raise ValueError(f"Missing weight for layer {layer_name}")
 
                 manually_loaded_keys = [weight_key]
 
-                layer_conf = state_dict.pop(f"{prefix}comfy_quant", None)
-                if layer_conf is not None:
-                    layer_conf = json.loads(layer_conf.numpy().tobytes())
-
-                if layer_conf is None:
+                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:
-                    self.quant_format = layer_conf.get("format", None)
-                    self._full_precision_mm_config = layer_conf.get("full_precision_matrix_mult", False)
-                    if not self._full_precision_mm:
-                        self._full_precision_mm = self._full_precision_mm_config
-
-                    if self.quant_format in MixedPrecisionOps._disabled:
-                        self._full_precision_mm = True
-
-                    if self.quant_format is None:
+                    quant_format = MixedPrecisionOps._layer_quant_config[layer_name].get("format", None)
+                    if quant_format is None:
                         raise ValueError(f"Unknown quantization format for layer {layer_name}")
 
-                    qconfig = QUANT_ALGOS[self.quant_format]
+                    qconfig = QUANT_ALGOS[quant_format]
                     self.layout_type = qconfig["comfy_tensor_layout"]
-                    layout_cls = get_layout_class(self.layout_type)
 
-                    # Load format-specific parameters
-                    if self.quant_format in ["float8_e4m3fn", "float8_e5m2"]:
-                        # FP8: single tensor scale
-                        scale = self._load_scale_param(state_dict, prefix, "weight_scale", device, manually_loaded_keys)
-
-                        params = layout_cls.Params(
-                            scale=scale,
-                            orig_dtype=MixedPrecisionOps._compute_dtype,
-                            orig_shape=(self.out_features, self.in_features),
-                        )
-
-                    elif self.quant_format == "nvfp4":
-                        # NVFP4: tensor_scale (weight_scale_2) + block_scale (weight_scale)
-                        tensor_scale = self._load_scale_param(state_dict, prefix, "weight_scale_2", device, manually_loaded_keys)
-                        block_scale = self._load_scale_param(state_dict, prefix, "weight_scale", device, manually_loaded_keys,
-                                                             dtype=torch.float8_e4m3fn)
-
-                        if tensor_scale is None or block_scale is None:
-                            raise ValueError(f"Missing NVFP4 scales for layer {layer_name}")
-
-                        params = layout_cls.Params(
-                            scale=tensor_scale,
-                            block_scale=block_scale,
-                            orig_dtype=MixedPrecisionOps._compute_dtype,
-                            orig_shape=(self.out_features, self.in_features),
-                        )
-                    else:
-                        raise ValueError(f"Unsupported quantization format: {self.quant_format}")
+                    weight_scale_key = f"{prefix}weight_scale"
+                    layout_params = {
+                        'scale': state_dict.pop(weight_scale_key, None),
+                        'orig_dtype': MixedPrecisionOps._compute_dtype,
+                        'block_size': qconfig.get("group_size", None),
+                    }
+                    if layout_params['scale'] is not None:
+                        manually_loaded_keys.append(weight_scale_key)
 
                     self.weight = torch.nn.Parameter(
-                        QuantizedTensor(weight.to(device=device, dtype=qconfig["storage_t"]), self.layout_type, params),
+                        QuantizedTensor(weight.to(device=device), self.layout_type, layout_params),
                         requires_grad=False
                     )
 
                     for param_name in qconfig["parameters"]:
-                        if param_name in {"weight_scale", "weight_scale_2"}:
-                            continue  # Already handled above
-
                         param_key = f"{prefix}{param_name}"
                         _v = state_dict.pop(param_key, None)
                         if _v is None:
                             continue
-                        self.register_parameter(param_name, torch.nn.Parameter(_v.to(device=device), requires_grad=False))
+                        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)
@@ -624,32 +628,6 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec
                     if key in missing_keys:
                         missing_keys.remove(key)
 
-            def state_dict(self, *args, destination=None, prefix="", **kwargs):
-                if destination is not None:
-                    sd = destination
-                else:
-                    sd = {}
-
-                if self.bias is not None:
-                    sd["{}bias".format(prefix)] = self.bias
-
-                if isinstance(self.weight, QuantizedTensor):
-                    sd_out = self.weight.state_dict("{}weight".format(prefix))
-                    for k in sd_out:
-                        sd[k] = sd_out[k]
-
-                    quant_conf = {"format": self.quant_format}
-                    if self._full_precision_mm_config:
-                        quant_conf["full_precision_matrix_mult"] = True
-                    sd["{}comfy_quant".format(prefix)] = torch.tensor(list(json.dumps(quant_conf).encode('utf-8')), dtype=torch.uint8)
-
-                    input_scale = getattr(self, 'input_scale', None)
-                    if input_scale is not None:
-                        sd["{}input_scale".format(prefix)] = input_scale
-                else:
-                    sd["{}weight".format(prefix)] = self.weight
-                return sd
-
             def _forward(self, input, weight, bias):
                 return torch.nn.functional.linear(input, weight, bias)
 
@@ -662,33 +640,13 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec
             def forward(self, input, *args, **kwargs):
                 run_every_op()
 
-                input_shape = input.shape
-                reshaped_3d = False
-
+                if self._full_precision_mm or 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
-                    not isinstance(input, QuantizedTensor) and not self._full_precision_mm and
-                    not getattr(self, 'comfy_force_cast_weights', False) and
-                    len(self.weight_function) == 0 and len(self.bias_function) == 0):
-
-                    # Reshape 3D tensors to 2D for quantization (needed for NVFP4 and others)
-                    input_reshaped = input.reshape(-1, input_shape[2]) if input.ndim == 3 else input
-
-                    # Fall back to non-quantized for non-2D tensors
-                    if input_reshaped.ndim == 2:
-                        reshaped_3d = input.ndim == 3
-                        # dtype is now implicit in the layout class
-                        scale = getattr(self, 'input_scale', None)
-                        if scale is not None:
-                            scale = comfy.model_management.cast_to_device(scale, input.device, None)
-                        input = QuantizedTensor.from_float(input_reshaped, self.layout_type, scale=scale)
-
-                output = self.forward_comfy_cast_weights(input)
-
-                # Reshape output back to 3D if input was 3D
-                if reshaped_3d:
-                    output = output.reshape((input_shape[0], input_shape[1], self.weight.shape[0]))
-
-                return output
+                    getattr(self, 'input_scale', None) is not None and
+                    not isinstance(input, QuantizedTensor)):
+                    input = QuantizedTensor.from_float(input, self.layout_type, scale=self.input_scale, dtype=self.weight.dtype)
+                return self._forward(input, self.weight, self.bias)
 
             def convert_weight(self, weight, inplace=False, **kwargs):
                 if isinstance(weight, QuantizedTensor):
@@ -698,8 +656,7 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec
 
             def set_weight(self, weight, inplace_update=False, seed=None, return_weight=False, **kwargs):
                 if getattr(self, 'layout_type', None) is not None:
-                    # dtype is now implicit in the layout class
-                    weight = QuantizedTensor.from_float(weight, self.layout_type, scale="recalculate", stochastic_rounding=seed, inplace_ops=True)
+                    weight = QuantizedTensor.from_float(weight, self.layout_type, scale=None, dtype=self.weight.dtype, stochastic_rounding=seed, inplace_ops=True)
                 else:
                     weight = weight.to(self.weight.dtype)
                 if return_weight:
@@ -708,35 +665,17 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec
                 assert inplace_update is False  # TODO: eventually remove the inplace_update stuff
                 self.weight = torch.nn.Parameter(weight, requires_grad=False)
 
-            def _apply(self, fn, recurse=True):  # This is to get torch.compile + moving weights to another device working
-                if recurse:
-                    for module in self.children():
-                        module._apply(fn)
-
-                for key, param in self._parameters.items():
-                    if param is None:
-                        continue
-                    self.register_parameter(key, torch.nn.Parameter(fn(param), requires_grad=False))
-                for key, buf in self._buffers.items():
-                    if buf is not None:
-                        self._buffers[key] = fn(buf)
-                return self
-
     return MixedPrecisionOps
 
-def pick_operations(weight_dtype, compute_dtype, load_device=None, disable_fast_fp8=False, fp8_optimizations=False, model_config=None):
+def pick_operations(weight_dtype, compute_dtype, load_device=None, disable_fast_fp8=False, fp8_optimizations=False, scaled_fp8=None, model_config=None):
     fp8_compute = comfy.model_management.supports_fp8_compute(load_device) # TODO: if we support more ops this needs to be more granular
-    nvfp4_compute = comfy.model_management.supports_nvfp4_compute(load_device)
 
-    if model_config and hasattr(model_config, 'quant_config') and model_config.quant_config:
-        logging.info("Using mixed precision operations")
-        disabled = set()
-        if not nvfp4_compute:
-            disabled.add("nvfp4")
-        if not fp8_compute:
-            disabled.add("float8_e4m3fn")
-            disabled.add("float8_e5m2")
-        return mixed_precision_ops(model_config.quant_config, compute_dtype, disabled=disabled)
+    if model_config and hasattr(model_config, 'layer_quant_config') and model_config.layer_quant_config:
+        logging.info(f"Using mixed precision operations: {len(model_config.layer_quant_config)} quantized layers")
+        return mixed_precision_ops(model_config.layer_quant_config, compute_dtype, full_precision_mm=not fp8_compute)
+
+    if scaled_fp8 is not None:
+        return scaled_fp8_ops(fp8_matrix_mult=fp8_compute and fp8_optimizations, scale_input=fp8_optimizations, override_dtype=scaled_fp8)
 
     if (
         fp8_compute and

comfy/quant_ops.py

@@ -1,141 +1,573 @@
 import torch
 import logging
-
-try:
-    import comfy_kitchen as ck
-    from comfy_kitchen.tensor import (
-        QuantizedTensor,
-        QuantizedLayout,
-        TensorCoreFP8Layout as _CKFp8Layout,
-        TensorCoreNVFP4Layout,  # Direct import, no wrapper needed
-        register_layout_op,
-        register_layout_class,
-        get_layout_class,
-    )
-    _CK_AVAILABLE = True
-    if torch.version.cuda is None:
-        ck.registry.disable("cuda")
-    else:
-        cuda_version = tuple(map(int, str(torch.version.cuda).split('.')))
-        if cuda_version < (13,):
-            ck.registry.disable("cuda")
-            logging.warning("WARNING: You need pytorch with cu130 or higher to use optimized CUDA operations.")
-
-    ck.registry.disable("triton")
-    for k, v in ck.list_backends().items():
-        logging.info(f"Found comfy_kitchen backend {k}: {v}")
-except ImportError as e:
-    logging.error(f"Failed to import comfy_kitchen, Error: {e}, fp8 and fp4 support will not be available.")
-    _CK_AVAILABLE = False
-
-    class QuantizedTensor:
-        pass
-
-    class _CKFp8Layout:
-        pass
-
-    class TensorCoreNVFP4Layout:
-        pass
-
-    def register_layout_class(name, cls):
-        pass
-
-    def get_layout_class(name):
-        return None
-
+from typing import Tuple, Dict
 import comfy.float
 
-# ==============================================================================
-# FP8 Layouts with Comfy-Specific Extensions
-# ==============================================================================
+_LAYOUT_REGISTRY = {}
+_GENERIC_UTILS = {}
 
-class _TensorCoreFP8LayoutBase(_CKFp8Layout):
-    FP8_DTYPE = None  # Must be overridden in subclass
+
+def register_layout_op(torch_op, layout_type):
+    """
+    Decorator to register a layout-specific operation handler.
+    Args:
+        torch_op: PyTorch operation (e.g., torch.ops.aten.linear.default)
+        layout_type: Layout class (e.g., TensorCoreFP8Layout)
+    Example:
+        @register_layout_op(torch.ops.aten.linear.default, TensorCoreFP8Layout)
+        def fp8_linear(func, args, kwargs):
+            # FP8-specific linear implementation
+            ...
+    """
+    def decorator(handler_func):
+        if torch_op not in _LAYOUT_REGISTRY:
+            _LAYOUT_REGISTRY[torch_op] = {}
+        _LAYOUT_REGISTRY[torch_op][layout_type] = handler_func
+        return handler_func
+    return decorator
+
+
+def register_generic_util(torch_op):
+    """
+    Decorator to register a generic utility that works for all layouts.
+    Args:
+        torch_op: PyTorch operation (e.g., torch.ops.aten.detach.default)
+
+    Example:
+        @register_generic_util(torch.ops.aten.detach.default)
+        def generic_detach(func, args, kwargs):
+            # Works for any layout
+            ...
+    """
+    def decorator(handler_func):
+        _GENERIC_UTILS[torch_op] = handler_func
+        return handler_func
+    return decorator
+
+
+def _get_layout_from_args(args):
+    for arg in args:
+        if isinstance(arg, QuantizedTensor):
+            return arg._layout_type
+        elif isinstance(arg, (list, tuple)):
+            for item in arg:
+                if isinstance(item, QuantizedTensor):
+                    return item._layout_type
+    return None
+
+
+def _move_layout_params_to_device(params, device):
+    new_params = {}
+    for k, v in params.items():
+        if isinstance(v, torch.Tensor):
+            new_params[k] = v.to(device=device)
+        else:
+            new_params[k] = v
+    return new_params
+
+
+def _copy_layout_params(params):
+    new_params = {}
+    for k, v in params.items():
+        if isinstance(v, torch.Tensor):
+            new_params[k] = v.clone()
+        else:
+            new_params[k] = v
+    return new_params
+
+def _copy_layout_params_inplace(src, dst, non_blocking=False):
+    for k, v in src.items():
+        if isinstance(v, torch.Tensor):
+            dst[k].copy_(v, non_blocking=non_blocking)
+        else:
+            dst[k] = v
+
+class QuantizedLayout:
+    """
+    Base class for quantization layouts.
+
+    A layout encapsulates the format-specific logic for quantization/dequantization
+    and provides a uniform interface for extracting raw tensors needed for computation.
+
+    New quantization formats should subclass this and implement the required methods.
+    """
+    @classmethod
+    def quantize(cls, tensor, **kwargs) -> Tuple[torch.Tensor, Dict]:
+        raise NotImplementedError(f"{cls.__name__} must implement quantize()")
+
+    @staticmethod
+    def dequantize(qdata, **layout_params) -> torch.Tensor:
+        raise NotImplementedError("TensorLayout must implement dequantize()")
 
