From f9618b7e4b5a5e022b587394e277cdf6e96ec574 Mon Sep 17 00:00:00 2001 From: Sasbom Date: Wed, 5 Nov 2025 07:40:32 +0100 Subject: [PATCH 1/9] Add QOL feature for changing the custom nodes folder location through cli args or a COMFYUI_CUSTOM_NODES_DIR environment variable. --- app/custom_node_manager.py | 16 ++++++++++---- comfy/cli_args.py | 2 ++ folder_paths.py | 11 +++++++--- main.py | 45 ++++++++++++++++++++++++-------------- 4 files changed, 50 insertions(+), 24 deletions(-) diff --git a/app/custom_node_manager.py b/app/custom_node_manager.py index 281febca9523..f408dcb1c89e 100644 --- a/app/custom_node_manager.py +++ b/app/custom_node_manager.py @@ -1,15 +1,15 @@ from __future__ import annotations -import os -import folder_paths import glob -from aiohttp import web import json import logging +import os from functools import lru_cache -from utils.json_util import merge_json_recursive +from aiohttp import web +import folder_paths +from utils.json_util import merge_json_recursive # Extra locale files to load into main.json EXTRA_LOCALE_FILES = [ @@ -32,6 +32,14 @@ def safe_load_json_file(file_path: str) -> dict: class CustomNodeManager: + """Manages and resolves folders for custom nodes. + + The custom_nodes folder location can be overridden using either + by passing in a folder through client argument --custom-nodes-directory, + or by setting COMFYUI_CUSTOM_NODES_DIR as an environment variable. + + The cli argument takes prescedence if passed in. + """ @lru_cache(maxsize=1) def build_translations(self): """Load all custom nodes translations during initialization. Translations are diff --git a/comfy/cli_args.py b/comfy/cli_args.py index 3947e62a8f00..0729d436b07b 100644 --- a/comfy/cli_args.py +++ b/comfy/cli_args.py @@ -1,6 +1,7 @@ import argparse import enum import os + import comfy.options @@ -47,6 +48,7 @@ def __call__(self, parser, namespace, values, option_string=None): parser.add_argument("--output-directory", type=str, default=None, help="Set the ComfyUI output directory. Overrides --base-directory.") parser.add_argument("--temp-directory", type=str, default=None, help="Set the ComfyUI temp directory (default is in the ComfyUI directory). Overrides --base-directory.") parser.add_argument("--input-directory", type=str, default=None, help="Set the ComfyUI input directory. Overrides --base-directory.") +parser.add_argument("--custom-nodes-directory", type=str, default=None, help="Set the ComfyUI custom_nodes directory. Overrides --base-directory and environment variable COMFYUI_CUSTOM_NODES_DIR.") parser.add_argument("--auto-launch", action="store_true", help="Automatically launch ComfyUI in the default browser.") parser.add_argument("--disable-auto-launch", action="store_true", help="Disable auto launching the browser.") parser.add_argument("--cuda-device", type=int, default=None, metavar="DEVICE_ID", help="Set the id of the cuda device this instance will use. All other devices will not be visible.") diff --git a/folder_paths.py b/folder_paths.py index f110d832bb23..8aa2b7be9e2a 100644 --- a/folder_paths.py +++ b/folder_paths.py @@ -1,11 +1,11 @@ from __future__ import annotations +import logging +import mimetypes import os import time -import mimetypes -import logging -from typing import Literal, List from collections.abc import Collection +from typing import List, Literal from comfy.cli_args import args @@ -115,6 +115,11 @@ def set_input_directory(input_dir: str) -> None: global input_directory input_directory = input_dir +def set_custom_nodes_directory(custom_nodes_dir: str) -> None: + #NOTE: this function doesn't need a "get_custom_nodes_directory", we have "get_folder_paths". + global folder_names_and_paths + folder_names_and_paths["custom_nodes"] = ([custom_nodes_dir], set()) + def get_output_directory() -> str: global output_directory return output_directory diff --git a/main.py b/main.py index e1b0f1620bbc..e415879bca5d 100644 --- a/main.py +++ b/main.py @@ -1,19 +1,21 @@ import comfy.options + comfy.options.enable_args_parsing() -import os import importlib.util -import folder_paths -import time -from comfy.cli_args import args -from app.logger import setup_logger import itertools -import utils.extra_config import logging +import os import sys +import time + +import folder_paths +import utils.extra_config +from app.logger import setup_logger +from comfy.cli_args import args +from comfy_api import feature_flags from comfy_execution.progress import get_progress_state from comfy_execution.utils import get_executing_context -from comfy_api import feature_flags if __name__ == "__main__": #NOTE: These do not do anything on core ComfyUI, they are for custom nodes. @@ -32,7 +34,7 @@ def apply_custom_paths(): for config_path in itertools.chain(*args.extra_model_paths_config): utils.extra_config.load_extra_path_config(config_path) - # --output-directory, --input-directory, --user-directory + # --output-directory, --input-directory, --user-directory, --custom-nodes-directory if args.output_directory: output_dir = os.path.abspath(args.output_directory) logging.info(f"Setting output directory to: {output_dir}") @@ -56,6 +58,16 @@ def apply_custom_paths(): logging.info(f"Setting user directory to: {user_dir}") folder_paths.set_user_directory(user_dir) + #NOTE: hierarchy of resolution should be cli args first, then environment. + if args.custom_nodes_directory: + custom_nodes_dir = os.path.abspath(args.custom_nodes_dir) + logging.info(f"Setting custom nodes directory to: {custom_nodes_dir}") + folder_paths.set_custom_nodes_directory(custom_nodes_dir) + elif os.environ.get("COMFYUI_CUSTOM_NODES_DIR", None): + custom_nodes_dir = os.path.abspath(os.environ.get("COMFYUI_CUSTOM_NODES_DIR")) + logging.info(f"Setting custom nodes directory to: {custom_nodes_dir} from environment variable COMFYUI_CUSTOM_NODES_DIR") + folder_paths.set_custom_nodes_directory(custom_nodes_dir) + def execute_prestartup_script(): if args.disable_all_custom_nodes and len(args.whitelist_custom_nodes) == 0: @@ -106,10 +118,9 @@ def execute_script(script_path): # Main code import asyncio +import gc import shutil import threading -import gc - if os.name == "nt": os.environ['MIMALLOC_PURGE_DELAY'] = '0' @@ -144,16 +155,16 @@ def execute_script(script_path): if 'torch' in sys.modules: logging.warning("WARNING: Potential Error in code: Torch already imported, torch should never be imported before this point.") +import app.logger +import comfy.model_management import comfy.utils - +import comfyui_version import execution +import hook_breaker_ac10a0 +import nodes import server from protocol import BinaryEventTypes -import nodes -import comfy.model_management -import comfyui_version -import app.logger -import hook_breaker_ac10a0 + def cuda_malloc_warning(): device = comfy.model_management.get_torch_device() @@ -293,7 +304,7 @@ def cleanup_temp(): def setup_database(): try: - from app.database.db import init_db, dependencies_available + from app.database.db import dependencies_available, init_db if dependencies_available(): init_db() except Exception as e: From 90b6ed7d095e64b08190c8bc7533718093d0f20d Mon Sep 17 00:00:00 2001 From: Sas van Gulik <87427129+Sasbom@users.noreply.github.com> Date: Wed, 5 Nov 2025 09:29:42 +0100 Subject: [PATCH 2/9] bugfix: fix typo in apply_directory for custom_nodes_directory --- main.