Merge branch 'master' into master

.ci/windows_base_files/README_VERY_IMPORTANT.txt

@@ -14,7 +14,7 @@ run_cpu.bat
 
 IF YOU GET A RED ERROR IN THE UI MAKE SURE YOU HAVE A MODEL/CHECKPOINT IN: ComfyUI\models\checkpoints
 
-You can download the stable diffusion 1.5 one from: https://huggingface.co/runwayml/stable-diffusion-v1-5/blob/main/v1-5-pruned-emaonly.ckpt
+You can download the stable diffusion 1.5 one from: https://huggingface.co/Comfy-Org/stable-diffusion-v1-5-archive/blob/main/v1-5-pruned-emaonly-fp16.safetensors
 
 
 RECOMMENDED WAY TO UPDATE:

.ci/windows_nightly_base_files/run_nvidia_gpu_fast.bat

@@ -0,0 +1,2 @@
+.\python_embeded\python.exe -s ComfyUI\main.py --windows-standalone-build --fast
+pause

.github/ISSUE_TEMPLATE/config.yml

@@ -1,5 +1,8 @@
 blank_issues_enabled: true
 contact_links:
+  - name: ComfyUI Frontend Issues
+    url: https://github.com/Comfy-Org/ComfyUI_frontend/issues
+    about: Issues related to the ComfyUI frontend (display issues, user interaction bugs), please go to the frontend repo to file the issue
   - name: ComfyUI Matrix Space
     url: https://app.element.io/#/room/%23comfyui_space%3Amatrix.org
     about: The ComfyUI Matrix Space is available for support and general discussion related to ComfyUI (Matrix is like Discord but open source).

.github/workflows/stale-issues.yml

@@ -0,0 +1,21 @@
+name: 'Close stale issues'
+on:
+  schedule:
+    # Run daily at 430 am PT
+    - cron: '30 11 * * *'
+permissions:
+  issues: write
+
+jobs:
+  stale:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/stale@v9
+        with:
+          stale-issue-message: "This issue is being marked stale because it has not had any activity for 30 days. Reply below within 7 days if your issue still isn't solved, and it will be left open. Otherwise, the issue will be closed automatically."
+          days-before-stale: 30
+          days-before-close: 7
+          stale-issue-label: 'Stale'
+          only-labels: 'User Support'
+          exempt-all-assignees: true
+          exempt-all-milestones: true

.github/workflows/windows_release_nightly_pytorch.yml

@@ -67,6 +67,7 @@ jobs:
             mkdir update
             cp -r ComfyUI/.ci/update_windows/* ./update/
             cp -r ComfyUI/.ci/windows_base_files/* ./
+            cp -r ComfyUI/.ci/windows_nightly_base_files/* ./
 
             echo "call update_comfyui.bat nopause
             ..\python_embeded\python.exe -s -m pip install --upgrade --pre torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/nightly/cu${{ inputs.cu }} -r ../ComfyUI/requirements.txt pygit2

README.md

@@ -1,8 +1,35 @@
-ComfyUI
+<div align="center">
-=======
+
-The most powerful and modular stable diffusion GUI and backend.
+# ComfyUI
------------
+**The most powerful and modular diffusion model GUI and backend.**
+
+
+[![Website][website-shield]][website-url]
+[![Dynamic JSON Badge][discord-shield]][discord-url]
+[![Matrix][matrix-shield]][matrix-url]
+<br>
+[![][github-release-shield]][github-release-link]
+[![][github-release-date-shield]][github-release-link]
+[![][github-downloads-shield]][github-downloads-link]
+[![][github-downloads-latest-shield]][github-downloads-link]
+
+[matrix-shield]: https://img.shields.io/badge/Matrix-000000?style=flat&logo=matrix&logoColor=white
+[matrix-url]: https://app.element.io/#/room/%23comfyui_space%3Amatrix.org
+[website-shield]: https://img.shields.io/badge/ComfyOrg-4285F4?style=flat
+[website-url]: https://www.comfy.org/
+<!-- Workaround to display total user from https://github.com/badges/shields/issues/4500#issuecomment-2060079995 -->
+[discord-shield]: https://img.shields.io/badge/dynamic/json?url=https%3A%2F%2Fdiscord.com%2Fapi%2Finvites%2Fcomfyorg%3Fwith_counts%3Dtrue&query=%24.approximate_member_count&logo=discord&logoColor=white&label=Discord&color=green&suffix=%20total
+[discord-url]: https://www.comfy.org/discord
+
+[github-release-shield]: https://img.shields.io/github/v/release/comfyanonymous/ComfyUI?style=flat&sort=semver
+[github-release-link]: https://github.com/comfyanonymous/ComfyUI/releases
+[github-release-date-shield]: https://img.shields.io/github/release-date/comfyanonymous/ComfyUI?style=flat
+[github-downloads-shield]: https://img.shields.io/github/downloads/comfyanonymous/ComfyUI/total?style=flat
+[github-downloads-latest-shield]: https://img.shields.io/github/downloads/comfyanonymous/ComfyUI/latest/total?style=flat&label=downloads%40latest
+[github-downloads-link]: https://github.com/comfyanonymous/ComfyUI/releases
+
 ![ComfyUI Screenshot](comfyui_screenshot.png)
+</div>
 
 This ui will let you design and execute advanced stable diffusion pipelines using a graph/nodes/flowchart based interface. For some workflow examples and see what ComfyUI can do you can check out:
 ### [ComfyUI Examples](https://comfyanonymous.github.io/ComfyUI_examples/)
@@ -48,6 +75,7 @@ Workflow examples can be found on the [Examples page](https://comfyanonymous.git
 |------------------------------------|--------------------------------------------------------------------------------------------------------------------|
 | Ctrl + Enter                       | Queue up current graph for generation                                                                              |
 | Ctrl + Shift + Enter               | Queue up current graph as first for generation                                                                     |
+| Ctrl + Alt + Enter                 | Cancel current generation                                                                                          |
 | Ctrl + Z/Ctrl + Y                  | Undo/Redo                                                                                                          |
 | Ctrl + S                           | Save workflow                                                                                                      |
 | Ctrl + O                           | Load workflow                                                                                                      |
@@ -70,6 +98,8 @@ Workflow examples can be found on the [Examples page](https://comfyanonymous.git
 | H                                  | Toggle visibility of history                                                                                       |
 | R                                  | Refresh graph                                                                                                      |
 | Double-Click LMB                   | Open node quick search palette                                                                                     |
+| Shift + Drag                       | Move multiple wires at once                                                                                        |
+| Ctrl + Alt + LMB                   | Disconnect all wires from clicked slot                                                                             |
 
 Ctrl can also be replaced with Cmd instead for macOS users
 
@@ -105,17 +135,17 @@ Put your VAE in: models/vae
 ### AMD GPUs (Linux only)
 AMD users can install rocm and pytorch with pip if you don't have it already installed, this is the command to install the stable version:
 
-```pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/rocm6.0```
+```pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/rocm6.1```
 
-This is the command to install the nightly with ROCm 6.0 which might have some performance improvements:
+This is the command to install the nightly with ROCm 6.2 which might have some performance improvements:
 
-```pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/rocm6.1```
+```pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/rocm6.2```
 
 ### NVIDIA
 
 Nvidia users should install stable pytorch using this command:
 
-```pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu121```
+```pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu124```
 
 This is the command to install pytorch nightly instead which might have performance improvements:
 
@@ -200,7 +230,7 @@ To use a textual inversion concepts/embeddings in a text prompt put them in the
 
 Use ```--preview-method auto``` to enable previews.
 
-The default installation includes a fast latent preview method that's low-resolution. To enable higher-quality previews with [TAESD](https://github.com/madebyollin/taesd), download the [taesd_decoder.pth](https://github.com/madebyollin/taesd/raw/main/taesd_decoder.pth) (for SD1.x and SD2.x) and [taesdxl_decoder.pth](https://github.com/madebyollin/taesd/raw/main/taesdxl_decoder.pth) (for SDXL) models and place them in the `models/vae_approx` folder. Once they're installed, restart ComfyUI to enable high-quality previews.
+The default installation includes a fast latent preview method that's low-resolution. To enable higher-quality previews with [TAESD](https://github.com/madebyollin/taesd), download the [taesd_decoder.pth, taesdxl_decoder.pth, taesd3_decoder.pth and taef1_decoder.pth](https://github.com/madebyollin/taesd/) and place them in the `models/vae_approx` folder. Once they're installed, restart ComfyUI and launch it with `--preview-method taesd` to enable high-quality previews.
 
 ## How to use TLS/SSL?
 Generate a self-signed certificate (not appropriate for shared/production use) and key by running the command: `openssl req -x509 -newkey rsa:4096 -keyout key.pem -out cert.pem -sha256 -days 3650 -nodes -subj "/C=XX/ST=StateName/L=CityName/O=CompanyName/OU=CompanySectionName/CN=CommonNameOrHostname"`
@@ -216,6 +246,47 @@ Use `--tls-keyfile key.pem --tls-certfile cert.pem` to enable TLS/SSL, the app w
 
 See also: [https://www.comfy.org/](https://www.comfy.org/)
 
+## Frontend Development
+
+As of August 15, 2024, we have transitioned to a new frontend, which is now hosted in a separate repository: [ComfyUI Frontend](https://github.com/Comfy-Org/ComfyUI_frontend). This repository now hosts the compiled JS (from TS/Vue) under the `web/` directory.
+
+### Reporting Issues and Requesting Features
+
+For any bugs, issues, or feature requests related to the frontend, please use the [ComfyUI Frontend repository](https://github.com/Comfy-Org/ComfyUI_frontend). This will help us manage and address frontend-specific concerns more efficiently.
+
+### Using the Latest Frontend
+
+The new frontend is now the default for ComfyUI. However, please note:
+
+1. The frontend in the main ComfyUI repository is updated weekly.
+2. Daily releases are available in the separate frontend repository.
+
+To use the most up-to-date frontend version:
+
+1. For the latest daily release, launch ComfyUI with this command line argument:
+
+   ```
+   --front-end-version Comfy-Org/ComfyUI_frontend@latest
+   ```
+
+2. For a specific version, replace `latest` with the desired version number:
+
+   ```
+   --front-end-version Comfy-Org/ComfyUI_frontend@1.2.2
+   ```
+
+This approach allows you to easily switch between the stable weekly release and the cutting-edge daily updates, or even specific versions for testing purposes.
+
+### Accessing the Legacy Frontend
+
+If you need to use the legacy frontend for any reason, you can access it using the following command line argument:
+
+```
+--front-end-version Comfy-Org/ComfyUI_legacy_frontend@latest
+```
+
+This will use a snapshot of the legacy frontend preserved in the [ComfyUI Legacy Frontend repository](https://github.com/Comfy-Org/ComfyUI_legacy_frontend).
+
 # QA
 
 ### Which GPU should I buy for this?

api_server/routes/internal/README.md

@@ -0,0 +1,3 @@
+# ComfyUI Internal Routes
+
+All routes under the `/internal` path are designated for **internal use by ComfyUI only**. These routes are not intended for use by external applications may change at any time without notice.

api_server/routes/internal/internal_routes.py

@@ -0,0 +1,44 @@
+from aiohttp import web
+from typing import Optional
+from folder_paths import models_dir, user_directory, output_directory
+from api_server.services.file_service import FileService
+import app.logger
+
+class InternalRoutes:
+    '''
+    The top level web router for internal routes: /internal/*
+    The endpoints here should NOT be depended upon. It is for ComfyUI frontend use only.
+    Check README.md for more information.
+    
+    '''
+    def __init__(self):
+        self.routes: web.RouteTableDef = web.RouteTableDef()
+        self._app: Optional[web.Application] = None
+        self.file_service = FileService({
+            "models": models_dir,
+            "user": user_directory,
+            "output": output_directory
+        })
+
+    def setup_routes(self):
+        @self.routes.get('/files')
+        async def list_files(request):
+            directory_key = request.query.get('directory', '')
+            try:
+                file_list = self.file_service.list_files(directory_key)
+                return web.json_response({"files": file_list})
+            except ValueError as e:
+                return web.json_response({"error": str(e)}, status=400)
+            except Exception as e:
+                return web.json_response({"error": str(e)}, status=500)
+
+        @self.routes.get('/logs')
+        async def get_logs(request):
+            return web.json_response(app.logger.get_logs())
+
+    def get_app(self):
+        if self._app is None:
+            self._app = web.Application()
+            self.setup_routes()
+            self._app.add_routes(self.routes)
+        return self._app

api_server/services/file_service.py

@@ -0,0 +1,13 @@
+from typing import Dict, List, Optional
+from api_server.utils.file_operations import FileSystemOperations, FileSystemItem
+
+class FileService:
+    def __init__(self, allowed_directories: Dict[str, str], file_system_ops: Optional[FileSystemOperations] = None):
+        self.allowed_directories: Dict[str, str] = allowed_directories
+        self.file_system_ops: FileSystemOperations = file_system_ops or FileSystemOperations()
+
+    def list_files(self, directory_key: str) -> List[FileSystemItem]:
+        if directory_key not in self.allowed_directories:
+            raise ValueError("Invalid directory key")
+        directory_path: str = self.allowed_directories[directory_key]
+        return self.file_system_ops.walk_directory(directory_path)

api_server/utils/file_operations.py

@@ -0,0 +1,42 @@
+import os
+from typing import List, Union, TypedDict, Literal
+from typing_extensions import TypeGuard
+class FileInfo(TypedDict):
+    name: str
+    path: str
+    type: Literal["file"]
+    size: int
+
+class DirectoryInfo(TypedDict):
+    name: str
+    path: str
+    type: Literal["directory"]
+
+FileSystemItem = Union[FileInfo, DirectoryInfo]
+
+def is_file_info(item: FileSystemItem) -> TypeGuard[FileInfo]:
+    return item["type"] == "file"
+
+class FileSystemOperations:
+    @staticmethod
+    def walk_directory(directory: str) -> List[FileSystemItem]:
+        file_list: List[FileSystemItem] = []
+        for root, dirs, files in os.walk(directory):
+            for name in files:
+                file_path = os.path.join(root, name)
+                relative_path = os.path.relpath(file_path, directory)
+                file_list.append({
+                    "name": name,
+                    "path": relative_path,
+                    "type": "file",
+                    "size": os.path.getsize(file_path)
+                })
+            for name in dirs:
+                dir_path = os.path.join(root, name)
+                relative_path = os.path.relpath(dir_path, directory)
+                file_list.append({
+                    "name": name,
+                    "path": relative_path,
+                    "type": "directory"
+                })
+        return file_list

app/frontend_management.py

@@ -8,7 +8,7 @@ import zipfile
 from dataclasses import dataclass
 from functools import cached_property
 from pathlib import Path
-from typing import TypedDict
+from typing import TypedDict, Optional
 
 import requests
 from typing_extensions import NotRequired
@@ -132,12 +132,13 @@ class FrontendManager:
         return match_result.group(1), match_result.group(2), match_result.group(3)
 
     @classmethod
-    def init_frontend_unsafe(cls, version_string: str) -> str:
+    def init_frontend_unsafe(cls, version_string: str, provider: Optional[FrontEndProvider] = None) -> str:
         """
         Initializes the frontend for the specified version.
 
         Args:
             version_string (str): The version string.
+            provider (FrontEndProvider, optional): The provider to use. Defaults to None.
 
         Returns:
             str: The path to the initialized frontend.
@@ -150,7 +151,7 @@ class FrontendManager:
             return cls.DEFAULT_FRONTEND_PATH
 
         repo_owner, repo_name, version = cls.parse_version_string(version_string)
-        provider = FrontEndProvider(repo_owner, repo_name)
+        provider = provider or FrontEndProvider(repo_owner, repo_name)
         release = provider.get_release(version)
 
         semantic_version = release["tag_name"].lstrip("v")
@@ -158,15 +159,21 @@ class FrontendManager:
             Path(cls.CUSTOM_FRONTENDS_ROOT) / provider.folder_name / semantic_version
         )
         if not os.path.exists(web_root):
-            os.makedirs(web_root, exist_ok=True)
+            try:
-            logging.info(
+                os.makedirs(web_root, exist_ok=True)
-                "Downloading frontend(%s) version(%s) to (%s)",
+                logging.info(
-                provider.folder_name,
+                    "Downloading frontend(%s) version(%s) to (%s)",
-                semantic_version,
+                    provider.folder_name,
-                web_root,
+                    semantic_version,
-            )
+                    web_root,
-            logging.debug(release)
+                )
-            download_release_asset_zip(release, destination_path=web_root)
+                logging.debug(release)
+                download_release_asset_zip(release, destination_path=web_root)
+            finally:
+                # Clean up the directory if it is empty, i.e. the download failed
+                if not os.listdir(web_root):
+                    os.rmdir(web_root)
+
         return web_root
 
     @classmethod

app/logger.py

@@ -0,0 +1,31 @@
+import logging
+from logging.handlers import MemoryHandler
+from collections import deque
+
+logs = None
+formatter = logging.Formatter("%(asctime)s - %(name)s - %(levelname)s - %(message)s")
+
+
+def get_logs():
+    return "\n".join([formatter.format(x) for x in logs])
+
+
+def setup_logger(verbose: bool = False, capacity: int = 300):
+    global logs
+    if logs:
+        return
+
+    # Setup default global logger
+    logger = logging.getLogger()
+    logger.setLevel(logging.DEBUG if verbose else logging.INFO)
+
+    stream_handler = logging.StreamHandler()
+    stream_handler.setFormatter(logging.Formatter("%(message)s"))
+    logger.addHandler(stream_handler)
+
+    # Create a memory handler with a deque as its buffer
+    logs = deque(maxlen=capacity)
+    memory_handler = MemoryHandler(capacity, flushLevel=logging.INFO)
+    memory_handler.buffer = logs
+    memory_handler.setFormatter(formatter)
+    logger.addHandler(memory_handler)

comfy/cli_args.py

@@ -92,6 +92,8 @@ class LatentPreviewMethod(enum.Enum):
 
 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.")
+
 cache_group = parser.add_mutually_exclusive_group()
 cache_group.add_argument("--cache-classic", action="store_true", help="Use the old style (aggressive) caching.")
 cache_group.add_argument("--cache-lru", type=int, default=0, help="Use LRU caching with a maximum of N node results cached. May use more RAM/VRAM.")
@@ -116,10 +118,14 @@ vram_group.add_argument("--lowvram", action="store_true", help="Split the unet i
 vram_group.add_argument("--novram", action="store_true", help="When lowvram isn't enough.")
 vram_group.add_argument("--cpu", action="store_true", help="To use the CPU for everything (slow).")
 
+parser.add_argument("--reserve-vram", type=float, default=None, help="Set the amount of vram in GB you want to reserve for use by your OS/other software. By default some amount is reverved depending on your OS.")
+
+
 parser.add_argument("--default-hashing-function", type=str, choices=['md5', 'sha1', 'sha256', 'sha512'], default='sha256', help="Allows you to choose the hash function to use for duplicate filename / contents comparison. Default is sha256.")
 
 parser.add_argument("--disable-smart-memory", action="store_true", help="Force ComfyUI to agressively offload to regular ram instead of keeping models in vram when it can.")
 parser.add_argument("--deterministic", action="store_true", help="Make pytorch use slower deterministic algorithms when it can. Note that this might not make images deterministic in all cases.")
+parser.add_argument("--fast", action="store_true", help="Enable some untested and potentially quality deteriorating optimizations.")
 
 parser.add_argument("--dont-print-server", action="store_true", help="Don't print server output.")
 parser.add_argument("--quick-test-for-ci", action="store_true", help="Quick test for CI.")
@@ -175,10 +181,3 @@ if args.windows_standalone_build:
 
 if args.disable_auto_launch:
     args.auto_launch = False
-
-import logging
-logging_level = logging.INFO
-if args.verbose:
-    logging_level = logging.DEBUG
-
-logging.basicConfig(format="%(message)s", level=logging_level)

comfy/clip_model.py

@@ -88,10 +88,11 @@ class CLIPTextModel_(torch.nn.Module):
         heads = config_dict["num_attention_heads"]
         intermediate_size = config_dict["intermediate_size"]
         intermediate_activation = config_dict["hidden_act"]
+        num_positions = config_dict["max_position_embeddings"]
         self.eos_token_id = config_dict["eos_token_id"]
 
         super().__init__()
-        self.embeddings = CLIPEmbeddings(embed_dim, dtype=dtype, device=device, operations=operations)
+        self.embeddings = CLIPEmbeddings(embed_dim, num_positions=num_positions, dtype=dtype, device=device, operations=operations)
         self.encoder = CLIPEncoder(num_layers, embed_dim, heads, intermediate_size, intermediate_activation, dtype, device, operations)
         self.final_layer_norm = operations.LayerNorm(embed_dim, dtype=dtype, device=device)
 
@@ -123,7 +124,6 @@ class CLIPTextModel(torch.nn.Module):
         self.text_model = CLIPTextModel_(config_dict, dtype, device, operations)
         embed_dim = config_dict["hidden_size"]
         self.text_projection = operations.Linear(embed_dim, embed_dim, bias=False, dtype=dtype, device=device)
-        self.text_projection.weight.copy_(torch.eye(embed_dim))
         self.dtype = dtype
 
     def get_input_embeddings(self):

comfy/controlnet.py

@@ -34,7 +34,7 @@ import comfy.t2i_adapter.adapter
 import comfy.ldm.cascade.controlnet
 import comfy.cldm.mmdit
 import comfy.ldm.hydit.controlnet
-import comfy.ldm.flux.controlnet_xlabs
+import comfy.ldm.flux.controlnet
 
 
 def broadcast_image_to(tensor, target_batch_size, batched_number):
@@ -148,7 +148,7 @@ class ControlBase:
                         elif self.strength_type == StrengthType.LINEAR_UP:
                             x *= (self.strength ** float(len(control_output) - i))
 
-                    if x.dtype != output_dtype:
+                    if output_dtype is not None and x.dtype != output_dtype:
                         x = x.to(output_dtype)
 
                 out[key].append(x)
@@ -206,7 +206,6 @@ class ControlNet(ControlBase):
         if self.manual_cast_dtype is not None:
             dtype = self.manual_cast_dtype
 
-        output_dtype = x_noisy.dtype
         if self.cond_hint is None or x_noisy.shape[2] * self.compression_ratio != self.cond_hint.shape[2] or x_noisy.shape[3] * self.compression_ratio != self.cond_hint.shape[3]:
             if self.cond_hint is not None:
                 del self.cond_hint
@@ -236,7 +235,7 @@ class ControlNet(ControlBase):
         x_noisy = self.model_sampling_current.calculate_input(t, x_noisy)
 
         control = self.control_model(x=x_noisy.to(dtype), hint=self.cond_hint, timesteps=timestep.to(dtype), context=context.to(dtype), **extra)
-        return self.control_merge(control, control_prev, output_dtype)
+        return self.control_merge(control, control_prev, output_dtype=None)
 
     def copy(self):
         c = ControlNet(None, global_average_pooling=self.global_average_pooling, load_device=self.load_device, manual_cast_dtype=self.manual_cast_dtype)
@@ -391,7 +390,8 @@ def controlnet_config(sd):
     else:
         operations = comfy.ops.disable_weight_init
 
-    return model_config, operations, load_device, unet_dtype, manual_cast_dtype
+    offload_device = comfy.model_management.unet_offload_device()
+    return model_config, operations, load_device, unet_dtype, manual_cast_dtype, offload_device
 
 def controlnet_load_state_dict(control_model, sd):
     missing, unexpected = control_model.load_state_dict(sd, strict=False)
@@ -405,12 +405,12 @@ def controlnet_load_state_dict(control_model, sd):
 
 def load_controlnet_mmdit(sd):
     new_sd = comfy.model_detection.convert_diffusers_mmdit(sd, "")
-    model_config, operations, load_device, unet_dtype, manual_cast_dtype = controlnet_config(new_sd)
+    model_config, operations, load_device, unet_dtype, manual_cast_dtype, offload_device = controlnet_config(new_sd)
     num_blocks = comfy.model_detection.count_blocks(new_sd, 'joint_blocks.{}.')
     for k in sd:
         new_sd[k] = sd[k]
 