     @classmethod
-    def quantize(cls, tensor, scale=None, stochastic_rounding=0, inplace_ops=False):
-        if cls.FP8_DTYPE is None:
-            raise NotImplementedError(f"{cls.__name__} must define FP8_DTYPE")
+    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)
+
+    def data_ptr(self):
+        return self._qdata.data_ptr()
+
+    def is_pinned(self):
+        return self._qdata.is_pinned()
+
+    def is_contiguous(self, *arg, **kwargs):
+        return self._qdata.is_contiguous(*arg, **kwargs)
+
+# ==============================================================================
+# Generic Utilities (Layout-Agnostic Operations)
+# ==============================================================================
+
+def _create_transformed_qtensor(qt, transform_fn):
+    new_data = transform_fn(qt._qdata)
+    new_params = _copy_layout_params(qt._layout_params)
+    return QuantizedTensor(new_data, qt._layout_type, new_params)
+
+
+def _handle_device_transfer(qt, target_device, target_dtype=None, target_layout=None, op_name="to"):
+    if target_dtype is not None and target_dtype != qt.dtype:
+        logging.warning(
+            f"QuantizedTensor: dtype conversion requested to {target_dtype}, "
+            f"but not supported for quantized tensors. Ignoring dtype."
+        )
+
+    if target_layout is not None and target_layout != torch.strided:
+        logging.warning(
+            f"QuantizedTensor: layout change requested to {target_layout}, "
+            f"but not supported. Ignoring layout."
+        )
+
+    # Handle device transfer
+    current_device = qt._qdata.device
+    if target_device is not None:
+        # Normalize device for comparison
+        if isinstance(target_device, str):
+            target_device = torch.device(target_device)
+        if isinstance(current_device, str):
+            current_device = torch.device(current_device)
+
+        if target_device != current_device:
+            logging.debug(f"QuantizedTensor.{op_name}: Moving from {current_device} to {target_device}")
+            new_q_data = qt._qdata.to(device=target_device)
+            new_params = _move_layout_params_to_device(qt._layout_params, target_device)
+            new_qt = QuantizedTensor(new_q_data, qt._layout_type, new_params)
+            logging.debug(f"QuantizedTensor.{op_name}: Created new tensor on {target_device}")
+            return new_qt
+
+    logging.debug(f"QuantizedTensor.{op_name}: No device change needed, returning original")
+    return qt
+
+
+@register_generic_util(torch.ops.aten.detach.default)
+def generic_detach(func, args, kwargs):
+    """Detach operation - creates a detached copy of the quantized tensor."""
+    qt = args[0]
+    if isinstance(qt, QuantizedTensor):
+        return _create_transformed_qtensor(qt, lambda x: x.detach())
+    return func(*args, **kwargs)
+
+
+@register_generic_util(torch.ops.aten.clone.default)
+def generic_clone(func, args, kwargs):
+    """Clone operation - creates a deep copy of the quantized tensor."""
+    qt = args[0]
+    if isinstance(qt, QuantizedTensor):
+        return _create_transformed_qtensor(qt, lambda x: x.clone())
+    return func(*args, **kwargs)
+
+
+@register_generic_util(torch.ops.aten._to_copy.default)
+def generic_to_copy(func, args, kwargs):
+    """Device/dtype transfer operation - handles .to(device) calls."""
+    qt = args[0]
+    if isinstance(qt, QuantizedTensor):
+        return _handle_device_transfer(
+            qt,
+            target_device=kwargs.get('device', None),
+            target_dtype=kwargs.get('dtype', None),
+            op_name="_to_copy"
+        )
+    return func(*args, **kwargs)
+
+
+@register_generic_util(torch.ops.aten.to.dtype_layout)
+def generic_to_dtype_layout(func, args, kwargs):
+    """Handle .to(device) calls using the dtype_layout variant."""
+    qt = args[0]
+    if isinstance(qt, QuantizedTensor):
+        return _handle_device_transfer(
+            qt,
+            target_device=kwargs.get('device', None),
+            target_dtype=kwargs.get('dtype', None),
+            target_layout=kwargs.get('layout', None),
+            op_name="to"
+        )
+    return func(*args, **kwargs)
+
+
+@register_generic_util(torch.ops.aten.copy_.default)
+def generic_copy_(func, args, kwargs):
+    qt_dest = args[0]
+    src = args[1]
+    non_blocking = args[2] if len(args) > 2 else False
+    if isinstance(qt_dest, QuantizedTensor):
+        if isinstance(src, QuantizedTensor):
+            # Copy from another quantized tensor
+            qt_dest._qdata.copy_(src._qdata, non_blocking=non_blocking)
+            qt_dest._layout_type = src._layout_type
+            _copy_layout_params_inplace(src._layout_params, qt_dest._layout_params, non_blocking=non_blocking)
+        else:
+            # Copy from regular tensor - just copy raw data
+            qt_dest._qdata.copy_(src)
+        return qt_dest
+    return func(*args, **kwargs)
+
+
+@register_generic_util(torch.ops.aten.to.dtype)
+def generic_to_dtype(func, args, kwargs):
+    """Handle .to(dtype) calls - dtype conversion only."""
+    src = args[0]
+    if isinstance(src, QuantizedTensor):
+        # For dtype-only conversion, just change the orig_dtype, no real cast is needed
+        target_dtype = args[1] if len(args) > 1 else kwargs.get('dtype')
+        src._layout_params["orig_dtype"] = target_dtype
+        return src
+    return func(*args, **kwargs)
+
+
+@register_generic_util(torch.ops.aten._has_compatible_shallow_copy_type.default)
+def generic_has_compatible_shallow_copy_type(func, args, kwargs):
+    return True
+
+
+@register_generic_util(torch.ops.aten.empty_like.default)
+def generic_empty_like(func, args, kwargs):
+    """Empty_like operation - creates an empty tensor with the same quantized structure."""
+    qt = args[0]
+    if isinstance(qt, QuantizedTensor):
+        # Create empty tensor with same shape and dtype as the quantized data
+        hp_dtype = kwargs.pop('dtype', qt._layout_params["orig_dtype"])
+        new_qdata = torch.empty_like(qt._qdata, **kwargs)
+
+        # Handle device transfer for layout params
+        target_device = kwargs.get('device', new_qdata.device)
+        new_params = _move_layout_params_to_device(qt._layout_params, target_device)
+
+        # Update orig_dtype if dtype is specified
+        new_params['orig_dtype'] = hp_dtype
+
+        return QuantizedTensor(new_qdata, qt._layout_type, new_params)
+    return func(*args, **kwargs)
+
+# ==============================================================================
+# FP8 Layout + Operation Handlers
+# ==============================================================================
+class TensorCoreFP8Layout(QuantizedLayout):
+    """
+    Storage format:
+    - qdata: FP8 tensor (torch.float8_e4m3fn or torch.float8_e5m2)
+    - scale: Scalar tensor (float32) for dequantization
+    - orig_dtype: Original dtype before quantization (for casting back)
+    """
+    @classmethod
+    def quantize(cls, tensor, scale=None, dtype=torch.float8_e4m3fn, stochastic_rounding=0, inplace_ops=False):
         orig_dtype = tensor.dtype
-        orig_shape = tuple(tensor.shape)
-
-        if isinstance(scale, str) and scale == "recalculate":
-            scale = torch.amax(tensor.abs()).to(dtype=torch.float32) / torch.finfo(cls.FP8_DTYPE).max
-            if tensor.dtype not in [torch.float32, torch.bfloat16]:  # Prevent scale from being too small
-                tensor_info = torch.finfo(tensor.dtype)
-                scale = (1.0 / torch.clamp((1.0 / scale), min=tensor_info.min, max=tensor_info.max))
 
         if scale is None:
-            scale = torch.ones((), device=tensor.device, dtype=torch.float32)
+            scale = torch.amax(tensor.abs()) / torch.finfo(dtype).max
+
         if not isinstance(scale, torch.Tensor):
-            scale = torch.tensor(scale, device=tensor.device, dtype=torch.float32)
+            scale = torch.tensor(scale)
+        scale = scale.to(device=tensor.device, dtype=torch.float32)
+
+        if inplace_ops:
+            tensor *= (1.0 / scale).to(tensor.dtype)
+        else:
+            tensor = tensor * (1.0 / scale).to(tensor.dtype)
 
         if stochastic_rounding > 0:
-            if inplace_ops:
-                tensor *= (1.0 / scale).to(tensor.dtype)
-            else:
-                tensor = tensor * (1.0 / scale).to(tensor.dtype)
-            qdata = comfy.float.stochastic_rounding(tensor, dtype=cls.FP8_DTYPE, seed=stochastic_rounding)
+            tensor = comfy.float.stochastic_rounding(tensor, dtype=dtype, seed=stochastic_rounding)
         else:
-            qdata = ck.quantize_per_tensor_fp8(tensor, scale, cls.FP8_DTYPE)
+            lp_amax = torch.finfo(dtype).max
+            torch.clamp(tensor, min=-lp_amax, max=lp_amax, out=tensor)
+            tensor = tensor.to(dtype, memory_format=torch.contiguous_format)
 
-        params = cls.Params(scale=scale.float(), orig_dtype=orig_dtype, orig_shape=orig_shape)
-        return qdata, params
+        layout_params = {
+            'scale': scale,
+            'orig_dtype': orig_dtype
+        }
+        return tensor, layout_params
 
+    @staticmethod
+    def dequantize(qdata, scale, orig_dtype, **kwargs):
+        plain_tensor = torch.ops.aten._to_copy.default(qdata, dtype=orig_dtype)
+        plain_tensor.mul_(scale)
+        return plain_tensor
 
-class TensorCoreFP8E4M3Layout(_TensorCoreFP8LayoutBase):
-    FP8_DTYPE = torch.float8_e4m3fn
-
-
-class TensorCoreFP8E5M2Layout(_TensorCoreFP8LayoutBase):
-    FP8_DTYPE = torch.float8_e5m2
-
-
-# Backward compatibility alias - default to E4M3
-TensorCoreFP8Layout = TensorCoreFP8E4M3Layout
-
-
-# ==============================================================================
-# Registry
-# ==============================================================================
-
-register_layout_class("TensorCoreFP8Layout", TensorCoreFP8Layout)
-register_layout_class("TensorCoreFP8E4M3Layout", TensorCoreFP8E4M3Layout)
-register_layout_class("TensorCoreFP8E5M2Layout", TensorCoreFP8E5M2Layout)
-register_layout_class("TensorCoreNVFP4Layout", TensorCoreNVFP4Layout)
+    @classmethod
+    def get_plain_tensors(cls, qtensor):
+        return qtensor._qdata, qtensor._layout_params['scale']
 
 QUANT_ALGOS = {
     "float8_e4m3fn": {
         "storage_t": torch.float8_e4m3fn,
         "parameters": {"weight_scale", "input_scale"},
-        "comfy_tensor_layout": "TensorCoreFP8E4M3Layout",
-    },
-    "float8_e5m2": {
-        "storage_t": torch.float8_e5m2,
-        "parameters": {"weight_scale", "input_scale"},
-        "comfy_tensor_layout": "TensorCoreFP8E5M2Layout",
-    },
-    "nvfp4": {
-        "storage_t": torch.uint8,
-        "parameters": {"weight_scale", "weight_scale_2", "input_scale"},
-        "comfy_tensor_layout": "TensorCoreNVFP4Layout",
-        "group_size": 16,
+        "comfy_tensor_layout": "TensorCoreFP8Layout",
     },
 }
 
+LAYOUTS = {
+    "TensorCoreFP8Layout": TensorCoreFP8Layout,
+}
 
-# ==============================================================================
-# Re-exports for backward compatibility
-# ==============================================================================
 
-__all__ = [
-    "QuantizedTensor",
-    "QuantizedLayout",
-    "TensorCoreFP8Layout",
-    "TensorCoreFP8E4M3Layout",
-    "TensorCoreFP8E5M2Layout",
-    "TensorCoreNVFP4Layout",
-    "QUANT_ALGOS",
-    "register_layout_op",
-]
+@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/sampler_helpers.py

@@ -122,20 +122,20 @@ def estimate_memory(model, noise_shape, conds):
     minimum_memory_required = model.model.memory_required([noise_shape[0]] + list(noise_shape[1:]), cond_shapes=cond_shapes_min)
     return memory_required, minimum_memory_required
 
-def prepare_sampling(model: ModelPatcher, noise_shape, conds, model_options=None, force_full_load=False):
+def prepare_sampling(model: ModelPatcher, noise_shape, conds, model_options=None):
     executor = comfy.patcher_extension.WrapperExecutor.new_executor(
         _prepare_sampling,
         comfy.patcher_extension.get_all_wrappers(comfy.patcher_extension.WrappersMP.PREPARE_SAMPLING, model_options, is_model_options=True)
     )
-    return executor.execute(model, noise_shape, conds, model_options=model_options, force_full_load=force_full_load)
+    return executor.execute(model, noise_shape, conds, model_options=model_options)
 
-def _prepare_sampling(model: ModelPatcher, noise_shape, conds, model_options=None, force_full_load=False):
+def _prepare_sampling(model: ModelPatcher, noise_shape, conds, model_options=None):
     real_model: BaseModel = None
     models, inference_memory = get_additional_models(conds, model.model_dtype())
     models += get_additional_models_from_model_options(model_options)
     models += model.get_nested_additional_models()  # TODO: does this require inference_memory update?
     memory_required, minimum_memory_required = estimate_memory(model, noise_shape, conds)
-    comfy.model_management.load_models_gpu([model] + models, memory_required=memory_required + inference_memory, minimum_memory_required=minimum_memory_required + inference_memory, force_full_load=force_full_load)
+    comfy.model_management.load_models_gpu([model] + models, memory_required=memory_required + inference_memory, minimum_memory_required=minimum_memory_required + inference_memory)
     real_model = model.model
 
     return real_model, conds, models

comfy/samplers.py

@@ -720,7 +720,7 @@ class Sampler:
         sigma = float(sigmas[0])
         return math.isclose(max_sigma, sigma, rel_tol=1e-05) or sigma > max_sigma
 
-KSAMPLER_NAMES = ["euler", "euler_cfg_pp", "euler_ancestral", "euler_ancestral_cfg_pp", "heun", "heunpp2", "exp_heun_2_x0", "exp_heun_2_x0_sde", "dpm_2", "dpm_2_ancestral",
+KSAMPLER_NAMES = ["euler", "euler_cfg_pp", "euler_ancestral", "euler_ancestral_cfg_pp", "heun", "heunpp2","dpm_2", "dpm_2_ancestral",
                   "lms", "dpm_fast", "dpm_adaptive", "dpmpp_2s_ancestral", "dpmpp_2s_ancestral_cfg_pp", "dpmpp_sde", "dpmpp_sde_gpu",
                   "dpmpp_2m", "dpmpp_2m_cfg_pp", "dpmpp_2m_sde", "dpmpp_2m_sde_gpu", "dpmpp_2m_sde_heun", "dpmpp_2m_sde_heun_gpu", "dpmpp_3m_sde", "dpmpp_3m_sde_gpu", "ddpm", "lcm",
                   "ipndm", "ipndm_v", "deis", "res_multistep", "res_multistep_cfg_pp", "res_multistep_ancestral", "res_multistep_ancestral_cfg_pp",
@@ -984,6 +984,9 @@ class CFGGuider:
         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
 
+        if denoise_mask is not None:
+            denoise_mask = comfy.sampler_helpers.prepare_mask(denoise_mask, noise.shape, device)
+
         noise = noise.to(device)
         latent_image = latent_image.to(device)
         sigmas = sigmas.to(device)
@@ -1010,24 +1013,6 @@ class CFGGuider:
         else:
             latent_shapes = [latent_image.shape]
 
-        if denoise_mask is not None:
-            if denoise_mask.is_nested:
-                denoise_masks = denoise_mask.unbind()
-                denoise_masks = denoise_masks[:len(latent_shapes)]
-            else:
-                denoise_masks = [denoise_mask]
-
-            for i in range(len(denoise_masks), len(latent_shapes)):
-                denoise_masks.append(torch.ones(latent_shapes[i]))
-
-            for i in range(len(denoise_masks)):
-                denoise_masks[i] = comfy.sampler_helpers.prepare_mask(denoise_masks[i], latent_shapes[i], self.model_patcher.load_device)
-
-            if len(denoise_masks) > 1:
-                denoise_mask, _ = comfy.utils.pack_latents(denoise_masks)
-            else:
-                denoise_mask = denoise_masks[0]
-
         self.conds = {}
         for k in self.original_conds:
             self.conds[k] = list(map(lambda a: a.copy(), self.original_conds[k]))

comfy/sd.py

@@ -53,10 +53,6 @@ import comfy.text_encoders.omnigen2
 import comfy.text_encoders.qwen_image
 import comfy.text_encoders.hunyuan_image
 import comfy.text_encoders.z_image
-import comfy.text_encoders.ovis
-import comfy.text_encoders.kandinsky5
-import comfy.text_encoders.jina_clip_2
-import comfy.text_encoders.newbie
 
 import comfy.model_patcher
 import comfy.lora
@@ -101,7 +97,7 @@ def load_lora_for_models(model, clip, lora, strength_model, strength_clip):
 
 
 class CLIP:
-    def __init__(self, target=None, embedding_directory=None, no_init=False, tokenizer_data={}, parameters=0, state_dict=[], model_options={}):
+    def __init__(self, target=None, embedding_directory=None, no_init=False, tokenizer_data={}, parameters=0, model_options={}):
         if no_init:
             return
         params = target.params.copy()
@@ -129,32 +125,9 @@ class CLIP:
 
         self.tokenizer = tokenizer(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data)
         self.patcher = comfy.model_patcher.ModelPatcher(self.cond_stage_model, load_device=load_device, offload_device=offload_device)
-        #Match torch.float32 hardcode upcast in TE implemention
-        self.patcher.set_model_compute_dtype(torch.float32)
         self.patcher.hook_mode = comfy.hooks.EnumHookMode.MinVram
         self.patcher.is_clip = True
         self.apply_hooks_to_conds = None
-        if len(state_dict) > 0:
-            if isinstance(state_dict, list):
-                for c in state_dict:
-                    m, u = self.load_sd(c)
-                    if len(m) > 0:
-                        logging.warning("clip missing: {}".format(m))
-
-                    if len(u) > 0:
-                        logging.debug("clip unexpected: {}".format(u))
-            else:
-                m, u = self.load_sd(state_dict, full_model=True)
-                if len(m) > 0:
-                    m_filter = list(filter(lambda a: ".logit_scale" not in a and ".transformer.text_projection.weight" not in a, m))
-                    if len(m_filter) > 0:
-                        logging.warning("clip missing: {}".format(m))
-                    else:
-                        logging.debug("clip missing: {}".format(m))
-
-                if len(u) > 0:
-                    logging.debug("clip unexpected {}:".format(u))
-
         if params['device'] == load_device:
             model_management.load_models_gpu([self.patcher], force_full_load=True)
         self.layer_idx = None
@@ -218,8 +191,7 @@ class CLIP:
             if unprojected:
                 self.cond_stage_model.set_clip_options({"projected_pooled": False})
 
-            self.load_model(tokens)
-            self.cond_stage_model.set_clip_options({"execution_device": self.patcher.load_device})
+            self.load_model()
             all_hooks.reset()
             self.patcher.patch_hooks(None)
             if show_pbar:
@@ -266,8 +238,7 @@ class CLIP:
         if return_pooled == "unprojected":
             self.cond_stage_model.set_clip_options({"projected_pooled": False})
 
-        self.load_model(tokens)
-        self.cond_stage_model.set_clip_options({"execution_device": self.patcher.load_device})
+        self.load_model()
         o = self.cond_stage_model.encode_token_weights(tokens)
         cond, pooled = o[:2]
         if return_dict:
@@ -299,11 +270,8 @@ class CLIP:
             sd_clip[k] = sd_tokenizer[k]
         return sd_clip
 