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/main.py b/main.py index e415879bca5d..7dd4fd2c70b2 100644 --- a/main.py +++ b/main.py @@ -60,7 +60,7 @@ def apply_custom_paths(): #NOTE: hierarchy of resolution should be cli args first, then environment. if args.custom_nodes_directory: - custom_nodes_dir = os.path.abspath(args.custom_nodes_dir) + custom_nodes_dir = os.path.abspath(args.custom_nodes_directory) logging.info(f"Setting custom nodes directory to: {custom_nodes_dir}") folder_paths.set_custom_nodes_directory(custom_nodes_dir) elif os.environ.get("COMFYUI_CUSTOM_NODES_DIR", None): From 764d6c43f3f8dfed62f2040dbba1a32ec0c9a2ad Mon Sep 17 00:00:00 2001 From: Sasbom Date: Wed, 5 Nov 2025 11:16:59 +0100 Subject: [PATCH 3/9] allow for PATH style ';' delimited custom_node directories. --- folder_paths.py | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/folder_paths.py b/folder_paths.py index 8aa2b7be9e2a..d98c8209d97b 100644 --- a/folder_paths.py +++ b/folder_paths.py @@ -117,8 +117,9 @@ def set_input_directory(input_dir: str) -> None: def set_custom_nodes_directory(custom_nodes_dir: str) -> None: #NOTE: this function doesn't need a "get_custom_nodes_directory", we have "get_folder_paths". + # PATH style ';' delimiters allow for more folders. global folder_names_and_paths - folder_names_and_paths["custom_nodes"] = ([custom_nodes_dir], set()) + folder_names_and_paths["custom_nodes"] = (custom_nodes_dir.split(";"), set()) def get_output_directory() -> str: global output_directory From 4a010caa0dd8c352671196e85d8649fbc732500c Mon Sep 17 00:00:00 2001 From: Sasbom Date: Wed, 5 Nov 2025 11:26:11 +0100 Subject: [PATCH 4/9] change delimiter type for seperate folders per platform. --- folder_paths.py | 11 +++++++++-- 1 file changed, 9 insertions(+), 2 deletions(-) diff --git a/folder_paths.py b/folder_paths.py index d98c8209d97b..2373e0ad08a0 100644 --- a/folder_paths.py +++ b/folder_paths.py @@ -3,6 +3,7 @@ import logging import mimetypes import os +import sys import time from collections.abc import Collection from typing import List, Literal @@ -117,9 +118,15 @@ def set_input_directory(input_dir: str) -> None: def set_custom_nodes_directory(custom_nodes_dir: str) -> None: #NOTE: this function doesn't need a "get_custom_nodes_directory", we have "get_folder_paths". - # PATH style ';' delimiters allow for more folders. + # PATH style ';' / ':' (platform dependant) delimiters allow for more folders. global folder_names_and_paths - folder_names_and_paths["custom_nodes"] = (custom_nodes_dir.split(";"), set()) + + delimiter = ":" + if sys.platform in ("win32", "cygwin"): + delimiter = ";" + + directories = custom_nodes_dir.split(delimiter) + folder_names_and_paths["custom_nodes"] = (directories, set()) def get_output_directory() -> str: global output_directory From 60163d5643e04f422775530d9a3752fe87375181 Mon Sep 17 00:00:00 2001 From: Sas van Gulik <87427129+Sasbom@users.noreply.github.com> Date: Thu, 6 Nov 2025 10:12:28 +0100 Subject: [PATCH 5/9] Remove environment variable. Removed environment variable fallback for custom nodes directory. --- main.py | 5 ----- 1 file changed, 5 deletions(-) diff --git a/main.py b/main.py index 7dd4fd2c70b2..16813e154205 100644 --- a/main.py +++ b/main.py @@ -58,15 +58,10 @@ def apply_custom_paths(): logging.info(f"Setting user directory to: {user_dir}") folder_paths.set_user_directory(user_dir) - #NOTE: hierarchy of resolution should be cli args first, then environment. if args.custom_nodes_directory: custom_nodes_dir = os.path.abspath(args.custom_nodes_directory) logging.info(f"Setting custom nodes directory to: {custom_nodes_dir}") folder_paths.set_custom_nodes_directory(custom_nodes_dir) - elif os.environ.get("COMFYUI_CUSTOM_NODES_DIR", None): - custom_nodes_dir = os.path.abspath(os.environ.get("COMFYUI_CUSTOM_NODES_DIR")) - logging.info(f"Setting custom nodes directory to: {custom_nodes_dir} from environment variable COMFYUI_CUSTOM_NODES_DIR") - folder_paths.set_custom_nodes_directory(custom_nodes_dir) def execute_prestartup_script(): From 41b875c0ed765e0219119bdd4653c9aad5283b82 Mon Sep 17 00:00:00 2001 From: Sas van Gulik <87427129+Sasbom@users.noreply.github.com> Date: Thu, 6 Nov 2025 10:20:20 +0100 Subject: [PATCH 6/9] Update documentation for custom nodes directory Clarified documentation on custom nodes directory argument, removed documentation on environment variable --- app/custom_node_manager.py | 9 +++++---- 1 file changed, 5 insertions(+), 4 deletions(-) diff --git a/app/custom_node_manager.py b/app/custom_node_manager.py index f408dcb1c89e..accae616befa 100644 --- a/app/custom_node_manager.py +++ b/app/custom_node_manager.py @@ -34,11 +34,12 @@ def safe_load_json_file(file_path: str) -> dict: class CustomNodeManager: """Manages and resolves folders for custom nodes. - The custom_nodes folder location can be overridden using either - by passing in a folder through client argument --custom-nodes-directory, - or by setting COMFYUI_CUSTOM_NODES_DIR as an environment variable. + The custom_nodes folder location can be overridden + by passing in a folder argument --custom-nodes-directory. - The cli argument takes prescedence if passed in. + The argument can take appropriately delimited path entries, e.g. + `... --custom-nodes-directory /path/to/custom1:/path/to/custom2` for linux / osx (:) + `... --custom-nodes-directory "X:\\path\\to\\custom1;X:\\path\\to\\custom2"` for windows (;) """ @lru_cache(maxsize=1) def build_translations(self): From a0b6979f07a030c944c6496d2e7885eb2cf17118 Mon Sep 17 00:00:00 2001 From: Sasbom Date: Mon, 17 Nov 2025 06:06:23 +0100 Subject: [PATCH 7/9] revert import reordering --- app/custom_node_manager.py | 17 ++++------------- comfy/cli_args.py | 1 - folder_paths.py | 10 ++++------ main.py | 16 +++++++--------- 4 files changed, 15 insertions(+), 29 deletions(-) diff --git a/app/custom_node_manager.py b/app/custom_node_manager.py index accae616befa..281febca9523 100644 --- a/app/custom_node_manager.py +++ b/app/custom_node_manager.py @@ -1,16 +1,16 @@ from __future__ import annotations +import os +import folder_paths import glob +from aiohttp import web import json import logging -import os from functools import lru_cache -from aiohttp import web - -import folder_paths from utils.json_util import merge_json_recursive + # Extra locale files to load into main.json EXTRA_LOCALE_FILES = [ "nodeDefs.json", @@ -32,15 +32,6 @@ def safe_load_json_file(file_path: str) -> dict: class CustomNodeManager: - """Manages and resolves folders for custom nodes. - - The custom_nodes folder location can be overridden - by passing in a folder argument --custom-nodes-directory. - - The argument can take appropriately delimited path entries, e.g. - `... --custom-nodes-directory /path/to/custom1:/path/to/custom2` for linux / osx (:) - `... --custom-nodes-directory "X:\\path\\to\\custom1;X:\\path\\to\\custom2"` for windows (;) - """ @lru_cache(maxsize=1) def build_translations(self): """Load all custom nodes translations during initialization. Translations are diff --git a/comfy/cli_args.py b/comfy/cli_args.py index a159cfda8c9f..d86d74267f48 100644 --- a/comfy/cli_args.py +++ b/comfy/cli_args.py @@ -1,7 +1,6 @@ import argparse import enum import os - import comfy.options diff --git a/folder_paths.py b/folder_paths.py index 2373e0ad08a0..98b7d340c609 100644 --- a/folder_paths.py +++ b/folder_paths.py @@ -1,12 +1,12 @@ from __future__ import annotations -import logging -import mimetypes import os -import sys import time +import mimetypes +import logging +import sys +from typing import Literal, List from collections.abc import Collection -from typing import List, Literal from comfy.cli_args import args @@ -117,8 +117,6 @@ def set_input_directory(input_dir: str) -> None: input_directory = input_dir def set_custom_nodes_directory(custom_nodes_dir: str) -> None: - #NOTE: this function doesn't need a "get_custom_nodes_directory", we have "get_folder_paths". - # PATH style ';' / ':' (platform dependant) delimiters allow for more folders. global folder_names_and_paths delimiter = ":" diff --git a/main.py b/main.py index 16813e154205..1b8b1f638e5b 100644 --- a/main.py +++ b/main.py @@ -1,21 +1,19 @@ import comfy.options - comfy.options.enable_args_parsing() +import os import importlib.util +import folder_paths +import time +from comfy.cli_args import args +from app.logger import setup_logger import itertools +import utils.extra_config import logging -import os import sys -import time - -import folder_paths -import utils.extra_config -from app.logger import setup_logger -from comfy.cli_args import args -from comfy_api import feature_flags from comfy_execution.progress import get_progress_state from comfy_execution.utils import get_executing_context +from comfy_api import feature_flags if __name__ == "__main__": #NOTE: These do not do anything on core ComfyUI, they are for custom nodes. From 1ec230e65c6c8bca70631993d2c5c0fabf13d401 Mon Sep 17 00:00:00 2001 From: Sasbom Date: Mon, 17 Nov 2025 06:09:21 +0100 Subject: [PATCH 8/9] revert imports pt 2 --- main.py | 13 +++++++------ 1 file changed, 7 insertions(+), 6 deletions(-) diff --git a/main.py b/main.py index 1b8b1f638e5b..a6b241da48f7 100644 --- a/main.py +++ b/main.py @@ -111,9 +111,9 @@ def execute_script(script_path): # Main code import asyncio -import gc import shutil import threading +import gc if os.name == "nt": os.environ['MIMALLOC_PURGE_DELAY'] = '0' @@ -148,15 +148,16 @@ def execute_script(script_path): if 'torch' in sys.modules: logging.warning("WARNING: Potential Error in code: Torch already imported, torch should never be imported before this point.") -import app.logger -import comfy.model_management import comfy.utils -import comfyui_version + import execution -import hook_breaker_ac10a0 -import nodes import server from protocol import BinaryEventTypes +import nodes +import comfy.model_management +import comfyui_version +import app.logger +import hook_breaker_ac10a0 def cuda_malloc_warning(): From 0ef5557d6ad874c348c86fdc0fca3b2a7d84f90a Mon Sep 17 00:00:00 2001 From: Sasbom Date: Wed, 5 Nov 2025 07:40:32 +0100 Subject: [PATCH 9/9] Add QOL feature for changing the custom nodes folder location through cli args. bugfix: fix typo in apply_directory for custom_nodes_directory allow for PATH style ';' delimited custom_node directories. change delimiter type for seperate folders per platform. feat(API-nodes): move Rodin3D nodes to new client; removed old api client.py (#10645) Fix qwen controlnet regression. (#10657) Enable pinned memory by default on Nvidia. (#10656) Removed the --fast pinned_memory flag. You can use --disable-pinned-memory to disable it. Please report if it causes any issues. Pinned mem also seems to work on AMD. (#10658) Remove environment variable. Removed environment variable fallback for custom nodes directory. Update documentation for custom nodes directory Clarified documentation on custom nodes directory argument, removed documentation on environment variable Clarify release cycle. (#10667) Tell users they need to upload their logs in bug reports. (#10671) mm: guard against double pin and unpin explicitly (#10672) As commented, if you let cuda be the one to detect double pin/unpinning it actually creates an asyc GPU error. Only unpin tensor if it was pinned by ComfyUI (#10677) Make ScaleROPE node work on Flux. (#10686) Add logging for model unloading. (#10692) Unload weights if vram usage goes up between runs. (#10690) ops: Put weight cast on the offload stream (#10697) This needs to be on the offload stream. This reproduced a black screen with low resolution images on a slow bus when using FP8. Update CI workflow to remove dead macOS runner. (#10704) * Update CI workflow to remove dead macOS runner. * revert * revert Don't pin tensor if not a torch.nn.parameter.Parameter (#10718) Update README.md for Intel Arc GPU installation, remove IPEX (#10729) IPEX is no longer needed for Intel Arc GPUs. Removing instruction to setup ipex. mm/mp: always unload re-used but modified models (#10724) The partial unloader path in model re-use flow skips straight to the actual unload without any check of the patching UUID. This means that if you do an upscale flow with a model patch on an existing model, it will not apply your patchings. Fix by delaying the partial_unload until after the uuid checks. This is done by making partial_unload a model of partial_load where extra_mem is -ve. qwen: reduce VRAM usage (#10725) Clean up a bunch of stacked and no-longer-needed tensors on the QWEN VRAM peak (currently FFN). With this I go from OOMing at B=37x1328x1328 to being able to succesfully run B=47 (RTX5090). Update Python 3.14 compatibility notes in README (#10730) Quantized Ops fixes (#10715) * offload support, bug fixes, remove mixins * add readme add PR template for API-Nodes (#10736) feat: add create_time dict to prompt field in /history and /queue (#10741) flux: reduce VRAM usage (#10737) Cleanup a bunch of stack tensors on Flux. This take me from B=19 to B=22 for 1600x1600 on RTX5090. Better instructions for the portable. (#10743) Use same code for chroma and flux blocks so that optimizations are shared. (#10746) Fix custom nodes import error. (#10747) This should fix the import errors but will break if the custom nodes actually try to use the class. revert import reordering revert imports pt 2 Add left padding support to tokenizers. (#10753) chore(api-nodes): mark OpenAIDalle2 and OpenAIDalle3 nodes as deprecated (#10757) Revert "chore(api-nodes): mark OpenAIDalle2 and OpenAIDalle3 nodes as deprecated (#10757)" (#10759) This reverts commit 9a0238256873711bd38ce0e0b1d15a617a1ee454. Change ROCm nightly install command to 7.1 (#10764) --- .github/ISSUE_TEMPLATE/bug-report.yml | 8 +- .github/PULL_REQUEST_TEMPLATE/api-node.md | 21 + .github/workflows/api-node-template.yml | 58 ++ .github/workflows/test-ci.yml | 20 +- QUANTIZATION.md | 168 +++ README.md | 17 +- comfy/cli_args.py | 4 +- comfy/ldm/chroma/layers.py | 124 +-- comfy/ldm/chroma/model.py | 7 +- comfy/ldm/chroma_radiance/model.py | 7 +- comfy/ldm/flux/layers.py | 64 +- comfy/ldm/flux/model.py | 22 +- comfy/ldm/qwen_image/controlnet.py | 2 +- comfy/ldm/qwen_image/model.py | 20 +- comfy/model_management.py | 47 +- comfy/model_patcher.py | 24 +- comfy/ops.py | 41 +- comfy/quant_ops.py | 39 +- comfy/sd1_clip.py | 17 +- comfy_api_nodes/apis/PixverseController.py | 17 - comfy_api_nodes/apis/PixverseDto.py | 57 - comfy_api_nodes/apis/client.py | 981 ------------------ comfy_api_nodes/nodes_rodin.py | 188 ++-- comfy_api_nodes/util/client.py | 4 +- comfy_api_nodes/util/download_helpers.py | 2 +- .../{apis => util}/request_logger.py | 4 +- comfy_api_nodes/util/upload_helpers.py | 2 +- folder_paths.py | 11 + main.py | 11 +- server.py | 2 + 30 files changed, 572 insertions(+), 1417 deletions(-) create mode 100644 .github/PULL_REQUEST_TEMPLATE/api-node.md create mode 100644 .github/workflows/api-node-template.yml create mode 100644 QUANTIZATION.md delete mode 100644 comfy_api_nodes/apis/PixverseController.py delete mode 100644 comfy_api_nodes/apis/PixverseDto.py delete mode 100644 comfy_api_nodes/apis/client.py rename comfy_api_nodes/{apis => util}/request_logger.py (100%) diff --git a/.github/ISSUE_TEMPLATE/bug-report.yml b/.github/ISSUE_TEMPLATE/bug-report.yml index 3cf2717b75fa..6556677e0e6b 100644 --- a/.github/ISSUE_TEMPLATE/bug-report.yml +++ b/.github/ISSUE_TEMPLATE/bug-report.yml @@ -8,13 +8,15 @@ body: Before submitting a **Bug Report**, please ensure the following: - **1:** You are running the latest version of ComfyUI. - - **2:** You have looked at the existing bug reports and made sure this isn't already reported. + - **2:** You have your ComfyUI logs and relevant workflow on hand and will post them in this bug report. - **3:** You confirmed that the bug is not caused by a custom node. You can disable all custom nodes by passing - `--disable-all-custom-nodes` command line argument. + `--disable-all-custom-nodes` command line argument. If you have custom node try updating them to the latest version. - **4:** This is an actual bug in ComfyUI, not just a support question. A bug is when you can specify exact steps to replicate what went wrong and others will be able to repeat your steps and see the same issue happen. - If unsure, ask on the [ComfyUI Matrix Space](https://app.element.io/#/room/%23comfyui_space%3Amatrix.org) or the [Comfy Org Discord](https://discord.gg/comfyorg) first. + ## Very Important + + Please make sure that you post ALL your ComfyUI logs in the bug report. A bug report without logs will likely be ignored. - type: checkboxes id: custom-nodes-test attributes: diff --git a/.github/PULL_REQUEST_TEMPLATE/api-node.md b/.github/PULL_REQUEST_TEMPLATE/api-node.md new file mode 100644 index 000000000000..f62744878629 --- /dev/null +++ b/.github/PULL_REQUEST_TEMPLATE/api-node.md @@ -0,0 +1,21 @@ + + +## API Node PR Checklist + +### Scope +- [ ] **Is API Node Change** + +### Pricing & Billing +- [ ] **Need pricing update** +- [ ] **No pricing update** + +If **Need pricing update**: +- [ ] Metronome rate cards updated +- [ ] Auto‑billing tests updated and passing + +### QA +- [ ] **QA done** +- [ ] **QA not required** + +### Comms +- [ ] Informed **@Kosinkadink** diff --git a/.github/workflows/api-node-template.yml b/.github/workflows/api-node-template.yml new file mode 100644 index 000000000000..0775f99793ef --- /dev/null +++ b/.github/workflows/api-node-template.yml @@ -0,0 +1,58 @@ +name: Append API Node PR template + +on: + pull_request_target: + types: [opened, reopened, synchronize, edited, ready_for_review] + paths: + - 'comfy_api_nodes/**' # only run if these files changed + +permissions: + contents: read + pull-requests: write + +jobs: + inject: + runs-on: ubuntu-latest + steps: + - name: Ensure template exists and append to PR body + uses: actions/github-script@v7 + with: + script: | + const { owner, repo } = context.repo; + const number = context.payload.pull_request.number; + const templatePath = '.github/PULL_REQUEST_TEMPLATE/api-node.md'; + const marker = ''; + + const { data: pr } = await github.rest.pulls.get({ owner, repo, pull_number: number }); + + let templateText; + try { + const res = await github.rest.repos.getContent({ + owner, + repo, + path: templatePath, + ref: pr.base.ref + }); + const buf = Buffer.from(res.data.content, res.data.encoding || 'base64'); + templateText = buf.toString('utf8'); + } catch (e) { + core.setFailed(`Required PR template not found at "${templatePath}" on ${pr.base.ref}. Please add it to the repo.`); + return; + } + + // Enforce the presence of the marker inside the template (for idempotence) + if (!templateText.includes(marker)) { + core.setFailed(`Template at "${templatePath}" does not contain the required marker:\n${marker}\nAdd it so we can detect duplicates safely.`); + return; + } + + // If the PR already contains the marker, do not append again. + const body = pr.body || ''; + if (body.includes(marker)) { + core.info('Template already present in PR body; nothing to inject.'); + return; + } + + const newBody = (body ? body + '\n\n' : '') + templateText + '\n'; + await github.rest.pulls.update({ owner, repo, pull_number: number, body: newBody }); + core.notice('API Node template appended to PR description.'); diff --git a/.github/workflows/test-ci.yml b/.github/workflows/test-ci.yml index 418dca0ab75e..1660ec8e35d3 100644 --- a/.github/workflows/test-ci.yml +++ b/.github/workflows/test-ci.yml @@ -21,14 +21,15 @@ jobs: fail-fast: false matrix: # os: [macos, linux, windows] - os: [macos, linux] - python_version: ["3.9", "3.10", "3.11", "3.12"] + # os: [macos, linux] + os: [linux] + python_version: ["3.10", "3.11", "3.12"] cuda_version: ["12.1"] torch_version: ["stable"] include: - - os: macos - runner_label: [self-hosted, macOS] - flags: "--use-pytorch-cross-attention" + # - os: macos + # runner_label: [self-hosted, macOS] + # flags: "--use-pytorch-cross-attention" - os: linux runner_label: [self-hosted, Linux] flags: "" @@ -73,14 +74,15 @@ jobs: strategy: fail-fast: false matrix: - os: [macos, linux] + # os: [macos, linux] + os: [linux] python_version: ["3.11"] cuda_version: ["12.1"] torch_version: ["nightly"] include: - - os: macos - runner_label: [self-hosted, macOS] - flags: "--use-pytorch-cross-attention" + # - os: macos + # runner_label: [self-hosted, macOS] + # flags: "--use-pytorch-cross-attention" - os: linux runner_label: [self-hosted, Linux] flags: "" diff --git a/QUANTIZATION.md b/QUANTIZATION.md new file mode 100644 index 000000000000..1693e13f32e2 --- /dev/null +++ b/QUANTIZATION.md @@ -0,0 +1,168 @@ +# The Comfy guide to Quantization + + +## How does quantization work? + +Quantization aims to map a high-precision value x_f to a lower precision format with minimal loss in accuracy. These smaller formats then serve to reduce the models memory footprint and increase throughput by using specialized hardware. + +When simply converting a value from FP16 to FP8 using the round-nearest method we might hit two issues: +- The dynamic range of FP16 (-65,504, 65,504) far exceeds FP8 formats like E4M3 (-448, 448) or E5M2 (-57,344, 57,344), potentially resulting in clipped values +- The original values are concentrated in a small range (e.g. -1,1) leaving many FP8-bits "unused" + +By using a scaling factor, we aim to map these values into the quantized-dtype range, making use of the full spectrum. One of the easiest approaches, and common, is using per-tensor absolute-maximum scaling. + +``` +absmax = max(abs(tensor)) +scale = amax / max_dynamic_range_low_precision + +# Quantization +tensor_q = (tensor / scale).to(low_precision_dtype) + +# De-Quantization +tensor_dq = tensor_q.to(fp16) * scale + +tensor_dq ~ tensor +``` + +Given that additional information (scaling factor) is needed to "interpret" the quantized values, we describe those as derived datatypes. + + +## Quantization in Comfy + +``` +QuantizedTensor (torch.Tensor subclass) + ↓ __torch_dispatch__ +Two-Level Registry (generic + layout handlers) + ↓ +MixedPrecisionOps + Metadata Detection +``` + +### Representation + +To represent these derived datatypes, ComfyUI uses a subclass of torch.Tensor to implements these using the `QuantizedTensor` class found in `comfy/quant_ops.py` + +A `Layout` class defines how a specific quantization format behaves: +- Required parameters +- Quantize method +- De-Quantize method + +```python +from comfy.quant_ops import QuantizedLayout + +class MyLayout(QuantizedLayout): + @classmethod + def quantize(cls, tensor, **kwargs): + # Convert to quantized format + qdata = ... + params = {'scale': ..., 'orig_dtype': tensor.dtype} + return qdata, params + + @staticmethod + def dequantize(qdata, scale, orig_dtype, **kwargs): + return qdata.to(orig_dtype) * scale +``` + +To then run operations using these QuantizedTensors we use two registry systems to define supported operations. +The first is a **generic registry** that handles operations common to all quantized formats (e.g., `.to()`, `.clone()`, `.reshape()`). + +The second registry is layout-specific and allows to implement fast-paths like nn.Linear. +```python +from comfy.quant_ops import register_layout_op + +@register_layout_op(torch.ops.aten.linear.default, MyLayout) +def my_linear(func, args, kwargs): + # Extract tensors, call optimized kernel + ... +``` +When `torch.nn.functional.linear()` is called with QuantizedTensor arguments, `__torch_dispatch__` automatically routes to the registered implementation. +For any unsupported operation, QuantizedTensor will fallback to call `dequantize` and dispatch using the high-precision implementation. + + +### Mixed Precision + +The `MixedPrecisionOps` class (lines 542-648 in `comfy/ops.py`) enables per-layer quantization decisions, allowing different layers in a model to use different precisions. This is activated when a model config contains a `layer_quant_config` dictionary that specifies which layers should be quantized and how. + +**Architecture:** + +```python +class MixedPrecisionOps(disable_weight_init): + _layer_quant_config = {} # Maps layer names to quantization configs + _compute_dtype = torch.bfloat16 # Default compute / dequantize precision +``` + +**Key mechanism:** + +The custom `Linear._load_from_state_dict()` method inspects each layer during model loading: +- If the layer name is **not** in `_layer_quant_config`: load weight as regular tensor in `_compute_dtype` +- If the layer name **is** in `_layer_quant_config`: + - Load weight as `QuantizedTensor` with the specified layout (e.g., `TensorCoreFP8Layout`) + - Load associated quantization parameters (scales, block_size, etc.) + +**Why it's needed:** + +Not all layers tolerate quantization equally. Sensitive operations like final projections can be kept in higher precision, while compute-heavy matmuls are quantized. This provides most of the performance benefits while maintaining quality. + +The system is selected in `pick_operations()` when `model_config.layer_quant_config` is present, making it the highest-priority operation mode. + + +## Checkpoint Format + +Quantized checkpoints are stored as standard safetensors files with quantized weight tensors and associated scaling parameters, plus a `_quantization_metadata` JSON entry describing the quantization scheme. + +The quantized checkpoint will contain the same layers as the original checkpoint but: +- The weights are stored as quantized values, sometimes using a different storage datatype. E.g. uint8 container for fp8. +- For each quantized weight a number of additional scaling parameters are stored alongside depending on the recipe. +- We store a metadata.json in the metadata of the final safetensor containing the `_quantization_metadata` describing which layers are quantized and what layout has been used. + +### Scaling Parameters details +We define 4 possible scaling parameters that should cover most recipes in the near-future: +- **weight_scale**: quantization scalers for the weights +- **weight_scale_2**: global scalers in the context of double scaling +- **pre_quant_scale**: scalers used for smoothing salient weights +- **input_scale**: quantization scalers for the activations + +| Format | Storage dtype | weight_scale | weight_scale_2 | pre_quant_scale | input_scale | +|--------|---------------|--------------|----------------|-----------------|-------------| +| float8_e4m3fn | float32 | float32 (scalar) | - | - | float32 (scalar) | + +You can find the defined formats in `comfy/quant_ops.py` (QUANT_ALGOS). + +### Quantization Metadata + +The metadata stored alongside the checkpoint contains: +- **format_version**: String to define a version of the standard +- **layers**: A dictionary mapping layer names to their quantization format. The format string maps to the definitions found in `QUANT_ALGOS`. + +Example: +```json +{ + "_quantization_metadata": { + "format_version": "1.0", + "layers": { + "model.layers.0.mlp.up_proj": "float8_e4m3fn", + "model.layers.0.mlp.down_proj": "float8_e4m3fn", + "model.layers.1.mlp.up_proj": "float8_e4m3fn" + } + } +} +``` + + +## Creating Quantized Checkpoints + +To create compatible checkpoints, use any quantization tool provided the output follows the checkpoint format described above and uses a layout defined in `QUANT_ALGOS`. + +### Weight Quantization + +Weight quantization is straightforward - compute the scaling factor directly from the weight tensor using the absolute maximum method described earlier. Each layer's weights are quantized independently and stored with their corresponding `weight_scale` parameter. + +### Calibration (for Activation Quantization) + +Activation quantization (e.g., for FP8 Tensor Core operations) requires `input_scale` parameters that cannot be determined from static weights alone. Since activation values depend on actual inputs, we use **post-training calibration (PTQ)**: + +1. **Collect statistics**: Run inference on N representative samples +2. **Track activations**: Record the absolute maximum (`amax`) of inputs to each quantized layer +3. **Compute scales**: Derive `input_scale` from collected