-    control_model = comfy.cldm.mmdit.ControlNet(num_blocks=num_blocks, operations=operations, device=load_device, dtype=unet_dtype, **model_config.unet_config)
+    control_model = comfy.cldm.mmdit.ControlNet(num_blocks=num_blocks, operations=operations, device=offload_device, dtype=unet_dtype, **model_config.unet_config)
     control_model = controlnet_load_state_dict(control_model, new_sd)
 
     latent_format = comfy.latent_formats.SD3()
@@ -420,9 +420,9 @@ def load_controlnet_mmdit(sd):
 
 
 def load_controlnet_hunyuandit(controlnet_data):
-    model_config, operations, load_device, unet_dtype, manual_cast_dtype = controlnet_config(controlnet_data)
+    model_config, operations, load_device, unet_dtype, manual_cast_dtype, offload_device = controlnet_config(controlnet_data)
 
-    control_model = comfy.ldm.hydit.controlnet.HunYuanControlNet(operations=operations, device=load_device, dtype=unet_dtype)
+    control_model = comfy.ldm.hydit.controlnet.HunYuanControlNet(operations=operations, device=offload_device, dtype=unet_dtype)
     control_model = controlnet_load_state_dict(control_model, controlnet_data)
 
     latent_format = comfy.latent_formats.SDXL()
@@ -430,14 +430,36 @@ def load_controlnet_hunyuandit(controlnet_data):
     control = ControlNet(control_model, compression_ratio=1, latent_format=latent_format, load_device=load_device, manual_cast_dtype=manual_cast_dtype, extra_conds=extra_conds, strength_type=StrengthType.CONSTANT)
     return control
 
-def load_controlnet_flux_xlabs(sd):
+def load_controlnet_flux_xlabs_mistoline(sd, mistoline=False):
-    model_config, operations, load_device, unet_dtype, manual_cast_dtype = controlnet_config(sd)
+    model_config, operations, load_device, unet_dtype, manual_cast_dtype, offload_device = controlnet_config(sd)
-    control_model = comfy.ldm.flux.controlnet_xlabs.ControlNetFlux(operations=operations, device=load_device, dtype=unet_dtype, **model_config.unet_config)
+    control_model = comfy.ldm.flux.controlnet.ControlNetFlux(mistoline=mistoline, operations=operations, device=offload_device, dtype=unet_dtype, **model_config.unet_config)
     control_model = controlnet_load_state_dict(control_model, sd)
     extra_conds = ['y', 'guidance']
     control = ControlNet(control_model, load_device=load_device, manual_cast_dtype=manual_cast_dtype, extra_conds=extra_conds)
     return control
 
+def load_controlnet_flux_instantx(sd):
+    new_sd = comfy.model_detection.convert_diffusers_mmdit(sd, "")
+    model_config, operations, load_device, unet_dtype, manual_cast_dtype, offload_device = controlnet_config(new_sd)
+    for k in sd:
+        new_sd[k] = sd[k]
+
+    num_union_modes = 0
+    union_cnet = "controlnet_mode_embedder.weight"
+    if union_cnet in new_sd:
+        num_union_modes = new_sd[union_cnet].shape[0]
+
+    control_model = comfy.ldm.flux.controlnet.ControlNetFlux(latent_input=True, num_union_modes=num_union_modes, operations=operations, device=offload_device, dtype=unet_dtype, **model_config.unet_config)
+    control_model = controlnet_load_state_dict(control_model, new_sd)
+
+    latent_format = comfy.latent_formats.Flux()
+    extra_conds = ['y', 'guidance']
+    control = ControlNet(control_model, compression_ratio=1, latent_format=latent_format, load_device=load_device, manual_cast_dtype=manual_cast_dtype, extra_conds=extra_conds)
+    return control
+
+def convert_mistoline(sd):
+    return comfy.utils.state_dict_prefix_replace(sd, {"single_controlnet_blocks.": "controlnet_single_blocks."})
+
 
 def load_controlnet(ckpt_path, model=None):
     controlnet_data = comfy.utils.load_torch_file(ckpt_path, safe_load=True)
@@ -500,11 +522,15 @@ def load_controlnet(ckpt_path, model=None):
         if len(leftover_keys) > 0:
             logging.warning("leftover keys: {}".format(leftover_keys))
         controlnet_data = new_sd
-    elif "controlnet_blocks.0.weight" in controlnet_data: #SD3 diffusers format
+    elif "controlnet_blocks.0.weight" in controlnet_data:
         if "double_blocks.0.img_attn.norm.key_norm.scale" in controlnet_data:
-            return load_controlnet_flux_xlabs(controlnet_data)
+            return load_controlnet_flux_xlabs_mistoline(controlnet_data)
-        else:
+        elif "pos_embed_input.proj.weight" in controlnet_data:
-            return load_controlnet_mmdit(controlnet_data)
+            return load_controlnet_mmdit(controlnet_data) #SD3 diffusers controlnet
+        elif "controlnet_x_embedder.weight" in controlnet_data:
+            return load_controlnet_flux_instantx(controlnet_data)
+    elif "controlnet_blocks.0.linear.weight" in controlnet_data: #mistoline flux
+        return load_controlnet_flux_xlabs_mistoline(convert_mistoline(controlnet_data), mistoline=True)
 
     pth_key = 'control_model.zero_convs.0.0.weight'
     pth = False
@@ -536,6 +562,7 @@ def load_controlnet(ckpt_path, model=None):
     if manual_cast_dtype is not None:
         controlnet_config["operations"] = comfy.ops.manual_cast
     controlnet_config["dtype"] = unet_dtype
+    controlnet_config["device"] = comfy.model_management.unet_offload_device()
     controlnet_config.pop("out_channels")
     controlnet_config["hint_channels"] = controlnet_data["{}input_hint_block.0.weight".format(prefix)].shape[1]
     control_model = comfy.cldm.cldm.ControlNet(**controlnet_config)

comfy/float.py

@@ -0,0 +1,66 @@
+import torch
+import math
+
+def calc_mantissa(abs_x, exponent, normal_mask, MANTISSA_BITS, EXPONENT_BIAS, generator=None):
+    mantissa_scaled = torch.where(
+        normal_mask,
+        (abs_x / (2.0 ** (exponent - EXPONENT_BIAS)) - 1.0) * (2**MANTISSA_BITS),
+        (abs_x / (2.0 ** (-EXPONENT_BIAS + 1 - MANTISSA_BITS)))
+    )
+
+    mantissa_scaled += torch.rand(mantissa_scaled.size(), dtype=mantissa_scaled.dtype, layout=mantissa_scaled.layout, device=mantissa_scaled.device, generator=generator)
+    return mantissa_scaled.floor() / (2**MANTISSA_BITS)
+
+#Not 100% sure about this
+def manual_stochastic_round_to_float8(x, dtype, generator=None):
+    if dtype == torch.float8_e4m3fn:
+        EXPONENT_BITS, MANTISSA_BITS, EXPONENT_BIAS = 4, 3, 7
+    elif dtype == torch.float8_e5m2:
+        EXPONENT_BITS, MANTISSA_BITS, EXPONENT_BIAS = 5, 2, 15
+    else:
+        raise ValueError("Unsupported dtype")
+
+    x = x.half()
+    sign = torch.sign(x)
+    abs_x = x.abs()
+    sign = torch.where(abs_x == 0, 0, sign)
+
+    # Combine exponent calculation and clamping
+    exponent = torch.clamp(
+        torch.floor(torch.log2(abs_x)) + EXPONENT_BIAS,
+        0, 2**EXPONENT_BITS - 1
+    )
+
+    # Combine mantissa calculation and rounding
+    normal_mask = ~(exponent == 0)
+
+    abs_x[:] = calc_mantissa(abs_x, exponent, normal_mask, MANTISSA_BITS, EXPONENT_BIAS, generator=generator)
+
+    sign *= torch.where(
+        normal_mask,
+        (2.0 ** (exponent - EXPONENT_BIAS)) * (1.0 + abs_x),
+        (2.0 ** (-EXPONENT_BIAS + 1)) * abs_x
+    )
+
+    return sign
+
+
+
+def stochastic_rounding(value, dtype, seed=0):
+    if dtype == torch.float32:
+        return value.to(dtype=torch.float32)
+    if dtype == torch.float16:
+        return value.to(dtype=torch.float16)
+    if dtype == torch.bfloat16:
+        return value.to(dtype=torch.bfloat16)
+    if dtype == torch.float8_e4m3fn or dtype == torch.float8_e5m2:
+        generator = torch.Generator(device=value.device)
+        generator.manual_seed(seed)
+        output = torch.empty_like(value, dtype=dtype)
+        num_slices = max(1, (value.numel() / (4096 * 4096)))
+        slice_size = max(1, round(value.shape[0] / num_slices))
+        for i in range(0, value.shape[0], slice_size):
+            output[i:i+slice_size].copy_(manual_stochastic_round_to_float8(value[i:i+slice_size], dtype, generator=generator))
+        return output
+
+    return value.to(dtype=dtype)

comfy/k_diffusion/sampling.py

@@ -9,6 +9,7 @@ from tqdm.auto import trange, tqdm
 from . import utils
 from . import deis
 import comfy.model_patcher
+import comfy.model_sampling
 
 def append_zero(x):
     return torch.cat([x, x.new_zeros([1])])
@@ -509,6 +510,9 @@ def sample_dpm_adaptive(model, x, sigma_min, sigma_max, extra_args=None, callbac
 
 @torch.no_grad()
 def sample_dpmpp_2s_ancestral(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None):
+    if isinstance(model.inner_model.inner_model.model_sampling, comfy.model_sampling.CONST):
+        return sample_dpmpp_2s_ancestral_RF(model, x, sigmas, extra_args, callback, disable, eta, s_noise, noise_sampler)
+
     """Ancestral sampling with DPM-Solver++(2S) second-order steps."""
     extra_args = {} if extra_args is None else extra_args
     noise_sampler = default_noise_sampler(x) if noise_sampler is None else noise_sampler
@@ -541,6 +545,55 @@ def sample_dpmpp_2s_ancestral(model, x, sigmas, extra_args=None, callback=None,
     return x
 
 
+@torch.no_grad()
+def sample_dpmpp_2s_ancestral_RF(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None):
+    """Ancestral sampling with DPM-Solver++(2S) second-order steps."""
+    extra_args = {} if extra_args is None else extra_args
+    noise_sampler = default_noise_sampler(x) if noise_sampler is None else noise_sampler
+    s_in = x.new_ones([x.shape[0]])
+    sigma_fn = lambda lbda: (lbda.exp() + 1) ** -1
+    lambda_fn = lambda sigma: ((1-sigma)/sigma).log()
+
+    # logged_x = x.unsqueeze(0)
+
+    for i in trange(len(sigmas) - 1, disable=disable):
+        denoised = model(x, sigmas[i] * s_in, **extra_args)
+        downstep_ratio = 1 + (sigmas[i+1]/sigmas[i] - 1) * eta
+        sigma_down = sigmas[i+1] * downstep_ratio
+        alpha_ip1 = 1 - sigmas[i+1]
+        alpha_down = 1 - sigma_down
+        renoise_coeff = (sigmas[i+1]**2 - sigma_down**2*alpha_ip1**2/alpha_down**2)**0.5
+        # sigma_down, sigma_up = get_ancestral_step(sigmas[i], sigmas[i + 1], eta=eta)
+        if callback is not None:
+            callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised})
+        if sigmas[i + 1] == 0:
+            # Euler method
+            d = to_d(x, sigmas[i], denoised)
+            dt = sigma_down - sigmas[i]
+            x = x + d * dt
+        else:
+            # DPM-Solver++(2S)
+            if sigmas[i] == 1.0:
+                sigma_s = 0.9999
+            else:
+                t_i, t_down = lambda_fn(sigmas[i]), lambda_fn(sigma_down)
+                r = 1 / 2
+                h = t_down - t_i
+                s = t_i + r * h
+                sigma_s = sigma_fn(s)
+            # sigma_s = sigmas[i+1]
+            sigma_s_i_ratio = sigma_s / sigmas[i]
+            u = sigma_s_i_ratio * x + (1 - sigma_s_i_ratio) * denoised
+            D_i = model(u, sigma_s * s_in, **extra_args)
+            sigma_down_i_ratio = sigma_down / sigmas[i]
+            x = sigma_down_i_ratio * x + (1 - sigma_down_i_ratio) * D_i
+            # print("sigma_i", sigmas[i], "sigma_ip1", sigmas[i+1],"sigma_down", sigma_down, "sigma_down_i_ratio", sigma_down_i_ratio, "sigma_s_i_ratio", sigma_s_i_ratio, "renoise_coeff", renoise_coeff)
+        # Noise addition
+        if sigmas[i + 1] > 0 and eta > 0:
+            x = (alpha_ip1/alpha_down) * x + noise_sampler(sigmas[i], sigmas[i + 1]) * s_noise * renoise_coeff
+        # logged_x = torch.cat((logged_x, x.unsqueeze(0)), dim=0)
+    return x
+
 @torch.no_grad()
 def sample_dpmpp_sde(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None, r=1 / 2):
     """DPM-Solver++ (stochastic)."""

comfy/latent_formats.py

@@ -141,6 +141,7 @@ class StableAudio1(LatentFormat):
     latent_channels = 64
 
 class Flux(SD3):
+    latent_channels = 16
     def __init__(self):
         self.scale_factor = 0.3611
         self.shift_factor = 0.1159
@@ -162,6 +163,7 @@ class Flux(SD3):
             [-0.0005, -0.0530, -0.0020],
             [-0.1273, -0.0932, -0.0680]
         ]
+        self.taesd_decoder_name = "taef1_decoder"
 
     def process_in(self, latent):
         return (latent - self.shift_factor) * self.scale_factor

comfy/ldm/common_dit.py

@@ -1,4 +1,5 @@
 import torch
+import comfy.ops
 
 def pad_to_patch_size(img, patch_size=(2, 2), padding_mode="circular"):
     if padding_mode == "circular" and torch.jit.is_tracing() or torch.jit.is_scripting():
@@ -6,3 +7,15 @@ def pad_to_patch_size(img, patch_size=(2, 2), padding_mode="circular"):
     pad_h = (patch_size[0] - img.shape[-2] % patch_size[0]) % patch_size[0]
     pad_w = (patch_size[1] - img.shape[-1] % patch_size[1]) % patch_size[1]
     return torch.nn.functional.pad(img, (0, pad_w, 0, pad_h), mode=padding_mode)
+
+try:
+    rms_norm_torch = torch.nn.functional.rms_norm
+except:
+    rms_norm_torch = None
+
+def rms_norm(x, weight, eps=1e-6):
+    if rms_norm_torch is not None:
+        return rms_norm_torch(x, weight.shape, weight=comfy.ops.cast_to(weight, dtype=x.dtype, device=x.device), eps=eps)
+    else:
+        rrms = torch.rsqrt(torch.mean(x**2, dim=-1, keepdim=True) + eps)
+        return (x * rrms) * comfy.ops.cast_to(weight, dtype=x.dtype, device=x.device)

comfy/ldm/flux/controlnet.py

@@ -0,0 +1,200 @@
+#Original code can be found on: https://github.com/XLabs-AI/x-flux/blob/main/src/flux/controlnet.py
+#modified to support different types of flux controlnets
+
+import torch
+import math
+from torch import Tensor, nn
+from einops import rearrange, repeat
+
+from .layers import (DoubleStreamBlock, EmbedND, LastLayer,
+                                 MLPEmbedder, SingleStreamBlock,
+                                 timestep_embedding)
+
+from .model import Flux
+import comfy.ldm.common_dit
+
+class MistolineCondDownsamplBlock(nn.Module):
+    def __init__(self, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.encoder = nn.Sequential(
+            operations.Conv2d(3, 16, 3, padding=1, dtype=dtype, device=device),
+            nn.SiLU(),
+            operations.Conv2d(16, 16, 1, dtype=dtype, device=device),
+            nn.SiLU(),
+            operations.Conv2d(16, 16, 3, padding=1, dtype=dtype, device=device),
+            nn.SiLU(),
+            operations.Conv2d(16, 16, 3, padding=1, stride=2, dtype=dtype, device=device),
+            nn.SiLU(),
+            operations.Conv2d(16, 16, 3, padding=1, dtype=dtype, device=device),
+            nn.SiLU(),
+            operations.Conv2d(16, 16, 3, padding=1, stride=2, dtype=dtype, device=device),
+            nn.SiLU(),
+            operations.Conv2d(16, 16, 3, padding=1, dtype=dtype, device=device),
+            nn.SiLU(),
+            operations.Conv2d(16, 16, 3, padding=1, stride=2, dtype=dtype, device=device),
+            nn.SiLU(),
+            operations.Conv2d(16, 16, 1, dtype=dtype, device=device),
+            nn.SiLU(),
+            operations.Conv2d(16, 16, 3, padding=1, dtype=dtype, device=device)
+        )
+
+    def forward(self, x):
+        return self.encoder(x)
+
+class MistolineControlnetBlock(nn.Module):
+    def __init__(self, hidden_size, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.linear = operations.Linear(hidden_size, hidden_size, dtype=dtype, device=device)
+        self.act = nn.SiLU()
+
+    def forward(self, x):
+        return self.act(self.linear(x))
+
+
+class ControlNetFlux(Flux):
+    def __init__(self, latent_input=False, num_union_modes=0, mistoline=False, image_model=None, dtype=None, device=None, operations=None, **kwargs):
+        super().__init__(final_layer=False, dtype=dtype, device=device, operations=operations, **kwargs)
+
+        self.main_model_double = 19
+        self.main_model_single = 38
+
+        self.mistoline = mistoline
+        # add ControlNet blocks
+        if self.mistoline:
+            control_block = lambda : MistolineControlnetBlock(self.hidden_size, dtype=dtype, device=device, operations=operations)
+        else:
+            control_block = lambda : operations.Linear(self.hidden_size, self.hidden_size, dtype=dtype, device=device)
+
+        self.controlnet_blocks = nn.ModuleList([])
+        for _ in range(self.params.depth):
+            self.controlnet_blocks.append(control_block())
+
+        self.controlnet_single_blocks = nn.ModuleList([])
+        for _ in range(self.params.depth_single_blocks):
+            self.controlnet_single_blocks.append(control_block())
+
+        self.num_union_modes = num_union_modes
+        self.controlnet_mode_embedder = None
+        if self.num_union_modes > 0:
+            self.controlnet_mode_embedder = operations.Embedding(self.num_union_modes, self.hidden_size, dtype=dtype, device=device)
+
+        self.gradient_checkpointing = False
+        self.latent_input = latent_input
+        self.pos_embed_input = operations.Linear(self.in_channels, self.hidden_size, bias=True, dtype=dtype, device=device)
+        if not self.latent_input:
+            if self.mistoline:
+                self.input_cond_block = MistolineCondDownsamplBlock(dtype=dtype, device=device, operations=operations)
+            else:
+                self.input_hint_block = nn.Sequential(
+                    operations.Conv2d(3, 16, 3, padding=1, dtype=dtype, device=device),
+                    nn.SiLU(),
+                    operations.Conv2d(16, 16, 3, padding=1, dtype=dtype, device=device),
+                    nn.SiLU(),
+                    operations.Conv2d(16, 16, 3, padding=1, stride=2, dtype=dtype, device=device),
+                    nn.SiLU(),
+                    operations.Conv2d(16, 16, 3, padding=1, dtype=dtype, device=device),
+                    nn.SiLU(),
+                    operations.Conv2d(16, 16, 3, padding=1, stride=2, dtype=dtype, device=device),
+                    nn.SiLU(),
+                    operations.Conv2d(16, 16, 3, padding=1, dtype=dtype, device=device),
+                    nn.SiLU(),
+                    operations.Conv2d(16, 16, 3, padding=1, stride=2, dtype=dtype, device=device),
+                    nn.SiLU(),
+                    operations.Conv2d(16, 16, 3, padding=1, dtype=dtype, device=device)
+                )
+
+    def forward_orig(
+        self,
+        img: Tensor,
+        img_ids: Tensor,
+        controlnet_cond: Tensor,
+        txt: Tensor,
+        txt_ids: Tensor,
+        timesteps: Tensor,
+        y: Tensor,
+        guidance: Tensor = None,
+        control_type: Tensor = None,
+    ) -> Tensor:
+        if img.ndim != 3 or txt.ndim != 3:
+            raise ValueError("Input img and txt tensors must have 3 dimensions.")
+
+        # running on sequences img
+        img = self.img_in(img)
+
+        controlnet_cond = self.pos_embed_input(controlnet_cond)
+        img = img + controlnet_cond
+        vec = self.time_in(timestep_embedding(timesteps, 256))
+        if self.params.guidance_embed:
+            vec = vec + self.guidance_in(timestep_embedding(guidance, 256))
+        vec = vec + self.vector_in(y)
+        txt = self.txt_in(txt)
+
+        if self.controlnet_mode_embedder is not None and len(control_type) > 0:
+            control_cond = self.controlnet_mode_embedder(torch.tensor(control_type, device=img.device), out_dtype=img.dtype).unsqueeze(0).repeat((txt.shape[0], 1, 1))
+            txt = torch.cat([control_cond, txt], dim=1)
+            txt_ids = torch.cat([txt_ids[:,:1], txt_ids], dim=1)
+
+        ids = torch.cat((txt_ids, img_ids), dim=1)
+        pe = self.pe_embedder(ids)
+
+        controlnet_double = ()
+
+        for i in range(len(self.double_blocks)):
+            img, txt = self.double_blocks[i](img=img, txt=txt, vec=vec, pe=pe)
+            controlnet_double = controlnet_double + (self.controlnet_blocks[i](img),)
+
+        img = torch.cat((txt, img), 1)
+
+        controlnet_single = ()
+
+        for i in range(len(self.single_blocks)):
+            img = self.single_blocks[i](img, vec=vec, pe=pe)
+            controlnet_single = controlnet_single + (self.controlnet_single_blocks[i](img[:, txt.shape[1] :, ...]),)
+
+        repeat = math.ceil(self.main_model_double / len(controlnet_double))
+        if self.latent_input:
+            out_input = ()
+            for x in controlnet_double:
+                    out_input += (x,) * repeat
+        else:
+            out_input = (controlnet_double * repeat)
+
+        out = {"input": out_input[:self.main_model_double]}
+        if len(controlnet_single) > 0:
+            repeat = math.ceil(self.main_model_single / len(controlnet_single))
+            out_output = ()
+            if self.latent_input:
+                for x in controlnet_single:
+                        out_output += (x,) * repeat
+            else:
+                out_output = (controlnet_single * repeat)
+            out["output"] = out_output[:self.main_model_single]
+        return out
+
+    def forward(self, x, timesteps, context, y, guidance=None, hint=None, **kwargs):
+        patch_size = 2
+        if self.latent_input:
+            hint = comfy.ldm.common_dit.pad_to_patch_size(hint, (patch_size, patch_size))
+        elif self.mistoline:
+            hint = hint * 2.0 - 1.0
+            hint = self.input_cond_block(hint)
+        else:
+            hint = hint * 2.0 - 1.0
+            hint = self.input_hint_block(hint)
+
+        hint = rearrange(hint, "b c (h ph) (w pw) -> b (h w) (c ph pw)", ph=patch_size, pw=patch_size)
+
+        bs, c, h, w = x.shape
+        x = comfy.ldm.common_dit.pad_to_patch_size(x, (patch_size, patch_size))
+
+        img = rearrange(x, "b c (h ph) (w pw) -> b (h w) (c ph pw)", ph=patch_size, pw=patch_size)
+
+        h_len = ((h + (patch_size // 2)) // patch_size)
+        w_len = ((w + (patch_size // 2)) // patch_size)
+        img_ids = torch.zeros((h_len, w_len, 3), device=x.device, dtype=x.dtype)
+        img_ids[..., 1] = img_ids[..., 1] + torch.linspace(0, h_len - 1, steps=h_len, device=x.device, dtype=x.dtype)[:, None]
+        img_ids[..., 2] = img_ids[..., 2] + torch.linspace(0, w_len - 1, steps=w_len, device=x.device, dtype=x.dtype)[None, :]
+        img_ids = repeat(img_ids, "h w c -> b (h w) c", b=bs)
+
+        txt_ids = torch.zeros((bs, context.shape[1], 3), device=x.device, dtype=x.dtype)
+        return self.forward_orig(img, img_ids, hint, context, txt_ids, timesteps, y, guidance, control_type=kwargs.get("control_type", []))