-    def load_model(self, tokens={}):
-        memory_used = 0
-        if hasattr(self.cond_stage_model, "memory_estimation_function"):
-            memory_used = self.cond_stage_model.memory_estimation_function(tokens, device=self.patcher.load_device)
-        model_management.load_models_gpu([self.patcher], memory_required=memory_used)
+    def load_model(self):
+        model_management.load_model_gpu(self.patcher)
         return self.patcher
 
     def get_key_patches(self):
@@ -326,7 +294,6 @@ class VAE:
         self.latent_channels = 4
         self.latent_dim = 2
         self.output_channels = 3
-        self.pad_channel_value = None
         self.process_input = lambda image: image * 2.0 - 1.0
         self.process_output = lambda image: torch.clamp((image + 1.0) / 2.0, min=0.0, max=1.0)
         self.working_dtypes = [torch.bfloat16, torch.float32]
@@ -441,7 +408,6 @@ class VAE:
                 self.memory_used_decode = lambda shape, dtype: (1000 * shape[2] * 2048) * model_management.dtype_size(dtype)
                 self.latent_channels = 64
                 self.output_channels = 2
-                self.pad_channel_value = "replicate"
                 self.upscale_ratio = 2048
                 self.downscale_ratio =  2048
                 self.latent_dim = 1
@@ -479,8 +445,8 @@ class VAE:
                 self.first_stage_model = comfy.ldm.lightricks.vae.causal_video_autoencoder.VideoVAE(version=version, config=vae_config)
                 self.latent_channels = 128
                 self.latent_dim = 3
-                self.memory_used_decode = lambda shape, dtype: (1200 * shape[2] * shape[3] * shape[4] * (8 * 8 * 8)) * model_management.dtype_size(dtype)
-                self.memory_used_encode = lambda shape, dtype: (80 * max(shape[2], 7) * shape[3] * shape[4]) * model_management.dtype_size(dtype)
+                self.memory_used_decode = lambda shape, dtype: (900 * shape[2] * shape[3] * shape[4] * (8 * 8 * 8)) * model_management.dtype_size(dtype)
+                self.memory_used_encode = lambda shape, dtype: (70 * max(shape[2], 7) * shape[3] * shape[4]) * model_management.dtype_size(dtype)
                 self.upscale_ratio = (lambda a: max(0, a * 8 - 7), 32, 32)
                 self.upscale_index_formula = (8, 32, 32)
                 self.downscale_ratio = (lambda a: max(0, math.floor((a + 7) / 8)), 32, 32)
@@ -502,7 +468,7 @@ class VAE:
                                                             decoder_config={'target': "comfy.ldm.hunyuan_video.vae_refiner.Decoder", 'params': ddconfig})
 
                 self.memory_used_encode = lambda shape, dtype: (1400 * 9 * shape[-2] * shape[-1]) * model_management.dtype_size(dtype)
-                self.memory_used_decode = lambda shape, dtype: (3600 * 4 * shape[-2] * shape[-1] * 16 * 16) * model_management.dtype_size(dtype)
+                self.memory_used_decode = lambda shape, dtype: (2800 * 4 * shape[-2] * shape[-1] * 16 * 16) * model_management.dtype_size(dtype)
             elif "decoder.conv_in.conv.weight" in sd:
                 ddconfig = {'double_z': True, 'z_channels': 4, 'resolution': 256, 'in_channels': 3, 'out_ch': 3, 'ch': 128, 'ch_mult': [1, 2, 4, 4], 'num_res_blocks': 2, 'attn_resolutions': [], 'dropout': 0.0}
                 ddconfig["conv3d"] = True
@@ -514,10 +480,8 @@ class VAE:
                 self.latent_dim = 3
                 self.latent_channels = ddconfig['z_channels'] = sd["decoder.conv_in.conv.weight"].shape[1]
                 self.first_stage_model = AutoencoderKL(ddconfig=ddconfig, embed_dim=sd['post_quant_conv.weight'].shape[1])
-                #This is likely to significantly over-estimate with single image or low frame counts as the
-                #implementation is able to completely skip caching. Rework if used as an image only VAE
-                self.memory_used_decode = lambda shape, dtype: (2800 * min(8, ((shape[2] - 1) * 4) + 1) * shape[3] * shape[4] * (8 * 8)) * model_management.dtype_size(dtype)
-                self.memory_used_encode = lambda shape, dtype: (1400 * min(9, shape[2]) * shape[3] * shape[4]) * model_management.dtype_size(dtype)
+                self.memory_used_decode = lambda shape, dtype: (1500 * shape[2] * shape[3] * shape[4] * (4 * 8 * 8)) * model_management.dtype_size(dtype)
+                self.memory_used_encode = lambda shape, dtype: (900 * max(shape[2], 2) * shape[3] * shape[4]) * model_management.dtype_size(dtype)
                 self.working_dtypes = [torch.bfloat16, torch.float16, torch.float32]
             elif "decoder.unpatcher3d.wavelets" in sd:
                 self.upscale_ratio = (lambda a: max(0, a * 8 - 7), 8, 8)
@@ -553,15 +517,11 @@ class VAE:
                     self.downscale_index_formula = (4, 8, 8)
                     self.latent_dim = 3
                     self.latent_channels = 16
-                    self.output_channels = sd["encoder.conv1.weight"].shape[1]
-                    self.pad_channel_value = 1.0
-                    ddconfig = {"dim": dim, "z_dim": self.latent_channels, "dim_mult": [1, 2, 4, 4], "num_res_blocks": 2, "attn_scales": [], "temperal_downsample": [False, True, True], "image_channels": self.output_channels, "dropout": 0.0}
+                    ddconfig = {"dim": dim, "z_dim": self.latent_channels, "dim_mult": [1, 2, 4, 4], "num_res_blocks": 2, "attn_scales": [], "temperal_downsample": [False, True, True], "dropout": 0.0}
                     self.first_stage_model = comfy.ldm.wan.vae.WanVAE(**ddconfig)
                     self.working_dtypes = [torch.bfloat16, torch.float16, torch.float32]
-                    self.memory_used_encode = lambda shape, dtype: (1500 if shape[2]<=4 else 6000) * shape[3] * shape[4] * model_management.dtype_size(dtype)
-                    self.memory_used_decode = lambda shape, dtype: (2200 if shape[2]<=4 else 7000) * shape[3] * shape[4] * (8*8) * model_management.dtype_size(dtype)
-
-
+                    self.memory_used_encode = lambda shape, dtype: 6000 * shape[3] * shape[4] * model_management.dtype_size(dtype)
+                    self.memory_used_decode = lambda shape, dtype: 7000 * shape[3] * shape[4] * (8 * 8) * model_management.dtype_size(dtype)
             # Hunyuan 3d v2 2.0 & 2.1
             elif "geo_decoder.cross_attn_decoder.ln_1.bias" in sd:
 
@@ -591,7 +551,6 @@ class VAE:
                 self.memory_used_decode = lambda shape, dtype: (shape[2] * shape[3] * 87000) * model_management.dtype_size(dtype)
                 self.latent_channels = 8
                 self.output_channels = 2
-                self.pad_channel_value = "replicate"
                 self.upscale_ratio = 4096
                 self.downscale_ratio = 4096
                 self.latent_dim = 2
@@ -700,28 +659,17 @@ class VAE:
             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 self.crop_input:
-            downscale_ratio = self.spacial_compression_encode()
+        if not self.crop_input:
+            return pixels
 
-            dims = pixels.shape[1:-1]
-            for d in range(len(dims)):
-                x = (dims[d] // downscale_ratio) * downscale_ratio
-                x_offset = (dims[d] % downscale_ratio) // 2
-                if x != dims[d]:
-                    pixels = pixels.narrow(d + 1, x_offset, x)
+        downscale_ratio = self.spacial_compression_encode()
 
-        if pixels.shape[-1] > self.output_channels:
-            pixels = pixels[..., :self.output_channels]
-        elif pixels.shape[-1] < self.output_channels:
-            if self.pad_channel_value is not None:
-                if isinstance(self.pad_channel_value, str):
-                    mode = self.pad_channel_value
-                    value = None
-                else:
-                    mode = "constant"
-                    value = self.pad_channel_value
-
-                pixels = torch.nn.functional.pad(pixels, (0, self.output_channels - pixels.shape[-1]), mode=mode, value=value)
+        dims = pixels.shape[1:-1]
+        for d in range(len(dims)):
+            x = (dims[d] // downscale_ratio) * downscale_ratio
+            x_offset = (dims[d] % downscale_ratio) // 2
+            if x != dims[d]:
+                pixels = pixels.narrow(d + 1, x_offset, x)
         return pixels
 
     def decode_tiled_(self, samples, tile_x=64, tile_y=64, overlap = 16):
@@ -792,8 +740,6 @@ class VAE:
         self.throw_exception_if_invalid()
         pixel_samples = None
         do_tile = False
-        if self.latent_dim == 2 and samples_in.ndim == 5:
-            samples_in = samples_in[:, :, 0]
         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)
@@ -1010,18 +956,16 @@ class CLIPType(Enum):
     QWEN_IMAGE = 18
     HUNYUAN_IMAGE = 19
     HUNYUAN_VIDEO_15 = 20
-    OVIS = 21
-    KANDINSKY5 = 22
-    KANDINSKY5_IMAGE = 23
-    NEWBIE = 24
 
 
 def load_clip(ckpt_paths, embedding_directory=None, clip_type=CLIPType.STABLE_DIFFUSION, model_options={}):
     clip_data = []
     for p in ckpt_paths:
         sd, metadata = comfy.utils.load_torch_file(p, safe_load=True, return_metadata=True)
-        if model_options.get("custom_operations", None) is None:
-            sd, metadata = comfy.utils.convert_old_quants(sd, model_prefix="", metadata=metadata)
+        if metadata is not None:
+            quant_metadata = metadata.get("_quantization_metadata", None)
+            if quant_metadata is not None:
+                sd["_quantization_metadata"] = quant_metadata
         clip_data.append(sd)
     return load_text_encoder_state_dicts(clip_data, embedding_directory=embedding_directory, clip_type=clip_type, model_options=model_options)
 
@@ -1043,9 +987,6 @@ class TEModel(Enum):
     MISTRAL3_24B = 14
     MISTRAL3_24B_PRUNED_FLUX2 = 15
     QWEN3_4B = 16
-    QWEN3_2B = 17
-    GEMMA_3_12B = 18
-    JINA_CLIP_2 = 19
 
 
 def detect_te_model(sd):
@@ -1055,13 +996,11 @@ def detect_te_model(sd):
         return TEModel.CLIP_H
     if "text_model.encoder.layers.0.mlp.fc1.weight" in sd:
         return TEModel.CLIP_L
-    if "model.encoder.layers.0.mixer.Wqkv.weight" in sd:
-        return TEModel.JINA_CLIP_2
     if "encoder.block.23.layer.1.DenseReluDense.wi_1.weight" in sd:
         weight = sd["encoder.block.23.layer.1.DenseReluDense.wi_1.weight"]
-        if weight.shape[0] == 10240:
+        if weight.shape[-1] == 4096:
             return TEModel.T5_XXL
-        elif weight.shape[0] == 5120:
+        elif weight.shape[-1] == 2048:
             return TEModel.T5_XL
     if 'encoder.block.23.layer.1.DenseReluDense.wi.weight' in sd:
         return TEModel.T5_XXL_OLD
@@ -1071,8 +1010,6 @@ def detect_te_model(sd):
             return TEModel.BYT5_SMALL_GLYPH
         return TEModel.T5_BASE
     if 'model.layers.0.post_feedforward_layernorm.weight' in sd:
-        if 'model.layers.47.self_attn.q_norm.weight' in sd:
-            return TEModel.GEMMA_3_12B
         if 'model.layers.0.self_attn.q_norm.weight' in sd:
             return TEModel.GEMMA_3_4B
         return TEModel.GEMMA_2_2B
@@ -1083,12 +1020,9 @@ def detect_te_model(sd):
         if weight.shape[0] == 512:
             return TEModel.QWEN25_7B
     if "model.layers.0.post_attention_layernorm.weight" in sd:
-        weight = sd['model.layers.0.post_attention_layernorm.weight']
         if 'model.layers.0.self_attn.q_norm.weight' in sd:
-            if weight.shape[0] == 2560:
-                return TEModel.QWEN3_4B
-            elif weight.shape[0] == 2048:
-                return TEModel.QWEN3_2B
+            return TEModel.QWEN3_4B
+        weight = sd['model.layers.0.post_attention_layernorm.weight']
         if weight.shape[0] == 5120:
             if "model.layers.39.post_attention_layernorm.weight" in sd:
                 return TEModel.MISTRAL3_24B
@@ -1144,7 +1078,7 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip
                 clip_target.clip = comfy.text_encoders.sd3_clip.sd3_clip(clip_l=False, clip_g=True, t5=False)
                 clip_target.tokenizer = comfy.text_encoders.sd3_clip.SD3Tokenizer
             elif clip_type == CLIPType.HIDREAM:
-                clip_target.clip = comfy.text_encoders.hidream.hidream_clip(clip_l=False, clip_g=True, t5=False, llama=False, dtype_t5=None, dtype_llama=None)
+                clip_target.clip = comfy.text_encoders.hidream.hidream_clip(clip_l=False, clip_g=True, t5=False, llama=False, dtype_t5=None, dtype_llama=None, t5xxl_scaled_fp8=None, llama_scaled_fp8=None)
                 clip_target.tokenizer = comfy.text_encoders.hidream.HiDreamTokenizer
             else:
                 clip_target.clip = sdxl_clip.SDXLRefinerClipModel
@@ -1168,7 +1102,7 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip
                 tokenizer_data["spiece_model"] = clip_data[0].get("spiece_model", None)
             elif clip_type == CLIPType.HIDREAM:
                 clip_target.clip = comfy.text_encoders.hidream.hidream_clip(**t5xxl_detect(clip_data),
-                                                                        clip_l=False, clip_g=False, t5=True, llama=False, dtype_llama=None)
+                                                                        clip_l=False, clip_g=False, t5=True, llama=False, dtype_llama=None, llama_scaled_fp8=None)
                 clip_target.tokenizer = comfy.text_encoders.hidream.HiDreamTokenizer
             else: #CLIPType.MOCHI
                 clip_target.clip = comfy.text_encoders.genmo.mochi_te(**t5xxl_detect(clip_data))
@@ -1197,7 +1131,7 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip
             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)
+                                                                        clip_l=False, clip_g=False, t5=False, llama=True, dtype_t5=None, t5xxl_scaled_fp8=None)
             clip_target.tokenizer = comfy.text_encoders.hidream.HiDreamTokenizer
         elif te_model == TEModel.QWEN25_3B:
             clip_target.clip = comfy.text_encoders.omnigen2.te(**llama_detect(clip_data))
@@ -1216,19 +1150,13 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip
         elif te_model == TEModel.QWEN3_4B:
             clip_target.clip = comfy.text_encoders.z_image.te(**llama_detect(clip_data))
             clip_target.tokenizer = comfy.text_encoders.z_image.ZImageTokenizer
-        elif te_model == TEModel.QWEN3_2B:
-            clip_target.clip = comfy.text_encoders.ovis.te(**llama_detect(clip_data))
-            clip_target.tokenizer = comfy.text_encoders.ovis.OvisTokenizer
-        elif te_model == TEModel.JINA_CLIP_2:
-            clip_target.clip = comfy.text_encoders.jina_clip_2.JinaClip2TextModelWrapper
-            clip_target.tokenizer = comfy.text_encoders.jina_clip_2.JinaClip2TokenizerWrapper
         else:
             # clip_l
             if clip_type == CLIPType.SD3:
                 clip_target.clip = comfy.text_encoders.sd3_clip.sd3_clip(clip_l=True, clip_g=False, t5=False)
                 clip_target.tokenizer = comfy.text_encoders.sd3_clip.SD3Tokenizer
             elif clip_type == CLIPType.HIDREAM:
-                clip_target.clip = comfy.text_encoders.hidream.hidream_clip(clip_l=True, clip_g=False, t5=False, llama=False, dtype_t5=None, dtype_llama=None)
+                clip_target.clip = comfy.text_encoders.hidream.hidream_clip(clip_l=True, clip_g=False, t5=False, llama=False, dtype_t5=None, dtype_llama=None, t5xxl_scaled_fp8=None, llama_scaled_fp8=None)
                 clip_target.tokenizer = comfy.text_encoders.hidream.HiDreamTokenizer
             else:
                 clip_target.clip = sd1_clip.SD1ClipModel
@@ -1271,27 +1199,6 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip
         elif clip_type == CLIPType.HUNYUAN_VIDEO_15:
             clip_target.clip = comfy.text_encoders.hunyuan_image.te(**llama_detect(clip_data))
             clip_target.tokenizer = comfy.text_encoders.hunyuan_video.HunyuanVideo15Tokenizer
-        elif clip_type == CLIPType.KANDINSKY5:
-            clip_target.clip = comfy.text_encoders.kandinsky5.te(**llama_detect(clip_data))
-            clip_target.tokenizer = comfy.text_encoders.kandinsky5.Kandinsky5Tokenizer
-        elif clip_type == CLIPType.KANDINSKY5_IMAGE:
-            clip_target.clip = comfy.text_encoders.kandinsky5.te(**llama_detect(clip_data))
-            clip_target.tokenizer = comfy.text_encoders.kandinsky5.Kandinsky5TokenizerImage
-        elif clip_type == CLIPType.LTXV:
-            clip_target.clip = comfy.text_encoders.lt.ltxav_te(**llama_detect(clip_data))
-            clip_target.tokenizer = comfy.text_encoders.lt.LTXAVGemmaTokenizer
-            tokenizer_data["spiece_model"] = clip_data[0].get("spiece_model", None)
-        elif clip_type == CLIPType.NEWBIE:
-            clip_target.clip = comfy.text_encoders.newbie.te(**llama_detect(clip_data))
-            clip_target.tokenizer = comfy.text_encoders.newbie.NewBieTokenizer
-            if "model.layers.0.self_attn.q_norm.weight" in clip_data[0]:
-                clip_data_gemma = clip_data[0]
-                clip_data_jina = clip_data[1]
-            else:
-                clip_data_gemma = clip_data[1]
-                clip_data_jina = clip_data[0]
-            tokenizer_data["gemma_spiece_model"] = clip_data_gemma.get("spiece_model", None)
-            tokenizer_data["jina_spiece_model"] = clip_data_jina.get("spiece_model", None)
         else:
             clip_target.clip = sdxl_clip.SDXLClipModel
             clip_target.tokenizer = sdxl_clip.SDXLTokenizer
@@ -1304,10 +1211,19 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip
 
     parameters = 0
     for c in clip_data:
+        if "_quantization_metadata" in c:
+            c.pop("_quantization_metadata")
         parameters += comfy.utils.calculate_parameters(c)
         tokenizer_data, model_options = comfy.text_encoders.long_clipl.model_options_long_clip(c, tokenizer_data, model_options)
 
-    clip = CLIP(clip_target, embedding_directory=embedding_directory, parameters=parameters, tokenizer_data=tokenizer_data, state_dict=clip_data, model_options=model_options)
+    clip = CLIP(clip_target, embedding_directory=embedding_directory, parameters=parameters, tokenizer_data=tokenizer_data, model_options=model_options)
+    for c in clip_data:
+        m, u = clip.load_sd(c)
+        if len(m) > 0:
+            logging.warning("clip missing: {}".format(m))
+
+        if len(u) > 0:
+            logging.debug("clip unexpected: {}".format(u))
     return clip
 
 def load_gligen(ckpt_path):
@@ -1366,10 +1282,6 @@ def load_state_dict_guess_config(sd, output_vae=True, output_clip=True, output_c
     weight_dtype = comfy.utils.weight_dtype(sd, diffusion_model_prefix)
     load_device = model_management.get_torch_device()
 