statistics +4. **Store in checkpoint**: Save `input_scale` parameters alongside weights + +The calibration dataset should be representative of your target use case. For diffusion models, this typically means a diverse set of prompts and generation parameters. \ No newline at end of file diff --git a/README.md b/README.md index 4204777e9008..c0384099dda3 100644 --- a/README.md +++ b/README.md @@ -112,10 +112,11 @@ Workflow examples can be found on the [Examples page](https://comfyanonymous.git ## Release Process -ComfyUI follows a weekly release cycle targeting Friday but this regularly changes because of model releases or large changes to the codebase. There are three interconnected repositories: +ComfyUI follows a weekly release cycle targeting Monday but this regularly changes because of model releases or large changes to the codebase. There are three interconnected repositories: 1. **[ComfyUI Core](https://github.com/comfyanonymous/ComfyUI)** - - Releases a new stable version (e.g., v0.7.0) + - Releases a new stable version (e.g., v0.7.0) roughly every week. + - Commits outside of the stable release tags may be very unstable and break many custom nodes. - Serves as the foundation for the desktop release 2. **[ComfyUI Desktop](https://github.com/Comfy-Org/desktop)** @@ -172,7 +173,7 @@ There is a portable standalone build for Windows that should work for running on ### [Direct link to download](https://github.com/comfyanonymous/ComfyUI/releases/latest/download/ComfyUI_windows_portable_nvidia.7z) -Simply download, extract with [7-Zip](https://7-zip.org) and run. Make sure you put your Stable Diffusion checkpoints/models (the huge ckpt/safetensors files) in: ComfyUI\models\checkpoints +Simply download, extract with [7-Zip](https://7-zip.org) or with the windows explorer on recent windows versions and run. For smaller models you normally only need to put the checkpoints (the huge ckpt/safetensors files) in: ComfyUI\models\checkpoints but many of the larger models have multiple files. Make sure to follow the instructions to know which subfolder to put them in ComfyUI\models\ If you have trouble extracting it, right click the file -> properties -> unblock @@ -199,7 +200,7 @@ comfy install ## Manual Install (Windows, Linux) -Python 3.14 will work if you comment out the `kornia` dependency in the requirements.txt file (breaks the canny node) but it is not recommended. +Python 3.14 works but you may encounter issues with the torch compile node. The free threaded variant is still missing some dependencies. Python 3.13 is very well supported. If you have trouble with some custom node dependencies on 3.13 you can try 3.12 @@ -220,7 +221,7 @@ AMD users can install rocm and pytorch with pip if you don't have it already ins This is the command to install the nightly with ROCm 7.0 which might have some performance improvements: -```pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/rocm7.0``` +```pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/rocm7.1``` ### AMD GPUs (Experimental: Windows and Linux), RDNA 3, 3.5 and 4 only. @@ -241,7 +242,7 @@ RDNA 4 (RX 9000 series): ### Intel GPUs (Windows and Linux) -(Option 1) Intel Arc GPU users can install native PyTorch with torch.xpu support using pip. More information can be found [here](https://pytorch.org/docs/main/notes/get_start_xpu.html) +Intel Arc GPU users can install native PyTorch with torch.xpu support using pip. More information can be found [here](https://pytorch.org/docs/main/notes/get_start_xpu.html) 1. To install PyTorch xpu, use the following command: @@ -251,10 +252,6 @@ This is the command to install the Pytorch xpu nightly which might have some per ```pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/xpu``` -(Option 2) Alternatively, Intel GPUs supported by Intel Extension for PyTorch (IPEX) can leverage IPEX for improved performance. - -1. visit [Installation](https://intel.github.io/intel-extension-for-pytorch/index.html#installation?platform=gpu) for more information. - ### NVIDIA Nvidia users should install stable pytorch using this command: diff --git a/comfy/cli_args.py b/comfy/cli_args.py index 3947e62a8f00..d86d74267f48 100644 --- a/comfy/cli_args.py +++ b/comfy/cli_args.py @@ -47,6 +47,7 @@ def __call__(self, parser, namespace, values, option_string=None): parser.add_argument("--output-directory", type=str, default=None, help="Set the ComfyUI output directory. Overrides --base-directory.") parser.add_argument("--temp-directory", type=str, default=None, help="Set the ComfyUI temp directory (default is in the ComfyUI directory). Overrides --base-directory.") parser.add_argument("--input-directory", type=str, default=None, help="Set the ComfyUI input directory. Overrides --base-directory.") +parser.add_argument("--custom-nodes-directory", type=str, default=None, help="Set the ComfyUI custom_nodes directory. Overrides --base-directory and environment variable COMFYUI_CUSTOM_NODES_DIR.") parser.add_argument("--auto-launch", action="store_true", help="Automatically launch ComfyUI in the default browser.") parser.add_argument("--disable-auto-launch", action="store_true", help="Disable auto launching the browser.") parser.add_argument("--cuda-device", type=int, default=None, metavar="DEVICE_ID", help="Set the id of the cuda device this instance will use. All other devices will not be visible.") @@ -145,10 +146,11 @@ class PerformanceFeature(enum.Enum): Fp8MatrixMultiplication = "fp8_matrix_mult" CublasOps = "cublas_ops" AutoTune = "autotune" - PinnedMem = "pinned_memory" parser.add_argument("--fast", nargs="*", type=PerformanceFeature, help="Enable some untested and potentially quality deteriorating optimizations. This is used to test new features so using it might crash your comfyui. --fast with no arguments enables everything. You can pass a list specific optimizations if you only want to enable specific ones. Current valid optimizations: {}".format(" ".join(map(lambda c: c.value, PerformanceFeature)))) +parser.add_argument("--disable-pinned-memory", action="store_true", help="Disable pinned memory use.") + parser.add_argument("--mmap-torch-files", action="store_true", help="Use mmap when loading ckpt/pt files.") parser.add_argument("--disable-mmap", action="store_true", help="Don't use mmap when loading safetensors.") diff --git a/comfy/ldm/chroma/layers.py b/comfy/ldm/chroma/layers.py index fc7110cce50e..2d568434871b 100644 --- a/comfy/ldm/chroma/layers.py +++ b/comfy/ldm/chroma/layers.py @@ -1,15 +1,15 @@ import torch from torch import Tensor, nn -from comfy.ldm.flux.math import attention from comfy.ldm.flux.layers import ( MLPEmbedder, RMSNorm, - QKNorm, - SelfAttention, ModulationOut, ) +# TODO: remove this in a few months +SingleStreamBlock = None +DoubleStreamBlock = None class ChromaModulationOut(ModulationOut): @@ -48,124 +48,6 @@ def forward(self, x: Tensor) -> Tensor: return x -class DoubleStreamBlock(nn.Module): - def __init__(self, hidden_size: int, num_heads: int, mlp_ratio: float, qkv_bias: bool = False, flipped_img_txt=False, dtype=None, device=None, operations=None): - super().