comfy/ldm/flux/controlnet_xlabs.py

@@ -1,104 +0,0 @@
-#Original code can be found on: https://github.com/XLabs-AI/x-flux/blob/main/src/flux/controlnet.py
-
-import torch
-from torch import Tensor, nn
-from einops import rearrange, repeat
-
-from .layers import (DoubleStreamBlock, EmbedND, LastLayer,
-                                 MLPEmbedder, SingleStreamBlock,
-                                 timestep_embedding)
-
-from .model import Flux
-import comfy.ldm.common_dit
-
-
-class ControlNetFlux(Flux):
-    def __init__(self, image_model=None, dtype=None, device=None, operations=None, **kwargs):
-        super().__init__(final_layer=False, dtype=dtype, device=device, operations=operations, **kwargs)
-
-        # add ControlNet blocks
-        self.controlnet_blocks = nn.ModuleList([])
-        for _ in range(self.params.depth):
-            controlnet_block = operations.Linear(self.hidden_size, self.hidden_size, dtype=dtype, device=device)
-            # controlnet_block = zero_module(controlnet_block)
-            self.controlnet_blocks.append(controlnet_block)
-        self.pos_embed_input = operations.Linear(self.in_channels, self.hidden_size, bias=True, dtype=dtype, device=device)
-        self.gradient_checkpointing = False
-        self.input_hint_block = nn.Sequential(
-            operations.Conv2d(3, 16, 3, padding=1, dtype=dtype, device=device),
-            nn.SiLU(),
-            operations.Conv2d(16, 16, 3, padding=1, dtype=dtype, device=device),
-            nn.SiLU(),
-            operations.Conv2d(16, 16, 3, padding=1, stride=2, dtype=dtype, device=device),
-            nn.SiLU(),
-            operations.Conv2d(16, 16, 3, padding=1, dtype=dtype, device=device),
-            nn.SiLU(),
-            operations.Conv2d(16, 16, 3, padding=1, stride=2, dtype=dtype, device=device),
-            nn.SiLU(),
-            operations.Conv2d(16, 16, 3, padding=1, dtype=dtype, device=device),
-            nn.SiLU(),
-            operations.Conv2d(16, 16, 3, padding=1, stride=2, dtype=dtype, device=device),
-            nn.SiLU(),
-            operations.Conv2d(16, 16, 3, padding=1, dtype=dtype, device=device)
-        )
-
-    def forward_orig(
-        self,
-        img: Tensor,
-        img_ids: Tensor,
-        controlnet_cond: Tensor,
-        txt: Tensor,
-        txt_ids: Tensor,
-        timesteps: Tensor,
-        y: Tensor,
-        guidance: Tensor = None,
-    ) -> Tensor:
-        if img.ndim != 3 or txt.ndim != 3:
-            raise ValueError("Input img and txt tensors must have 3 dimensions.")
-
-        # running on sequences img
-        img = self.img_in(img)
-        controlnet_cond = self.input_hint_block(controlnet_cond)
-        controlnet_cond = rearrange(controlnet_cond, "b c (h ph) (w pw) -> b (h w) (c ph pw)", ph=2, pw=2)
-        controlnet_cond = self.pos_embed_input(controlnet_cond)
-        img = img + controlnet_cond
-        vec = self.time_in(timestep_embedding(timesteps, 256))
-        if self.params.guidance_embed:
-            vec = vec + self.guidance_in(timestep_embedding(guidance, 256))
-        vec = vec + self.vector_in(y)
-        txt = self.txt_in(txt)
-
-        ids = torch.cat((txt_ids, img_ids), dim=1)
-        pe = self.pe_embedder(ids)
-
-        block_res_samples = ()
-
-        for block in self.double_blocks:
-            img, txt = block(img=img, txt=txt, vec=vec, pe=pe)
-            block_res_samples = block_res_samples + (img,)
-
-        controlnet_block_res_samples = ()
-        for block_res_sample, controlnet_block in zip(block_res_samples, self.controlnet_blocks):
-            block_res_sample = controlnet_block(block_res_sample)
-            controlnet_block_res_samples = controlnet_block_res_samples + (block_res_sample,)
-
-        return {"output": (controlnet_block_res_samples * 10)[:19]}
-
-    def forward(self, x, timesteps, context, y, guidance=None, hint=None, **kwargs):
-        hint = hint * 2.0 - 1.0
-
-        bs, c, h, w = x.shape
-        patch_size = 2
-        x = comfy.ldm.common_dit.pad_to_patch_size(x, (patch_size, patch_size))
-
-        img = rearrange(x, "b c (h ph) (w pw) -> b (h w) (c ph pw)", ph=patch_size, pw=patch_size)
-
-        h_len = ((h + (patch_size // 2)) // patch_size)
-        w_len = ((w + (patch_size // 2)) // patch_size)
-        img_ids = torch.zeros((h_len, w_len, 3), device=x.device, dtype=x.dtype)
-        img_ids[..., 1] = img_ids[..., 1] + torch.linspace(0, h_len - 1, steps=h_len, device=x.device, dtype=x.dtype)[:, None]
-        img_ids[..., 2] = img_ids[..., 2] + torch.linspace(0, w_len - 1, steps=w_len, device=x.device, dtype=x.dtype)[None, :]
-        img_ids = repeat(img_ids, "h w c -> b (h w) c", b=bs)
-
-        txt_ids = torch.zeros((bs, context.shape[1], 3), device=x.device, dtype=x.dtype)
-        return self.forward_orig(img, img_ids, hint, context, txt_ids, timesteps, y, guidance)

comfy/ldm/flux/layers.py

@@ -6,6 +6,7 @@ from torch import Tensor, nn
 
 from .math import attention, rope
 import comfy.ops
+import comfy.ldm.common_dit
 
 
 class EmbedND(nn.Module):
@@ -63,10 +64,7 @@ class RMSNorm(torch.nn.Module):
         self.scale = nn.Parameter(torch.empty((dim), dtype=dtype, device=device))
 
     def forward(self, x: Tensor):
-        x_dtype = x.dtype
+        return comfy.ldm.common_dit.rms_norm(x, self.scale, 1e-6)
-        x = x.float()
-        rrms = torch.rsqrt(torch.mean(x**2, dim=-1, keepdim=True) + 1e-6)
-        return (x * rrms).to(dtype=x_dtype) * comfy.ops.cast_to(self.scale, dtype=x_dtype, device=x.device)
 
 
 class QKNorm(torch.nn.Module):
@@ -178,7 +176,7 @@ class DoubleStreamBlock(nn.Module):
         txt += txt_mod2.gate * self.txt_mlp((1 + txt_mod2.scale) * self.txt_norm2(txt) + txt_mod2.shift)
 
         if txt.dtype == torch.float16:
-            txt = txt.clip(-65504, 65504)
+            txt = torch.nan_to_num(txt, nan=0.0, posinf=65504, neginf=-65504)
 
         return img, txt
 
@@ -233,7 +231,7 @@ class SingleStreamBlock(nn.Module):
         output = self.linear2(torch.cat((attn, self.mlp_act(mlp)), 2))
         x += mod.gate * output
         if x.dtype == torch.float16:
-            x = x.clip(-65504, 65504)
+            x = torch.nan_to_num(x, nan=0.0, posinf=65504, neginf=-65504)
         return x
 
 

comfy/ldm/flux/model.py

@@ -114,19 +114,28 @@ class Flux(nn.Module):
         ids = torch.cat((txt_ids, img_ids), dim=1)
         pe = self.pe_embedder(ids)
 
-        for i in range(len(self.double_blocks)):
+        for i, block in enumerate(self.double_blocks):
-            img, txt = self.double_blocks[i](img=img, txt=txt, vec=vec, pe=pe)
+            img, txt = block(img=img, txt=txt, vec=vec, pe=pe)
 
-            if control is not None: #Controlnet
+            if control is not None: # Controlnet
-                control_o = control.get("output")
+                control_i = control.get("input")
-                if i < len(control_o):
+                if i < len(control_i):
-                    add = control_o[i]
+                    add = control_i[i]
                     if add is not None:
                         img += add
 
         img = torch.cat((txt, img), 1)
-        for block in self.single_blocks:
+
+        for i, block in enumerate(self.single_blocks):
             img = block(img, vec=vec, pe=pe)
+
+            if control is not None: # Controlnet
+                control_o = control.get("output")
+                if i < len(control_o):
+                    add = control_o[i]
+                    if add is not None:
+                        img[:, txt.shape[1] :, ...] += add
+
         img = img[:, txt.shape[1] :, ...]
 
         img = self.final_layer(img, vec)  # (N, T, patch_size ** 2 * out_channels)

comfy/ldm/hydit/models.py

@@ -372,7 +372,7 @@ class HunYuanDiT(nn.Module):
         for layer, block in enumerate(self.blocks):
             if layer > self.depth // 2:
                 if controls is not None:
-                    skip = skips.pop() + controls.pop()
+                    skip = skips.pop() + controls.pop().to(dtype=x.dtype)
                 else:
                     skip = skips.pop()
                 x = block(x, c, text_states, freqs_cis_img, skip)   # (N, L, D)

comfy/ldm/modules/diffusionmodules/mmdit.py

@@ -351,29 +351,9 @@ class RMSNorm(torch.nn.Module):
         else:
             self.register_parameter("weight", None)
 
-    def _norm(self, x):
-        """
-        Apply the RMSNorm normalization to the input tensor.
-        Args:
-            x (torch.Tensor): The input tensor.
-        Returns:
-            torch.Tensor: The normalized tensor.
-        """
-        return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)
-
     def forward(self, x):
-        """
+        return comfy.ldm.common_dit.rms_norm(x, self.weight, self.eps)
-        Forward pass through the RMSNorm layer.
+
-        Args:
-            x (torch.Tensor): The input tensor.
-        Returns:
-            torch.Tensor: The output tensor after applying RMSNorm.
-        """
-        x = self._norm(x)
-        if self.learnable_scale:
-            return x * self.weight.to(device=x.device, dtype=x.dtype)
-        else:
-            return x
 
 
 class SwiGLUFeedForward(nn.Module):

comfy/ldm/modules/diffusionmodules/openaimodel.py

@@ -841,6 +841,11 @@ class UNetModel(nn.Module):
         t_emb = timestep_embedding(timesteps, self.model_channels, repeat_only=False).to(x.dtype)
         emb = self.time_embed(t_emb)
 
+        if "emb_patch" in transformer_patches:
+            patch = transformer_patches["emb_patch"]
+            for p in patch:
+                emb = p(emb, self.model_channels, transformer_options)
+
         if self.num_classes is not None:
             assert y.shape[0] == x.shape[0]
             emb = emb + self.label_emb(y)

comfy/lora.py

@@ -16,8 +16,12 @@
     along with this program.  If not, see <https://www.gnu.org/licenses/>.
 """
 
+from __future__ import annotations
 import comfy.utils
+import comfy.model_management
+import comfy.model_base
 import logging
+import torch
 
 LORA_CLIP_MAP = {
     "mlp.fc1": "mlp_fc1",
@@ -318,7 +322,256 @@ def model_lora_keys_unet(model, key_map={}):
         for k in diffusers_keys:
             if k.endswith(".weight"):
                 to = diffusers_keys[k]
-                key_lora = "transformer.{}".format(k[:-len(".weight")]) #simpletrainer and probably regular diffusers flux lora format
+                key_map["transformer.{}".format(k[:-len(".weight")])] = to #simpletrainer and probably regular diffusers flux lora format
-                key_map[key_lora] = to
+                key_map["lycoris_{}".format(k[:-len(".weight")].replace(".", "_"))] = to #simpletrainer lycoris
+                key_map["lora_transformer_{}".format(k[:-len(".weight")].replace(".", "_"))] = to #onetrainer
 
     return key_map
+
+
+def weight_decompose(dora_scale, weight, lora_diff, alpha, strength, intermediate_dtype):
+    dora_scale = comfy.model_management.cast_to_device(dora_scale, weight.device, intermediate_dtype)
+    lora_diff *= alpha
+    weight_calc = weight + lora_diff.type(weight.dtype)
+    weight_norm = (
+        weight_calc.transpose(0, 1)
+        .reshape(weight_calc.shape[1], -1)
+        .norm(dim=1, keepdim=True)
+        .reshape(weight_calc.shape[1], *[1] * (weight_calc.dim() - 1))
+        .transpose(0, 1)
+    )
+
+    weight_calc *= (dora_scale / weight_norm).type(weight.dtype)
+    if strength != 1.0:
+        weight_calc -= weight
+        weight += strength * (weight_calc)
+    else:
+        weight[:] = weight_calc
+    return weight
+
+def pad_tensor_to_shape(tensor: torch.Tensor, new_shape: list[int]) -> torch.Tensor:
+    """
+    Pad a tensor to a new shape with zeros.
+
+    Args:
+        tensor (torch.Tensor): The original tensor to be padded.
+        new_shape (List[int]): The desired shape of the padded tensor.
+
+    Returns:
+        torch.Tensor: A new tensor padded with zeros to the specified shape.
+
+    Note:
+        If the new shape is smaller than the original tensor in any dimension,
+        the original tensor will be truncated in that dimension.
+    """
+    if any([new_shape[i] < tensor.shape[i] for i in range(len(new_shape))]):
+        raise ValueError("The new shape must be larger than the original tensor in all dimensions")
+
+    if len(new_shape) != len(tensor.shape):
+        raise ValueError("The new shape must have the same number of dimensions as the original tensor")
+
+    # Create a new tensor filled with zeros
+    padded_tensor = torch.zeros(new_shape, dtype=tensor.dtype, device=tensor.device)
+
+    # Create slicing tuples for both tensors
+    orig_slices = tuple(slice(0, dim) for dim in tensor.shape)
+    new_slices = tuple(slice(0, dim) for dim in tensor.shape)
+
+    # Copy the original tensor into the new tensor
+    padded_tensor[new_slices] = tensor[orig_slices]
+
+    return padded_tensor
+
+def calculate_weight(patches, weight, key, intermediate_dtype=torch.float32):
+    for p in patches:
+        strength = p[0]
+        v = p[1]
+        strength_model = p[2]
+        offset = p[3]
+        function = p[4]
+        if function is None:
+            function = lambda a: a
+
+        old_weight = None
+        if offset is not None:
+            old_weight = weight
+            weight = weight.narrow(offset[0], offset[1], offset[2])
+
+        if strength_model != 1.0:
+            weight *= strength_model
+
+        if isinstance(v, list):
+            v = (calculate_weight(v[1:], v[0].clone(), key, intermediate_dtype=intermediate_dtype), )
+
+        if len(v) == 1:
+            patch_type = "diff"
+        elif len(v) == 2:
+            patch_type = v[0]
+            v = v[1]
+
+        if patch_type == "diff":
+            diff: torch.Tensor = v[0]
+            # An extra flag to pad the weight if the diff's shape is larger than the weight
+            do_pad_weight = len(v) > 1 and v[1]['pad_weight']
+            if do_pad_weight and diff.shape != weight.shape:
+                logging.info("Pad weight {} from {} to shape: {}".format(key, weight.shape, diff.shape))
+                weight = pad_tensor_to_shape(weight, diff.shape)
+
+            if strength != 0.0:
+                if diff.shape != weight.shape:
+                    logging.warning("WARNING SHAPE MISMATCH {} WEIGHT NOT MERGED {} != {}".format(key, diff.shape, weight.shape))
+                else:
+                    weight += function(strength * comfy.model_management.cast_to_device(diff, weight.device, weight.dtype))
+        elif patch_type == "lora": #lora/locon
+            mat1 = comfy.model_management.cast_to_device(v[0], weight.device, intermediate_dtype)
+            mat2 = comfy.model_management.cast_to_device(v[1], weight.device, intermediate_dtype)
+            dora_scale = v[4]
+            if v[2] is not None:
+                alpha = v[2] / mat2.shape[0]
+            else:
+                alpha = 1.0
+
+            if v[3] is not None:
+                #locon mid weights, hopefully the math is fine because I didn't properly test it
+                mat3 = comfy.model_management.cast_to_device(v[3], weight.device, intermediate_dtype)
+                final_shape = [mat2.shape[1], mat2.shape[0], mat3.shape[2], mat3.shape[3]]
+                mat2 = torch.mm(mat2.transpose(0, 1).flatten(start_dim=1), mat3.transpose(0, 1).flatten(start_dim=1)).reshape(final_shape).transpose(0, 1)
+            try:
+                lora_diff = torch.mm(mat1.flatten(start_dim=1), mat2.flatten(start_dim=1)).reshape(weight.shape)
+                if dora_scale is not None:
+                    weight = function(weight_decompose(dora_scale, weight, lora_diff, alpha, strength, intermediate_dtype))
+                else:
+                    weight += function(((strength * alpha) * lora_diff).type(weight.dtype))
+            except Exception as e:
+                logging.error("ERROR {} {} {}".format(patch_type, key, e))
+        elif patch_type == "lokr":
+            w1 = v[0]
+            w2 = v[1]
+            w1_a = v[3]
+            w1_b = v[4]
+            w2_a = v[5]
+            w2_b = v[6]
+            t2 = v[7]
+            dora_scale = v[8]
+            dim = None
+
+            if w1 is None:
+                dim = w1_b.shape[0]
+                w1 = torch.mm(comfy.model_management.cast_to_device(w1_a, weight.device, intermediate_dtype),
+                                comfy.model_management.cast_to_device(w1_b, weight.device, intermediate_dtype))
+            else:
+                w1 = comfy.model_management.cast_to_device(w1, weight.device, intermediate_dtype)
+
+            if w2 is None:
+                dim = w2_b.shape[0]
+                if t2 is None:
+                    w2 = torch.mm(comfy.model_management.cast_to_device(w2_a, weight.device, intermediate_dtype),
+                                    comfy.model_management.cast_to_device(w2_b, weight.device, intermediate_dtype))
+                else:
+                    w2 = torch.einsum('i j k l, j r, i p -> p r k l',
+                                        comfy.model_management.cast_to_device(t2, weight.device, intermediate_dtype),
+                                        comfy.model_management.cast_to_device(w2_b, weight.device, intermediate_dtype),
+                                        comfy.model_management.cast_to_device(w2_a, weight.device, intermediate_dtype))
+            else:
+                w2 = comfy.model_management.cast_to_device(w2, weight.device, intermediate_dtype)
+
+            if len(w2.shape) == 4:
+                w1 = w1.unsqueeze(2).unsqueeze(2)
+            if v[2] is not None and dim is not None:
+                alpha = v[2] / dim
+            else:
+                alpha = 1.0
+
+            try:
+                lora_diff = torch.kron(w1, w2).reshape(weight.shape)
+                if dora_scale is not None:
+                    weight = function(weight_decompose(dora_scale, weight, lora_diff, alpha, strength, intermediate_dtype))
+                else:
+                    weight += function(((strength * alpha) * lora_diff).type(weight.dtype))
+            except Exception as e:
+                logging.error("ERROR {} {} {}".format(patch_type, key, e))
+        elif patch_type == "loha":
+            w1a = v[0]
+            w1b = v[1]
+            if v[2] is not None:
+                alpha = v[2] / w1b.shape[0]
+            else:
+                alpha = 1.0
+
+            w2a = v[3]
+            w2b = v[4]
+            dora_scale = v[7]
+            if v[5] is not None: #cp decomposition
+                t1 = v[5]
+                t2 = v[6]
+                m1 = torch.einsum('i j k l, j r, i p -> p r k l',
+                                    comfy.model_management.cast_to_device(t1, weight.device, intermediate_dtype),
+                                    comfy.model_management.cast_to_device(w1b, weight.device, intermediate_dtype),
+                                    comfy.model_management.cast_to_device(w1a, weight.device, intermediate_dtype))
+
+                m2 = torch.einsum('i j k l, j r, i p -> p r k l',
+                                    comfy.model_management.cast_to_device(t2, weight.device, intermediate_dtype),
+                                    comfy.model_management.cast_to_device(w2b, weight.device, intermediate_dtype),
+                                    comfy.model_management.cast_to_device(w2a, weight.device, intermediate_dtype))
+            else:
+                m1 = torch.mm(comfy.model_management.cast_to_device(w1a, weight.device, intermediate_dtype),
+                                comfy.model_management.cast_to_device(w1b, weight.device, intermediate_dtype))
+                m2 = torch.mm(comfy.model_management.cast_to_device(w2a, weight.device, intermediate_dtype),
+                                comfy.model_management.cast_to_device(w2b, weight.device, intermediate_dtype))
+
+            try:
+                lora_diff = (m1 * m2).reshape(weight.shape)
+                if dora_scale is not None:
+                    weight = function(weight_decompose(dora_scale, weight, lora_diff, alpha, strength, intermediate_dtype))
+                else:
+                    weight += function(((strength * alpha) * lora_diff).type(weight.dtype))
+            except Exception as e:
+                logging.error("ERROR {} {} {}".format(patch_type, key, e))
+        elif patch_type == "glora":
+            dora_scale = v[5]
+
+            old_glora = False
+            if v[3].shape[1] == v[2].shape[0] == v[0].shape[0] == v[1].shape[1]:
+                rank = v[0].shape[0]
+                old_glora = True
+
+            if v[3].shape[0] == v[2].shape[1] == v[0].shape[1] == v[1].shape[0]:
+                if old_glora and v[1].shape[0] == weight.shape[0] and weight.shape[0] == weight.shape[1]:
+                    pass
+                else:
+                    old_glora = False
+                    rank = v[1].shape[0]
+
+            a1 = comfy.model_management.cast_to_device(v[0].flatten(start_dim=1), weight.device, intermediate_dtype)
+            a2 = comfy.model_management.cast_to_device(v[1].flatten(start_dim=1), weight.device, intermediate_dtype)
+            b1 = comfy.model_management.cast_to_device(v[2].flatten(start_dim=1), weight.device, intermediate_dtype)
+            b2 = comfy.model_management.cast_to_device(v[3].flatten(start_dim=1), weight.device, intermediate_dtype)
+
+            if v[4] is not None:
+                alpha = v[4] / rank
+            else:
+                alpha = 1.0
+
+            try:
+                if old_glora:
+                    lora_diff = (torch.mm(b2, b1) + torch.mm(torch.mm(weight.flatten(start_dim=1).to(dtype=intermediate_dtype), a2), a1)).reshape(weight.shape) #old lycoris glora
+                else:
+                    if weight.dim() > 2:
+                        lora_diff = torch.einsum("o i ..., i j -> o j ...", torch.einsum("o i ..., i j -> o j ...", weight.to(dtype=intermediate_dtype), a1), a2).reshape(weight.shape)
+                    else:
+                        lora_diff = torch.mm(torch.mm(weight.to(dtype=intermediate_dtype), a1), a2).reshape(weight.shape)
+                    lora_diff += torch.mm(b1, b2).reshape(weight.shape)
+
+                if dora_scale is not None:
+                    weight = function(weight_decompose(dora_scale, weight, lora_diff, alpha, strength, intermediate_dtype))
+                else:
+                    weight += function(((strength * alpha) * lora_diff).type(weight.dtype))
+            except Exception as e:
+                logging.error("ERROR {} {} {}".format(patch_type, key, e))
+        else:
+            logging.warning("patch type not recognized {} {}".format(patch_type, key))
+
+        if old_weight is not None:
+            weight = old_weight
+
+    return weight

comfy/model_base.py

@@ -96,10 +96,7 @@ class BaseModel(torch.nn.Module):
 
         if not unet_config.get("disable_unet_model_creation", False):
             if model_config.custom_operations is None:
-                if self.manual_cast_dtype is not None:
+                operations = comfy.ops.pick_operations(unet_config.get("dtype", None), self.manual_cast_dtype)
-                    operations = comfy.ops.manual_cast
-                else:
-                    operations = comfy.ops.disable_weight_init
             else:
                 operations = model_config.custom_operations
             self.diffusion_model = unet_model(**unet_config, device=device, operations=operations)

comfy/model_detection.py

@@ -473,8 +473,14 @@ def unet_config_from_diffusers_unet(state_dict, dtype=None):
             'context_dim': 768, 'num_head_channels': 64, 'transformer_depth_output': [0, 0, 1, 1, 1, 1],
             'use_temporal_attention': False, 'use_temporal_resblock': False}
     