-    custom_operations = model_options.get("custom_operations", None)
-    if custom_operations is None:
-        sd, metadata = comfy.utils.convert_old_quants(sd, diffusion_model_prefix, metadata=metadata)
-
     model_config = model_detection.model_config_from_unet(sd, diffusion_model_prefix, metadata=metadata)
     if model_config is None:
         logging.warning("Warning, This is not a checkpoint file, trying to load it as a diffusion model only.")
@@ -1378,22 +1290,18 @@ def load_state_dict_guess_config(sd, output_vae=True, output_clip=True, output_c
             return None
         return (diffusion_model, None, VAE(sd={}), None)  # The VAE object is there to throw an exception if it's actually used'
 
+
     unet_weight_dtype = list(model_config.supported_inference_dtypes)
-    if model_config.quant_config is not None:
+    if model_config.scaled_fp8 is not None:
         weight_dtype = None
 
-    if custom_operations is not None:
-        model_config.custom_operations = custom_operations
-
+    model_config.custom_operations = model_options.get("custom_operations", None)
     unet_dtype = model_options.get("dtype", model_options.get("weight_dtype", None))
 
     if unet_dtype is None:
         unet_dtype = model_management.unet_dtype(model_params=parameters, supported_dtypes=unet_weight_dtype, weight_dtype=weight_dtype)
 
-    if model_config.quant_config is not None:
-        manual_cast_dtype = model_management.unet_manual_cast(None, load_device, model_config.supported_inference_dtypes)
-    else:
-        manual_cast_dtype = model_management.unet_manual_cast(unet_dtype, load_device, model_config.supported_inference_dtypes)
+    manual_cast_dtype = model_management.unet_manual_cast(unet_dtype, load_device, model_config.supported_inference_dtypes)
     model_config.set_inference_dtype(unet_dtype, manual_cast_dtype)
 
     if model_config.clip_vision_prefix is not None:
@@ -1411,33 +1319,22 @@ def load_state_dict_guess_config(sd, output_vae=True, output_clip=True, output_c
         vae = VAE(sd=vae_sd, metadata=metadata)
 
     if output_clip:
-        if te_model_options.get("custom_operations", None) is None:
-            scaled_fp8_list = []
-            for k in list(sd.keys()):  # Convert scaled fp8 to mixed ops
-                if k.endswith(".scaled_fp8"):
-                    scaled_fp8_list.append(k[:-len("scaled_fp8")])
-
-            if len(scaled_fp8_list) > 0:
-                out_sd = {}
-                for k in sd:
-                    skip = False
-                    for pref in scaled_fp8_list:
-                        skip = skip or k.startswith(pref)
-                    if not skip:
-                        out_sd[k] = sd[k]
-
-                for pref in scaled_fp8_list:
-                    quant_sd, qmetadata = comfy.utils.convert_old_quants(sd, pref, metadata={})
-                    for k in quant_sd:
-                        out_sd[k] = quant_sd[k]
-                    sd = out_sd
-
         clip_target = model_config.clip_target(state_dict=sd)
         if clip_target is not None:
             clip_sd = model_config.process_clip_state_dict(sd)
             if len(clip_sd) > 0:
                 parameters = comfy.utils.calculate_parameters(clip_sd)
-                clip = CLIP(clip_target, embedding_directory=embedding_directory, tokenizer_data=clip_sd, parameters=parameters, state_dict=clip_sd, model_options=te_model_options)
+                clip = CLIP(clip_target, embedding_directory=embedding_directory, tokenizer_data=clip_sd, parameters=parameters, model_options=te_model_options)
+                m, u = clip.load_sd(clip_sd, full_model=True)
+                if len(m) > 0:
+                    m_filter = list(filter(lambda a: ".logit_scale" not in a and ".transformer.text_projection.weight" not in a, m))
+                    if len(m_filter) > 0:
+                        logging.warning("clip missing: {}".format(m))
+                    else:
+                        logging.debug("clip missing: {}".format(m))
+
+                if len(u) > 0:
+                    logging.debug("clip unexpected {}:".format(u))
             else:
                 logging.warning("no CLIP/text encoder weights in checkpoint, the text encoder model will not be loaded.")
 
@@ -1484,9 +1381,6 @@ def load_diffusion_model_state_dict(sd, model_options={}, metadata=None):
     if len(temp_sd) > 0:
         sd = temp_sd
 
-    custom_operations = model_options.get("custom_operations", None)
-    if custom_operations is None:
-        sd, metadata = comfy.utils.convert_old_quants(sd, "", metadata=metadata)
     parameters = comfy.utils.calculate_parameters(sd)
     weight_dtype = comfy.utils.weight_dtype(sd)
 
@@ -1517,7 +1411,7 @@ def load_diffusion_model_state_dict(sd, model_options={}, metadata=None):
 
     offload_device = model_management.unet_offload_device()
     unet_weight_dtype = list(model_config.supported_inference_dtypes)
-    if model_config.quant_config is not None:
+    if model_config.scaled_fp8 is not None:
         weight_dtype = None
 
     if dtype is None:
@@ -1525,15 +1419,12 @@ def load_diffusion_model_state_dict(sd, model_options={}, metadata=None):
     else:
         unet_dtype = dtype
 
-    if model_config.quant_config is not None:
+    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)
-
-    if custom_operations is not None:
-        model_config.custom_operations = custom_operations
-
+    model_config.custom_operations = model_options.get("custom_operations", model_config.custom_operations)
     if model_options.get("fp8_optimizations", False):
         model_config.optimizations["fp8"] = True
 
@@ -1572,9 +1463,6 @@ def save_checkpoint(output_path, model, clip=None, vae=None, clip_vision=None, m
     if vae is not None:
         vae_sd = vae.get_sd()
 
-    if metadata is None:
-        metadata = {}
-
     model_management.load_models_gpu(load_models, force_patch_weights=True)
     clip_vision_sd = clip_vision.get_sd() if clip_vision is not None else None
     sd = model.model.state_dict_for_saving(clip_sd, vae_sd, clip_vision_sd)

comfy/sd1_clip.py

@@ -107,17 +107,29 @@ class SDClipModel(torch.nn.Module, ClipTokenWeightEncoder):
             config[k] = v
 
         operations = model_options.get("custom_operations", None)
-        quant_config = model_options.get("quantization_metadata", None)
+        scaled_fp8 = None
+        quantization_metadata = model_options.get("quantization_metadata", None)
 
         if operations is None:
-            if quant_config is not None:
-                operations = comfy.ops.mixed_precision_ops(quant_config, dtype, full_precision_mm=True)
-                logging.info("Using MixedPrecisionOps for text encoder")
+            layer_quant_config = None
+            if quantization_metadata is not None:
+                layer_quant_config = json.loads(quantization_metadata).get("layers", None)
+
+            if layer_quant_config is not None:
+                operations = comfy.ops.mixed_precision_ops(layer_quant_config, dtype, full_precision_mm=True)
+                logging.info(f"Using MixedPrecisionOps for text encoder: {len(layer_quant_config)} quantized layers")
             else:
-                operations = comfy.ops.manual_cast
+                # Fallback to scaled_fp8_ops for backward compatibility
+                scaled_fp8 = model_options.get("scaled_fp8", None)
+                if scaled_fp8 is not None:
+                    operations = comfy.ops.scaled_fp8_ops(fp8_matrix_mult=False, override_dtype=scaled_fp8)
+                else:
+                    operations = comfy.ops.manual_cast
 
         self.operations = operations
         self.transformer = model_class(config, dtype, device, self.operations)
+        if scaled_fp8 is not None:
+            self.transformer.scaled_fp8 = torch.nn.Parameter(torch.tensor([], dtype=scaled_fp8))
 
         self.num_layers = self.transformer.num_layers
 
@@ -135,7 +147,6 @@ class SDClipModel(torch.nn.Module, ClipTokenWeightEncoder):
         self.layer_norm_hidden_state = layer_norm_hidden_state
         self.return_projected_pooled = return_projected_pooled
         self.return_attention_masks = return_attention_masks
-        self.execution_device = None
 
         if layer == "hidden":
             assert layer_idx is not None
@@ -152,7 +163,6 @@ class SDClipModel(torch.nn.Module, ClipTokenWeightEncoder):
     def set_clip_options(self, options):
         layer_idx = options.get("layer", self.layer_idx)
         self.return_projected_pooled = options.get("projected_pooled", self.return_projected_pooled)
-        self.execution_device = options.get("execution_device", self.execution_device)
         if isinstance(self.layer, list) or self.layer == "all":
             pass
         elif layer_idx is None or abs(layer_idx) > self.num_layers:
@@ -165,7 +175,6 @@ class SDClipModel(torch.nn.Module, ClipTokenWeightEncoder):
         self.layer = self.options_default[0]
         self.layer_idx = self.options_default[1]
         self.return_projected_pooled = self.options_default[2]
-        self.execution_device = None
 
     def process_tokens(self, tokens, device):
         end_token = self.special_tokens.get("end", None)
@@ -249,11 +258,7 @@ class SDClipModel(torch.nn.Module, ClipTokenWeightEncoder):
         return torch.cat(embeds_out), torch.tensor(attention_masks, device=device, dtype=torch.long), num_tokens, embeds_info
 
     def forward(self, tokens):
-        if self.execution_device is None:
-            device = self.transformer.get_input_embeddings().weight.device
-        else:
-            device = self.execution_device
-
+        device = self.transformer.get_input_embeddings().weight.device
         embeds, attention_mask, num_tokens, embeds_info = self.process_tokens(tokens, device)
 
         attention_mask_model = None
@@ -466,7 +471,7 @@ def load_embed(embedding_name, embedding_directory, embedding_size, embed_key=No
     return embed_out
 
 class SDTokenizer:
-    def __init__(self, tokenizer_path=None, max_length=77, pad_with_end=True, embedding_directory=None, embedding_size=768, embedding_key='clip_l', tokenizer_class=CLIPTokenizer, has_start_token=True, has_end_token=True, pad_to_max_length=True, min_length=None, pad_token=None, end_token=None, min_padding=None, pad_left=False, disable_weights=False, tokenizer_data={}, tokenizer_args={}):
+    def __init__(self, tokenizer_path=None, max_length=77, pad_with_end=True, embedding_directory=None, embedding_size=768, embedding_key='clip_l', tokenizer_class=CLIPTokenizer, has_start_token=True, has_end_token=True, pad_to_max_length=True, min_length=None, pad_token=None, end_token=None, min_padding=None, pad_left=False, tokenizer_data={}, tokenizer_args={}):
         if tokenizer_path is None:
             tokenizer_path = os.path.join(os.path.dirname(os.path.realpath(__file__)), "sd1_tokenizer")
         self.tokenizer = tokenizer_class.from_pretrained(tokenizer_path, **tokenizer_args)
@@ -513,8 +518,6 @@ class SDTokenizer:
         self.embedding_size = embedding_size
         self.embedding_key = embedding_key
 
-        self.disable_weights = disable_weights
-
     def _try_get_embedding(self, embedding_name:str):
         '''
         Takes a potential embedding name and tries to retrieve it.
@@ -549,7 +552,7 @@ class SDTokenizer:
         min_padding = tokenizer_options.get("{}_min_padding".format(self.embedding_key), self.min_padding)
 
         text = escape_important(text)
-        if kwargs.get("disable_weights", self.disable_weights):
+        if kwargs.get("disable_weights", False):
             parsed_weights = [(text, 1.0)]
         else:
             parsed_weights = token_weights(text, 1.0)

comfy/supported_models.py

@@ -21,14 +21,12 @@ import comfy.text_encoders.ace
 import comfy.text_encoders.omnigen2
 import comfy.text_encoders.qwen_image
 import comfy.text_encoders.hunyuan_image
-import comfy.text_encoders.kandinsky5
 import comfy.text_encoders.z_image
 
 from . import supported_models_base
 from . import latent_formats
 
 from . import diffusers_convert
-import comfy.model_management
 
 class SD15(supported_models_base.BASE):
     unet_config = {
@@ -542,7 +540,7 @@ class SD3(supported_models_base.BASE):
     unet_extra_config = {}
     latent_format = latent_formats.SD3
 
-    memory_usage_factor = 1.6
+    memory_usage_factor = 1.2
 
     text_encoder_key_prefix = ["text_encoders."]
 
@@ -836,21 +834,6 @@ class LTXV(supported_models_base.BASE):
         t5_detect = comfy.text_encoders.sd3_clip.t5_xxl_detect(state_dict, "{}t5xxl.transformer.".format(pref))
         return supported_models_base.ClipTarget(comfy.text_encoders.lt.LTXVT5Tokenizer, comfy.text_encoders.lt.ltxv_te(**t5_detect))
 
-class LTXAV(LTXV):
-    unet_config = {
-        "image_model": "ltxav",
-    }
-
-    latent_format = latent_formats.LTXAV
-
-    def __init__(self, unet_config):
-        super().__init__(unet_config)
-        self.memory_usage_factor = 0.061  # TODO
-
-    def get_model(self, state_dict, prefix="", device=None):
-        out = model_base.LTXAV(self, device=device)
-        return out
-
 class HunyuanVideo(supported_models_base.BASE):
     unet_config = {
         "image_model": "hunyuan_video",
@@ -981,7 +964,7 @@ class CosmosT2IPredict2(supported_models_base.BASE):
 
     def __init__(self, unet_config):
         super().__init__(unet_config)
-        self.memory_usage_factor = (unet_config.get("model_channels", 2048) / 2048) * 0.95
+        self.memory_usage_factor = (unet_config.get("model_channels", 2048) / 2048) * 0.9
 
     def get_model(self, state_dict, prefix="", device=None):
         out = model_base.CosmosPredict2(self, device=device)
@@ -1042,15 +1025,7 @@ class ZImage(Lumina2):
         "shift": 3.0,
     }
 
-    memory_usage_factor = 2.0
-
-    supported_inference_dtypes = [torch.bfloat16, torch.float32]
-
-    def __init__(self, unet_config):
-        super().__init__(unet_config)
-        if comfy.model_management.extended_fp16_support():
-            self.supported_inference_dtypes = self.supported_inference_dtypes.copy()
-            self.supported_inference_dtypes.insert(1, torch.float16)
+    memory_usage_factor = 1.7
 
     def clip_target(self, state_dict={}):
         pref = self.text_encoder_key_prefix[0]
@@ -1311,7 +1286,7 @@ class ChromaRadiance(Chroma):
     latent_format = comfy.latent_formats.ChromaRadiance
 
     # Pixel-space model, no spatial compression for model input.
-    memory_usage_factor = 0.044
+    memory_usage_factor = 0.038
 
     def get_model(self, state_dict, prefix="", device=None):
         return model_base.ChromaRadiance(self, device=device)
@@ -1354,7 +1329,7 @@ class Omnigen2(supported_models_base.BASE):
         "shift": 2.6,
     }
 
-    memory_usage_factor = 1.95 #TODO
+    memory_usage_factor = 1.65 #TODO
 
     unet_extra_config = {}
     latent_format = latent_formats.Flux
@@ -1419,7 +1394,7 @@ class HunyuanImage21(HunyuanVideo):
 
     latent_format = latent_formats.HunyuanImage21
 
-    memory_usage_factor = 8.7
+    memory_usage_factor = 7.7
 
     supported_inference_dtypes = [torch.bfloat16, torch.float32]
 
@@ -1497,60 +1472,7 @@ class HunyuanVideo15_SR_Distilled(HunyuanVideo):
         hunyuan_detect = comfy.text_encoders.hunyuan_video.llama_detect(state_dict, "{}qwen25_7b.transformer.".format(pref))
         return supported_models_base.ClipTarget(comfy.text_encoders.hunyuan_video.HunyuanVideo15Tokenizer, comfy.text_encoders.hunyuan_image.te(**hunyuan_detect))
 
+models = [LotusD, Stable_Zero123, SD15_instructpix2pix, SD15, SD20, SD21UnclipL, SD21UnclipH, SDXL_instructpix2pix, SDXLRefiner, SDXL, SSD1B, KOALA_700M, KOALA_1B, Segmind_Vega, SD_X4Upscaler, Stable_Cascade_C, Stable_Cascade_B, SV3D_u, SV3D_p, SD3, StableAudio, AuraFlow, PixArtAlpha, PixArtSigma, HunyuanDiT, HunyuanDiT1, FluxInpaint, Flux, FluxSchnell, GenmoMochi, LTXV, HunyuanVideo15_SR_Distilled, HunyuanVideo15, HunyuanImage21Refiner, HunyuanImage21, HunyuanVideoSkyreelsI2V, HunyuanVideoI2V, HunyuanVideo, CosmosT2V, CosmosI2V, CosmosT2IPredict2, CosmosI2VPredict2, ZImage, Lumina2, WAN22_T2V, WAN21_T2V, WAN21_I2V, WAN21_FunControl2V, WAN21_Vace, WAN21_Camera, WAN22_Camera, WAN22_S2V, WAN21_HuMo, WAN22_Animate, Hunyuan3Dv2mini, Hunyuan3Dv2, Hunyuan3Dv2_1, HiDream, Chroma, ChromaRadiance, ACEStep, Omnigen2, QwenImage, Flux2]
 