__init__() - - mlp_hidden_dim = int(hidden_size * mlp_ratio) - self.num_heads = num_heads - self.hidden_size = hidden_size - self.img_norm1 = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device) - self.img_attn = SelfAttention(dim=hidden_size, num_heads=num_heads, qkv_bias=qkv_bias, dtype=dtype, device=device, operations=operations) - - self.img_norm2 = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device) - self.img_mlp = nn.Sequential( - operations.Linear(hidden_size, mlp_hidden_dim, bias=True, dtype=dtype, device=device), - nn.GELU(approximate="tanh"), - operations.Linear(mlp_hidden_dim, hidden_size, bias=True, dtype=dtype, device=device), - ) - - self.txt_norm1 = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device) - self.txt_attn = SelfAttention(dim=hidden_size, num_heads=num_heads, qkv_bias=qkv_bias, dtype=dtype, device=device, operations=operations) - - self.txt_norm2 = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device) - self.txt_mlp = nn.Sequential( - operations.Linear(hidden_size, mlp_hidden_dim, bias=True, dtype=dtype, device=device), - nn.GELU(approximate="tanh"), - operations.Linear(mlp_hidden_dim, hidden_size, bias=True, dtype=dtype, device=device), - ) - self.flipped_img_txt = flipped_img_txt - - def forward(self, img: Tensor, txt: Tensor, pe: Tensor, vec: Tensor, attn_mask=None, transformer_options={}): - (img_mod1, img_mod2), (txt_mod1, txt_mod2) = vec - - # prepare image for attention - img_modulated = torch.addcmul(img_mod1.shift, 1 + img_mod1.scale, self.img_norm1(img)) - img_qkv = self.img_attn.qkv(img_modulated) - img_q, img_k, img_v = img_qkv.view(img_qkv.shape[0], img_qkv.shape[1], 3, self.num_heads, -1).permute(2, 0, 3, 1, 4) - img_q, img_k = self.img_attn.norm(img_q, img_k, img_v) - - # prepare txt for attention - txt_modulated = torch.addcmul(txt_mod1.shift, 1 + txt_mod1.scale, self.txt_norm1(txt)) - txt_qkv = self.txt_attn.qkv(txt_modulated) - txt_q, txt_k, txt_v = txt_qkv.view(txt_qkv.shape[0], txt_qkv.shape[1], 3, self.num_heads, -1).permute(2, 0, 3, 1, 4) - txt_q, txt_k = self.txt_attn.norm(txt_q, txt_k, txt_v) - - # run actual attention - attn = attention(torch.cat((txt_q, img_q), dim=2), - torch.cat((txt_k, img_k), dim=2), - torch.cat((txt_v, img_v), dim=2), - pe=pe, mask=attn_mask, transformer_options=transformer_options) - - txt_attn, img_attn = attn[:, : txt.shape[1]], attn[:, txt.shape[1] :] - - # calculate the img bloks - img.addcmul_(img_mod1.gate, self.img_attn.proj(img_attn)) - img.addcmul_(img_mod2.gate, self.img_mlp(torch.addcmul(img_mod2.shift, 1 + img_mod2.scale, self.img_norm2(img)))) - - # calculate the txt bloks - txt.addcmul_(txt_mod1.gate, self.txt_attn.proj(txt_attn)) - txt.addcmul_(txt_mod2.gate, self.txt_mlp(torch.addcmul(txt_mod2.shift, 1 + txt_mod2.scale, self.txt_norm2(txt)))) - - if txt.dtype == torch.float16: - txt = torch.nan_to_num(txt, nan=0.0, posinf=65504, neginf=-65504) - - return img, txt - - -class SingleStreamBlock(nn.Module): - """ - A DiT block with parallel linear layers as described in - https://arxiv.org/abs/2302.05442 and adapted modulation interface. - """ - - def __init__( - self, - hidden_size: int, - num_heads: int, - mlp_ratio: float = 4.0, - qk_scale: float = None, - dtype=None, - device=None, - operations=None - ): - super().__init__() - self.hidden_dim = hidden_size - self.num_heads = num_heads - head_dim = hidden_size // num_heads - self.scale = qk_scale or head_dim**-0.5 - - self.mlp_hidden_dim = int(hidden_size * mlp_ratio) - # qkv and mlp_in - self.linear1 = operations.Linear(hidden_size, hidden_size * 3 + self.mlp_hidden_dim, dtype=dtype, device=device) - # proj and mlp_out - self.linear2 = operations.Linear(hidden_size + self.mlp_hidden_dim, hidden_size, dtype=dtype, device=device) - - self.norm = QKNorm(head_dim, dtype=dtype, device=device, operations=operations) - - self.hidden_size = hidden_size - self.pre_norm = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device) - - self.mlp_act = nn.GELU(approximate="tanh") - - def forward(self, x: Tensor, pe: Tensor, vec: Tensor, attn_mask=None, transformer_options={}) -> Tensor: - mod = vec - x_mod = torch.addcmul(mod.shift, 1 + mod.scale, self.pre_norm(x)) - qkv, mlp = torch.split(self.linear1(x_mod), [3 * self.hidden_size, self.mlp_hidden_dim], dim=-1) - - q, k, v = qkv.view(qkv.shape[0], qkv.shape[1], 3, self.num_heads, -1).permute(2, 0, 3, 1, 4) - q, k = self.norm(q, k, v) - - # compute attention - attn = attention(q, k, v, pe=pe, mask=attn_mask, transformer_options=transformer_options) - # compute activation in mlp stream, cat again and run second linear layer - output = self.linear2(torch.cat((attn, self.mlp_act(mlp)), 2)) - x.addcmul_(mod.gate, output) - if x.dtype == torch.float16: - x = torch.nan_to_num(x, nan=0.0, posinf=65504, neginf=-65504) - return x - - class LastLayer(nn.Module): def __init__(self, hidden_size: int, patch_size: int, out_channels: int, dtype=None, device=None, operations=None): super().__init__() diff --git a/comfy/ldm/chroma/model.py b/comfy/ldm/chroma/model.py index ad1c523fe1d3..67bf70eb10b4 100644 --- a/comfy/ldm/chroma/model.py +++ b/comfy/ldm/chroma/model.py @@ -11,12 +11,12 @@ from comfy.ldm.flux.layers import ( EmbedND, timestep_embedding, + DoubleStreamBlock, + SingleStreamBlock, ) from .layers import ( - DoubleStreamBlock, LastLayer, - SingleStreamBlock, Approximator, ChromaModulationOut, ) @@ -90,6 +90,7 @@ def __init__(self, image_model=None, final_layer=True, dtype=None, device=None, self.num_heads, mlp_ratio=params.mlp_ratio, qkv_bias=params.qkv_bias, + modulation=False, dtype=dtype, device=device, operations=operations ) for _ in range(params.depth) @@ -98,7 +99,7 @@ def __init__(self, image_model=None, final_layer=True, dtype=None, device=None, self.single_blocks = nn.ModuleList( [ - SingleStreamBlock(self.hidden_size, self.num_heads, mlp_ratio=params.mlp_ratio, dtype=dtype, device=device, operations=operations) + SingleStreamBlock(self.hidden_size, self.num_heads, mlp_ratio=params.mlp_ratio, modulation=False, dtype=dtype, device=device, operations=operations) for _ in range(params.depth_single_blocks) ] ) diff --git a/comfy/ldm/chroma_radiance/model.py b/comfy/ldm/chroma_radiance/model.py index 7d7be80f5943..e643b4414629 100644 --- a/comfy/ldm/chroma_radiance/model.py +++ b/comfy/ldm/chroma_radiance/model.py @@ -10,12 +10,10 @@ from einops import repeat import comfy.ldm.common_dit -from comfy.ldm.flux.layers import EmbedND +from comfy.ldm.flux.layers import EmbedND, DoubleStreamBlock, SingleStreamBlock from comfy.ldm.chroma.model import Chroma, ChromaParams from comfy.ldm.chroma.layers import ( - DoubleStreamBlock, - SingleStreamBlock, Approximator, ) from .layers import ( @@ -89,7 +87,6 @@ def __init__(self, image_model=None, final_layer=True, dtype=None, device=None, dtype=dtype, device=device, operations=operations ) - self.double_blocks = nn.ModuleList( [ DoubleStreamBlock( @@ -97,6 +94,7 @@ def __init__(self, image_model=None, final_layer=True, dtype=None, device=None, self.num_heads, mlp_ratio=params.mlp_ratio, qkv_bias=params.qkv_bias, + modulation=False, dtype=dtype, device=device, operations=operations ) for _ in range(params.depth) @@ -109,6 +107,7 @@ def __init__(self, image_model=None, final_layer=True, dtype=None, device=None, self.hidden_size, self.num_heads, mlp_ratio=params.mlp_ratio, + modulation=False, dtype=dtype, device=device, operations=operations, ) for _ in range(params.depth_single_blocks) diff --git a/comfy/ldm/flux/layers.py b/comfy/ldm/flux/layers.py index a3eab0470d95..23150a7125a2 100644 --- a/comfy/ldm/flux/layers.py +++ b/comfy/ldm/flux/layers.py @@ -130,13 +130,17 @@ def apply_mod(tensor, m_mult, m_add=None, modulation_dims=None): class DoubleStreamBlock(nn.Module): - def __init__(self, hidden_size: int, num_heads: int, mlp_ratio: float, qkv_bias: bool = False, flipped_img_txt=False, dtype=None, device=None, operations=None): + def __init__(self, hidden_size: int, num_heads: int, mlp_ratio: float, qkv_bias: bool = False, flipped_img_txt=False, modulation=True, dtype=None, device=None, operations=None): super().