+    SD15_diffusers_inpaint = {'use_checkpoint': False, 'image_size': 32, 'out_channels': 4, 'use_spatial_transformer': True, 'legacy': False, 'adm_in_channels': None,
+            'dtype': dtype, 'in_channels': 9, 'model_channels': 320, 'num_res_blocks': [2, 2, 2, 2], 'transformer_depth': [1, 1, 1, 1, 1, 1, 0, 0],
+            'channel_mult': [1, 2, 4, 4], 'transformer_depth_middle': 1, 'use_linear_in_transformer': False, 'context_dim': 768, 'num_heads': 8,
+            'transformer_depth_output': [1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0],
+            'use_temporal_attention': False, 'use_temporal_resblock': False}  
 
-    supported_models = [SDXL, SDXL_refiner, SD21, SD15, SD21_uncliph, SD21_unclipl, SDXL_mid_cnet, SDXL_small_cnet, SDXL_diffusers_inpaint, SSD_1B, Segmind_Vega, KOALA_700M, KOALA_1B, SD09_XS, SD_XS, SDXL_diffusers_ip2p]
+
+    supported_models = [SDXL, SDXL_refiner, SD21, SD15, SD21_uncliph, SD21_unclipl, SDXL_mid_cnet, SDXL_small_cnet, SDXL_diffusers_inpaint, SSD_1B, Segmind_Vega, KOALA_700M, KOALA_1B, SD09_XS, SD_XS, SDXL_diffusers_ip2p, SD15_diffusers_inpaint]
 
     for unet_config in supported_models:
         matches = True

comfy/model_management.py

@@ -44,9 +44,15 @@ cpu_state = CPUState.GPU
 
 total_vram = 0
 
-lowvram_available = True
 xpu_available = False
+torch_version = ""
+try:
+    torch_version = torch.version.__version__
+    xpu_available = (int(torch_version[0]) < 2 or (int(torch_version[0]) == 2 and int(torch_version[2]) <= 4)) and torch.xpu.is_available()
+except:
+    pass
 
+lowvram_available = True
 if args.deterministic:
     logging.info("Using deterministic algorithms for pytorch")
     torch.use_deterministic_algorithms(True, warn_only=True)
@@ -66,10 +72,10 @@ if args.directml is not None:
 
 try:
     import intel_extension_for_pytorch as ipex
-    if torch.xpu.is_available():
+    _ = torch.xpu.device_count()
-        xpu_available = True
+    xpu_available = torch.xpu.is_available()
 except:
-    pass
+    xpu_available = xpu_available or (hasattr(torch, "xpu") and torch.xpu.is_available())
 
 try:
     if torch.backends.mps.is_available():
@@ -189,7 +195,6 @@ VAE_DTYPES = [torch.float32]
 
 try:
     if is_nvidia():
-        torch_version = torch.version.__version__
         if int(torch_version[0]) >= 2:
             if ENABLE_PYTORCH_ATTENTION == False and args.use_split_cross_attention == False and args.use_quad_cross_attention == False:
                 ENABLE_PYTORCH_ATTENTION = True
@@ -315,17 +320,15 @@ class LoadedModel:
             self.model_use_more_vram(use_more_vram)
         else:
             try:
-                if lowvram_model_memory > 0 and load_weights:
+                self.real_model = self.model.patch_model(device_to=patch_model_to, lowvram_model_memory=lowvram_model_memory, load_weights=load_weights, force_patch_weights=force_patch_weights)
-                    self.real_model = self.model.patch_model_lowvram(device_to=patch_model_to, lowvram_model_memory=lowvram_model_memory, force_patch_weights=force_patch_weights)
-                else:
-                    self.real_model = self.model.patch_model(device_to=patch_model_to, patch_weights=load_weights)
             except Exception as e:
                 self.model.unpatch_model(self.model.offload_device)
                 self.model_unload()
                 raise e
 
-        if is_intel_xpu() and not args.disable_ipex_optimize:
+        if is_intel_xpu() and not args.disable_ipex_optimize and self.real_model is not None:
-            self.real_model = ipex.optimize(self.real_model.eval(), graph_mode=True, concat_linear=True)
+            with torch.no_grad():
+                self.real_model = ipex.optimize(self.real_model.eval(), inplace=True, graph_mode=True, concat_linear=True)
 
         self.weights_loaded = True
         return self.real_model
@@ -367,8 +370,21 @@ def offloaded_memory(loaded_models, device):
             offloaded_mem += m.model_offloaded_memory()
     return offloaded_mem
 
+WINDOWS = any(platform.win32_ver())
+
+EXTRA_RESERVED_VRAM = 400 * 1024 * 1024
+if WINDOWS:
+    EXTRA_RESERVED_VRAM = 600 * 1024 * 1024 #Windows is higher because of the shared vram issue
+
+if args.reserve_vram is not None:
+    EXTRA_RESERVED_VRAM = args.reserve_vram * 1024 * 1024 * 1024
+    logging.debug("Reserving {}MB vram for other applications.".format(EXTRA_RESERVED_VRAM / (1024 * 1024)))
+
+def extra_reserved_memory():
+    return EXTRA_RESERVED_VRAM
+
 def minimum_inference_memory():
-    return (1024 * 1024 * 1024) * 1.2
+    return (1024 * 1024 * 1024) * 0.8 + extra_reserved_memory()
 
 def unload_model_clones(model, unload_weights_only=True, force_unload=True):
     to_unload = []
@@ -392,6 +408,8 @@ def unload_model_clones(model, unload_weights_only=True, force_unload=True):
     if not force_unload:
         if unload_weights_only and unload_weight == False:
             return None
+    else:
+        unload_weight = True
 
     for i in to_unload:
         logging.debug("unload clone {} {}".format(i, unload_weight))
@@ -408,7 +426,7 @@ def free_memory(memory_required, device, keep_loaded=[]):
         shift_model = current_loaded_models[i]
         if shift_model.device == device:
             if shift_model not in keep_loaded:
-                can_unload.append((sys.getrefcount(shift_model.model), shift_model.model_memory(), i))
+                can_unload.append((-shift_model.model_offloaded_memory(), sys.getrefcount(shift_model.model), shift_model.model_memory(), i))
                 shift_model.currently_used = False
 
     for x in sorted(can_unload):
@@ -439,11 +457,11 @@ def load_models_gpu(models, memory_required=0, force_patch_weights=False, minimu
     global vram_state
 
     inference_memory = minimum_inference_memory()
-    extra_mem = max(inference_memory, memory_required + 300 * 1024 * 1024)
+    extra_mem = max(inference_memory, memory_required + extra_reserved_memory())
     if minimum_memory_required is None:
         minimum_memory_required = extra_mem
     else:
-        minimum_memory_required = max(inference_memory, minimum_memory_required + 300 * 1024 * 1024)
+        minimum_memory_required = max(inference_memory, minimum_memory_required + extra_reserved_memory())
 
     models = set(models)
 
@@ -553,7 +571,9 @@ def loaded_models(only_currently_used=False):
 def cleanup_models(keep_clone_weights_loaded=False):
     to_delete = []
     for i in range(len(current_loaded_models)):
-        if sys.getrefcount(current_loaded_models[i].model) <= 2:
+        #TODO: very fragile function needs improvement
+        num_refs = sys.getrefcount(current_loaded_models[i].model)
+        if num_refs <= 2:
             if not keep_clone_weights_loaded:
                 to_delete = [i] + to_delete
             #TODO: find a less fragile way to do this.
@@ -660,6 +680,7 @@ def unet_manual_cast(weight_dtype, inference_device, supported_dtypes=[torch.flo
     if bf16_supported and weight_dtype == torch.bfloat16:
         return None
 
+    fp16_supported = should_use_fp16(inference_device, prioritize_performance=True)
     for dt in supported_dtypes:
         if dt == torch.float16 and fp16_supported:
             return torch.float16
@@ -875,7 +896,8 @@ def pytorch_attention_flash_attention():
 def force_upcast_attention_dtype():
     upcast = args.force_upcast_attention
     try:
-        if platform.mac_ver()[0] in ['14.5']: #black image bug on OSX Sonoma 14.5
+        macos_version = tuple(int(n) for n in platform.mac_ver()[0].split("."))
+        if (14, 5) <= macos_version < (14, 7):  # black image bug on recent versions of MacOS
             upcast = True
     except:
         pass
@@ -971,23 +993,23 @@ def should_use_fp16(device=None, model_params=0, prioritize_performance=True, ma
     if torch.version.hip:
         return True
 
-    props = torch.cuda.get_device_properties("cuda")
+    props = torch.cuda.get_device_properties(device)
     if props.major >= 8:
         return True
 
     if props.major < 6:
         return False
 
-    fp16_works = False
+    #FP16 is confirmed working on a 1080 (GP104) and on latest pytorch actually seems faster than fp32
-    #FP16 is confirmed working on a 1080 (GP104) but it's a bit slower than FP32 so it should only be enabled
-    #when the model doesn't actually fit on the card
-    #TODO: actually test if GP106 and others have the same type of behavior
     nvidia_10_series = ["1080", "1070", "titan x", "p3000", "p3200", "p4000", "p4200", "p5000", "p5200", "p6000", "1060", "1050", "p40", "p100", "p6", "p4"]
     for x in nvidia_10_series:
         if x in props.name.lower():
-            fp16_works = True
+            if WINDOWS or manual_cast:
+                return True
+            else:
+                return False #weird linux behavior where fp32 is faster
 
-    if fp16_works or manual_cast:
+    if manual_cast:
         free_model_memory = maximum_vram_for_weights(device)
         if (not prioritize_performance) or model_params * 4 > free_model_memory:
             return True
@@ -1027,7 +1049,7 @@ def should_use_bf16(device=None, model_params=0, prioritize_performance=True, ma
     if is_intel_xpu():
         return True
 
-    props = torch.cuda.get_device_properties("cuda")
+    props = torch.cuda.get_device_properties(device)
     if props.major >= 8:
         return True
 
@@ -1040,6 +1062,16 @@ def should_use_bf16(device=None, model_params=0, prioritize_performance=True, ma
 
     return False
 
+def supports_fp8_compute(device=None):
+    props = torch.cuda.get_device_properties(device)
+    if props.major >= 9:
+        return True
+    if props.major < 8:
+        return False
+    if props.minor < 9:
+        return False
+    return True
+
 def soft_empty_cache(force=False):
     global cpu_state
     if cpu_state == CPUState.MPS:

comfy/model_patcher.py

@@ -22,32 +22,26 @@ import inspect
 import logging
 import uuid
 import collections
+import math
 
 import comfy.utils
+import comfy.float
 import comfy.model_management
+import comfy.lora
 from comfy.types import UnetWrapperFunction
 
-
+def string_to_seed(data):
-def weight_decompose(dora_scale, weight, lora_diff, alpha, strength):
+    crc = 0xFFFFFFFF
-    dora_scale = comfy.model_management.cast_to_device(dora_scale, weight.device, torch.float32)
+    for byte in data:
-    lora_diff *= alpha
+        if isinstance(byte, str):
-    weight_calc = weight + lora_diff.type(weight.dtype)
+            byte = ord(byte)
-    weight_norm = (
+        crc ^= byte
-        weight_calc.transpose(0, 1)
+        for _ in range(8):
-        .reshape(weight_calc.shape[1], -1)
+            if crc & 1:
-        .norm(dim=1, keepdim=True)
+                crc = (crc >> 1) ^ 0xEDB88320
-        .reshape(weight_calc.shape[1], *[1] * (weight_calc.dim() - 1))
+            else:
-        .transpose(0, 1)
+                crc >>= 1
-    )
+    return crc ^ 0xFFFFFFFF
-
-    weight_calc *= (dora_scale / weight_norm).type(weight.dtype)
-    if strength != 1.0:
-        weight_calc -= weight
-        weight += strength * (weight_calc)
-    else:
-        weight[:] = weight_calc
-    return weight
-
 
 def set_model_options_patch_replace(model_options, patch, name, block_name, number, transformer_index=None):
     to = model_options["transformer_options"].copy()
@@ -90,12 +84,11 @@ def wipe_lowvram_weight(m):
     m.bias_function = None
 
 class LowVramPatch:
-    def __init__(self, key, model_patcher):
+    def __init__(self, key, patches):
         self.key = key
-        self.model_patcher = model_patcher
+        self.patches = patches
     def __call__(self, weight):
-        return self.model_patcher.calculate_weight(self.model_patcher.patches[self.key], weight, self.key)
+        return comfy.lora.calculate_weight(self.patches[self.key], weight, self.key, intermediate_dtype=weight.dtype)
-
 
 class ModelPatcher:
     def __init__(self, model, load_device, offload_device, size=0, weight_inplace_update=False):
@@ -327,47 +320,36 @@ class ModelPatcher:
             temp_weight = comfy.model_management.cast_to_device(weight, device_to, torch.float32, copy=True)
         else:
             temp_weight = weight.to(torch.float32, copy=True)
-        out_weight = self.calculate_weight(self.patches[key], temp_weight, key).to(weight.dtype)
+        out_weight = comfy.lora.calculate_weight(self.patches[key], temp_weight, key)
+        out_weight = comfy.float.stochastic_rounding(out_weight, weight.dtype, seed=string_to_seed(key))
         if inplace_update:
             comfy.utils.copy_to_param(self.model, key, out_weight)
         else:
             comfy.utils.set_attr_param(self.model, key, out_weight)
 
-    def patch_model(self, device_to=None, patch_weights=True):
+    def load(self, device_to=None, lowvram_model_memory=0, force_patch_weights=False, full_load=False):
-        for k in self.object_patches:
-            old = comfy.utils.set_attr(self.model, k, self.object_patches[k])
-            if k not in self.object_patches_backup:
-                self.object_patches_backup[k] = old
-
-        if patch_weights:
-            model_sd = self.model_state_dict()
-            for key in self.patches:
-                if key not in model_sd:
-                    logging.warning("could not patch. key doesn't exist in model: {}".format(key))
-                    continue
-
-                self.patch_weight_to_device(key, device_to)
-
-            if device_to is not None:
-                self.model.to(device_to)
-                self.model.device = device_to
-                self.model.model_loaded_weight_memory = self.model_size()
-
-        return self.model
-
-    def lowvram_load(self, device_to=None, lowvram_model_memory=0, force_patch_weights=False, full_load=False):
         mem_counter = 0
         patch_counter = 0
         lowvram_counter = 0
+        loading = []
         for n, m in self.model.named_modules():
+            if hasattr(m, "comfy_cast_weights") or hasattr(m, "weight"):
+                loading.append((comfy.model_management.module_size(m), n, m))
+
+        load_completely = []
+        loading.sort(reverse=True)
+        for x in loading:
+            n = x[1]
+            m = x[2]
+            module_mem = x[0]
+
             lowvram_weight = False
 
             if not full_load and hasattr(m, "comfy_cast_weights"):
-                module_mem = comfy.model_management.module_size(m)
                 if mem_counter + module_mem >= lowvram_model_memory:
                     lowvram_weight = True
                     lowvram_counter += 1
-                    if m.comfy_cast_weights:
+                    if hasattr(m, "prev_comfy_cast_weights"): #Already lowvramed
                         continue
 
             weight_key = "{}.weight".format(n)
@@ -378,13 +360,13 @@ class ModelPatcher:
                     if force_patch_weights:
                         self.patch_weight_to_device(weight_key)
                     else:
-                        m.weight_function = LowVramPatch(weight_key, self)
+                        m.weight_function = LowVramPatch(weight_key, self.patches)
                         patch_counter += 1
                 if bias_key in self.patches:
                     if force_patch_weights:
                         self.patch_weight_to_device(bias_key)
                     else:
-                        m.bias_function = LowVramPatch(bias_key, self)
+                        m.bias_function = LowVramPatch(bias_key, self.patches)
                         patch_counter += 1
 
                 m.prev_comfy_cast_weights = m.comfy_cast_weights
@@ -395,205 +377,56 @@ class ModelPatcher:
                         wipe_lowvram_weight(m)
 
                 if hasattr(m, "weight"):
-                    mem_counter += comfy.model_management.module_size(m)
+                    mem_counter += module_mem
-                    param = list(m.parameters())
+                    load_completely.append((module_mem, n, m))
-                    if len(param) > 0:
-                        weight = param[0]
-                        if weight.device == device_to:
-                            continue
 
-                    weight_to = None
+        load_completely.sort(reverse=True)
-                    if full_load:#TODO
+        for x in load_completely:
-                        weight_to = device_to
+            n = x[1]
-                    self.patch_weight_to_device(weight_key, device_to=weight_to) #TODO: speed this up without OOM
+            m = x[2]
-                    self.patch_weight_to_device(bias_key, device_to=weight_to)
+            weight_key = "{}.weight".format(n)
-                    m.to(device_to)
+            bias_key = "{}.bias".format(n)
-                    logging.debug("lowvram: loaded module regularly {} {}".format(n, m))
+            if hasattr(m, "comfy_patched_weights"):
+                if m.comfy_patched_weights == True:
+                    continue
+
+            self.patch_weight_to_device(weight_key, device_to=device_to)
+            self.patch_weight_to_device(bias_key, device_to=device_to)
+            logging.debug("lowvram: loaded module regularly {} {}".format(n, m))
+            m.comfy_patched_weights = True
+
+        for x in load_completely:
+            x[2].to(device_to)
 
         if lowvram_counter > 0:
             logging.info("loaded partially {} {} {}".format(lowvram_model_memory / (1024 * 1024), mem_counter / (1024 * 1024), patch_counter))
             self.model.model_lowvram = True
         else:
-            logging.info("loaded completely {} {}".format(lowvram_model_memory / (1024 * 1024), mem_counter / (1024 * 1024)))
+            logging.info("loaded completely {} {} {}".format(lowvram_model_memory / (1024 * 1024), mem_counter / (1024 * 1024), full_load))
             self.model.model_lowvram = False
+            if full_load:
+                self.model.to(device_to)
+                mem_counter = self.model_size()
+
         self.model.lowvram_patch_counter += patch_counter
         self.model.device = device_to
         self.model.model_loaded_weight_memory = mem_counter
 
+    def patch_model(self, device_to=None, lowvram_model_memory=0, load_weights=True, force_patch_weights=False):
+        for k in self.object_patches:
+            old = comfy.utils.set_attr(self.model, k, self.object_patches[k])
+            if k not in self.object_patches_backup:
+                self.object_patches_backup[k] = old
 
-    def patch_model_lowvram(self, device_to=None, lowvram_model_memory=0, force_patch_weights=False):
+        if lowvram_model_memory == 0:
-        self.patch_model(device_to, patch_weights=False)
+            full_load = True
-        self.lowvram_load(device_to, lowvram_model_memory=lowvram_model_memory, force_patch_weights=force_patch_weights)
+        else:
+            full_load = False
+
+        if load_weights:
+            self.load(device_to, lowvram_model_memory=lowvram_model_memory, force_patch_weights=force_patch_weights, full_load=full_load)
         return self.model
 