-class Kandinsky5(supported_models_base.BASE):
-    unet_config = {
-        "image_model": "kandinsky5",
-    }
-
-    sampling_settings = {
-        "shift": 10.0,
-    }
-
-    unet_extra_config = {}
-    latent_format = latent_formats.HunyuanVideo
-
-    memory_usage_factor = 1.25 #TODO
-
-    supported_inference_dtypes = [torch.bfloat16, torch.float32]
-
-    vae_key_prefix = ["vae."]
-    text_encoder_key_prefix = ["text_encoders."]
-
-    def get_model(self, state_dict, prefix="", device=None):
-        out = model_base.Kandinsky5(self, device=device)
-        return out
-
-    def clip_target(self, state_dict={}):
-        pref = self.text_encoder_key_prefix[0]
-        hunyuan_detect = comfy.text_encoders.hunyuan_video.llama_detect(state_dict, "{}qwen25_7b.transformer.".format(pref))
-        return supported_models_base.ClipTarget(comfy.text_encoders.kandinsky5.Kandinsky5Tokenizer, comfy.text_encoders.kandinsky5.te(**hunyuan_detect))
-
-
-class Kandinsky5Image(Kandinsky5):
-    unet_config = {
-        "image_model": "kandinsky5",
-        "model_dim": 2560,
-        "visual_embed_dim": 64,
-    }
-
-    sampling_settings = {
-        "shift": 3.0,
-    }
-
-    latent_format = latent_formats.Flux
-    memory_usage_factor = 1.25 #TODO
-
-    def get_model(self, state_dict, prefix="", device=None):
-        out = model_base.Kandinsky5Image(self, device=device)
-        return out
-
-    def clip_target(self, state_dict={}):
-        pref = self.text_encoder_key_prefix[0]
-        hunyuan_detect = comfy.text_encoders.hunyuan_video.llama_detect(state_dict, "{}qwen25_7b.transformer.".format(pref))
-        return supported_models_base.ClipTarget(comfy.text_encoders.kandinsky5.Kandinsky5TokenizerImage, comfy.text_encoders.kandinsky5.te(**hunyuan_detect))
-
-
-models = [LotusD, Stable_Zero123, SD15_instructpix2pix, SD15, SD20, SD21UnclipL, SD21UnclipH, SDXL_instructpix2pix, SDXLRefiner, SDXL, SSD1B, KOALA_700M, KOALA_1B, Segmind_Vega, SD_X4Upscaler, Stable_Cascade_C, Stable_Cascade_B, SV3D_u, SV3D_p, SD3, StableAudio, AuraFlow, PixArtAlpha, PixArtSigma, HunyuanDiT, HunyuanDiT1, FluxInpaint, Flux, FluxSchnell, GenmoMochi, LTXV, LTXAV, HunyuanVideo15_SR_Distilled, HunyuanVideo15, HunyuanImage21Refiner, HunyuanImage21, HunyuanVideoSkyreelsI2V, HunyuanVideoI2V, HunyuanVideo, CosmosT2V, CosmosI2V, CosmosT2IPredict2, CosmosI2VPredict2, ZImage, Lumina2, WAN22_T2V, WAN21_T2V, WAN21_I2V, WAN21_FunControl2V, WAN21_Vace, WAN21_Camera, WAN22_Camera, WAN22_S2V, WAN21_HuMo, WAN22_Animate, Hunyuan3Dv2mini, Hunyuan3Dv2, Hunyuan3Dv2_1, HiDream, Chroma, ChromaRadiance, ACEStep, Omnigen2, QwenImage, Flux2, Kandinsky5Image, Kandinsky5]
 
 models += [SVD_img2vid]

comfy/supported_models_base.py

@@ -17,7 +17,6 @@
 """
 
 import torch
-import logging
 from . import model_base
 from . import utils
 from . import latent_formats
@@ -50,7 +49,8 @@ class BASE:
 
     manual_cast_dtype = None
     custom_operations = None
-    quant_config = None  # quantization configuration for mixed precision
+    scaled_fp8 = None
+    layer_quant_config = None  # Per-layer quantization configuration for mixed precision
     optimizations = {"fp8": False}
 
     @classmethod
@@ -118,7 +118,3 @@ class BASE:
     def set_inference_dtype(self, dtype, manual_cast_dtype):
         self.unet_config['dtype'] = dtype
         self.manual_cast_dtype = manual_cast_dtype
-
-    def __getattr__(self, name):
-        logging.warning("\nWARNING, you accessed {} from the model config object which doesn't exist. Please fix your code.\n".format(name))
-        return None

comfy/taesd/taehv.py

@@ -154,8 +154,7 @@ class TAEHV(nn.Module):
             self._show_progress_bar = value
 
     def encode(self, x, **kwargs):
-        if self.patch_size > 1:
-            x = F.pixel_unshuffle(x, self.patch_size)
+        if self.patch_size > 1: x = F.pixel_unshuffle(x, self.patch_size)
         x = x.movedim(2, 1)  # [B, C, T, H, W] -> [B, T, C, H, W]
         if x.shape[1] % 4 != 0:
             # pad at end to multiple of 4
@@ -168,6 +167,5 @@ class TAEHV(nn.Module):
     def decode(self, x, **kwargs):
         x = self.process_in(x).movedim(2, 1)  # [B, C, T, H, W] -> [B, T, C, H, W]
         x = apply_model_with_memblocks(self.decoder, x, self.parallel, self.show_progress_bar)
-        if self.patch_size > 1:
-            x = F.pixel_shuffle(x, self.patch_size)
+        if self.patch_size > 1: x = F.pixel_shuffle(x, self.patch_size)
         return x[:, self.frames_to_trim:].movedim(2, 1)

comfy/text_encoders/cosmos.py

@@ -7,10 +7,10 @@ from transformers import T5TokenizerFast
 class T5XXLModel(sd1_clip.SDClipModel):
     def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None, attention_mask=True, model_options={}):
         textmodel_json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "t5_old_config_xxl.json")
-        t5xxl_quantization_metadata = model_options.get("t5xxl_quantization_metadata", None)
-        if t5xxl_quantization_metadata is not None:
+        t5xxl_scaled_fp8 = model_options.get("t5xxl_scaled_fp8", None)
+        if t5xxl_scaled_fp8 is not None:
             model_options = model_options.copy()
-            model_options["quantization_metadata"] = t5xxl_quantization_metadata
+            model_options["scaled_fp8"] = t5xxl_scaled_fp8
 
         super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype, special_tokens={"end": 1, "pad": 0}, model_class=comfy.text_encoders.t5.T5, enable_attention_masks=attention_mask, return_attention_masks=attention_mask, zero_out_masked=attention_mask, model_options=model_options)
 
@@ -30,13 +30,13 @@ class CosmosT5Tokenizer(sd1_clip.SD1Tokenizer):
         super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, clip_name="t5xxl", tokenizer=T5XXLTokenizer)
 
 
-def te(dtype_t5=None, t5_quantization_metadata=None):
+def te(dtype_t5=None, t5xxl_scaled_fp8=None):
     class CosmosTEModel_(CosmosT5XXL):
         def __init__(self, device="cpu", dtype=None, model_options={}):
-            if t5_quantization_metadata is not None:
+            if t5xxl_scaled_fp8 is not None and "t5xxl_scaled_fp8" not in model_options:
                 model_options = model_options.copy()
-                model_options["t5xxl_quantization_metadata"] = t5_quantization_metadata
-            if dtype_t5 is not None:
+                model_options["t5xxl_scaled_fp8"] = t5xxl_scaled_fp8
+            if dtype is None:
                 dtype = dtype_t5
             super().__init__(device=device, dtype=dtype, model_options=model_options)
     return CosmosTEModel_

comfy/text_encoders/flux.py

@@ -63,12 +63,12 @@ class FluxClipModel(torch.nn.Module):
         else:
             return self.t5xxl.load_sd(sd)
 
-def flux_clip(dtype_t5=None, t5_quantization_metadata=None):
+def flux_clip(dtype_t5=None, t5xxl_scaled_fp8=None):
     class FluxClipModel_(FluxClipModel):
         def __init__(self, device="cpu", dtype=None, model_options={}):
-            if t5_quantization_metadata is not None:
+            if t5xxl_scaled_fp8 is not None and "t5xxl_scaled_fp8" not in model_options:
                 model_options = model_options.copy()
-                model_options["t5xxl_quantization_metadata"] = t5_quantization_metadata
+                model_options["t5xxl_scaled_fp8"] = t5xxl_scaled_fp8
             super().__init__(dtype_t5=dtype_t5, device=device, dtype=dtype, model_options=model_options)
     return FluxClipModel_
 
@@ -159,13 +159,15 @@ class Flux2TEModel(sd1_clip.SD1ClipModel):
         out = out.reshape(out.shape[0], out.shape[1], -1)
         return out, pooled, extra
 
-def flux2_te(dtype_llama=None, llama_quantization_metadata=None, pruned=False):
+def flux2_te(dtype_llama=None, llama_scaled_fp8=None, llama_quantization_metadata=None, pruned=False):
     class Flux2TEModel_(Flux2TEModel):
         def __init__(self, device="cpu", dtype=None, model_options={}):
+            if llama_scaled_fp8 is not None and "scaled_fp8" not in model_options:
+                model_options = model_options.copy()
+                model_options["scaled_fp8"] = llama_scaled_fp8
             if dtype_llama is not None:
                 dtype = dtype_llama
             if llama_quantization_metadata is not None:
-                model_options = model_options.copy()
                 model_options["quantization_metadata"] = llama_quantization_metadata
             if pruned:
                 model_options = model_options.copy()

comfy/text_encoders/genmo.py

@@ -26,13 +26,13 @@ class MochiT5Tokenizer(sd1_clip.SD1Tokenizer):
         super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, clip_name="t5xxl", tokenizer=T5XXLTokenizer)
 
 
-def mochi_te(dtype_t5=None, t5_quantization_metadata=None):
+def mochi_te(dtype_t5=None, t5xxl_scaled_fp8=None):
     class MochiTEModel_(MochiT5XXL):
         def __init__(self, device="cpu", dtype=None, model_options={}):
-            if t5_quantization_metadata is not None:
+            if t5xxl_scaled_fp8 is not None and "t5xxl_scaled_fp8" not in model_options:
                 model_options = model_options.copy()
-                model_options["t5xxl_quantization_metadata"] = t5_quantization_metadata
-            if dtype_t5 is not None:
+                model_options["t5xxl_scaled_fp8"] = t5xxl_scaled_fp8
+            if dtype is None:
                 dtype = dtype_t5
             super().__init__(device=device, dtype=dtype, model_options=model_options)
     return MochiTEModel_

comfy/text_encoders/hidream.py

@@ -142,14 +142,14 @@ class HiDreamTEModel(torch.nn.Module):
             return self.llama.load_sd(sd)
 
 
-def hidream_clip(clip_l=True, clip_g=True, t5=True, llama=True, dtype_t5=None, dtype_llama=None, t5_quantization_metadata=None, llama_quantization_metadata=None):
+def hidream_clip(clip_l=True, clip_g=True, t5=True, llama=True, dtype_t5=None, dtype_llama=None, t5xxl_scaled_fp8=None, llama_scaled_fp8=None):
     class HiDreamTEModel_(HiDreamTEModel):
         def __init__(self, device="cpu", dtype=None, model_options={}):
-            if t5_quantization_metadata is not None:
+            if t5xxl_scaled_fp8 is not None and "t5xxl_scaled_fp8" not in model_options:
                 model_options = model_options.copy()
-                model_options["t5xxl_quantization_metadata"] = t5_quantization_metadata
-            if llama_quantization_metadata is not None:
+                model_options["t5xxl_scaled_fp8"] = t5xxl_scaled_fp8
+            if llama_scaled_fp8 is not None and "llama_scaled_fp8" not in model_options:
                 model_options = model_options.copy()
-                model_options["llama_quantization_metadata"] = llama_quantization_metadata
+                model_options["llama_scaled_fp8"] = llama_scaled_fp8
             super().__init__(clip_l=clip_l, clip_g=clip_g, t5=t5, llama=llama, dtype_t5=dtype_t5, dtype_llama=dtype_llama, device=device, dtype=dtype, model_options=model_options)
     return HiDreamTEModel_

comfy/text_encoders/hunyuan_image.py

@@ -40,10 +40,10 @@ class HunyuanImageTokenizer(QwenImageTokenizer):
 
 class Qwen25_7BVLIModel(sd1_clip.SDClipModel):
     def __init__(self, device="cpu", layer="hidden", layer_idx=-3, dtype=None, attention_mask=True, model_options={}):
-        llama_quantization_metadata = model_options.get("llama_quantization_metadata", None)
-        if llama_quantization_metadata is not None:
+        llama_scaled_fp8 = model_options.get("qwen_scaled_fp8", None)
+        if llama_scaled_fp8 is not None:
             model_options = model_options.copy()
-            model_options["quantization_metadata"] = llama_quantization_metadata
+            model_options["scaled_fp8"] = llama_scaled_fp8
         super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config={}, dtype=dtype, special_tokens={"pad": 151643}, layer_norm_hidden_state=False, model_class=comfy.text_encoders.llama.Qwen25_7BVLI, enable_attention_masks=attention_mask, return_attention_masks=attention_mask, model_options=model_options)
 
 
@@ -91,12 +91,12 @@ class HunyuanImageTEModel(QwenImageTEModel):
         else:
             return super().load_sd(sd)
 
-def te(byt5=True, dtype_llama=None, llama_quantization_metadata=None):
+def te(byt5=True, dtype_llama=None, llama_scaled_fp8=None):
     class QwenImageTEModel_(HunyuanImageTEModel):
         def __init__(self, device="cpu", dtype=None, model_options={}):
-            if llama_quantization_metadata is not None:
+            if llama_scaled_fp8 is not None and "scaled_fp8" not in model_options:
                 model_options = model_options.copy()
-                model_options["llama_quantization_metadata"] = llama_quantization_metadata
+                model_options["qwen_scaled_fp8"] = llama_scaled_fp8
             if dtype_llama is not None:
                 dtype = dtype_llama
             super().__init__(byt5=byt5, device=device, dtype=dtype, model_options=model_options)

comfy/text_encoders/hunyuan_video.py

@@ -6,7 +6,7 @@ from transformers import LlamaTokenizerFast
 import torch
 import os
 import numbers
-import comfy.utils
+
 
 def llama_detect(state_dict, prefix=""):
     out = {}
@@ -14,9 +14,12 @@ def llama_detect(state_dict, prefix=""):
     if t5_key in state_dict:
         out["dtype_llama"] = state_dict[t5_key].dtype
 
-    quant = comfy.utils.detect_layer_quantization(state_dict, prefix)
-    if quant is not None:
-        out["llama_quantization_metadata"] = quant
+    scaled_fp8_key = "{}scaled_fp8".format(prefix)
+    if scaled_fp8_key in state_dict:
+        out["llama_scaled_fp8"] = state_dict[scaled_fp8_key].dtype
+
+    if "_quantization_metadata" in state_dict:
+        out["llama_quantization_metadata"] = state_dict["_quantization_metadata"]
 
     return out
 
@@ -28,10 +31,10 @@ class LLAMA3Tokenizer(sd1_clip.SDTokenizer):
 
 class LLAMAModel(sd1_clip.SDClipModel):
     def __init__(self, device="cpu", layer="hidden", layer_idx=-3, dtype=None, attention_mask=True, model_options={}, special_tokens={"start": 128000, "pad": 128258}):
-        llama_quantization_metadata = model_options.get("llama_quantization_metadata", None)
-        if llama_quantization_metadata is not None:
+        llama_scaled_fp8 = model_options.get("llama_scaled_fp8", None)
+        if llama_scaled_fp8 is not None:
             model_options = model_options.copy()
-            model_options["quantization_metadata"] = llama_quantization_metadata
+            model_options["scaled_fp8"] = llama_scaled_fp8
 
         textmodel_json_config = {}
         vocab_size = model_options.get("vocab_size", None)
@@ -158,11 +161,11 @@ class HunyuanVideoClipModel(torch.nn.Module):
             return self.llama.load_sd(sd)
 
 
-def hunyuan_video_clip(dtype_llama=None, llama_quantization_metadata=None):
+def hunyuan_video_clip(dtype_llama=None, llama_scaled_fp8=None):
     class HunyuanVideoClipModel_(HunyuanVideoClipModel):
         def __init__(self, device="cpu", dtype=None, model_options={}):
-            if llama_quantization_metadata is not None:
+            if llama_scaled_fp8 is not None and "llama_scaled_fp8" not in model_options:
                 model_options = model_options.copy()
-                model_options["llama_quantization_metadata"] = llama_quantization_metadata
+                model_options["llama_scaled_fp8"] = llama_scaled_fp8
             super().__init__(dtype_llama=dtype_llama, device=device, dtype=dtype, model_options=model_options)
     return HunyuanVideoClipModel_