-    def calculate_weight(self, patches, weight, key):
-        for p in patches:
-            strength = p[0]
-            v = p[1]
-            strength_model = p[2]
-            offset = p[3]
-            function = p[4]
-            if function is None:
-                function = lambda a: a
-
-            old_weight = None
-            if offset is not None:
-                old_weight = weight
-                weight = weight.narrow(offset[0], offset[1], offset[2])
-
-            if strength_model != 1.0:
-                weight *= strength_model
-
-            if isinstance(v, list):
-                v = (self.calculate_weight(v[1:], v[0].clone(), key), )
-
-            if len(v) == 1:
-                patch_type = "diff"
-            elif len(v) == 2:
-                patch_type = v[0]
-                v = v[1]
-
-            if patch_type == "diff":
-                w1 = v[0]
-                if strength != 0.0:
-                    if w1.shape != weight.shape:
-                        logging.warning("WARNING SHAPE MISMATCH {} WEIGHT NOT MERGED {} != {}".format(key, w1.shape, weight.shape))
-                    else:
-                        weight += function(strength * comfy.model_management.cast_to_device(w1, weight.device, weight.dtype))
-            elif patch_type == "lora": #lora/locon
-                mat1 = comfy.model_management.cast_to_device(v[0], weight.device, torch.float32)
-                mat2 = comfy.model_management.cast_to_device(v[1], weight.device, torch.float32)
-                dora_scale = v[4]
-                if v[2] is not None:
-                    alpha = v[2] / mat2.shape[0]
-                else:
-                    alpha = 1.0
-
-                if v[3] is not None:
-                    #locon mid weights, hopefully the math is fine because I didn't properly test it
-                    mat3 = comfy.model_management.cast_to_device(v[3], weight.device, torch.float32)
-                    final_shape = [mat2.shape[1], mat2.shape[0], mat3.shape[2], mat3.shape[3]]
-                    mat2 = torch.mm(mat2.transpose(0, 1).flatten(start_dim=1), mat3.transpose(0, 1).flatten(start_dim=1)).reshape(final_shape).transpose(0, 1)
-                try:
-                    lora_diff = torch.mm(mat1.flatten(start_dim=1), mat2.flatten(start_dim=1)).reshape(weight.shape)
-                    if dora_scale is not None:
-                        weight = function(weight_decompose(dora_scale, weight, lora_diff, alpha, strength))
-                    else:
-                        weight += function(((strength * alpha) * lora_diff).type(weight.dtype))
-                except Exception as e:
-                    logging.error("ERROR {} {} {}".format(patch_type, key, e))
-            elif patch_type == "lokr":
-                w1 = v[0]
-                w2 = v[1]
-                w1_a = v[3]
-                w1_b = v[4]
-                w2_a = v[5]
-                w2_b = v[6]
-                t2 = v[7]
-                dora_scale = v[8]
-                dim = None
-
-                if w1 is None:
-                    dim = w1_b.shape[0]
-                    w1 = torch.mm(comfy.model_management.cast_to_device(w1_a, weight.device, torch.float32),
-                                  comfy.model_management.cast_to_device(w1_b, weight.device, torch.float32))
-                else:
-                    w1 = comfy.model_management.cast_to_device(w1, weight.device, torch.float32)
-
-                if w2 is None:
-                    dim = w2_b.shape[0]
-                    if t2 is None:
-                        w2 = torch.mm(comfy.model_management.cast_to_device(w2_a, weight.device, torch.float32),
-                                      comfy.model_management.cast_to_device(w2_b, weight.device, torch.float32))
-                    else:
-                        w2 = torch.einsum('i j k l, j r, i p -> p r k l',
-                                          comfy.model_management.cast_to_device(t2, weight.device, torch.float32),
-                                          comfy.model_management.cast_to_device(w2_b, weight.device, torch.float32),
-                                          comfy.model_management.cast_to_device(w2_a, weight.device, torch.float32))
-                else:
-                    w2 = comfy.model_management.cast_to_device(w2, weight.device, torch.float32)
-
-                if len(w2.shape) == 4:
-                    w1 = w1.unsqueeze(2).unsqueeze(2)
-                if v[2] is not None and dim is not None:
-                    alpha = v[2] / dim
-                else:
-                    alpha = 1.0
-
-                try:
-                    lora_diff = torch.kron(w1, w2).reshape(weight.shape)
-                    if dora_scale is not None:
-                        weight = function(weight_decompose(dora_scale, weight, lora_diff, alpha, strength))
-                    else:
-                        weight += function(((strength * alpha) * lora_diff).type(weight.dtype))
-                except Exception as e:
-                    logging.error("ERROR {} {} {}".format(patch_type, key, e))
-            elif patch_type == "loha":
-                w1a = v[0]
-                w1b = v[1]
-                if v[2] is not None:
-                    alpha = v[2] / w1b.shape[0]
-                else:
-                    alpha = 1.0
-
-                w2a = v[3]
-                w2b = v[4]
-                dora_scale = v[7]
-                if v[5] is not None: #cp decomposition
-                    t1 = v[5]
-                    t2 = v[6]
-                    m1 = torch.einsum('i j k l, j r, i p -> p r k l',
-                                      comfy.model_management.cast_to_device(t1, weight.device, torch.float32),
-                                      comfy.model_management.cast_to_device(w1b, weight.device, torch.float32),
-                                      comfy.model_management.cast_to_device(w1a, weight.device, torch.float32))
-
-                    m2 = torch.einsum('i j k l, j r, i p -> p r k l',
-                                      comfy.model_management.cast_to_device(t2, weight.device, torch.float32),
-                                      comfy.model_management.cast_to_device(w2b, weight.device, torch.float32),
-                                      comfy.model_management.cast_to_device(w2a, weight.device, torch.float32))
-                else:
-                    m1 = torch.mm(comfy.model_management.cast_to_device(w1a, weight.device, torch.float32),
-                                  comfy.model_management.cast_to_device(w1b, weight.device, torch.float32))
-                    m2 = torch.mm(comfy.model_management.cast_to_device(w2a, weight.device, torch.float32),
-                                  comfy.model_management.cast_to_device(w2b, weight.device, torch.float32))
-
-                try:
-                    lora_diff = (m1 * m2).reshape(weight.shape)
-                    if dora_scale is not None:
-                        weight = function(weight_decompose(dora_scale, weight, lora_diff, alpha, strength))
-                    else:
-                        weight += function(((strength * alpha) * lora_diff).type(weight.dtype))
-                except Exception as e:
-                    logging.error("ERROR {} {} {}".format(patch_type, key, e))
-            elif patch_type == "glora":
-                if v[4] is not None:
-                    alpha = v[4] / v[0].shape[0]
-                else:
-                    alpha = 1.0
-
-                dora_scale = v[5]
-
-                a1 = comfy.model_management.cast_to_device(v[0].flatten(start_dim=1), weight.device, torch.float32)
-                a2 = comfy.model_management.cast_to_device(v[1].flatten(start_dim=1), weight.device, torch.float32)
-                b1 = comfy.model_management.cast_to_device(v[2].flatten(start_dim=1), weight.device, torch.float32)
-                b2 = comfy.model_management.cast_to_device(v[3].flatten(start_dim=1), weight.device, torch.float32)
-
-                try:
-                    lora_diff = (torch.mm(b2, b1) + torch.mm(torch.mm(weight.flatten(start_dim=1), a2), a1)).reshape(weight.shape)
-                    if dora_scale is not None:
-                        weight = function(weight_decompose(dora_scale, weight, lora_diff, alpha, strength))
-                    else:
-                        weight += function(((strength * alpha) * lora_diff).type(weight.dtype))
-                except Exception as e:
-                    logging.error("ERROR {} {} {}".format(patch_type, key, e))
-            else:
-                logging.warning("patch type not recognized {} {}".format(patch_type, key))
-
-            if old_weight is not None:
-                weight = old_weight
-
-        return weight
-
     def unpatch_model(self, device_to=None, unpatch_weights=True):
         if unpatch_weights:
             if self.model.model_lowvram:
@@ -619,6 +452,10 @@ class ModelPatcher:
                 self.model.device = device_to
             self.model.model_loaded_weight_memory = 0
 
+            for m in self.model.modules():
+                if hasattr(m, "comfy_patched_weights"):
+                    del m.comfy_patched_weights
+
         keys = list(self.object_patches_backup.keys())
         for k in keys:
             comfy.utils.set_attr(self.model, k, self.object_patches_backup[k])
@@ -628,40 +465,47 @@ class ModelPatcher:
     def partially_unload(self, device_to, memory_to_free=0):
         memory_freed = 0
         patch_counter = 0
+        unload_list = []
 
-        for n, m in list(self.model.named_modules())[::-1]:
+        for n, m in self.model.named_modules():
-            if memory_to_free < memory_freed:
-                break
-
             shift_lowvram = False
             if hasattr(m, "comfy_cast_weights"):
                 module_mem = comfy.model_management.module_size(m)
-                weight_key = "{}.weight".format(n)
+                unload_list.append((module_mem, n, m))
-                bias_key = "{}.bias".format(n)
 
+        unload_list.sort()
+        for unload in unload_list:
+            if memory_to_free < memory_freed:
+                break
+            module_mem = unload[0]
+            n = unload[1]
+            m = unload[2]
+            weight_key = "{}.weight".format(n)
+            bias_key = "{}.bias".format(n)
 
-                if m.weight is not None and m.weight.device != device_to:
+            if hasattr(m, "comfy_patched_weights") and m.comfy_patched_weights == True:
-                    for key in [weight_key, bias_key]:
+                for key in [weight_key, bias_key]:
-                        bk = self.backup.get(key, None)
+                    bk = self.backup.get(key, None)
-                        if bk is not None:
+                    if bk is not None:
-                            if bk.inplace_update:
+                        if bk.inplace_update:
-                                comfy.utils.copy_to_param(self.model, key, bk.weight)
+                            comfy.utils.copy_to_param(self.model, key, bk.weight)
-                            else:
+                        else:
-                                comfy.utils.set_attr_param(self.model, key, bk.weight)
+                            comfy.utils.set_attr_param(self.model, key, bk.weight)
-                            self.backup.pop(key)
+                        self.backup.pop(key)
 
-                    m.to(device_to)
+                m.to(device_to)
-                    if weight_key in self.patches:
+                if weight_key in self.patches:
-                        m.weight_function = LowVramPatch(weight_key, self)
+                    m.weight_function = LowVramPatch(weight_key, self.patches)
-                        patch_counter += 1
+                    patch_counter += 1
-                    if bias_key in self.patches:
+                if bias_key in self.patches:
-                        m.bias_function = LowVramPatch(bias_key, self)
+                    m.bias_function = LowVramPatch(bias_key, self.patches)
-                        patch_counter += 1
+                    patch_counter += 1
 
-                    m.prev_comfy_cast_weights = m.comfy_cast_weights
+                m.prev_comfy_cast_weights = m.comfy_cast_weights
-                    m.comfy_cast_weights = True
+                m.comfy_cast_weights = True
-                    memory_freed += module_mem
+                m.comfy_patched_weights = False
-                    logging.debug("freed {}".format(n))
+                memory_freed += module_mem
+                logging.debug("freed {}".format(n))
 
         self.model.model_lowvram = True
         self.model.lowvram_patch_counter += patch_counter
@@ -670,15 +514,19 @@ class ModelPatcher:
 
     def partially_load(self, device_to, extra_memory=0):
         self.unpatch_model(unpatch_weights=False)
-        self.patch_model(patch_weights=False)
+        self.patch_model(load_weights=False)
         full_load = False
         if self.model.model_lowvram == False:
             return 0
         if self.model.model_loaded_weight_memory + extra_memory > self.model_size():
             full_load = True
         current_used = self.model.model_loaded_weight_memory
-        self.lowvram_load(device_to, lowvram_model_memory=current_used + extra_memory, full_load=full_load)
+        self.load(device_to, lowvram_model_memory=current_used + extra_memory, full_load=full_load)
         return self.model.model_loaded_weight_memory - current_used
 
     def current_loaded_device(self):
         return self.model.device
+
+    def calculate_weight(self, patches, weight, key, intermediate_dtype=torch.float32):
+        print("WARNING the ModelPatcher.calculate_weight function is deprecated, please use: comfy.lora.calculate_weight instead")
+        return comfy.lora.calculate_weight(patches, weight, key, intermediate_dtype=intermediate_dtype)

comfy/ops.py

@@ -18,29 +18,42 @@
 
 import torch
 import comfy.model_management
+from comfy.cli_args import args
 
+def cast_to(weight, dtype=None, device=None, non_blocking=False, copy=False):
+    if device is None or weight.device == device:
+        if not copy:
+            if dtype is None or weight.dtype == dtype:
+                return weight
+        return weight.to(dtype=dtype, copy=copy)
 
-def cast_to(weight, dtype=None, device=None, non_blocking=False):
+    r = torch.empty_like(weight, dtype=dtype, device=device)
-    return weight.to(device=device, dtype=dtype, non_blocking=non_blocking)
+    r.copy_(weight, non_blocking=non_blocking)
+    return r
 
-def cast_to_input(weight, input, non_blocking=False):
+def cast_to_input(weight, input, non_blocking=False, copy=True):
-    return cast_to(weight, input.dtype, input.device, non_blocking=non_blocking)
+    return cast_to(weight, input.dtype, input.device, non_blocking=non_blocking, copy=copy)
 
-def cast_bias_weight(s, input=None, dtype=None, device=None):
+def cast_bias_weight(s, input=None, dtype=None, device=None, bias_dtype=None):
     if input is not None:
         if dtype is None:
             dtype = input.dtype
+        if bias_dtype is None:
+            bias_dtype = dtype
         if device is None:
             device = input.device
 
     bias = None
-    non_blocking = comfy.model_management.device_should_use_non_blocking(device)
+    non_blocking = comfy.model_management.device_supports_non_blocking(device)
     if s.bias is not None:
-        bias = cast_to(s.bias, dtype, device, non_blocking=non_blocking)
+        has_function = s.bias_function is not None
-        if s.bias_function is not None:
+        bias = cast_to(s.bias, bias_dtype, device, non_blocking=non_blocking, copy=has_function)
+        if has_function:
             bias = s.bias_function(bias)
-    weight = cast_to(s.weight, dtype, device, non_blocking=non_blocking)
+
-    if s.weight_function is not None:
+    has_function = s.weight_function is not None
+    weight = cast_to(s.weight, dtype, device, non_blocking=non_blocking, copy=has_function)
+    if has_function:
         weight = s.weight_function(weight)
     return weight, bias
 
@@ -238,3 +251,59 @@ class manual_cast(disable_weight_init):
 
     class Embedding(disable_weight_init.Embedding):
         comfy_cast_weights = True
+
+
+def fp8_linear(self, input):
+    dtype = self.weight.dtype
+    if dtype not in [torch.float8_e4m3fn]:
+        return None
+
+    if len(input.shape) == 3:
+        inn = input.reshape(-1, input.shape[2]).to(dtype)
+        non_blocking = comfy.model_management.device_supports_non_blocking(input.device)
+        w, bias = cast_bias_weight(self, input, dtype=dtype, bias_dtype=input.dtype)
+        w = w.t()
+
+        scale_weight = self.scale_weight
+        scale_input = self.scale_input
+        if scale_weight is None:
+            scale_weight = torch.ones((1), device=input.device, dtype=torch.float32)
+            if scale_input is None:
+                scale_input = scale_weight
+        if scale_input is None:
+            scale_input = torch.ones((1), device=input.device, dtype=torch.float32)
+
+        if bias is not None:
+            o = torch._scaled_mm(inn, w, out_dtype=input.dtype, bias=bias, scale_a=scale_input, scale_b=scale_weight)
+        else:
+            o = torch._scaled_mm(inn, w, out_dtype=input.dtype, scale_a=scale_input, scale_b=scale_weight)
+
+        if isinstance(o, tuple):
+            o = o[0]
+
+        return o.reshape((-1, input.shape[1], self.weight.shape[0]))
+    return None
+
+class fp8_ops(manual_cast):
+    class Linear(manual_cast.Linear):
+        def reset_parameters(self):
+            self.scale_weight = None
+            self.scale_input = None
+            return None
+
+        def forward_comfy_cast_weights(self, input):
+            out = fp8_linear(self, input)
+            if out is not None:
+                return out
+
+            weight, bias = cast_bias_weight(self, input)
+            return torch.nn.functional.linear(input, weight, bias)
+
+
+def pick_operations(weight_dtype, compute_dtype, load_device=None):
+    if compute_dtype is None or weight_dtype == compute_dtype:
+        return disable_weight_init
+    if args.fast:
+        if comfy.model_management.supports_fp8_compute(load_device):
+            return fp8_ops
+    return manual_cast

comfy/sd.py

@@ -24,6 +24,7 @@ import comfy.text_encoders.sa_t5
 import comfy.text_encoders.aura_t5
 import comfy.text_encoders.hydit
 import comfy.text_encoders.flux
+import comfy.text_encoders.long_clipl
 
 import comfy.model_patcher
 import comfy.lora
@@ -62,7 +63,7 @@ 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):
+    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()
@@ -71,9 +72,14 @@ class CLIP:
 
         load_device = model_management.text_encoder_device()
         offload_device = model_management.text_encoder_offload_device()
-        dtype = model_management.text_encoder_dtype(load_device)
+        dtype = model_options.get("dtype", None)
+        if dtype is None:
+            dtype = model_management.text_encoder_dtype(load_device)
+
         params['dtype'] = dtype
         params['device'] = model_management.text_encoder_initial_device(load_device, offload_device, parameters * model_management.dtype_size(dtype))
+        params['model_options'] = model_options
+
         self.cond_stage_model = clip(**(params))
 
         for dt in self.cond_stage_model.dtypes:
@@ -394,11 +400,14 @@ class CLIPType(Enum):
     HUNYUAN_DIT = 5
     FLUX = 6
 
-def load_clip(ckpt_paths, embedding_directory=None, clip_type=CLIPType.STABLE_DIFFUSION):
+def load_clip(ckpt_paths, embedding_directory=None, clip_type=CLIPType.STABLE_DIFFUSION, model_options={}):
     clip_data = []
     for p in ckpt_paths:
         clip_data.append(comfy.utils.load_torch_file(p, safe_load=True))
+    return load_text_encoder_state_dicts(clip_data, embedding_directory=embedding_directory, clip_type=clip_type, model_options=model_options)
 
+def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip_type=CLIPType.STABLE_DIFFUSION, model_options={}):
+    clip_data = state_dicts
     class EmptyClass:
         pass
 
@@ -435,8 +444,13 @@ def load_clip(ckpt_paths, embedding_directory=None, clip_type=CLIPType.STABLE_DI
             clip_target.clip = comfy.text_encoders.sa_t5.SAT5Model
             clip_target.tokenizer = comfy.text_encoders.sa_t5.SAT5Tokenizer
         else:
-            clip_target.clip = sd1_clip.SD1ClipModel
+            w = clip_data[0].get("text_model.embeddings.position_embedding.weight", None)
-            clip_target.tokenizer = sd1_clip.SD1Tokenizer
+            if w is not None and w.shape[0] == 248:
+                clip_target.clip = comfy.text_encoders.long_clipl.LongClipModel
+                clip_target.tokenizer = comfy.text_encoders.long_clipl.LongClipTokenizer
+            else:
+                clip_target.clip = sd1_clip.SD1ClipModel
+                clip_target.tokenizer = sd1_clip.SD1Tokenizer
     elif len(clip_data) == 2:
         if clip_type == CLIPType.SD3:
             clip_target.clip = comfy.text_encoders.sd3_clip.sd3_clip(clip_l=True, clip_g=True, t5=False)
@@ -464,7 +478,7 @@ def load_clip(ckpt_paths, embedding_directory=None, clip_type=CLIPType.STABLE_DI
     for c in clip_data:
         parameters += comfy.utils.calculate_parameters(c)
 
-    clip = CLIP(clip_target, embedding_directory=embedding_directory, parameters=parameters)
+    clip = CLIP(clip_target, embedding_directory=embedding_directory, parameters=parameters, model_options=model_options)
     for c in clip_data:
         m, u = clip.load_sd(c)
         if len(m) > 0:
@@ -506,14 +520,14 @@ def load_checkpoint(config_path=None, ckpt_path=None, output_vae=True, output_cl
 
     return (model, clip, vae)
 
-def load_checkpoint_guess_config(ckpt_path, output_vae=True, output_clip=True, output_clipvision=False, embedding_directory=None, output_model=True, model_options={}):
+def load_checkpoint_guess_config(ckpt_path, output_vae=True, output_clip=True, output_clipvision=False, embedding_directory=None, output_model=True, model_options={}, te_model_options={}):
     sd = comfy.utils.load_torch_file(ckpt_path)
-    out = load_state_dict_guess_config(sd, output_vae, output_clip, output_clipvision, embedding_directory, output_model, model_options)
+    out = load_state_dict_guess_config(sd, output_vae, output_clip, output_clipvision, embedding_directory, output_model, model_options, te_model_options=te_model_options)
     if out is None:
         raise RuntimeError("ERROR: Could not detect model type of: {}".format(ckpt_path))
     return out
 
-def load_state_dict_guess_config(sd, output_vae=True, output_clip=True, output_clipvision=False, embedding_directory=None, output_model=True, model_options={}):
+def load_state_dict_guess_config(sd, output_vae=True, output_clip=True, output_clipvision=False, embedding_directory=None, output_model=True, model_options={}, te_model_options={}):
     clip = None
     clipvision = None
     vae = None
@@ -563,7 +577,7 @@ def load_state_dict_guess_config(sd, output_vae=True, output_clip=True, output_c
             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)
+                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))
@@ -665,10 +679,13 @@ def save_checkpoint(output_path, model, clip=None, vae=None, clip_vision=None, m
     if clip is not None:
         load_models.append(clip.load_model())
         clip_sd = clip.get_sd()
+    vae_sd = None
+    if vae is not None:
+        vae_sd = vae.get_sd()
 
     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.get_sd(), clip_vision_sd)
+    sd = model.model.state_dict_for_saving(clip_sd, vae_sd, clip_vision_sd)
     for k in extra_keys:
         sd[k] = extra_keys[k]
 

comfy/sd1_clip.py

@@ -75,7 +75,6 @@ class ClipTokenWeightEncoder:
         return r
 
 class SDClipModel(torch.nn.Module, ClipTokenWeightEncoder):
-    """Uses the CLIP transformer encoder for text (from huggingface)"""
     LAYERS = [
         "last",
         "pooled",
@@ -84,7 +83,7 @@ class SDClipModel(torch.nn.Module, ClipTokenWeightEncoder):
     def __init__(self, version="openai/clip-vit-large-patch14", device="cpu", max_length=77,
                  freeze=True, layer="last", layer_idx=None, textmodel_json_config=None, dtype=None, model_class=comfy.clip_model.CLIPTextModel,
                  special_tokens={"start": 49406, "end": 49407, "pad": 49407}, layer_norm_hidden_state=True, enable_attention_masks=False, zero_out_masked=False,
-                 return_projected_pooled=True, return_attention_masks=False):  # clip-vit-base-patch32
+                 return_projected_pooled=True, return_attention_masks=False, model_options={}):  # clip-vit-base-patch32
         super().__init__()
         assert layer in self.LAYERS
 
@@ -94,7 +93,11 @@ class SDClipModel(torch.nn.Module, ClipTokenWeightEncoder):
         with open(textmodel_json_config) as f:
             config = json.load(f)
 
-        self.operations = comfy.ops.manual_cast
+        operations = model_options.get("custom_operations", None)
+        if operations is None:
+            operations = comfy.ops.manual_cast
+
+        self.operations = operations
         self.transformer = model_class(config, dtype, device, self.operations)
         self.num_layers = self.transformer.num_layers
 
@@ -552,8 +555,12 @@ class SD1Tokenizer:
     def state_dict(self):
         return {}
 
+class SD1CheckpointClipModel(SDClipModel):
+    def __init__(self, device="cpu", dtype=None, model_options={}):
+        super().__init__(device=device, return_projected_pooled=False, dtype=dtype, model_options=model_options)
+
 class SD1ClipModel(torch.nn.Module):
-    def __init__(self, device="cpu", dtype=None, clip_name="l", clip_model=SDClipModel, name=None, **kwargs):
+    def __init__(self, device="cpu", dtype=None, model_options={}, clip_name="l", clip_model=SD1CheckpointClipModel, name=None, **kwargs):
         super().__init__()
 
         if name is not None:
@@ -563,7 +570,7 @@ class SD1ClipModel(torch.nn.Module):
             self.clip_name = clip_name
             self.clip = "clip_{}".format(self.clip_name)
 