comfy/text_encoders/jina_clip_2.py

@@ -1,219 +0,0 @@
-# Jina CLIP v2 and Jina Embeddings v3 both use their modified XLM-RoBERTa architecture. Reference implementation:
-# Jina CLIP v2 (both text and vision): https://huggingface.co/jinaai/jina-clip-implementation/blob/39e6a55ae971b59bea6e44675d237c99762e7ee2/modeling_clip.py
-# Jina XLM-RoBERTa (text only): http://huggingface.co/jinaai/xlm-roberta-flash-implementation/blob/2b6bc3f30750b3a9648fe9b63448c09920efe9be/modeling_xlm_roberta.py
-
-from dataclasses import dataclass
-
-import torch
-from torch import nn as nn
-from torch.nn import functional as F
-
-import comfy.model_management
-import comfy.ops
-from comfy import sd1_clip
-from .spiece_tokenizer import SPieceTokenizer
-
-class JinaClip2Tokenizer(sd1_clip.SDTokenizer):
-    def __init__(self, embedding_directory=None, tokenizer_data={}):
-        tokenizer = tokenizer_data.get("spiece_model", None)
-        # The official NewBie uses max_length=8000, but Jina Embeddings v3 actually supports 8192
-        super().__init__(tokenizer, pad_with_end=False, embedding_size=1024, embedding_key='jina_clip_2', tokenizer_class=SPieceTokenizer, has_start_token=True, has_end_token=True, pad_to_max_length=False, max_length=8192, min_length=1, pad_token=1, end_token=2, tokenizer_args={"add_bos": True, "add_eos": True}, tokenizer_data=tokenizer_data)
-
-    def state_dict(self):
-        return {"spiece_model": self.tokenizer.serialize_model()}
-
-class JinaClip2TokenizerWrapper(sd1_clip.SD1Tokenizer):
-    def __init__(self, embedding_directory=None, tokenizer_data={}):
-        super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, tokenizer=JinaClip2Tokenizer, name="jina_clip_2")
-
-# https://huggingface.co/jinaai/jina-embeddings-v3/blob/343dbf534c76fe845f304fa5c2d1fd87e1e78918/config.json
-@dataclass
-class XLMRobertaConfig:
-    vocab_size: int = 250002
-    type_vocab_size: int = 1
-    hidden_size: int = 1024
-    num_hidden_layers: int = 24
-    num_attention_heads: int = 16
-    rotary_emb_base: float = 20000.0
-    intermediate_size: int = 4096
-    hidden_act: str = "gelu"
-    hidden_dropout_prob: float = 0.1
-    attention_probs_dropout_prob: float = 0.1
-    layer_norm_eps: float = 1e-05
-    bos_token_id: int = 0
-    eos_token_id: int = 2
-    pad_token_id: int = 1
-
-class XLMRobertaEmbeddings(nn.Module):
-    def __init__(self, config, device=None, dtype=None, ops=None):
-        super().__init__()
-        embed_dim = config.hidden_size
-        self.word_embeddings = ops.Embedding(config.vocab_size, embed_dim, padding_idx=config.pad_token_id, device=device, dtype=dtype)
-        self.token_type_embeddings = ops.Embedding(config.type_vocab_size, embed_dim, device=device, dtype=dtype)
-
-    def forward(self, input_ids=None, embeddings=None):
-        if input_ids is not None and embeddings is None:
-            embeddings = self.word_embeddings(input_ids)
-
-        if embeddings is not None:
-            token_type_ids = torch.zeros(embeddings.shape[1], device=embeddings.device, dtype=torch.int32)
-            token_type_embeddings = self.token_type_embeddings(token_type_ids)
-            embeddings = embeddings + token_type_embeddings
-        return embeddings
-
-class RotaryEmbedding(nn.Module):
-    def __init__(self, dim, base, device=None):
-        super().__init__()
-        inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2, device=device, dtype=torch.float32) / dim))
-        self.register_buffer("inv_freq", inv_freq, persistent=False)
-        self._seq_len_cached = 0
-        self._cos_cached = None
-        self._sin_cached = None
-
-    def _update_cos_sin_cache(self, seqlen, device=None, dtype=None):
-        if seqlen > self._seq_len_cached or self._cos_cached is None or self._cos_cached.device != device or self._cos_cached.dtype != dtype:
-            self._seq_len_cached = seqlen
-            t = torch.arange(seqlen, device=device, dtype=torch.float32)
-            freqs = torch.outer(t, self.inv_freq.to(device=t.device))
-            emb = torch.cat((freqs, freqs), dim=-1)
-            self._cos_cached = emb.cos().to(dtype)
-            self._sin_cached = emb.sin().to(dtype)
-
-    def forward(self, q, k):
-        batch, seqlen, heads, head_dim = q.shape
-        self._update_cos_sin_cache(seqlen, device=q.device, dtype=q.dtype)
-
-        cos = self._cos_cached[:seqlen].view(1, seqlen, 1, head_dim)
-        sin = self._sin_cached[:seqlen].view(1, seqlen, 1, head_dim)
-
-        def rotate_half(x):
-            size = x.shape[-1] // 2
-            x1, x2 = x[..., :size], x[..., size:]
-            return torch.cat((-x2, x1), dim=-1)
-
-        q_embed = (q * cos) + (rotate_half(q) * sin)
-        k_embed = (k * cos) + (rotate_half(k) * sin)
-        return q_embed, k_embed
-
-class MHA(nn.Module):
-    def __init__(self, config, device=None, dtype=None, ops=None):
-        super().__init__()
-        embed_dim = config.hidden_size
-        self.num_heads = config.num_attention_heads
-        self.head_dim = embed_dim // config.num_attention_heads
-
-        self.rotary_emb = RotaryEmbedding(self.head_dim, config.rotary_emb_base, device=device)
-        self.Wqkv = ops.Linear(embed_dim, 3 * embed_dim, device=device, dtype=dtype)
-        self.out_proj = ops.Linear(embed_dim, embed_dim, device=device, dtype=dtype)
-
-    def forward(self, x, mask=None, optimized_attention=None):
-        qkv = self.Wqkv(x)
-        batch_size, seq_len, _ = qkv.shape
-        qkv = qkv.view(batch_size, seq_len, 3, self.num_heads, self.head_dim)
-        q, k, v = qkv.unbind(2)
-
-        q, k = self.rotary_emb(q, k)
-
-        # NHD -> HND
-        q = q.transpose(1, 2)
-        k = k.transpose(1, 2)
-        v = v.transpose(1, 2)
-
-        out = optimized_attention(q, k, v, heads=self.num_heads, mask=mask, skip_reshape=True)
-        return self.out_proj(out)
-
-class MLP(nn.Module):
-    def __init__(self, config, device=None, dtype=None, ops=None):
-        super().__init__()
-        self.fc1 = ops.Linear(config.hidden_size, config.intermediate_size, device=device, dtype=dtype)
-        self.activation = F.gelu
-        self.fc2 = ops.Linear(config.intermediate_size, config.hidden_size, device=device, dtype=dtype)
-
-    def forward(self, x):
-        x = self.fc1(x)
-        x = self.activation(x)
-        x = self.fc2(x)
-        return x
-
-class Block(nn.Module):
-    def __init__(self, config, device=None, dtype=None, ops=None):
-        super().__init__()
-        self.mixer = MHA(config, device=device, dtype=dtype, ops=ops)
-        self.dropout1 = nn.Dropout(config.hidden_dropout_prob)
-        self.norm1 = ops.LayerNorm(config.hidden_size, eps=config.layer_norm_eps, device=device, dtype=dtype)
-        self.mlp = MLP(config, device=device, dtype=dtype, ops=ops)
-        self.dropout2 = nn.Dropout(config.hidden_dropout_prob)
-        self.norm2 = ops.LayerNorm(config.hidden_size, eps=config.layer_norm_eps, device=device, dtype=dtype)
-
-    def forward(self, hidden_states, mask=None, optimized_attention=None):
-        mixer_out = self.mixer(hidden_states, mask=mask, optimized_attention=optimized_attention)
-        hidden_states = self.norm1(self.dropout1(mixer_out) + hidden_states)
-        mlp_out = self.mlp(hidden_states)
-        hidden_states = self.norm2(self.dropout2(mlp_out) + hidden_states)
-        return hidden_states
-
-class XLMRobertaEncoder(nn.Module):
-    def __init__(self, config, device=None, dtype=None, ops=None):
-        super().__init__()
-        self.layers = nn.ModuleList([Block(config, device=device, dtype=dtype, ops=ops) for _ in range(config.num_hidden_layers)])
-
-    def forward(self, hidden_states, attention_mask=None):
-        optimized_attention = comfy.ldm.modules.attention.optimized_attention_for_device(hidden_states.device, mask=attention_mask is not None, small_input=True)
-        for layer in self.layers:
-            hidden_states = layer(hidden_states, mask=attention_mask, optimized_attention=optimized_attention)
-        return hidden_states
-
-class XLMRobertaModel_(nn.Module):
-    def __init__(self, config, device=None, dtype=None, ops=None):
-        super().__init__()
-        self.embeddings = XLMRobertaEmbeddings(config, device=device, dtype=dtype, ops=ops)
-        self.emb_ln = ops.LayerNorm(config.hidden_size, eps=config.layer_norm_eps, device=device, dtype=dtype)
-        self.emb_drop = nn.Dropout(config.hidden_dropout_prob)
-        self.encoder = XLMRobertaEncoder(config, device=device, dtype=dtype, ops=ops)
-
-    def forward(self, input_ids, attention_mask=None, embeds=None, num_tokens=None, intermediate_output=None, final_layer_norm_intermediate=True, dtype=None, embeds_info=[]):
-        x = self.embeddings(input_ids=input_ids, embeddings=embeds)
-        x = self.emb_ln(x)
-        x = self.emb_drop(x)
-
-        mask = None
-        if attention_mask is not None:
-            mask = 1.0 - attention_mask.to(x.dtype).reshape((attention_mask.shape[0], 1, 1, attention_mask.shape[-1]))
-            mask = mask.masked_fill(mask.to(torch.bool), -torch.finfo(x.dtype).max)
-
-        sequence_output = self.encoder(x, attention_mask=mask)
-
-        # Mean pool, see https://huggingface.co/jinaai/jina-clip-implementation/blob/39e6a55ae971b59bea6e44675d237c99762e7ee2/hf_model.py
-        pooled_output = None
-        if attention_mask is None:
-            pooled_output = sequence_output.mean(dim=1)
-        else:
-            attention_mask = attention_mask.to(sequence_output.dtype)
-            pooled_output = (sequence_output * attention_mask.unsqueeze(-1)).sum(dim=1) / attention_mask.sum(dim=-1, keepdim=True)
-
-        # Intermediate output is not yet implemented, use None for placeholder
-        return sequence_output, None, pooled_output
-
-class XLMRobertaModel(nn.Module):
-    def __init__(self, config_dict, dtype, device, operations):
-        super().__init__()
-        self.config = XLMRobertaConfig(**config_dict)
-        self.model = XLMRobertaModel_(self.config, device=device, dtype=dtype, ops=operations)
-        self.num_layers = self.config.num_hidden_layers
-
-    def get_input_embeddings(self):
-        return self.model.embeddings.word_embeddings
-
-    def set_input_embeddings(self, embeddings):
-        self.model.embeddings.word_embeddings = embeddings
-
-    def forward(self, *args, **kwargs):
-        return self.model(*args, **kwargs)
-
-class JinaClip2TextModel(sd1_clip.SDClipModel):
-    def __init__(self, device="cpu", dtype=None, model_options={}):
-        super().__init__(device=device, dtype=dtype, textmodel_json_config={}, model_class=XLMRobertaModel, special_tokens={"start": 0, "end": 2, "pad": 1}, enable_attention_masks=True, return_attention_masks=True, model_options=model_options)
-
-class JinaClip2TextModelWrapper(sd1_clip.SD1ClipModel):
-    def __init__(self, device="cpu", dtype=None, model_options={}):
-        super().__init__(device=device, dtype=dtype, clip_model=JinaClip2TextModel, name="jina_clip_2", model_options=model_options)

comfy/text_encoders/kandinsky5.py

@@ -1,68 +0,0 @@
-from comfy import sd1_clip
-from .qwen_image import QwenImageTokenizer, QwenImageTEModel
-from .llama import Qwen25_7BVLI
-
-
-class Kandinsky5Tokenizer(QwenImageTokenizer):
-    def __init__(self, embedding_directory=None, tokenizer_data={}):
-        super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data)
-        self.llama_template = "<|im_start|>system\nYou are a prompt engineer. Describe the video in detail.\nDescribe how the camera moves or shakes, describe the zoom and view angle, whether it follows the objects.\nDescribe the location of the video, main characters or objects and their action.\nDescribe the dynamism of the video and presented actions.\nName the visual style of the video: whether it is a professional footage, user generated content, some kind of animation, video game or screen content.\nDescribe the visual effects, postprocessing and transitions if they are presented in the video.\nPay attention to the order of key actions shown in the scene.<|im_end|>\n<|im_start|>user\n{}<|im_end|>"
-        self.clip_l = sd1_clip.SDTokenizer(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data)
-
-    def tokenize_with_weights(self, text:str, return_word_ids=False, **kwargs):
-        out = super().tokenize_with_weights(text, return_word_ids, **kwargs)
-        out["l"] = self.clip_l.tokenize_with_weights(text, return_word_ids, **kwargs)
-
-        return out
-
-
-class Kandinsky5TokenizerImage(Kandinsky5Tokenizer):
-    def __init__(self, embedding_directory=None, tokenizer_data={}):
-        super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data)
-        self.llama_template = "<|im_start|>system\nYou are a promt engineer. Describe the image by detailing the color, shape, size, texture, quantity, text, spatial relationships of the objects and background:<|im_end|>\n<|im_start|>user\n{}<|im_end|>"
-
-
-class Qwen25_7BVLIModel(sd1_clip.SDClipModel):
-    def __init__(self, device="cpu", layer="hidden", layer_idx=-1, dtype=None, attention_mask=True, model_options={}):
-        llama_quantization_metadata = model_options.get("llama_quantization_metadata", None)
-        if llama_quantization_metadata is not None:
-            model_options = model_options.copy()
-            model_options["quantization_metadata"] = llama_quantization_metadata
-        super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config={}, dtype=dtype, special_tokens={"pad": 151643}, layer_norm_hidden_state=False, model_class=Qwen25_7BVLI, enable_attention_masks=attention_mask, return_attention_masks=attention_mask, model_options=model_options)
-
-
-class Kandinsky5TEModel(QwenImageTEModel):
-    def __init__(self, device="cpu", dtype=None, model_options={}):
-        super(QwenImageTEModel, self).__init__(device=device, dtype=dtype, name="qwen25_7b", clip_model=Qwen25_7BVLIModel, model_options=model_options)
-        self.clip_l = sd1_clip.SDClipModel(device=device, dtype=dtype, return_projected_pooled=False, model_options=model_options)
-
-    def encode_token_weights(self, token_weight_pairs):
-        cond, p, extra = super().encode_token_weights(token_weight_pairs, template_end=-1)
-        l_out, l_pooled = self.clip_l.encode_token_weights(token_weight_pairs["l"])
-
-        return cond, l_pooled, extra
-
-    def set_clip_options(self, options):
-        super().set_clip_options(options)
-        self.clip_l.set_clip_options(options)
-
-    def reset_clip_options(self):
-        super().reset_clip_options()
-        self.clip_l.reset_clip_options()
-
-    def load_sd(self, sd):
-        if "text_model.encoder.layers.1.mlp.fc1.weight" in sd:
-            return self.clip_l.load_sd(sd)
-        else:
-            return super().load_sd(sd)
-
-def te(dtype_llama=None, llama_quantization_metadata=None):
-    class Kandinsky5TEModel_(Kandinsky5TEModel):
-        def __init__(self, device="cpu", dtype=None, model_options={}):
-            if llama_quantization_metadata is not None:
-                model_options = model_options.copy()
-                model_options["llama_quantization_metadata"] = llama_quantization_metadata
-            if dtype_llama is not None:
-                dtype = dtype_llama
-            super().__init__(device=device, dtype=dtype, model_options=model_options)
-    return Kandinsky5TEModel_

comfy/text_encoders/llama.py

@@ -3,12 +3,13 @@ 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
 import comfy.ldm.common_dit
-import comfy.clip_model
 
+import comfy.model_management
 from . import qwen_vl
 
 @dataclass
@@ -99,28 +100,6 @@ class Qwen3_4BConfig:
     rope_scale = None
     final_norm: bool = True
 
-@dataclass
-class Ovis25_2BConfig:
-    vocab_size: int = 151936
-    hidden_size: int = 2048
-    intermediate_size: int = 6144
-    num_hidden_layers: int = 28
-    num_attention_heads: int = 16
-    num_key_value_heads: int = 8
-    max_position_embeddings: int = 40960
-    rms_norm_eps: float = 1e-6
-    rope_theta: float = 1000000.0
-    transformer_type: str = "llama"
-    head_dim = 128
-    rms_norm_add = False
-    mlp_activation = "silu"
-    qkv_bias = False
-    rope_dims = None
-    q_norm = "gemma3"
-    k_norm = "gemma3"
-    rope_scale = None
-    final_norm: bool = True
-
 @dataclass
 class Qwen25_7BVLI_Config:
     vocab_size: int = 152064
@@ -176,7 +155,7 @@ class Gemma3_4B_Config:
     num_key_value_heads: int = 4
     max_position_embeddings: int = 131072
     rms_norm_eps: float = 1e-6
-    rope_theta = [1000000.0, 10000.0]
+    rope_theta = [10000.0, 1000000.0]
     transformer_type: str = "gemma3"
     head_dim = 256
     rms_norm_add = True
@@ -185,35 +164,10 @@ class Gemma3_4B_Config:
     rope_dims = None
     q_norm = "gemma3"
     k_norm = "gemma3"
-    sliding_attention = [1024, 1024, 1024, 1024, 1024, False]
-    rope_scale = [8.0, 1.0]
+    sliding_attention = [False, False, False, False, False, 1024]
+    rope_scale = [1.0, 8.0]
     final_norm: bool = True
 
-@dataclass
-class Gemma3_12B_Config:
-    vocab_size: int = 262208
-    hidden_size: int = 3840
-    intermediate_size: int = 15360
-    num_hidden_layers: int = 48
-    num_attention_heads: int = 16
-    num_key_value_heads: int = 8
-    max_position_embeddings: int = 131072
-    rms_norm_eps: float = 1e-6
-    rope_theta = [1000000.0, 10000.0]
-    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 = [1024, 1024, 1024, 1024, 1024, False]
-    rope_scale = [8.0, 1.0]
-    final_norm: bool = True
-    vision_config = {"num_channels": 3, "hidden_act": "gelu_pytorch_tanh", "hidden_size": 1152, "image_size": 896, "intermediate_size": 4304, "model_type": "siglip_vision_model", "num_attention_heads": 16, "num_hidden_layers": 27, "patch_size": 14}
-    mm_tokens_per_image = 256
-
 class RMSNorm(nn.Module):
     def __init__(self, dim: int, eps: float = 1e-5, add=False, device=None, dtype=None):
         super().__init__()
@@ -394,7 +348,7 @@ 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:
+        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
@@ -411,12 +365,7 @@ class TransformerBlockGemma2(nn.Module):
         if self.transformer_type == 'gemma3':
             if self.sliding_attention:
                 if x.shape[1] > self.sliding_attention:
-                    sliding_mask = torch.full((x.shape[1], x.shape[1]), float("-inf"), device=x.device, dtype=x.dtype)
-                    sliding_mask.tril_(diagonal=-self.sliding_attention)
-                    if attention_mask is not None:
-                        attention_mask = attention_mask + sliding_mask
-                    else:
-                        attention_mask = sliding_mask
+                    logging.warning("Warning: sliding attention not implemented, results may be incorrect")
                 freqs_cis = freqs_cis[1]
             else:
                 freqs_cis = freqs_cis[0]
@@ -546,41 +495,6 @@ class Llama2_(nn.Module):
 
         return x, intermediate
 
-
-class Gemma3MultiModalProjector(torch.nn.Module):
-    def __init__(self, config, dtype, device, operations):
-        super().__init__()
-
-        self.mm_input_projection_weight = nn.Parameter(
-            torch.empty(config.vision_config["hidden_size"], config.hidden_size, device=device, dtype=dtype)
-        )
-
-        self.mm_soft_emb_norm = RMSNorm(config.vision_config["hidden_size"], eps=config.rms_norm_eps, add=config.rms_norm_add, device=device, dtype=dtype)
-
-        self.patches_per_image = int(config.vision_config["image_size"] // config.vision_config["patch_size"])
-        self.tokens_per_side = int(config.mm_tokens_per_image**0.5)
-        self.kernel_size = self.patches_per_image // self.tokens_per_side
-        self.avg_pool = nn.AvgPool2d(kernel_size=self.kernel_size, stride=self.kernel_size)
-
-    def forward(self, vision_outputs: torch.Tensor):
-        batch_size, _, seq_length = vision_outputs.shape
-
-        reshaped_vision_outputs = vision_outputs.transpose(1, 2)
-        reshaped_vision_outputs = reshaped_vision_outputs.reshape(
-            batch_size, seq_length, self.patches_per_image, self.patches_per_image
-        )
-        reshaped_vision_outputs = reshaped_vision_outputs.contiguous()
-
-        pooled_vision_outputs = self.avg_pool(reshaped_vision_outputs)
-        pooled_vision_outputs = pooled_vision_outputs.flatten(2)
-        pooled_vision_outputs = pooled_vision_outputs.transpose(1, 2)
-
-        normed_vision_outputs = self.mm_soft_emb_norm(pooled_vision_outputs)
-
-        projected_vision_outputs = torch.matmul(normed_vision_outputs, comfy.model_management.cast_to_device(self.mm_input_projection_weight, device=normed_vision_outputs.device, dtype=normed_vision_outputs.dtype))
-        return projected_vision_outputs.type_as(vision_outputs)
-
-
 class BaseLlama:
     def get_input_embeddings(self):
         return self.model.embed_tokens
@@ -628,15 +542,6 @@ class Qwen3_4B(BaseLlama, torch.nn.Module):
         self.model = Llama2_(config, device=device, dtype=dtype, ops=operations)
         self.dtype = dtype
 
-class Ovis25_2B(BaseLlama, torch.nn.Module):
-    def __init__(self, config_dict, dtype, device, operations):
-        super().__init__()
-        config = Ovis25_2BConfig(**config_dict)
-        self.num_layers = config.num_hidden_layers
-
-        self.model = Llama2_(config, device=device, dtype=dtype, ops=operations)
-        self.dtype = dtype
-
 class Qwen25_7BVLI(BaseLlama, torch.nn.Module):
     def __init__(self, config_dict, dtype, device, operations):
         super().__init__()
@@ -697,21 +602,3 @@ class Gemma3_4B(BaseLlama, torch.nn.Module):
 
         self.model = Llama2_(config, device=device, dtype=dtype, ops=operations)
         self.dtype = dtype
-
-class Gemma3_12B(BaseLlama, torch.nn.Module):
-    def __init__(self, config_dict, dtype, device, operations):
-        super().__init__()
-        config = Gemma3_12B_Config(**config_dict)
-        self.num_layers = config.num_hidden_layers
-
-        self.model = Llama2_(config, device=device, dtype=dtype, ops=operations)
-        self.multi_modal_projector = Gemma3MultiModalProjector(config, dtype, device, operations)
-        self.vision_model = comfy.clip_model.CLIPVision(config.vision_config, dtype, device, operations)
-        self.dtype = dtype
-        self.image_size = config.vision_config["image_size"]
-
-    def preprocess_embed(self, embed, device):
-        if embed["type"] == "image":
-            image = comfy.clip_model.clip_preprocess(embed["data"], size=self.image_size, mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5], crop=True)
-            return self.multi_modal_projector(self.vision_model(image.to(device, dtype=torch.float32))[0]), None
-        return None, None

comfy/text_encoders/lt.py

@@ -1,11 +1,7 @@
 from comfy import sd1_clip
 import os
 from transformers import T5TokenizerFast
-from .spiece_tokenizer import SPieceTokenizer
 import comfy.text_encoders.genmo
-from comfy.ldm.lightricks.embeddings_connector import Embeddings1DConnector
-import torch
-import comfy.utils
 
 class T5XXLTokenizer(sd1_clip.SDTokenizer):
     def __init__(self, embedding_directory=None, tokenizer_data={}):
@@ -20,123 +16,3 @@ class LTXVT5Tokenizer(sd1_clip.SD1Tokenizer):
 
 def ltxv_te(*args, **kwargs):
     return comfy.text_encoders.genmo.mochi_te(*args, **kwargs)
-
-
-class Gemma3_12BTokenizer(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=3840, embedding_key='gemma3_12b', 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 LTXAVGemmaTokenizer(sd1_clip.SD1Tokenizer):
-    def __init__(self, embedding_directory=None, tokenizer_data={}):
-        super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, name="gemma3_12b", tokenizer=Gemma3_12BTokenizer)
-
-class Gemma3_12BModel(sd1_clip.SDClipModel):
-    def __init__(self, device="cpu", layer="all", layer_idx=None, dtype=None, attention_mask=True, model_options={}):
-        llama_quantization_metadata = model_options.get("llama_quantization_metadata", None)
-        if llama_quantization_metadata is not None:
-            model_options = model_options.copy()
-            model_options["quantization_metadata"] = llama_quantization_metadata
-
-        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_12B, enable_attention_masks=attention_mask, return_attention_masks=attention_mask, model_options=model_options)
-
-    def tokenize_with_weights(self, text, return_word_ids=False, llama_template="{}", image_embeds=None, **kwargs):
-        text = llama_template.format(text)
-        text_tokens = super().tokenize_with_weights(text, return_word_ids)
-        embed_count = 0
-        for k in text_tokens:
-            tt = text_tokens[k]
-            for r in tt:
-                for i in range(len(r)):
-                    if r[i][0] == 262144:
-                        if image_embeds is not None and embed_count < image_embeds.shape[0]:
-                            r[i] = ({"type": "embedding", "data": image_embeds[embed_count], "original_type": "image"},) + r[i][1:]
-                            embed_count += 1
-        return text_tokens
-
-class LTXAVTEModel(torch.nn.Module):
-    def __init__(self, dtype_llama=None, device="cpu", dtype=None, model_options={}):
-        super().__init__()
-        self.dtypes = set()
-        self.dtypes.add(dtype)
-
-        self.gemma3_12b = Gemma3_12BModel(device=device, dtype=dtype_llama, model_options=model_options, layer="all", layer_idx=None)
-        self.dtypes.add(dtype_llama)
-
-        operations = self.gemma3_12b.operations # TODO
-        self.text_embedding_projection = operations.Linear(3840 * 49, 3840, bias=False, dtype=dtype, device=device)
-
-        self.audio_embeddings_connector = Embeddings1DConnector(
-            split_rope=True,
-            double_precision_rope=True,
-            dtype=dtype,
-            device=device,
-            operations=operations,
-        )
-
-        self.video_embeddings_connector = Embeddings1DConnector(
-            split_rope=True,
-            double_precision_rope=True,
-            dtype=dtype,
-            device=device,
-            operations=operations,
-        )
-
-    def set_clip_options(self, options):
-        self.execution_device = options.get("execution_device", self.execution_device)
-        self.gemma3_12b.set_clip_options(options)
-
-    def reset_clip_options(self):
-        self.gemma3_12b.reset_clip_options()
-        self.execution_device = None
-
-    def encode_token_weights(self, token_weight_pairs):
-        token_weight_pairs = token_weight_pairs["gemma3_12b"]
-
-        out, pooled, extra = self.gemma3_12b.encode_token_weights(token_weight_pairs)
-        out_device = out.device
-        if comfy.model_management.should_use_bf16(self.execution_device):
-            out = out.to(device=self.execution_device, dtype=torch.bfloat16)
-        out = out.movedim(1, -1).to(self.execution_device)
-        out = 8.0 * (out - out.mean(dim=(1, 2), keepdim=True)) / (out.amax(dim=(1, 2), keepdim=True) - out.amin(dim=(1, 2), keepdim=True) + 1e-6)
-        out = out.reshape((out.shape[0], out.shape[1], -1))
-        out = self.text_embedding_projection(out)
-        out = out.float()
-        out_vid = self.video_embeddings_connector(out)[0]
-        out_audio = self.audio_embeddings_connector(out)[0]
-        out = torch.concat((out_vid, out_audio), dim=-1)
-
-        return out.to(out_device), pooled
-
-    def load_sd(self, sd):
-        if "model.layers.47.self_attn.q_norm.weight" in sd:
-            return self.gemma3_12b.load_sd(sd)
-        else:
-            sdo = comfy.utils.state_dict_prefix_replace(sd, {"text_embedding_projection.aggregate_embed.weight": "text_embedding_projection.weight", "model.diffusion_model.video_embeddings_connector.": "video_embeddings_connector.", "model.diffusion_model.audio_embeddings_connector.": "audio_embeddings_connector."}, filter_keys=True)
-            if len(sdo) == 0:
-                sdo = sd
-
-            return self.load_state_dict(sdo, strict=False)
-
-    def memory_estimation_function(self, token_weight_pairs, device=None):
-        constant = 6.0
-        if comfy.model_management.should_use_bf16(device):
-            constant /= 2.0
-
-        token_weight_pairs = token_weight_pairs.get("gemma3_12b", [])
-        num_tokens = sum(map(lambda a: len(a), token_weight_pairs))
-        return num_tokens * constant * 1024 * 1024
-
-def ltxav_te(dtype_llama=None, llama_quantization_metadata=None):
-    class LTXAVTEModel_(LTXAVTEModel):
-        def __init__(self, device="cpu", dtype=None, model_options={}):
-            if llama_quantization_metadata is not None:
-                model_options = model_options.copy()
-                model_options["llama_quantization_metadata"] = llama_quantization_metadata
-            if dtype_llama is not None:
-                dtype = dtype_llama
-            super().__init__(dtype_llama=dtype_llama, device=device, dtype=dtype, model_options=model_options)
-    return LTXAVTEModel_

comfy/text_encoders/lumina2.py

@@ -14,7 +14,7 @@ class Gemma2BTokenizer(sd1_clip.SDTokenizer):
 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}, disable_weights=True, tokenizer_data=tokenizer_data)
+        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()}
@@ -33,11 +33,6 @@ class Gemma2_2BModel(sd1_clip.SDClipModel):
 
 class Gemma3_4BModel(sd1_clip.SDClipModel):
     def __init__(self, device="cpu", layer="hidden", layer_idx=-2, dtype=None, attention_mask=True, model_options={}):
-        llama_quantization_metadata = model_options.get("llama_quantization_metadata", None)
-        if llama_quantization_metadata is not None:
-            model_options = model_options.copy()
-            model_options["quantization_metadata"] = llama_quantization_metadata
-
         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):
@@ -45,7 +40,7 @@ class LuminaModel(sd1_clip.SD1ClipModel):
         super().__init__(device=device, dtype=dtype, name=name, clip_model=clip_model, model_options=model_options)
 
 
-def te(dtype_llama=None, llama_quantization_metadata=None, model_type="gemma2_2b"):
+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":
@@ -53,9 +48,9 @@ def te(dtype_llama=None, llama_quantization_metadata=None, model_type="gemma2_2b
 
     class LuminaTEModel_(LuminaModel):
         def __init__(self, device="cpu", dtype=None, model_options={}):
-            if llama_quantization_metadata is not None:
+            if llama_scaled_fp8 is not None and "scaled_fp8" not in model_options:
                 model_options = model_options.copy()
-                model_options["quantization_metadata"] = llama_quantization_metadata
+                model_options["scaled_fp8"] = llama_scaled_fp8
             if dtype_llama is not None:
                 dtype = dtype_llama
             super().__init__(device=device, dtype=dtype, name=model_type, model_options=model_options, clip_model=model)

comfy/text_encoders/newbie.py

@@ -1,62 +0,0 @@
-import torch
-
-import comfy.model_management
-import comfy.text_encoders.jina_clip_2
-import comfy.text_encoders.lumina2
-
-class NewBieTokenizer:
-    def __init__(self, embedding_directory=None, tokenizer_data={}):
-        self.gemma = comfy.text_encoders.lumina2.Gemma3_4BTokenizer(embedding_directory=embedding_directory, tokenizer_data={"spiece_model": tokenizer_data["gemma_spiece_model"]})
-        self.jina = comfy.text_encoders.jina_clip_2.JinaClip2Tokenizer(embedding_directory=embedding_directory, tokenizer_data={"spiece_model": tokenizer_data["jina_spiece_model"]})
-
-    def tokenize_with_weights(self, text:str, return_word_ids=False, **kwargs):
-        out = {}
-        out["gemma"] = self.gemma.tokenize_with_weights(text, return_word_ids, **kwargs)
-        out["jina"] = self.jina.tokenize_with_weights(text, return_word_ids, **kwargs)
-        return out
-
-    def untokenize(self, token_weight_pair):
-        raise NotImplementedError
-
-    def state_dict(self):
-        return {}
-
-class NewBieTEModel(torch.nn.Module):
-    def __init__(self, dtype_gemma=None, device="cpu", dtype=None, model_options={}):
-        super().__init__()
-        dtype_gemma = comfy.model_management.pick_weight_dtype(dtype_gemma, dtype, device)
-        self.gemma = comfy.text_encoders.lumina2.Gemma3_4BModel(device=device, dtype=dtype_gemma, model_options=model_options)
-        self.jina = comfy.text_encoders.jina_clip_2.JinaClip2TextModel(device=device, dtype=dtype, model_options=model_options)
-        self.dtypes = {dtype, dtype_gemma}
-
-    def set_clip_options(self, options):
-        self.gemma.set_clip_options(options)
-        self.jina.set_clip_options(options)
-
-    def reset_clip_options(self):
-        self.gemma.reset_clip_options()
-        self.jina.reset_clip_options()
-
-    def encode_token_weights(self, token_weight_pairs):
-        token_weight_pairs_gemma = token_weight_pairs["gemma"]
-        token_weight_pairs_jina = token_weight_pairs["jina"]
-
-        gemma_out, gemma_pooled, gemma_extra = self.gemma.encode_token_weights(token_weight_pairs_gemma)
-        jina_out, jina_pooled, jina_extra = self.jina.encode_token_weights(token_weight_pairs_jina)
-
-        return gemma_out, jina_pooled, gemma_extra
-
-    def load_sd(self, sd):
-        if "model.layers.0.self_attn.q_norm.weight" in sd:
-            return self.gemma.load_sd(sd)
-        else:
-            return self.jina.load_sd(sd)
-
-def te(dtype_llama=None, llama_quantization_metadata=None):
-    class NewBieTEModel_(NewBieTEModel):
-        def __init__(self, device="cpu", dtype=None, model_options={}):
-            if llama_quantization_metadata is not None:
-                model_options = model_options.copy()
-                model_options["llama_quantization_metadata"] = llama_quantization_metadata
-            super().__init__(dtype_gemma=dtype_llama, device=device, dtype=dtype, model_options=model_options)
-    return NewBieTEModel_

comfy/text_encoders/omnigen2.py

@@ -32,12 +32,12 @@ class Omnigen2Model(sd1_clip.SD1ClipModel):
         super().__init__(device=device, dtype=dtype, name="qwen25_3b", clip_model=Qwen25_3BModel, model_options=model_options)
 
 
-def te(dtype_llama=None, llama_quantization_metadata=None):
+def te(dtype_llama=None, llama_scaled_fp8=None):
     class Omnigen2TEModel_(Omnigen2Model):
         def __init__(self, device="cpu", dtype=None, model_options={}):
-            if llama_quantization_metadata is not None:
+            if llama_scaled_fp8 is not None and "scaled_fp8" not in model_options:
                 model_options = model_options.copy()
-                model_options["quantization_metadata"] = llama_quantization_metadata
+                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)

comfy/text_encoders/ovis.py

@@ -1,66 +0,0 @@
-from transformers import Qwen2Tokenizer
-import comfy.text_encoders.llama
-from comfy import sd1_clip
-import os
-import torch
-import numbers
-
-class Qwen3Tokenizer(sd1_clip.SDTokenizer):
-    def __init__(self, embedding_directory=None, tokenizer_data={}):
-        tokenizer_path = os.path.join(os.path.dirname(os.path.realpath(__file__)), "qwen25_tokenizer")
-        super().__init__(tokenizer_path, pad_with_end=False, embedding_size=2048, embedding_key='qwen3_2b', tokenizer_class=Qwen2Tokenizer, has_start_token=False, has_end_token=False, pad_to_max_length=False, max_length=99999999, min_length=284, pad_token=151643, tokenizer_data=tokenizer_data)
-
-
-class OvisTokenizer(sd1_clip.SD1Tokenizer):
-    def __init__(self, embedding_directory=None, tokenizer_data={}):
-        super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, name="qwen3_2b", tokenizer=Qwen3Tokenizer)
-        self.llama_template = "<|im_start|>user\nDescribe the image by detailing the color, quantity, text, shape, size, texture, spatial relationships of the objects and background: {}<|im_end|>\n<|im_start|>assistant\n<think>\n\n</think>\n\n"
-
-    def tokenize_with_weights(self, text, return_word_ids=False, llama_template=None, **kwargs):
-        if llama_template is None:
-            llama_text = self.llama_template.format(text)
-        else:
-            llama_text = llama_template.format(text)
-
-        tokens = super().tokenize_with_weights(llama_text, return_word_ids=return_word_ids, disable_weights=True, **kwargs)
-        return tokens
-
-class Ovis25_2BModel(sd1_clip.SDClipModel):
-    def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None, attention_mask=True, model_options={}):
-        super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config={}, dtype=dtype, special_tokens={"pad": 151643}, layer_norm_hidden_state=False, model_class=comfy.text_encoders.llama.Ovis25_2B, enable_attention_masks=attention_mask, return_attention_masks=False, zero_out_masked=True, model_options=model_options)
-
-
-class OvisTEModel(sd1_clip.SD1ClipModel):
-    def __init__(self, device="cpu", dtype=None, model_options={}):
-        super().__init__(device=device, dtype=dtype, name="qwen3_2b", clip_model=Ovis25_2BModel, model_options=model_options)
-
-    def encode_token_weights(self, token_weight_pairs, template_end=-1):
-        out, pooled = super().encode_token_weights(token_weight_pairs)
-        tok_pairs = token_weight_pairs["qwen3_2b"][0]
-        count_im_start = 0
-        if template_end == -1:
-            for i, v in enumerate(tok_pairs):
-                elem = v[0]
-                if not torch.is_tensor(elem):
-                    if isinstance(elem, numbers.Integral):
-                        if elem == 4004 and count_im_start < 1:
-                            template_end = i
-                            count_im_start += 1
-
-            if out.shape[1] > (template_end + 1):
-                if tok_pairs[template_end + 1][0] == 25:
-                    template_end += 1
-
-        out = out[:, template_end:]
-        return out, pooled, {}
-
-
-def te(dtype_llama=None, llama_quantization_metadata=None):
-    class OvisTEModel_(OvisTEModel):
-        def __init__(self, device="cpu", dtype=None, model_options={}):
-            if dtype_llama is not None:
-                dtype = dtype_llama
-            if llama_quantization_metadata is not None:
-                model_options["quantization_metadata"] = llama_quantization_metadata
-            super().__init__(device=device, dtype=dtype, model_options=model_options)
-    return OvisTEModel_

comfy/text_encoders/pixart_t5.py

@@ -30,13 +30,13 @@ class PixArtTokenizer(sd1_clip.SD1Tokenizer):
     def __init__(self, embedding_directory=None, tokenizer_data={}):
         super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, clip_name="t5xxl", tokenizer=T5XXLTokenizer)
 