-        setattr(self, self.clip, clip_model(device=device, dtype=dtype, **kwargs))
+        setattr(self, self.clip, clip_model(device=device, dtype=dtype, model_options=model_options, **kwargs))
 
         self.dtypes = set()
         if dtype is not None:

comfy/sdxl_clip.py

@@ -3,14 +3,14 @@ import torch
 import os
 
 class SDXLClipG(sd1_clip.SDClipModel):
-    def __init__(self, device="cpu", max_length=77, freeze=True, layer="penultimate", layer_idx=None, dtype=None):
+    def __init__(self, device="cpu", max_length=77, freeze=True, layer="penultimate", layer_idx=None, dtype=None, model_options={}):
         if layer == "penultimate":
             layer="hidden"
             layer_idx=-2
 
         textmodel_json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "clip_config_bigg.json")
         super().__init__(device=device, freeze=freeze, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype,
-                         special_tokens={"start": 49406, "end": 49407, "pad": 0}, layer_norm_hidden_state=False)
+                         special_tokens={"start": 49406, "end": 49407, "pad": 0}, layer_norm_hidden_state=False, return_projected_pooled=True, model_options=model_options)
 
     def load_sd(self, sd):
         return super().load_sd(sd)
@@ -38,10 +38,10 @@ class SDXLTokenizer:
         return {}
 
 class SDXLClipModel(torch.nn.Module):
-    def __init__(self, device="cpu", dtype=None):
+    def __init__(self, device="cpu", dtype=None, model_options={}):
         super().__init__()
-        self.clip_l = sd1_clip.SDClipModel(layer="hidden", layer_idx=-2, device=device, dtype=dtype, layer_norm_hidden_state=False)
+        self.clip_l = sd1_clip.SDClipModel(layer="hidden", layer_idx=-2, device=device, dtype=dtype, layer_norm_hidden_state=False, model_options=model_options)
-        self.clip_g = SDXLClipG(device=device, dtype=dtype)
+        self.clip_g = SDXLClipG(device=device, dtype=dtype, model_options=model_options)
         self.dtypes = set([dtype])
 
     def set_clip_options(self, options):
@@ -66,8 +66,8 @@ class SDXLClipModel(torch.nn.Module):
             return self.clip_l.load_sd(sd)
 
 class SDXLRefinerClipModel(sd1_clip.SD1ClipModel):
-    def __init__(self, device="cpu", dtype=None):
+    def __init__(self, device="cpu", dtype=None, model_options={}):
-        super().__init__(device=device, dtype=dtype, clip_name="g", clip_model=SDXLClipG)
+        super().__init__(device=device, dtype=dtype, clip_name="g", clip_model=SDXLClipG, model_options=model_options)
 
 
 class StableCascadeClipGTokenizer(sd1_clip.SDTokenizer):
@@ -79,14 +79,14 @@ class StableCascadeTokenizer(sd1_clip.SD1Tokenizer):
         super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, clip_name="g", tokenizer=StableCascadeClipGTokenizer)
 
 class StableCascadeClipG(sd1_clip.SDClipModel):
-    def __init__(self, device="cpu", max_length=77, freeze=True, layer="hidden", layer_idx=-1, dtype=None):
+    def __init__(self, device="cpu", max_length=77, freeze=True, layer="hidden", layer_idx=-1, dtype=None, model_options={}):
         textmodel_json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "clip_config_bigg.json")
         super().__init__(device=device, freeze=freeze, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype,
-                         special_tokens={"start": 49406, "end": 49407, "pad": 49407}, layer_norm_hidden_state=False, enable_attention_masks=True)
+                         special_tokens={"start": 49406, "end": 49407, "pad": 49407}, layer_norm_hidden_state=False, enable_attention_masks=True, return_projected_pooled=True, model_options=model_options)
 
     def load_sd(self, sd):
         return super().load_sd(sd)
 
 class StableCascadeClipModel(sd1_clip.SD1ClipModel):
-    def __init__(self, device="cpu", dtype=None):
+    def __init__(self, device="cpu", dtype=None, model_options={}):
-        super().__init__(device=device, dtype=dtype, clip_name="g", clip_model=StableCascadeClipG)
+        super().__init__(device=device, dtype=dtype, clip_name="g", clip_model=StableCascadeClipG, model_options=model_options)

comfy/supported_models.py

@@ -654,6 +654,7 @@ class Flux(supported_models_base.BASE):
     def clip_target(self, state_dict={}):
         pref = self.text_encoder_key_prefix[0]
         t5_key = "{}t5xxl.transformer.encoder.final_layer_norm.weight".format(pref)
+        dtype_t5 = None
         if t5_key in state_dict:
             dtype_t5 = state_dict[t5_key].dtype
         return supported_models_base.ClipTarget(comfy.text_encoders.flux.FluxTokenizer, comfy.text_encoders.flux.flux_clip(dtype_t5=dtype_t5))

comfy/text_encoders/aura_t5.py

@@ -4,9 +4,9 @@ import comfy.text_encoders.t5
 import os
 
 class PT5XlModel(sd1_clip.SDClipModel):
-    def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None):
+    def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None, model_options={}):
         textmodel_json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "t5_pile_config_xl.json")
-        super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype, special_tokens={"end": 2, "pad": 1}, model_class=comfy.text_encoders.t5.T5, enable_attention_masks=True, zero_out_masked=True)
+        super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype, special_tokens={"end": 2, "pad": 1}, model_class=comfy.text_encoders.t5.T5, enable_attention_masks=True, zero_out_masked=True, model_options=model_options)
 
 class PT5XlTokenizer(sd1_clip.SDTokenizer):
     def __init__(self, embedding_directory=None, tokenizer_data={}):
@@ -18,5 +18,5 @@ class AuraT5Tokenizer(sd1_clip.SD1Tokenizer):
         super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, clip_name="pile_t5xl", tokenizer=PT5XlTokenizer)
 
 class AuraT5Model(sd1_clip.SD1ClipModel):
-    def __init__(self, device="cpu", dtype=None, **kwargs):
+    def __init__(self, device="cpu", dtype=None, model_options={}, **kwargs):
-        super().__init__(device=device, dtype=dtype, name="pile_t5xl", clip_model=PT5XlModel, **kwargs)
+        super().__init__(device=device, dtype=dtype, model_options=model_options, name="pile_t5xl", clip_model=PT5XlModel, **kwargs)

comfy/text_encoders/flux.py

@@ -6,9 +6,9 @@ import torch
 import os
 
 class T5XXLModel(sd1_clip.SDClipModel):
-    def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None):
+    def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None, model_options={}):
         textmodel_json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "t5_config_xxl.json")
-        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)
+        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, model_options=model_options)
 
 class T5XXLTokenizer(sd1_clip.SDTokenizer):
     def __init__(self, embedding_directory=None, tokenizer_data={}):
@@ -35,11 +35,11 @@ class FluxTokenizer:
 
 
 class FluxClipModel(torch.nn.Module):
-    def __init__(self, dtype_t5=None, device="cpu", dtype=None):
+    def __init__(self, dtype_t5=None, device="cpu", dtype=None, model_options={}):
         super().__init__()
         dtype_t5 = comfy.model_management.pick_weight_dtype(dtype_t5, dtype, device)
-        self.clip_l = sd1_clip.SDClipModel(device=device, dtype=dtype, return_projected_pooled=False)
+        self.clip_l = sd1_clip.SDClipModel(device=device, dtype=dtype, return_projected_pooled=False, model_options=model_options)
-        self.t5xxl = T5XXLModel(device=device, dtype=dtype_t5)
+        self.t5xxl = T5XXLModel(device=device, dtype=dtype_t5, model_options=model_options)
         self.dtypes = set([dtype, dtype_t5])
 
     def set_clip_options(self, options):
@@ -66,6 +66,6 @@ class FluxClipModel(torch.nn.Module):
 
 def flux_clip(dtype_t5=None):
     class FluxClipModel_(FluxClipModel):
-        def __init__(self, device="cpu", dtype=None):
+        def __init__(self, device="cpu", dtype=None, model_options={}):
-            super().__init__(dtype_t5=dtype_t5, device=device, dtype=dtype)
+            super().__init__(dtype_t5=dtype_t5, device=device, dtype=dtype, model_options=model_options)
     return FluxClipModel_

comfy/text_encoders/hydit.py

@@ -7,9 +7,9 @@ import os
 import torch
 
 class HyditBertModel(sd1_clip.SDClipModel):
-    def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None):
+    def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None, model_options={}):
         textmodel_json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "hydit_clip.json")
-        super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype, special_tokens={"start": 101, "end": 102, "pad": 0}, model_class=BertModel, enable_attention_masks=True, return_attention_masks=True)
+        super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype, special_tokens={"start": 101, "end": 102, "pad": 0}, model_class=BertModel, enable_attention_masks=True, return_attention_masks=True, model_options=model_options)
 
 class HyditBertTokenizer(sd1_clip.SDTokenizer):
     def __init__(self, embedding_directory=None, tokenizer_data={}):
@@ -18,9 +18,9 @@ class HyditBertTokenizer(sd1_clip.SDTokenizer):
 
 
 class MT5XLModel(sd1_clip.SDClipModel):
-    def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None):
+    def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None, model_options={}):
         textmodel_json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "mt5_config_xl.json")
-        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=True, return_attention_masks=True)
+        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=True, return_attention_masks=True, model_options=model_options)
 
 class MT5XLTokenizer(sd1_clip.SDTokenizer):
     def __init__(self, embedding_directory=None, tokenizer_data={}):
@@ -50,10 +50,10 @@ class HyditTokenizer:
         return {"mt5xl.spiece_model": self.mt5xl.state_dict()["spiece_model"]}
 
 class HyditModel(torch.nn.Module):
-    def __init__(self, device="cpu", dtype=None):
+    def __init__(self, device="cpu", dtype=None, model_options={}):
         super().__init__()
-        self.hydit_clip = HyditBertModel(dtype=dtype)
+        self.hydit_clip = HyditBertModel(dtype=dtype, model_options=model_options)
-        self.mt5xl = MT5XLModel(dtype=dtype)
+        self.mt5xl = MT5XLModel(dtype=dtype, model_options=model_options)
 
         self.dtypes = set()
         if dtype is not None:

comfy/text_encoders/long_clipl.json

@@ -0,0 +1,25 @@
+{
+  "_name_or_path": "openai/clip-vit-large-patch14",
+  "architectures": [
+    "CLIPTextModel"
+  ],
+  "attention_dropout": 0.0,
+  "bos_token_id": 0,
+  "dropout": 0.0,
+  "eos_token_id": 49407,
+  "hidden_act": "quick_gelu",
+  "hidden_size": 768,
+  "initializer_factor": 1.0,
+  "initializer_range": 0.02,
+  "intermediate_size": 3072,
+  "layer_norm_eps": 1e-05,
+  "max_position_embeddings": 248,
+  "model_type": "clip_text_model",
+  "num_attention_heads": 12,
+  "num_hidden_layers": 12,
+  "pad_token_id": 1,
+  "projection_dim": 768,
+  "torch_dtype": "float32",
+  "transformers_version": "4.24.0",
+  "vocab_size": 49408
+}

comfy/text_encoders/long_clipl.py

@@ -0,0 +1,19 @@
+from comfy import sd1_clip
+import os
+
+class LongClipTokenizer_(sd1_clip.SDTokenizer):
+    def __init__(self, embedding_directory=None, tokenizer_data={}):
+        super().__init__(max_length=248, embedding_directory=embedding_directory, tokenizer_data=tokenizer_data)
+
+class LongClipModel_(sd1_clip.SDClipModel):
+    def __init__(self, device="cpu", dtype=None, model_options={}):
+        textmodel_json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "long_clipl.json")
+        super().__init__(device=device, textmodel_json_config=textmodel_json_config, return_projected_pooled=False, dtype=dtype, model_options=model_options)
+
+class LongClipTokenizer(sd1_clip.SD1Tokenizer):
+    def __init__(self, embedding_directory=None, tokenizer_data={}):
+        super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, tokenizer=LongClipTokenizer_)
+
+class LongClipModel(sd1_clip.SD1ClipModel):
+    def __init__(self, device="cpu", dtype=None, model_options={}, **kwargs):
+        super().__init__(device=device, dtype=dtype, model_options=model_options, clip_model=LongClipModel_, **kwargs)

comfy/text_encoders/sa_t5.py

@@ -4,9 +4,9 @@ import comfy.text_encoders.t5
 import os
 
 class T5BaseModel(sd1_clip.SDClipModel):
-    def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None):
+    def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None, model_options={}):
         textmodel_json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "t5_config_base.json")
-        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=True, zero_out_masked=True)
+        super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype, model_options=model_options, special_tokens={"end": 1, "pad": 0}, model_class=comfy.text_encoders.t5.T5, enable_attention_masks=True, zero_out_masked=True)
 
 class T5BaseTokenizer(sd1_clip.SDTokenizer):
     def __init__(self, embedding_directory=None, tokenizer_data={}):
@@ -18,5 +18,5 @@ class SAT5Tokenizer(sd1_clip.SD1Tokenizer):
         super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, clip_name="t5base", tokenizer=T5BaseTokenizer)
 
 class SAT5Model(sd1_clip.SD1ClipModel):
-    def __init__(self, device="cpu", dtype=None, **kwargs):
+    def __init__(self, device="cpu", dtype=None, model_options={}, **kwargs):
-        super().__init__(device=device, dtype=dtype, name="t5base", clip_model=T5BaseModel, **kwargs)
+        super().__init__(device=device, dtype=dtype, model_options=model_options, name="t5base", clip_model=T5BaseModel, **kwargs)

comfy/text_encoders/sd2_clip.py

@@ -2,13 +2,13 @@ from comfy import sd1_clip
 import os
 
 class SD2ClipHModel(sd1_clip.SDClipModel):
-    def __init__(self, arch="ViT-H-14", device="cpu", max_length=77, freeze=True, layer="penultimate", layer_idx=None, dtype=None):
+    def __init__(self, arch="ViT-H-14", device="cpu", max_length=77, freeze=True, layer="penultimate", layer_idx=None, dtype=None, model_options={}):
         if layer == "penultimate":
             layer="hidden"
             layer_idx=-2
 
         textmodel_json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "sd2_clip_config.json")
-        super().__init__(device=device, freeze=freeze, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype, special_tokens={"start": 49406, "end": 49407, "pad": 0})
+        super().__init__(device=device, freeze=freeze, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype, special_tokens={"start": 49406, "end": 49407, "pad": 0}, return_projected_pooled=True, model_options=model_options)
 
 class SD2ClipHTokenizer(sd1_clip.SDTokenizer):
     def __init__(self, tokenizer_path=None, embedding_directory=None, tokenizer_data={}):
@@ -19,5 +19,5 @@ class SD2Tokenizer(sd1_clip.SD1Tokenizer):
         super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, clip_name="h", tokenizer=SD2ClipHTokenizer)
 
 class SD2ClipModel(sd1_clip.SD1ClipModel):
-    def __init__(self, device="cpu", dtype=None, **kwargs):
+    def __init__(self, device="cpu", dtype=None, model_options={}, **kwargs):
-        super().__init__(device=device, dtype=dtype, clip_name="h", clip_model=SD2ClipHModel, **kwargs)
+        super().__init__(device=device, dtype=dtype, model_options=model_options, clip_name="h", clip_model=SD2ClipHModel, **kwargs)

comfy/text_encoders/sd3_clip.py

@@ -8,14 +8,14 @@ import comfy.model_management
 import logging
 
 class T5XXLModel(sd1_clip.SDClipModel):
-    def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None):
+    def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None, model_options={}):
         textmodel_json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "t5_config_xxl.json")
-        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)
+        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, model_options=model_options)
 
 class T5XXLTokenizer(sd1_clip.SDTokenizer):
     def __init__(self, embedding_directory=None, tokenizer_data={}):
         tokenizer_path = os.path.join(os.path.dirname(os.path.realpath(__file__)), "t5_tokenizer")
-        super().__init__(tokenizer_path, pad_with_end=False, embedding_size=4096, embedding_key='t5xxl', tokenizer_class=T5TokenizerFast, has_start_token=False, pad_to_max_length=False, max_length=99999999, min_length=77)
+        super().__init__(tokenizer_path, embedding_directory=embedding_directory, pad_with_end=False, embedding_size=4096, embedding_key='t5xxl', tokenizer_class=T5TokenizerFast, has_start_token=False, pad_to_max_length=False, max_length=99999999, min_length=77)
 
 
 class SD3Tokenizer:
@@ -38,24 +38,24 @@ class SD3Tokenizer:
         return {}
 
 class SD3ClipModel(torch.nn.Module):
-    def __init__(self, clip_l=True, clip_g=True, t5=True, dtype_t5=None, device="cpu", dtype=None):
+    def __init__(self, clip_l=True, clip_g=True, t5=True, dtype_t5=None, device="cpu", dtype=None, model_options={}):
         super().__init__()
         self.dtypes = set()
         if clip_l:
-            self.clip_l = sd1_clip.SDClipModel(layer="hidden", layer_idx=-2, device=device, dtype=dtype, layer_norm_hidden_state=False, return_projected_pooled=False)
+            self.clip_l = sd1_clip.SDClipModel(layer="hidden", layer_idx=-2, device=device, dtype=dtype, layer_norm_hidden_state=False, return_projected_pooled=False, model_options=model_options)
             self.dtypes.add(dtype)
         else:
             self.clip_l = None
 
         if clip_g:
-            self.clip_g = sdxl_clip.SDXLClipG(device=device, dtype=dtype)
+            self.clip_g = sdxl_clip.SDXLClipG(device=device, dtype=dtype, model_options=model_options)
             self.dtypes.add(dtype)
         else:
             self.clip_g = None
 
         if t5:
             dtype_t5 = comfy.model_management.pick_weight_dtype(dtype_t5, dtype, device)
-            self.t5xxl = T5XXLModel(device=device, dtype=dtype_t5)
+            self.t5xxl = T5XXLModel(device=device, dtype=dtype_t5, model_options=model_options)
             self.dtypes.add(dtype_t5)
         else:
             self.t5xxl = None
@@ -132,6 +132,6 @@ class SD3ClipModel(torch.nn.Module):
 
 def sd3_clip(clip_l=True, clip_g=True, t5=True, dtype_t5=None):
     class SD3ClipModel_(SD3ClipModel):
-        def __init__(self, device="cpu", dtype=None):
+        def __init__(self, device="cpu", dtype=None, model_options={}):
-            super().__init__(clip_l=clip_l, clip_g=clip_g, t5=t5, dtype_t5=dtype_t5, device=device, dtype=dtype)
+            super().__init__(clip_l=clip_l, clip_g=clip_g, t5=t5, dtype_t5=dtype_t5, device=device, dtype=dtype, model_options=model_options)
     return SD3ClipModel_

comfy/utils.py

@@ -528,6 +528,8 @@ def flux_to_diffusers(mmdit_config, output_prefix=""):
         ("guidance_in.out_layer.weight", "time_text_embed.guidance_embedder.linear_2.weight"),
         ("final_layer.adaLN_modulation.1.bias", "norm_out.linear.bias", swap_scale_shift),
         ("final_layer.adaLN_modulation.1.weight", "norm_out.linear.weight", swap_scale_shift),
+        ("pos_embed_input.bias", "controlnet_x_embedder.bias"),
+        ("pos_embed_input.weight", "controlnet_x_embedder.weight"),
     }
 
     for k in MAP_BASIC:

comfy_execution/caching.py

@@ -56,6 +56,8 @@ class CacheKeySetID(CacheKeySet):
         for node_id in node_ids:
             if node_id in self.keys:
                 continue
+            if not self.dynprompt.has_node(node_id):
+                continue
             node = self.dynprompt.get_node(node_id)
             self.keys[node_id] = (node_id, node["class_type"])
             self.subcache_keys[node_id] = (node_id, node["class_type"])
@@ -74,6 +76,8 @@ class CacheKeySetInputSignature(CacheKeySet):
         for node_id in node_ids:
             if node_id in self.keys:
                 continue
+            if not self.dynprompt.has_node(node_id):
+                continue
             node = self.dynprompt.get_node(node_id)
             self.keys[node_id] = self.get_node_signature(self.dynprompt, node_id)
             self.subcache_keys[node_id] = (node_id, node["class_type"])
@@ -87,6 +91,9 @@ class CacheKeySetInputSignature(CacheKeySet):
         return to_hashable(signature)
 
     def get_immediate_node_signature(self, dynprompt, node_id, ancestor_order_mapping):
+        if not dynprompt.has_node(node_id):
+            # This node doesn't exist -- we can't cache it.
+            return [float("NaN")]
         node = dynprompt.get_node(node_id)
         class_type = node["class_type"]
         class_def = nodes.NODE_CLASS_MAPPINGS[class_type]
@@ -112,6 +119,8 @@ class CacheKeySetInputSignature(CacheKeySet):
         return ancestors, order_mapping
 
     def get_ordered_ancestry_internal(self, dynprompt, node_id, ancestors, order_mapping):
+        if not dynprompt.has_node(node_id):
+            return
         inputs = dynprompt.get_node(node_id)["inputs"]
         input_keys = sorted(inputs.keys())
         for key in input_keys:

comfy_execution/graph.py

@@ -176,19 +176,41 @@ class ExecutionList(TopologicalSort):
                 "current_inputs": []
             }
             return None, error_details, ex
-        next_node = available[0]
+
+        self.staged_node_id = self.ux_friendly_pick_node(available)
+        return self.staged_node_id, None, None
+
+    def ux_friendly_pick_node(self, node_list):
         # If an output node is available, do that first.
         # Technically this has no effect on the overall length of execution, but it feels better as a user
         # for a PreviewImage to display a result as soon as it can
         # Some other heuristics could probably be used here to improve the UX further.
-        for node_id in available:
+        def is_output(node_id):
             class_type = self.dynprompt.get_node(node_id)["class_type"]
             class_def = nodes.NODE_CLASS_MAPPINGS[class_type]
             if hasattr(class_def, 'OUTPUT_NODE') and class_def.OUTPUT_NODE == True:
-                next_node = node_id
+                return True
-                break
+            return False
-        self.staged_node_id = next_node
+
-        return self.staged_node_id, None, None
+        for node_id in node_list:
+            if is_output(node_id):
+                return node_id
+
+        #This should handle the VAEDecode -> preview case
+        for node_id in node_list:
+            for blocked_node_id in self.blocking[node_id]:
+                if is_output(blocked_node_id):
+                    return node_id
+
+        #This should handle the VAELoader -> VAEDecode -> preview case
+        for node_id in node_list:
+            for blocked_node_id in self.blocking[node_id]:
+                for blocked_node_id1 in self.blocking[blocked_node_id]:
+                    if is_output(blocked_node_id1):
+                        return node_id
+
+        #TODO: this function should be improved
+        return node_list[0]
 
     def unstage_node_execution(self):
         assert self.staged_node_id is not None

comfy_extras/nodes_lora_extract.py

@@ -0,0 +1,91 @@
+import torch
+import comfy.model_management
+import comfy.utils
+import folder_paths
+import os
+import logging
+
+CLAMP_QUANTILE = 0.99
+
+def extract_lora(diff, rank):
+    conv2d = (len(diff.shape) == 4)
+    kernel_size = None if not conv2d else diff.size()[2:4]
+    conv2d_3x3 = conv2d and kernel_size != (1, 1)
+    out_dim, in_dim = diff.size()[0:2]
+    rank = min(rank, in_dim, out_dim)
+
+    if conv2d:
+        if conv2d_3x3:
+            diff = diff.flatten(start_dim=1)
+        else:
+            diff = diff.squeeze()
+
+
+    U, S, Vh = torch.linalg.svd(diff.float())
+    U = U[:, :rank]
+    S = S[:rank]
+    U = U @ torch.diag(S)
+    Vh = Vh[:rank, :]
+
+    dist = torch.cat([U.flatten(), Vh.flatten()])
+    hi_val = torch.quantile(dist, CLAMP_QUANTILE)
+    low_val = -hi_val
+
+    U = U.clamp(low_val, hi_val)
+    Vh = Vh.clamp(low_val, hi_val)
+    if conv2d:
+        U = U.reshape(out_dim, rank, 1, 1)
+        Vh = Vh.reshape(rank, in_dim, kernel_size[0], kernel_size[1])
+    return (U, Vh)
+
+class LoraSave:
+    def __init__(self):
+        self.output_dir = folder_paths.get_output_directory()
+
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": {"filename_prefix": ("STRING", {"default": "loras/ComfyUI_extracted_lora"}),
+                              "rank": ("INT", {"default": 8, "min": 1, "max": 1024, "step": 1}),
+                            },
+                "optional": {"model_diff": ("MODEL",),},
+    }
+    RETURN_TYPES = ()
+    FUNCTION = "save"
+    OUTPUT_NODE = True
+
+    CATEGORY = "_for_testing"
+
+    def save(self, filename_prefix, rank, model_diff=None):
+        if model_diff is None:
+            return {}
+
+        full_output_folder, filename, counter, subfolder, filename_prefix = folder_paths.get_save_image_path(filename_prefix, self.output_dir)
+
+        output_sd = {}
+        prefix_key = "diffusion_model."
+        stored = set()
+
+        comfy.model_management.load_models_gpu([model_diff], force_patch_weights=True)
+        sd = model_diff.model_state_dict(filter_prefix=prefix_key)
+
+        for k in sd:
+            if k.endswith(".weight"):
+                weight_diff = sd[k]
+                if weight_diff.ndim < 2:
+                    continue
+                try:
+                    out = extract_lora(weight_diff, rank)
+                    output_sd["{}.lora_up.weight".format(k[:-7])] = out[0].contiguous().half().cpu()
+                    output_sd["{}.lora_down.weight".format(k[:-7])] = out[1].contiguous().half().cpu()
+                except:
+                    logging.warning("Could not generate lora weights for key {}, is the weight difference a zero?".format(k))
+
+        output_checkpoint = f"{filename}_{counter:05}_.safetensors"
+        output_checkpoint = os.path.join(full_output_folder, output_checkpoint)
+
+        comfy.utils.save_torch_file(output_sd, output_checkpoint, metadata=None)
+        return {}
+
+NODE_CLASS_MAPPINGS = {
+    "LoraSave": LoraSave
+}

comfy_extras/nodes_model_merging.py

@@ -333,6 +333,25 @@ class VAESave:
         comfy.utils.save_torch_file(vae.get_sd(), output_checkpoint, metadata=metadata)
         return {}
 