-def pixart_te(dtype_t5=None, t5_quantization_metadata=None):
+def pixart_te(dtype_t5=None, t5xxl_scaled_fp8=None):
     class PixArtTEModel_(PixArtT5XXL):
         def __init__(self, device="cpu", dtype=None, model_options={}):
-            if t5_quantization_metadata is not None:
+            if t5xxl_scaled_fp8 is not None and "t5xxl_scaled_fp8" not in model_options:
                 model_options = model_options.copy()
-                model_options["t5xxl_quantization_metadata"] = t5_quantization_metadata
-            if dtype_t5 is not None:
+                model_options["t5xxl_scaled_fp8"] = t5xxl_scaled_fp8
+            if dtype is None:
                 dtype = dtype_t5
             super().__init__(device=device, dtype=dtype, model_options=model_options)
     return PixArtTEModel_

comfy/text_encoders/qwen_image.py

@@ -85,12 +85,12 @@ class QwenImageTEModel(sd1_clip.SD1ClipModel):
         return out, pooled, extra
 
 
-def te(dtype_llama=None, llama_quantization_metadata=None):
+def te(dtype_llama=None, llama_scaled_fp8=None):
     class QwenImageTEModel_(QwenImageTEModel):
         def __init__(self, device="cpu", dtype=None, model_options={}):
-            if llama_quantization_metadata is not None:
+            if llama_scaled_fp8 is not None and "scaled_fp8" not in model_options:
                 model_options = model_options.copy()
-                model_options["quantization_metadata"] = llama_quantization_metadata
+                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)

comfy/text_encoders/sd3_clip.py

@@ -6,15 +6,14 @@ import torch
 import os
 import comfy.model_management
 import logging
-import comfy.utils
 
 class T5XXLModel(sd1_clip.SDClipModel):
     def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None, attention_mask=False, model_options={}):
         textmodel_json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "t5_config_xxl.json")
-        t5xxl_quantization_metadata = model_options.get("t5xxl_quantization_metadata", None)
-        if t5xxl_quantization_metadata is not None:
+        t5xxl_scaled_fp8 = model_options.get("t5xxl_scaled_fp8", None)
+        if t5xxl_scaled_fp8 is not None:
             model_options = model_options.copy()
-            model_options["quantization_metadata"] = t5xxl_quantization_metadata
+            model_options["scaled_fp8"] = t5xxl_scaled_fp8
 
         model_options = {**model_options, "model_name": "t5xxl"}
         super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype, special_tokens={"end": 1, "pad": 0}, model_class=comfy.text_encoders.t5.T5, enable_attention_masks=attention_mask, return_attention_masks=attention_mask, model_options=model_options)
@@ -26,9 +25,9 @@ def t5_xxl_detect(state_dict, prefix=""):
     if t5_key in state_dict:
         out["dtype_t5"] = state_dict[t5_key].dtype
 
-    quant = comfy.utils.detect_layer_quantization(state_dict, prefix)
-    if quant is not None:
-        out["t5_quantization_metadata"] = quant
+    scaled_fp8_key = "{}scaled_fp8".format(prefix)
+    if scaled_fp8_key in state_dict:
+        out["t5xxl_scaled_fp8"] = state_dict[scaled_fp8_key].dtype
 
     return out
 
@@ -157,11 +156,11 @@ class SD3ClipModel(torch.nn.Module):
         else:
             return self.t5xxl.load_sd(sd)
 
-def sd3_clip(clip_l=True, clip_g=True, t5=True, dtype_t5=None, t5_quantization_metadata=None, t5_attention_mask=False):
+def sd3_clip(clip_l=True, clip_g=True, t5=True, dtype_t5=None, t5xxl_scaled_fp8=None, t5_attention_mask=False):
     class SD3ClipModel_(SD3ClipModel):
         def __init__(self, device="cpu", dtype=None, model_options={}):
-            if t5_quantization_metadata is not None:
+            if t5xxl_scaled_fp8 is not None and "t5xxl_scaled_fp8" not in model_options:
                 model_options = model_options.copy()
-                model_options["t5xxl_quantization_metadata"] = t5_quantization_metadata
+                model_options["t5xxl_scaled_fp8"] = t5xxl_scaled_fp8
             super().__init__(clip_l=clip_l, clip_g=clip_g, t5=t5, dtype_t5=dtype_t5, t5_attention_mask=t5_attention_mask, device=device, dtype=dtype, model_options=model_options)
     return SD3ClipModel_

comfy/text_encoders/wan.py

@@ -25,12 +25,12 @@ class WanT5Model(sd1_clip.SD1ClipModel):
     def __init__(self, device="cpu", dtype=None, model_options={}, **kwargs):
         super().__init__(device=device, dtype=dtype, model_options=model_options, name="umt5xxl", clip_model=UMT5XXlModel, **kwargs)
 
-def te(dtype_t5=None, t5_quantization_metadata=None):
+def te(dtype_t5=None, t5xxl_scaled_fp8=None):
     class WanTEModel(WanT5Model):
         def __init__(self, device="cpu", dtype=None, model_options={}):
-            if t5_quantization_metadata is not None:
+            if t5xxl_scaled_fp8 is not None and "scaled_fp8" not in model_options:
                 model_options = model_options.copy()
-                model_options["quantization_metadata"] = t5_quantization_metadata
+                model_options["scaled_fp8"] = t5xxl_scaled_fp8
             if dtype_t5 is not None:
                 dtype = dtype_t5
             super().__init__(device=device, dtype=dtype, model_options=model_options)

comfy/text_encoders/z_image.py

@@ -34,9 +34,12 @@ class ZImageTEModel(sd1_clip.SD1ClipModel):
         super().__init__(device=device, dtype=dtype, name="qwen3_4b", clip_model=Qwen3_4BModel, model_options=model_options)
 
 
-def te(dtype_llama=None, llama_quantization_metadata=None):
+def te(dtype_llama=None, llama_scaled_fp8=None, llama_quantization_metadata=None):
     class ZImageTEModel_(ZImageTEModel):
         def __init__(self, device="cpu", dtype=None, model_options={}):
+            if llama_scaled_fp8 is not None and "scaled_fp8" not in model_options:
+                model_options = model_options.copy()
+                model_options["scaled_fp8"] = llama_scaled_fp8
             if dtype_llama is not None:
                 dtype = dtype_llama
             if llama_quantization_metadata is not None:

comfy/utils.py

@@ -29,7 +29,6 @@ import itertools
 from torch.nn.functional import interpolate
 from einops import rearrange
 from comfy.cli_args import args
-import json
 
 MMAP_TORCH_FILES = args.mmap_torch_files
 DISABLE_MMAP = args.disable_mmap
@@ -53,7 +52,7 @@ if hasattr(torch.serialization, "add_safe_globals"):  # TODO: this was added in
     ALWAYS_SAFE_LOAD = True
     logging.info("Checkpoint files will always be loaded safely.")
 else:
-    logging.warning("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 as older versions of pytorch are no longer supported.")
+    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 load_torch_file(ckpt, safe_load=False, device=None, return_metadata=False):
     if device is None:
@@ -803,17 +802,12 @@ def safetensors_header(safetensors_path, max_size=100*1024*1024):
             return None
         return f.read(length_of_header)
 
-ATTR_UNSET={}
-
 def set_attr(obj, attr, value):
     attrs = attr.split(".")
     for name in attrs[:-1]:
         obj = getattr(obj, name)
-    prev = getattr(obj, attrs[-1], ATTR_UNSET)
-    if value is ATTR_UNSET:
-        delattr(obj, attrs[-1])
-    else:
-        setattr(obj, attrs[-1], value)
+    prev = getattr(obj, attrs[-1])
+    setattr(obj, attrs[-1], value)
     return prev
 
 def set_attr_param(obj, attr, value):
@@ -1198,72 +1192,5 @@ def unpack_latents(combined_latent, latent_shapes):
             combined_latent = combined_latent[:, :, cut:]
             output_tensors.append(tens.reshape([tens.shape[0]] + list(shape)[1:]))
     else:
-        output_tensors = [combined_latent]
+        output_tensors = combined_latent
     return output_tensors
-
-def detect_layer_quantization(state_dict, prefix):
-    for k in state_dict:
-        if k.startswith(prefix) and k.endswith(".comfy_quant"):
-            logging.info("Found quantization metadata version 1")
-            return {"mixed_ops": True}
-    return None
-
-def convert_old_quants(state_dict, model_prefix="", metadata={}):
-    if metadata is None:
-        metadata = {}
-
-    quant_metadata = None
-    if "_quantization_metadata" not in metadata:
-        scaled_fp8_key = "{}scaled_fp8".format(model_prefix)
-
-        if scaled_fp8_key in state_dict:
-            scaled_fp8_weight = state_dict[scaled_fp8_key]
-            scaled_fp8_dtype = scaled_fp8_weight.dtype
-            if scaled_fp8_dtype == torch.float32:
-                scaled_fp8_dtype = torch.float8_e4m3fn
-
-            if scaled_fp8_weight.nelement() == 2:
-                full_precision_matrix_mult = True
-            else:
-                full_precision_matrix_mult = False
-
-            out_sd = {}
-            layers = {}
-            for k in list(state_dict.keys()):
-                if k == scaled_fp8_key:
-                    continue
-                if not k.startswith(model_prefix):
-                    out_sd[k] = state_dict[k]
-                    continue
-                k_out = k
-                w = state_dict.pop(k)
-                layer = None
-                if k_out.endswith(".scale_weight"):
-                    layer = k_out[:-len(".scale_weight")]
-                    k_out = "{}.weight_scale".format(layer)
-
-                if layer is not None:
-                    layer_conf = {"format": "float8_e4m3fn"}  # TODO: check if anyone did some non e4m3fn scaled checkpoints
-                    if full_precision_matrix_mult:
-                        layer_conf["full_precision_matrix_mult"] = full_precision_matrix_mult
-                    layers[layer] = layer_conf
-
-                if k_out.endswith(".scale_input"):
-                    layer = k_out[:-len(".scale_input")]
-                    k_out = "{}.input_scale".format(layer)
-                    if w.item() == 1.0:
-                        continue
-
-                out_sd[k_out] = w
-
-            state_dict = out_sd
-            quant_metadata = {"layers": layers}
-    else:
-        quant_metadata = json.loads(metadata["_quantization_metadata"])
-
-    if quant_metadata is not None:
-        layers = quant_metadata["layers"]
-        for k, v in layers.items():
-            state_dict["{}.comfy_quant".format(k)] = torch.tensor(list(json.dumps(v).encode('utf-8')), dtype=torch.uint8)
-
-    return state_dict, metadata

comfy_api/feature_flags.py

@@ -5,20 +5,19 @@ This module handles capability negotiation between frontend and backend,
 allowing graceful protocol evolution while maintaining backward compatibility.
 """
 
-from typing import Any
+from typing import Any, Dict
 
 from comfy.cli_args import args
 
 # Default server capabilities
-SERVER_FEATURE_FLAGS: dict[str, Any] = {
+SERVER_FEATURE_FLAGS: Dict[str, Any] = {
     "supports_preview_metadata": True,
     "max_upload_size": args.max_upload_size * 1024 * 1024, # Convert MB to bytes
-    "extension": {"manager": {"supports_v4": True}},
 }
 
 
 def get_connection_feature(
-    sockets_metadata: dict[str, dict[str, Any]],
+    sockets_metadata: Dict[str, Dict[str, Any]],
     sid: str,
     feature_name: str,
     default: Any = False
@@ -42,7 +41,7 @@ def get_connection_feature(
 
 
 def supports_feature(
-    sockets_metadata: dict[str, dict[str, Any]],
+    sockets_metadata: Dict[str, Dict[str, Any]],
     sid: str,
     feature_name: str
 ) -> bool:
@@ -60,7 +59,7 @@ def supports_feature(
     return get_connection_feature(sockets_metadata, sid, feature_name, False) is True
 
 
-def get_server_features() -> dict[str, Any]:
+def get_server_features() -> Dict[str, Any]:
     """
     Get the server's feature flags.
 

comfy_api/internal/api_registry.py

@@ -1,4 +1,4 @@
-from typing import NamedTuple
+from typing import Type, List, NamedTuple
 from comfy_api.internal.singleton import ProxiedSingleton
 from packaging import version as packaging_version
 
@@ -10,7 +10,7 @@ class ComfyAPIBase(ProxiedSingleton):
 
 class ComfyAPIWithVersion(NamedTuple):
     version: str
-    api_class: type[ComfyAPIBase]
+    api_class: Type[ComfyAPIBase]
 
 
 def parse_version(version_str: str) -> packaging_version.Version:
@@ -23,16 +23,16 @@ def parse_version(version_str: str) -> packaging_version.Version:
     return packaging_version.parse(version_str)
 
 
-registered_versions: list[ComfyAPIWithVersion] = []
+registered_versions: List[ComfyAPIWithVersion] = []
 
 
-def register_versions(versions: list[ComfyAPIWithVersion]):
+def register_versions(versions: List[ComfyAPIWithVersion]):
     versions.sort(key=lambda x: parse_version(x.version))
     global registered_versions
     registered_versions = versions
 
 
-def get_all_versions() -> list[ComfyAPIWithVersion]:
+def get_all_versions() -> List[ComfyAPIWithVersion]:
     """
     Returns a list of all registered ComfyAPI versions.
     """

comfy_api/internal/async_to_sync.py

@@ -8,7 +8,7 @@ import os
 import textwrap
 import threading
 from enum import Enum
-from typing import Optional, get_origin, get_args, get_type_hints
+from typing import Optional, Type, get_origin, get_args, get_type_hints
 
 
 class TypeTracker:
@@ -193,7 +193,7 @@ class AsyncToSyncConverter:
         return result_container["result"]
 
     @classmethod
-    def create_sync_class(cls, async_class: type, thread_pool_size=10) -> type:
+    def create_sync_class(cls, async_class: Type, thread_pool_size=10) -> Type:
         """
         Creates a new class with synchronous versions of all async methods.
 
@@ -563,7 +563,7 @@ class AsyncToSyncConverter:
 
     @classmethod
     def _generate_imports(
-        cls, async_class: type, type_tracker: TypeTracker
+        cls, async_class: Type, type_tracker: TypeTracker
     ) -> list[str]:
         """Generate import statements for the stub file."""
         imports = []
@@ -628,7 +628,7 @@ class AsyncToSyncConverter:
         return imports
 
     @classmethod
-    def _get_class_attributes(cls, async_class: type) -> list[tuple[str, type]]:
+    def _get_class_attributes(cls, async_class: Type) -> list[tuple[str, Type]]:
         """Extract class attributes that are classes themselves."""
         class_attributes = []
 
@@ -654,7 +654,7 @@ class AsyncToSyncConverter:
     def _generate_inner_class_stub(
         cls,
         name: str,
-        attr: type,
+        attr: Type,
         indent: str = "    ",
         type_tracker: Optional[TypeTracker] = None,
     ) -> list[str]:
@@ -782,7 +782,7 @@ class AsyncToSyncConverter:
         return processed
 
     @classmethod
-    def generate_stub_file(cls, async_class: type, sync_class: type) -> None:
+    def generate_stub_file(cls, async_class: Type, sync_class: Type) -> None:
         """
         Generate a .pyi stub file for the sync class to help IDEs with type checking.
         """
@@ -988,7 +988,7 @@ class AsyncToSyncConverter:
             logging.error(traceback.format_exc())
 
 
-def create_sync_class(async_class: type, thread_pool_size=10) -> type:
+def create_sync_class(async_class: Type, thread_pool_size=10) -> Type:
     """
     Creates a sync version of an async class
 

comfy_api/internal/singleton.py

@@ -1,4 +1,4 @@
-from typing import TypeVar
+from typing import Type, TypeVar
 
 class SingletonMetaclass(type):
     T = TypeVar("T", bound="SingletonMetaclass")
@@ -11,13 +11,13 @@ class SingletonMetaclass(type):
             )
         return cls._instances[cls]
 
-    def inject_instance(cls: type[T], instance: T) -> None:
+    def inject_instance(cls: Type[T], instance: T) -> None:
         assert cls not in SingletonMetaclass._instances, (
             "Cannot inject instance after first instantiation"
         )
         SingletonMetaclass._instances[cls] = instance
 
-    def get_instance(cls: type[T], *args, **kwargs) -> T:
+    def get_instance(cls: Type[T], *args, **kwargs) -> T:
         """
         Gets the singleton instance of the class, creating it if it doesn't exist.
         """