+class ModelSave:
+    def __init__(self):
+        self.output_dir = folder_paths.get_output_directory()
+
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": { "model": ("MODEL",),
+                              "filename_prefix": ("STRING", {"default": "diffusion_models/ComfyUI"}),},
+                "hidden": {"prompt": "PROMPT", "extra_pnginfo": "EXTRA_PNGINFO"},}
+    RETURN_TYPES = ()
+    FUNCTION = "save"
+    OUTPUT_NODE = True
+
+    CATEGORY = "advanced/model_merging"
+
+    def save(self, model, filename_prefix, prompt=None, extra_pnginfo=None):
+        save_checkpoint(model, filename_prefix=filename_prefix, output_dir=self.output_dir, prompt=prompt, extra_pnginfo=extra_pnginfo)
+        return {}
+
 NODE_CLASS_MAPPINGS = {
     "ModelMergeSimple": ModelMergeSimple,
     "ModelMergeBlocks": ModelMergeBlocks,
@@ -344,4 +363,9 @@ NODE_CLASS_MAPPINGS = {
     "CLIPMergeAdd": CLIPAdd,
     "CLIPSave": CLIPSave,
     "VAESave": VAESave,
+    "ModelSave": ModelSave,
+}
+
+NODE_DISPLAY_NAME_MAPPINGS = {
+    "CheckpointSave": "Save Checkpoint",
 }

custom_nodes/example_node.py.example

@@ -23,6 +23,12 @@ class Example:
         Assumed to be False if not present.
     CATEGORY (`str`):
         The category the node should appear in the UI.
+    DEPRECATED (`bool`):
+        Indicates whether the node is deprecated. Deprecated nodes are hidden by default in the UI, but remain
+        functional in existing workflows that use them.
+    EXPERIMENTAL (`bool`):
+        Indicates whether the node is experimental. Experimental nodes are marked as such in the UI and may be subject to
+        significant changes or removal in future versions. Use with caution in production workflows.
     execute(s) -> tuple || None:
         The entry point method. The name of this method must be the same as the value of property `FUNCTION`.
         For example, if `FUNCTION = "execute"` then this method's name must be `execute`, if `FUNCTION = "foo"` then it must be `foo`.

execution.py

@@ -47,7 +47,8 @@ class IsChangedCache:
             self.is_changed[node_id] = node["is_changed"]
             return self.is_changed[node_id]
 
-        input_data_all, _ = get_input_data(node["inputs"], class_def, node_id, self.outputs_cache)
+        # Intentionally do not use cached outputs here. We only want constants in IS_CHANGED
+        input_data_all, _ = get_input_data(node["inputs"], class_def, node_id, None)
         try:
             is_changed = _map_node_over_list(class_def, input_data_all, "IS_CHANGED")
             node["is_changed"] = [None if isinstance(x, ExecutionBlocker) else x for x in is_changed]
@@ -491,6 +492,7 @@ class PromptExecutor:
                     break
 
                 result, error, ex = execute(self.server, dynamic_prompt, self.caches, node_id, extra_data, executed, prompt_id, execution_list, pending_subgraph_results)
+                self.success = result != ExecutionResult.FAILURE
                 if result == ExecutionResult.FAILURE:
                     self.handle_execution_error(prompt_id, dynamic_prompt.original_prompt, current_outputs, executed, error, ex)
                     break

folder_paths.py

@@ -17,7 +17,7 @@ folder_names_and_paths["configs"] = ([os.path.join(models_dir, "configs")], [".y
 folder_names_and_paths["loras"] = ([os.path.join(models_dir, "loras")], supported_pt_extensions)
 folder_names_and_paths["vae"] = ([os.path.join(models_dir, "vae")], supported_pt_extensions)
 folder_names_and_paths["clip"] = ([os.path.join(models_dir, "clip")], supported_pt_extensions)
-folder_names_and_paths["unet"] = ([os.path.join(models_dir, "unet")], supported_pt_extensions)
+folder_names_and_paths["diffusion_models"] = ([os.path.join(models_dir, "unet"), os.path.join(models_dir, "diffusion_models")], supported_pt_extensions)
 folder_names_and_paths["clip_vision"] = ([os.path.join(models_dir, "clip_vision")], supported_pt_extensions)
 folder_names_and_paths["style_models"] = ([os.path.join(models_dir, "style_models")], supported_pt_extensions)
 folder_names_and_paths["embeddings"] = ([os.path.join(models_dir, "embeddings")], supported_pt_extensions)
@@ -44,6 +44,10 @@ user_directory = os.path.join(os.path.dirname(os.path.realpath(__file__)), "user
 
 filename_list_cache: dict[str, tuple[list[str], dict[str, float], float]] = {}
 
+def map_legacy(folder_name: str) -> str:
+    legacy = {"unet": "diffusion_models"}
+    return legacy.get(folder_name, folder_name)
+
 if not os.path.exists(input_directory):
     try:
         os.makedirs(input_directory)
@@ -128,6 +132,7 @@ def exists_annotated_filepath(name) -> bool:
 
 def add_model_folder_path(folder_name: str, full_folder_path: str) -> None:
     global folder_names_and_paths
+    folder_name = map_legacy(folder_name)
     if folder_name in folder_names_and_paths:
         folder_names_and_paths[folder_name][0].append(full_folder_path)
     else:
@@ -135,6 +140,7 @@ def add_model_folder_path(folder_name: str, full_folder_path: str) -> None:
 
 def get_folder_paths(folder_name: str) -> list[str]:
     global folder_names_and_paths
+    folder_name = map_legacy(folder_name)
     return folder_names_and_paths[folder_name][0][:]
 
 def recursive_search(directory: str, excluded_dir_names: list[str] | None=None) -> tuple[list[str], dict[str, float]]:
@@ -181,6 +187,7 @@ def filter_files_extensions(files: Collection[str], extensions: Collection[str])
 
 def get_full_path(folder_name: str, filename: str) -> str | None:
     global folder_names_and_paths
+    folder_name = map_legacy(folder_name)
     if folder_name not in folder_names_and_paths:
         return None
     folders = folder_names_and_paths[folder_name]
@@ -195,6 +202,7 @@ def get_full_path(folder_name: str, filename: str) -> str | None:
     return None
 
 def get_filename_list_(folder_name: str) -> tuple[list[str], dict[str, float], float]:
+    folder_name = map_legacy(folder_name)
     global folder_names_and_paths
     output_list = set()
     folders = folder_names_and_paths[folder_name]
@@ -209,6 +217,7 @@ def get_filename_list_(folder_name: str) -> tuple[list[str], dict[str, float], f
 def cached_filename_list_(folder_name: str) -> tuple[list[str], dict[str, float], float] | None:
     global filename_list_cache
     global folder_names_and_paths
+    folder_name = map_legacy(folder_name)
     if folder_name not in filename_list_cache:
         return None
     out = filename_list_cache[folder_name]
@@ -228,6 +237,7 @@ def cached_filename_list_(folder_name: str) -> tuple[list[str], dict[str, float]
     return out
 
 def get_filename_list(folder_name: str) -> list[str]:
+    folder_name = map_legacy(folder_name)
     out = cached_filename_list_(folder_name)
     if out is None:
         out = get_filename_list_(folder_name)

latent_preview.py

@@ -9,7 +9,7 @@ import folder_paths
 import comfy.utils
 import logging
 
-MAX_PREVIEW_RESOLUTION = 512
+MAX_PREVIEW_RESOLUTION = args.preview_size
 
 def preview_to_image(latent_image):
         latents_ubyte = (((latent_image + 1.0) / 2.0).clamp(0, 1)  # change scale from -1..1 to 0..1

main.py

@@ -6,6 +6,10 @@ import importlib.util
 import folder_paths
 import time
 from comfy.cli_args import args
+from app.logger import setup_logger
+
+
+setup_logger(verbose=args.verbose)
 
 
 def execute_prestartup_script():
@@ -242,6 +246,8 @@ if __name__ == "__main__":
     folder_paths.add_model_folder_path("checkpoints", os.path.join(folder_paths.get_output_directory(), "checkpoints"))
     folder_paths.add_model_folder_path("clip", os.path.join(folder_paths.get_output_directory(), "clip"))
     folder_paths.add_model_folder_path("vae", os.path.join(folder_paths.get_output_directory(), "vae"))
+    folder_paths.add_model_folder_path("diffusion_models", os.path.join(folder_paths.get_output_directory(), "diffusion_models"))
+    folder_paths.add_model_folder_path("loras", os.path.join(folder_paths.get_output_directory(), "loras"))
 
     if args.input_directory:
         input_dir = os.path.abspath(args.input_directory)

nodes.py

@@ -665,6 +665,8 @@ class VAELoader:
         sd1_taesd_dec = False
         sd3_taesd_enc = False
         sd3_taesd_dec = False
+        f1_taesd_enc = False
+        f1_taesd_dec = False
 
         for v in approx_vaes:
             if v.startswith("taesd_decoder."):
@@ -679,12 +681,18 @@ class VAELoader:
                 sd3_taesd_dec = True
             elif v.startswith("taesd3_encoder."):
                 sd3_taesd_enc = True
+            elif v.startswith("taef1_encoder."):
+                f1_taesd_dec = True
+            elif v.startswith("taef1_decoder."):
+                f1_taesd_enc = True
         if sd1_taesd_dec and sd1_taesd_enc:
             vaes.append("taesd")
         if sdxl_taesd_dec and sdxl_taesd_enc:
             vaes.append("taesdxl")
         if sd3_taesd_dec and sd3_taesd_enc:
             vaes.append("taesd3")
+        if f1_taesd_dec and f1_taesd_enc:
+            vaes.append("taef1")
         return vaes
 
     @staticmethod
@@ -712,6 +720,9 @@ class VAELoader:
         elif name == "taesd3":
             sd["vae_scale"] = torch.tensor(1.5305)
             sd["vae_shift"] = torch.tensor(0.0609)
+        elif name == "taef1":
+            sd["vae_scale"] = torch.tensor(0.3611)
+            sd["vae_shift"] = torch.tensor(0.1159)
         return sd
 
     @classmethod
@@ -724,7 +735,7 @@ class VAELoader:
 
     #TODO: scale factor?
     def load_vae(self, vae_name):
-        if vae_name in ["taesd", "taesdxl", "taesd3"]:
+        if vae_name in ["taesd", "taesdxl", "taesd3", "taef1"]:
             sd = self.load_taesd(vae_name)
         else:
             vae_path = folder_paths.get_full_path("vae", vae_name)
@@ -844,7 +855,7 @@ class ControlNetApplyAdvanced:
 class UNETLoader:
     @classmethod
     def INPUT_TYPES(s):
-        return {"required": { "unet_name": (folder_paths.get_filename_list("unet"), ),
+        return {"required": { "unet_name": (folder_paths.get_filename_list("diffusion_models"), ),
                               "weight_dtype": (["default", "fp8_e4m3fn", "fp8_e5m2"],)
                              }}
     RETURN_TYPES = ("MODEL",)
@@ -859,7 +870,7 @@ class UNETLoader:
         elif weight_dtype == "fp8_e5m2":
             model_options["dtype"] = torch.float8_e5m2
 
-        unet_path = folder_paths.get_full_path("unet", unet_name)
+        unet_path = folder_paths.get_full_path("diffusion_models", unet_name)
         model = comfy.sd.load_diffusion_model(unet_path, model_options=model_options)
         return (model,)
 
@@ -2090,6 +2101,7 @@ def init_builtin_extra_nodes():
         "nodes_controlnet.py",
         "nodes_hunyuan.py",
         "nodes_flux.py",
+        "nodes_lora_extract.py",
     ]
 
     import_failed = []
@@ -2118,3 +2130,5 @@ def init_extra_nodes(init_custom_nodes=True):
         else:
             logging.warning("Please do a: pip install -r requirements.txt")
         logging.warning("")
+    
+    return import_failed

notebooks/comfyui_colab.ipynb

@@ -79,7 +79,7 @@
         "#!wget -c https://huggingface.co/comfyanonymous/clip_vision_g/resolve/main/clip_vision_g.safetensors -P ./models/clip_vision/\n",
         "\n",
         "# SD1.5\n",
-        "!wget -c https://huggingface.co/runwayml/stable-diffusion-v1-5/resolve/main/v1-5-pruned-emaonly.ckpt -P ./models/checkpoints/\n",
+        "!wget -c https://huggingface.co/Comfy-Org/stable-diffusion-v1-5-archive/resolve/main/v1-5-pruned-emaonly-fp16.safetensors -P ./models/checkpoints/\n",
         "\n",
         "# SD2\n",
         "#!wget -c https://huggingface.co/stabilityai/stable-diffusion-2-1-base/resolve/main/v2-1_512-ema-pruned.safetensors -P ./models/checkpoints/\n",

script_examples/basic_api_example.py

@@ -43,7 +43,7 @@ prompt_text = """
     "4": {
         "class_type": "CheckpointLoaderSimple",
         "inputs": {
-            "ckpt_name": "v1-5-pruned-emaonly.ckpt"
+            "ckpt_name": "v1-5-pruned-emaonly.safetensors"
         }
     },
     "5": {

script_examples/websockets_api_example.py

@@ -41,15 +41,14 @@ def get_images(ws, prompt):
             continue #previews are binary data
 
     history = get_history(prompt_id)[prompt_id]
-    for o in history['outputs']:
+    for node_id in history['outputs']:
-        for node_id in history['outputs']:
+        node_output = history['outputs'][node_id]
-            node_output = history['outputs'][node_id]
+        images_output = []
-            if 'images' in node_output:
+        if 'images' in node_output:
-                images_output = []
+            for image in node_output['images']:
-                for image in node_output['images']:
+                image_data = get_image(image['filename'], image['subfolder'], image['type'])
-                    image_data = get_image(image['filename'], image['subfolder'], image['type'])
+                images_output.append(image_data)
-                    images_output.append(image_data)
+        output_images[node_id] = images_output
-            output_images[node_id] = images_output
 
     return output_images
 
@@ -85,7 +84,7 @@ prompt_text = """
     "4": {
         "class_type": "CheckpointLoaderSimple",
         "inputs": {
-            "ckpt_name": "v1-5-pruned-emaonly.ckpt"
+            "ckpt_name": "v1-5-pruned-emaonly.safetensors"
         }
     },
     "5": {

script_examples/websockets_api_example_ws_images.py

@@ -81,7 +81,7 @@ prompt_text = """
     "4": {
         "class_type": "CheckpointLoaderSimple",
         "inputs": {
-            "ckpt_name": "v1-5-pruned-emaonly.ckpt"
+            "ckpt_name": "v1-5-pruned-emaonly.safetensors"
         }
     },
     "5": {

server.py

@@ -29,6 +29,7 @@ from app.frontend_management import FrontendManager
 from app.user_manager import UserManager
 from model_filemanager import download_model, DownloadModelStatus
 from typing import Optional
+from api_server.routes.internal.internal_routes import InternalRoutes
 
 class BinaryEventTypes:
     PREVIEW_IMAGE = 1
@@ -40,6 +41,21 @@ async def send_socket_catch_exception(function, message):
     except (aiohttp.ClientError, aiohttp.ClientPayloadError, ConnectionResetError) as err:
         logging.warning("send error: {}".format(err))
 
+def get_comfyui_version():
+    comfyui_version = "unknown"
+    repo_path = os.path.dirname(os.path.realpath(__file__))
+    try:
+        import pygit2
+        repo = pygit2.Repository(repo_path)
+        comfyui_version = repo.describe(describe_strategy=pygit2.GIT_DESCRIBE_TAGS)
+    except Exception:
+        try:
+            import subprocess
+            comfyui_version = subprocess.check_output(["git", "describe", "--tags"], cwd=repo_path).decode('utf-8')
+        except Exception as e:
+            logging.warning(f"Failed to get ComfyUI version: {e}")
+    return comfyui_version.strip()
+
 @web.middleware
 async def cache_control(request: web.Request, handler):
     response: web.Response = await handler(request)
@@ -72,6 +88,7 @@ class PromptServer():
         mimetypes.types_map['.js'] = 'application/javascript; charset=utf-8'
 
         self.user_manager = UserManager()
+        self.internal_routes = InternalRoutes()
         self.supports = ["custom_nodes_from_web"]
         self.prompt_queue = None
         self.loop = loop
@@ -139,6 +156,14 @@ class PromptServer():
             embeddings = folder_paths.get_filename_list("embeddings")
             return web.json_response(list(map(lambda a: os.path.splitext(a)[0], embeddings)))
 
+        @routes.get("/models/{folder}")
+        async def get_models(request):
+            folder = request.match_info.get("folder", None)
+            if not folder in folder_paths.folder_names_and_paths:
+                return web.Response(status=404)
+            files = folder_paths.get_filename_list(folder)
+            return web.json_response(files)
+
         @routes.get("/extensions")
         async def get_extensions(request):
             files = glob.glob(os.path.join(
@@ -390,16 +415,20 @@ class PromptServer():
             return web.json_response(dt["__metadata__"])
 
         @routes.get("/system_stats")
-        async def get_queue(request):
+        async def system_stats(request):
             device = comfy.model_management.get_torch_device()
             device_name = comfy.model_management.get_torch_device_name(device)
             vram_total, torch_vram_total = comfy.model_management.get_total_memory(device, torch_total_too=True)
             vram_free, torch_vram_free = comfy.model_management.get_free_memory(device, torch_free_too=True)
+
             system_stats = {
                 "system": {
                     "os": os.name,
+                    "comfyui_version": get_comfyui_version(),
                     "python_version": sys.version,
-                    "embedded_python": os.path.split(os.path.split(sys.executable)[0])[1] == "python_embeded"
+                    "pytorch_version": comfy.model_management.torch_version,
+                    "embedded_python": os.path.split(os.path.split(sys.executable)[0])[1] == "python_embeded",
+                    "argv": sys.argv
                 },
                 "devices": [
                     {
@@ -442,6 +471,11 @@ class PromptServer():
 
             if hasattr(obj_class, 'OUTPUT_TOOLTIPS'):
                 info['output_tooltips'] = obj_class.OUTPUT_TOOLTIPS
+
+            if getattr(obj_class, "DEPRECATED", False):
+                info['deprecated'] = True
+            if getattr(obj_class, "EXPERIMENTAL", False):
+                info['experimental'] = True
             return info
 
         @routes.get("/object_info")
@@ -570,7 +604,9 @@ class PromptServer():
         @routes.post("/internal/models/download")
         async def download_handler(request):
             async def report_progress(filename: str, status: DownloadModelStatus):
-                await self.send_json("download_progress", status.to_dict())
+                payload = status.to_dict()
+                payload['download_path'] = filename
+                await self.send_json("download_progress", payload)
 
             data = await request.json()
             url = data.get('url')
@@ -597,6 +633,7 @@ class PromptServer():
 
     def add_routes(self):
         self.user_manager.add_routes(self.routes)
+        self.app.add_subapp('/internal', self.internal_routes.get_app())
 
         # Prefix every route with /api for easier matching for delegation.
         # This is very useful for frontend dev server, which need to forward

tests-unit/app_test/frontend_manager_test.py

@@ -1,6 +1,7 @@
 import argparse
 import pytest
 from requests.exceptions import HTTPError
+from unittest.mock import patch
 
 from app.frontend_management import (
     FrontendManager,
@@ -83,6 +84,35 @@ def test_init_frontend_invalid_provider():
     with pytest.raises(HTTPError):
         FrontendManager.init_frontend_unsafe(version_string)
 
+@pytest.fixture
+def mock_os_functions():
+    with patch('app.frontend_management.os.makedirs') as mock_makedirs, \
+         patch('app.frontend_management.os.listdir') as mock_listdir, \
+         patch('app.frontend_management.os.rmdir') as mock_rmdir:
+        mock_listdir.return_value = []  # Simulate empty directory
+        yield mock_makedirs, mock_listdir, mock_rmdir
+
+@pytest.fixture
+def mock_download():
+    with patch('app.frontend_management.download_release_asset_zip') as mock:
+        mock.side_effect = Exception("Download failed")  # Simulate download failure
+        yield mock
+
+def test_finally_block(mock_os_functions, mock_download, mock_provider):
+    # Arrange
+    mock_makedirs, mock_listdir, mock_rmdir = mock_os_functions
+    version_string = 'test-owner/test-repo@1.0.0'
+
+    # Act & Assert
+    with pytest.raises(Exception):
+        FrontendManager.init_frontend_unsafe(version_string, mock_provider)
+
+    # Assert
+    mock_makedirs.assert_called_once()
+    mock_download.assert_called_once()
+    mock_listdir.assert_called_once()
+    mock_rmdir.assert_called_once()
+
 
 def test_parse_version_string():
     version_string = "owner/repo@1.0.0"

tests-unit/server/routes/internal_routes_test.py

@@ -0,0 +1,115 @@
+import pytest
+from aiohttp import web
+from unittest.mock import MagicMock, patch
+from api_server.routes.internal.internal_routes import InternalRoutes
+from api_server.services.file_service import FileService
+from folder_paths import models_dir, user_directory, output_directory
+
+
+@pytest.fixture
+def internal_routes():
+    return InternalRoutes()
+
+@pytest.fixture
+def aiohttp_client_factory(aiohttp_client, internal_routes):
+    async def _get_client():
+        app = internal_routes.get_app()
+        return await aiohttp_client(app)
+    return _get_client
+
+@pytest.mark.asyncio
+async def test_list_files_valid_directory(aiohttp_client_factory, internal_routes):
+    mock_file_list = [
+        {"name": "file1.txt", "path": "file1.txt", "type": "file", "size": 100},
+        {"name": "dir1", "path": "dir1", "type": "directory"}
+    ]
+    internal_routes.file_service.list_files = MagicMock(return_value=mock_file_list)
+    client = await aiohttp_client_factory()
+    resp = await client.get('/files?directory=models')
+    assert resp.status == 200
+    data = await resp.json()
+    assert 'files' in data
+    assert len(data['files']) == 2
+    assert data['files'] == mock_file_list
+
+    # Check other valid directories
+    resp = await client.get('/files?directory=user')
+    assert resp.status == 200
+    resp = await client.get('/files?directory=output')
+    assert resp.status == 200
+
+@pytest.mark.asyncio
+async def test_list_files_invalid_directory(aiohttp_client_factory, internal_routes):
+    internal_routes.file_service.list_files = MagicMock(side_effect=ValueError("Invalid directory key"))
+    client = await aiohttp_client_factory()
+    resp = await client.get('/files?directory=invalid')
+    assert resp.status == 400
+    data = await resp.json()
+    assert 'error' in data
+    assert data['error'] == "Invalid directory key"
+
+@pytest.mark.asyncio
+async def test_list_files_exception(aiohttp_client_factory, internal_routes):
+    internal_routes.file_service.list_files = MagicMock(side_effect=Exception("Unexpected error"))
+    client = await aiohttp_client_factory()
+    resp = await client.get('/files?directory=models')
+    assert resp.status == 500
+    data = await resp.json()
+    assert 'error' in data
+    assert data['error'] == "Unexpected error"
+
+@pytest.mark.asyncio
+async def test_list_files_no_directory_param(aiohttp_client_factory, internal_routes):
+    mock_file_list = []
+    internal_routes.file_service.list_files = MagicMock(return_value=mock_file_list)
+    client = await aiohttp_client_factory()
+    resp = await client.get('/files')
+    assert resp.status == 200
+    data = await resp.json()
+    assert 'files' in data
+    assert len(data['files']) == 0
+
+def test_setup_routes(internal_routes):
+    internal_routes.setup_routes()
+    routes = internal_routes.routes
+    assert any(route.method == 'GET' and str(route.path) == '/files' for route in routes)
+
+def test_get_app(internal_routes):
+    app = internal_routes.get_app()
+    assert isinstance(app, web.Application)
+    assert internal_routes._app is not None
+
+def test_get_app_reuse(internal_routes):
+    app1 = internal_routes.get_app()
+    app2 = internal_routes.get_app()
+    assert app1 is app2
+
+@pytest.mark.asyncio
+async def test_routes_added_to_app(aiohttp_client_factory, internal_routes):
+    client = await aiohttp_client_factory()
+    try:
+        resp = await client.get('/files')
+        print(f"Response received: status {resp.status}")
+    except Exception as e:
+        print(f"Exception occurred during GET request: {e}")
+        raise
+
+    assert resp.status != 404, "Route /files does not exist"
+
+@pytest.mark.asyncio
+async def test_file_service_initialization():
+    with patch('api_server.routes.internal.internal_routes.FileService') as MockFileService:
+        # Create a mock instance
+        mock_file_service_instance = MagicMock(spec=FileService)
+        MockFileService.return_value = mock_file_service_instance
+        internal_routes = InternalRoutes()
+
+        # Check if FileService was initialized with the correct parameters
+        MockFileService.assert_called_once_with({
+            "models": models_dir,
+            "user": user_directory,
+            "output": output_directory
+        })
+
+        # Verify that the file_service attribute of InternalRoutes is set
+        assert internal_routes.file_service == mock_file_service_instance

tests-unit/server/services/file_service_test.py

@@ -0,0 +1,54 @@
+import pytest
+from unittest.mock import MagicMock
+from api_server.services.file_service import FileService
+
+@pytest.fixture
+def mock_file_system_ops():
+    return MagicMock()
+
+@pytest.fixture
+def file_service(mock_file_system_ops):
+    allowed_directories = {
+        "models": "/path/to/models",
+        "user": "/path/to/user",
+        "output": "/path/to/output"
+    }
+    return FileService(allowed_directories, file_system_ops=mock_file_system_ops)
+
+def test_list_files_valid_directory(file_service, mock_file_system_ops):
+    mock_file_system_ops.walk_directory.return_value = [
+        {"name": "file1.txt", "path": "file1.txt", "type": "file", "size": 100},
+        {"name": "dir1", "path": "dir1", "type": "directory"}
+    ]
+    
+    result = file_service.list_files("models")
+    
+    assert len(result) == 2
+    assert result[0]["name"] == "file1.txt"
+    assert result[1]["name"] == "dir1"
+    mock_file_system_ops.walk_directory.assert_called_once_with("/path/to/models")
+
+def test_list_files_invalid_directory(file_service):
+    # Does not support walking directories outside of the allowed directories
+    with pytest.raises(ValueError, match="Invalid directory key"):
+        file_service.list_files("invalid_key")
+
+def test_list_files_empty_directory(file_service, mock_file_system_ops):
+    mock_file_system_ops.walk_directory.return_value = []
+    
+    result = file_service.list_files("models")
+    
+    assert len(result) == 0
+    mock_file_system_ops.walk_directory.assert_called_once_with("/path/to/models")
+
+@pytest.mark.parametrize("directory_key", ["models", "user", "output"])
+def test_list_files_all_allowed_directories(file_service, mock_file_system_ops, directory_key):
+    mock_file_system_ops.walk_directory.return_value = [
+        {"name": f"file_{directory_key}.txt", "path": f"file_{directory_key}.txt", "type": "file", "size": 100}
+    ]
+    
+    result = file_service.list_files(directory_key)
+    
+    assert len(result) == 1
+    assert result[0]["name"] == f"file_{directory_key}.txt"
+    mock_file_system_ops.walk_directory.assert_called_once_with(f"/path/to/{directory_key}")

tests-unit/server/utils/file_operations_test.py

@@ -0,0 +1,42 @@
+import pytest
+from typing import List
+from api_server.utils.file_operations import FileSystemOperations, FileSystemItem, is_file_info
+
+@pytest.fixture
+def temp_directory(tmp_path):
+    # Create a temporary directory structure
+    dir1 = tmp_path / "dir1"
+    dir2 = tmp_path / "dir2"
+    dir1.mkdir()
+    dir2.mkdir()
+    (dir1 / "file1.txt").write_text("content1")
+    (dir2 / "file2.txt").write_text("content2")
+    (tmp_path / "file3.txt").write_text("content3")
+    return tmp_path
+
+def test_walk_directory(temp_directory):
+    result: List[FileSystemItem] = FileSystemOperations.walk_directory(str(temp_directory))
+    
+    assert len(result) == 5  # 2 directories and 3 files
+    
+    files = [item for item in result if item['type'] == 'file']
+    dirs = [item for item in result if item['type'] == 'directory']
+    
+    assert len(files) == 3
+    assert len(dirs) == 2
+    
+    file_names = {file['name'] for file in files}
+    assert file_names == {'file1.txt', 'file2.txt', 'file3.txt'}
+    
+    dir_names = {dir['name'] for dir in dirs}
+    assert dir_names == {'dir1', 'dir2'}
+
+def test_walk_directory_empty(tmp_path):
+    result = FileSystemOperations.walk_directory(str(tmp_path))
+    assert len(result) == 0
+
+def test_walk_directory_file_size(temp_directory):
+    result: List[FileSystemItem] = FileSystemOperations.walk_directory(str(temp_directory))
+    files = [item for item in result if is_file_info(item)]
+    for file in files:
+        assert file['size'] > 0  # Assuming all files have some content

tests/inference/test_execution.py

@@ -95,17 +95,16 @@ class ComfyClient:
                     pass # Probably want to store this off for testing
 
         history = self.get_history(prompt_id)[prompt_id]
-        for o in history['outputs']:
+        for node_id in history['outputs']:
-            for node_id in history['outputs']:
+            node_output = history['outputs'][node_id]
-                node_output = history['outputs'][node_id]
+            result.outputs[node_id] = node_output
-                result.outputs[node_id] = node_output
+            images_output = []
-                if 'images' in node_output:
+            if 'images' in node_output:
-                    images_output = []
+                for image in node_output['images']:
-                    for image in node_output['images']:
+                    image_data = self.get_image(image['filename'], image['subfolder'], image['type'])
-                        image_data = self.get_image(image['filename'], image['subfolder'], image['type'])
+                    image_obj = Image.open(BytesIO(image_data))
-                        image_obj = Image.open(BytesIO(image_data))
+                    images_output.append(image_obj)
-                        images_output.append(image_obj)
+                node_output['image_objects'] = images_output
-                    node_output['image_objects'] = images_output
 
         return result
 
@@ -357,6 +356,25 @@ class TestExecution:
             assert 'prompt_id' in e.args[0], f"Did not get back a proper error message: {e}"
             assert e.args[0]['node_id'] == generator.id, "Error should have been on the generator node"
 
+    def test_missing_node_error(self, client: ComfyClient, builder: GraphBuilder):
+        g = builder
+        input1 = g.node("StubImage", content="BLACK", height=512, width=512, batch_size=1)
+        input2 = g.node("StubImage", id="removeme", content="WHITE", height=512, width=512, batch_size=1)
+        input3 = g.node("StubImage", content="WHITE", height=512, width=512, batch_size=1)
+        mask = g.node("StubMask", value=0.5, height=512, width=512, batch_size=1)
+        mix1 = g.node("TestLazyMixImages", image1=input1.out(0), image2=input2.out(0), mask=mask.out(0))
+        mix2 = g.node("TestLazyMixImages", image1=input1.out(0), image2=input3.out(0), mask=mask.out(0))
+        # We have multiple outputs. The first is invalid, but the second is valid
+        g.node("SaveImage", images=mix1.out(0))
+        g.node("SaveImage", images=mix2.out(0))
+        g.remove_node("removeme")
+        
+        client.run(g)
+
+        # Add back in the missing node to make sure the error doesn't break the server
+        input2 = g.node("StubImage", id="removeme", content="WHITE", height=512, width=512, batch_size=1)
+        client.run(g)
+
     def test_custom_is_changed(self, client: ComfyClient, builder: GraphBuilder):
         g = builder
         # Creating the nodes in this specific order previously caused a bug
@@ -450,8 +468,8 @@ class TestExecution:
         g = builder
         input1 = g.node("StubImage", content="BLACK", height=512, width=512, batch_size=1)
 
-        output1 = g.node("PreviewImage", images=input1.out(0))
+        output1 = g.node("SaveImage", images=input1.out(0))
-        output2 = g.node("PreviewImage", images=input1.out(0))
+        output2 = g.node("SaveImage", images=input1.out(0))
 
         result = client.run(g)
         images1 = result.get_images(output1)
@@ -459,3 +477,22 @@ class TestExecution:
         assert len(images1) == 1, "Should have 1 image"
         assert len(images2) == 1, "Should have 1 image"
 
+
+    # This tests that only constant outputs are used in the call to `IS_CHANGED`
+    def test_is_changed_with_outputs(self, client: ComfyClient, builder: GraphBuilder):
+        g = builder
+        input1 = g.node("StubConstantImage", value=0.5, height=512, width=512, batch_size=1)
+        test_node = g.node("TestIsChangedWithConstants", image=input1.out(0), value=0.5)
+
+        output = g.node("PreviewImage", images=test_node.out(0))
+
+        result = client.run(g)
+        images = result.get_images(output)
+        assert len(images) == 1, "Should have 1 image"
+        assert numpy.array(images[0]).min() == 63 and numpy.array(images[0]).max() == 63, "Image should have value 0.25"
+
+        result = client.run(g)
+        images = result.get_images(output)
+        assert len(images) == 1, "Should have 1 image"
+        assert numpy.array(images[0]).min() == 63 and numpy.array(images[0]).max() == 63, "Image should have value 0.25"
+        assert not result.did_run(test_node), "The execution should have been cached"

tests/inference/test_inference.py

@@ -109,15 +109,14 @@ class ComfyClient:
                 continue #previews are binary data
 
         history = self.get_history(prompt_id)[prompt_id]
-        for o in history['outputs']:
+        for node_id in history['outputs']:
-            for node_id in history['outputs']:
+            node_output = history['outputs'][node_id]
-                node_output = history['outputs'][node_id]
+            images_output = []
-                if 'images' in node_output:
+            if 'images' in node_output:
-                    images_output = []
+                for image in node_output['images']:
-                    for image in node_output['images']:
+                    image_data = self.get_image(image['filename'], image['subfolder'], image['type'])
-                        image_data = self.get_image(image['filename'], image['subfolder'], image['type'])
+                    images_output.append(image_data)
-                        images_output.append(image_data)
+            output_images[node_id] = images_output
-                output_images[node_id] = images_output
 
         return output_images
 

tests/inference/testing_nodes/testing-pack/specific_tests.py

@@ -95,6 +95,31 @@ class TestCustomIsChanged:
         else:
             return False
 
+class TestIsChangedWithConstants:
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "image": ("IMAGE",),
+                "value": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 10.0}),
+            },
+        }
+
+    RETURN_TYPES = ("IMAGE",)
+    FUNCTION = "custom_is_changed"
+
+    CATEGORY = "Testing/Nodes"
+
+    def custom_is_changed(self, image, value):
+        return (image * value,)
+    
+    @classmethod
+    def IS_CHANGED(cls, image, value):
+        if image is None:
+            return value
+        else:
+            return image.mean().item() * value
+
 class TestCustomValidation1:
     @classmethod
     def INPUT_TYPES(cls):
@@ -312,6 +337,7 @@ TEST_NODE_CLASS_MAPPINGS = {
     "TestLazyMixImages": TestLazyMixImages,
     "TestVariadicAverage": TestVariadicAverage,
     "TestCustomIsChanged": TestCustomIsChanged,
+    "TestIsChangedWithConstants": TestIsChangedWithConstants,
     "TestCustomValidation1": TestCustomValidation1,
     "TestCustomValidation2": TestCustomValidation2,
     "TestCustomValidation3": TestCustomValidation3,
@@ -325,6 +351,7 @@ TEST_NODE_DISPLAY_NAME_MAPPINGS = {
     "TestLazyMixImages": "Lazy Mix Images",
     "TestVariadicAverage": "Variadic Average",
     "TestCustomIsChanged": "Custom IsChanged",
+    "TestIsChangedWithConstants": "IsChanged With Constants",
     "TestCustomValidation1": "Custom Validation 1",
     "TestCustomValidation2": "Custom Validation 2",
     "TestCustomValidation3": "Custom Validation 3",

tests/inference/testing_nodes/testing-pack/stubs.py

@@ -28,6 +28,28 @@ class StubImage:
         elif content == "NOISE":
             return (torch.rand(batch_size, height, width, 3),)
 
+class StubConstantImage:
+    def __init__(self):
+        pass
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "value": ("FLOAT", {"default": 0.5, "min": 0.0, "max": 1.0, "step": 0.01}),
+                "height": ("INT", {"default": 512, "min": 1, "max": 1024 ** 3, "step": 1}),
+                "width": ("INT", {"default": 512, "min": 1, "max": 4096 ** 3, "step": 1}),
+                "batch_size": ("INT", {"default": 1, "min": 1, "max": 1024 ** 3, "step": 1}),
+            },
+        }
+
+    RETURN_TYPES = ("IMAGE",)
+    FUNCTION = "stub_constant_image"
+
+    CATEGORY = "Testing/Stub Nodes"
+
+    def stub_constant_image(self, value, height, width, batch_size):
+        return (torch.ones(batch_size, height, width, 3) * value,)
+
 class StubMask:
     def __init__(self):
         pass
@@ -93,12 +115,14 @@ class StubFloat:
 
 TEST_STUB_NODE_CLASS_MAPPINGS = {
     "StubImage": StubImage,
+    "StubConstantImage": StubConstantImage,
     "StubMask": StubMask,
     "StubInt": StubInt,
     "StubFloat": StubFloat,
 }
 TEST_STUB_NODE_DISPLAY_NAME_MAPPINGS = {
     "StubImage": "Stub Image",
+    "StubConstantImage": "Stub Constant Image",
     "StubMask": "Stub Mask",
     "StubInt": "Stub Int",
     "StubFloat": "Stub Float",

web/assets/userSelection-BifVfRyx.js

@@ -1,138 +0,0 @@
-var __async = (__this, __arguments, generator) => {
-  return new Promise((resolve, reject) => {
-    var fulfilled = (value) => {
-      try {
-        step(generator.next(value));
-      } catch (e) {
-        reject(e);
-      }
-    };
-    var rejected = (value) => {
-      try {
-        step(generator.throw(value));
-      } catch (e) {
-        reject(e);
-      }
-    };
-    var step = (x) => x.done ? resolve(x.value) : Promise.resolve(x.value).then(fulfilled, rejected);
-    step((generator = generator.apply(__this, __arguments)).next());
-  });
-};
-import { j as createSpinner, g as api, $ as $el } from "./index-DIiqwEjy.js";
-class UserSelectionScreen {
-  show(users, user) {
-    return __async(this, null, function* () {
-      const userSelection = document.getElementById("comfy-user-selection");
-      userSelection.style.display = "";
-      return new Promise((resolve) => {
-        const input = userSelection.getElementsByTagName("input")[0];
-        const select = userSelection.getElementsByTagName("select")[0];
-        const inputSection = input.closest("section");
-        const selectSection = select.closest("section");
-        const form = userSelection.getElementsByTagName("form")[0];
-        const error = userSelection.getElementsByClassName("comfy-user-error")[0];
-        const button = userSelection.getElementsByClassName(
-          "comfy-user-button-next"
-        )[0];
-        let inputActive = null;
-        input.addEventListener("focus", () => {
-          inputSection.classList.add("selected");
-          selectSection.classList.remove("selected");
-          inputActive = true;
-        });
-        select.addEventListener("focus", () => {
-          inputSection.classList.remove("selected");
-          selectSection.classList.add("selected");
-          inputActive = false;
-          select.style.color = "";
-        });
-        select.addEventListener("blur", () => {
-          if (!select.value) {
-            select.style.color = "var(--descrip-text)";
-          }
-        });
-        form.addEventListener("submit", (e) => __async(this, null, function* () {
-          var _a, _b, _c;
-          e.preventDefault();
-          if (inputActive == null) {
-            error.textContent = "Please enter a username or select an existing user.";
-          } else if (inputActive) {
-            const username = input.value.trim();
-            if (!username) {
-              error.textContent = "Please enter a username.";
-              return;
-            }
-            input.disabled = select.disabled = // @ts-expect-error
-            input.readonly = // @ts-expect-error
-            select.readonly = true;
-            const spinner = createSpinner();
-            button.prepend(spinner);
-            try {
-              const resp = yield api.createUser(username);
-              if (resp.status >= 300) {
-                let message = "Error creating user: " + resp.status + " " + resp.statusText;
-                try {
-                  const res = yield resp.json();
-                  if (res.error) {
-                    message = res.error;
-                  }
-                } catch (error2) {
-                }
-                throw new Error(message);
-              }
-              resolve({ username, userId: yield resp.json(), created: true });
-            } catch (err) {
-              spinner.remove();
-              error.textContent = (_c = (_b = (_a = err.message) != null ? _a : err.statusText) != null ? _b : err) != null ? _c : "An unknown error occurred.";
-              input.disabled = select.disabled = // @ts-expect-error
-              input.readonly = // @ts-expect-error
-              select.readonly = false;
-              return;
-            }
-          } else if (!select.value) {
-            error.textContent = "Please select an existing user.";
-            return;
-          } else {
-            resolve({
-              username: users[select.value],
-              userId: select.value,
-              created: false
-            });
-          }
-        }));
-        if (user) {
-          const name = localStorage["Comfy.userName"];
-          if (name) {
-            input.value = name;
-          }
-        }
-        if (input.value) {
-          input.focus();
-        }
-        const userIds = Object.keys(users != null ? users : {});
-        if (userIds.length) {
-          for (const u of userIds) {
-            $el("option", { textContent: users[u], value: u, parent: select });
-          }
-          select.style.color = "var(--descrip-text)";
-          if (select.value) {
-            select.focus();
-          }
-        } else {
-          userSelection.classList.add("no-users");
-          input.focus();
-        }
-      }).then((r) => {
-        userSelection.remove();
-        return r;
-      });
-    });
-  }
-}
-window.comfyAPI = window.comfyAPI || {};
-window.comfyAPI.userSelection = window.comfyAPI.userSelection || {};
-window.comfyAPI.userSelection.UserSelectionScreen = UserSelectionScreen;
-export {
-  UserSelectionScreen
-};
-//# sourceMappingURL=userSelection-BifVfRyx.js.map

web/assets/userSelection-CyXKCVy3.js

@@ -0,0 +1,120 @@
+var __defProp = Object.defineProperty;
+var __name = (target, value) => __defProp(target, "name", { value, configurable: true });
+import { j as createSpinner, h as api, $ as $el } from "./index-CI3N807S.js";
+class UserSelectionScreen {
+  static {
+    __name(this, "UserSelectionScreen");
+  }
+  async show(users, user) {
+    const userSelection = document.getElementById("comfy-user-selection");
+    userSelection.style.display = "";
+    return new Promise((resolve) => {
+      const input = userSelection.getElementsByTagName("input")[0];
+      const select = userSelection.getElementsByTagName("select")[0];
+      const inputSection = input.closest("section");
+      const selectSection = select.closest("section");
+      const form = userSelection.getElementsByTagName("form")[0];
+      const error = userSelection.getElementsByClassName("comfy-user-error")[0];
+      const button = userSelection.getElementsByClassName(
+        "comfy-user-button-next"
+      )[0];
+      let inputActive = null;
+      input.addEventListener("focus", () => {
+        inputSection.classList.add("selected");
+        selectSection.classList.remove("selected");
+        inputActive = true;
+      });
+      select.addEventListener("focus", () => {
+        inputSection.classList.remove("selected");
+        selectSection.classList.add("selected");
+        inputActive = false;
+        select.style.color = "";
+      });
+      select.addEventListener("blur", () => {
+        if (!select.value) {
+          select.style.color = "var(--descrip-text)";
+        }
+      });
+      form.addEventListener("submit", async (e) => {
+        e.preventDefault();
+        if (inputActive == null) {
+          error.textContent = "Please enter a username or select an existing user.";
+        } else if (inputActive) {
+          const username = input.value.trim();
+          if (!username) {
+            error.textContent = "Please enter a username.";
+            return;
+          }
+          input.disabled = select.disabled = // @ts-expect-error
+          input.readonly = // @ts-expect-error
+          select.readonly = true;
+          const spinner = createSpinner();
+          button.prepend(spinner);
+          try {
+            const resp = await api.createUser(username);
+            if (resp.status >= 300) {
+              let message = "Error creating user: " + resp.status + " " + resp.statusText;
+              try {
+                const res = await resp.json();
+                if (res.error) {
+                  message = res.error;
+                }
+              } catch (error2) {
+              }
+              throw new Error(message);
+            }
+            resolve({ username, userId: await resp.json(), created: true });
+          } catch (err) {
+            spinner.remove();
+            error.textContent = err.message ?? err.statusText ?? err ?? "An unknown error occurred.";
+            input.disabled = select.disabled = // @ts-expect-error
+            input.readonly = // @ts-expect-error
+            select.readonly = false;
+            return;
+          }
+        } else if (!select.value) {
+          error.textContent = "Please select an existing user.";
+          return;
+        } else {
+          resolve({
+            username: users[select.value],
+            userId: select.value,
+            created: false
+          });
+        }
+      });
+      if (user) {
+        const name = localStorage["Comfy.userName"];
+        if (name) {
+          input.value = name;
+        }
+      }
+      if (input.value) {
+        input.focus();
+      }
+      const userIds = Object.keys(users ?? {});
+      if (userIds.length) {
+        for (const u of userIds) {
+          $el("option", { textContent: users[u], value: u, parent: select });
+        }
+        select.style.color = "var(--descrip-text)";
+        if (select.value) {
+          select.focus();
+        }
+      } else {
+        userSelection.classList.add("no-users");
+        input.focus();
+      }
+    }).then((r) => {
+      userSelection.remove();
+      return r;
+    });
+  }
+}
+window.comfyAPI = window.comfyAPI || {};
+window.comfyAPI.userSelection = window.comfyAPI.userSelection || {};
+window.comfyAPI.userSelection.UserSelectionScreen = UserSelectionScreen;
+export {
+  UserSelectionScreen
+};
+//# sourceMappingURL=userSelection-CyXKCVy3.js.map

web/index.html

@@ -12,10 +12,10 @@
 				font-family: 'Roboto Mono', 'Noto Color Emoji';
 			}
 		</style> -->
-		<link rel="stylesheet" type="text/css" href="/user.css" />
+		<link rel="stylesheet" type="text/css" href="user.css" />
-		<link rel="stylesheet" type="text/css" href="/materialdesignicons.min.css" />
+		<link rel="stylesheet" type="text/css" href="materialdesignicons.min.css" />
-		<script type="module" crossorigin src="/assets/index-DIiqwEjy.js"></script>
+		<script type="module" crossorigin src="./assets/index-CI3N807S.js"></script>
-		<link rel="stylesheet" crossorigin href="/assets/index-CwWW6Xjy.css">
+		<link rel="stylesheet" crossorigin href="./assets/index-_5czGnTA.css">
 	</head>
 	<body class="litegraph">
 		<div id="vue-app"></div>