__init__() mlp_hidden_dim = int(hidden_size * mlp_ratio) self.num_heads = num_heads self.hidden_size = hidden_size - self.img_mod = Modulation(hidden_size, double=True, dtype=dtype, device=device, operations=operations) + self.modulation = modulation + + if self.modulation: + self.img_mod = Modulation(hidden_size, double=True, dtype=dtype, device=device, operations=operations) + self.img_norm1 = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device) self.img_attn = SelfAttention(dim=hidden_size, num_heads=num_heads, qkv_bias=qkv_bias, dtype=dtype, device=device, operations=operations) @@ -147,7 +151,9 @@ def __init__(self, hidden_size: int, num_heads: int, mlp_ratio: float, qkv_bias: operations.Linear(mlp_hidden_dim, hidden_size, bias=True, dtype=dtype, device=device), ) - self.txt_mod = Modulation(hidden_size, double=True, dtype=dtype, device=device, operations=operations) + if self.modulation: + self.txt_mod = Modulation(hidden_size, double=True, dtype=dtype, device=device, operations=operations) + self.txt_norm1 = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device) self.txt_attn = SelfAttention(dim=hidden_size, num_heads=num_heads, qkv_bias=qkv_bias, dtype=dtype, device=device, operations=operations) @@ -160,46 +166,65 @@ def __init__(self, hidden_size: int, num_heads: int, mlp_ratio: float, qkv_bias: self.flipped_img_txt = flipped_img_txt def forward(self, img: Tensor, txt: Tensor, vec: Tensor, pe: Tensor, attn_mask=None, modulation_dims_img=None, modulation_dims_txt=None, transformer_options={}): - img_mod1, img_mod2 = self.img_mod(vec) - txt_mod1, txt_mod2 = self.txt_mod(vec) + if self.modulation: + img_mod1, img_mod2 = self.img_mod(vec) + txt_mod1, txt_mod2 = self.txt_mod(vec) + else: + (img_mod1, img_mod2), (txt_mod1, txt_mod2) = vec # prepare image for attention img_modulated = self.img_norm1(img) img_modulated = apply_mod(img_modulated, (1 + img_mod1.scale), img_mod1.shift, modulation_dims_img) img_qkv = self.img_attn.qkv(img_modulated) + del img_modulated img_q, img_k, img_v = img_qkv.view(img_qkv.shape[0], img_qkv.shape[1], 3, self.num_heads, -1).permute(2, 0, 3, 1, 4) + del img_qkv img_q, img_k = self.img_attn.norm(img_q, img_k, img_v) # prepare txt for attention txt_modulated = self.txt_norm1(txt) txt_modulated = apply_mod(txt_modulated, (1 + txt_mod1.scale), txt_mod1.shift, modulation_dims_txt) txt_qkv = self.txt_attn.qkv(txt_modulated) + del txt_modulated txt_q, txt_k, txt_v = txt_qkv.view(txt_qkv.shape[0], txt_qkv.shape[1], 3, self.num_heads, -1).permute(2, 0, 3, 1, 4) + del txt_qkv txt_q, txt_k = self.txt_attn.norm(txt_q, txt_k, txt_v) if self.flipped_img_txt: + q = torch.cat((img_q, txt_q), dim=2) + del img_q, txt_q + k = torch.cat((img_k, txt_k), dim=2) + del img_k, txt_k + v = torch.cat((img_v, txt_v), dim=2) + del img_v, txt_v # run actual attention - attn = attention(torch.cat((img_q, txt_q), dim=2), - torch.cat((img_k, txt_k), dim=2), - torch.cat((img_v, txt_v), dim=2), + attn = attention(q, k, v, pe=pe, mask=attn_mask, transformer_options=transformer_options) + del q, k, v img_attn, txt_attn = attn[:, : img.shape[1]], attn[:, img.shape[1]:] else: + q = torch.cat((txt_q, img_q), dim=2) + del txt_q, img_q + k = torch.cat((txt_k, img_k), dim=2) + del txt_k, img_k + v = torch.cat((txt_v, img_v), dim=2) + del txt_v, img_v # run actual attention - attn = attention(torch.cat((txt_q, img_q), dim=2), - torch.cat((txt_k, img_k), dim=2), - torch.cat((txt_v, img_v), dim=2), + attn = attention(q, k, v, pe=pe, mask=attn_mask, transformer_options=transformer_options) + del q, k, v txt_attn, img_attn = attn[:, : txt.shape[1]], attn[:, txt.shape[1]:] # calculate the img bloks img += apply_mod(self.img_attn.proj(img_attn), img_mod1.gate, None, modulation_dims_img) + del img_attn img += apply_mod(self.img_mlp(apply_mod(self.img_norm2(img), (1 + img_mod2.scale), img_mod2.shift, modulation_dims_img)), img_mod2.gate, None, modulation_dims_img) # calculate the txt bloks txt += apply_mod(self.txt_attn.proj(txt_attn), txt_mod1.gate, None, modulation_dims_txt) + del txt_attn txt += apply_mod(self.txt_mlp(apply_mod(self.txt_norm2(txt), (1 + txt_mod2.scale), txt_mod2.shift, modulation_dims_txt)), txt_mod2.gate, None, modulation_dims_txt) if txt.dtype == torch.float16: @@ -220,6 +245,7 @@ def __init__( num_heads: int, mlp_ratio: float = 4.0, qk_scale: float = None, + modulation=True, dtype=None, device=None, operations=None @@ -242,19 +268,29 @@ def __init__( self.pre_norm = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device) self.mlp_act = nn.GELU(approximate="tanh") - self.modulation = Modulation(hidden_size, double=False, dtype=dtype, device=device, operations=operations) + if modulation: + self.modulation = Modulation(hidden_size, double=False, dtype=dtype, device=device, operations=operations) + else: + self.modulation = None def forward(self, x: Tensor, vec: Tensor, pe: Tensor, attn_mask=None, modulation_dims=None, transformer_options={}) -> Tensor: - mod, _ = self.modulation(vec) + if self.modulation: + mod, _ = self.modulation(vec) + else: + mod = vec + qkv, mlp = torch.split(self.linear1(apply_mod(self.pre_norm(x), (1 + mod.scale), mod.shift, modulation_dims)), [3 * self.hidden_size, self.mlp_hidden_dim], dim=-1) q, k, v = qkv.view(qkv.shape[0], qkv.shape[1], 3, self.num_heads, -1).permute(2, 0, 3, 1, 4) + del qkv q, k = self.norm(q, k, v) # compute attention attn = attention(q, k, v, pe=pe, mask=attn_mask, transformer_options=transformer_options) + del q, k, v # compute activation in mlp stream, cat again and run second linear layer - output = self.linear2(torch.cat((attn, self.mlp_act(mlp)), 2)) + mlp = self.mlp_act(mlp) + output = self.linear2(torch.cat((attn, mlp), 2)) x += apply_mod(output, mod.gate, None, modulation_dims) if x.dtype == torch.float16: x = torch.nan_to_num(x, nan=0.0, posinf=65504, neginf=-65504) diff --git a/comfy/ldm/flux/model.py b/comfy/ldm/flux/model.py index 14f90cea55e6..b9d36f2024d7 100644 --- a/comfy/ldm/flux/model.py +++ b/comfy/ldm/flux/model.py @@ -210,7 +210,7 @@ def block_wrap(args): img = self.final_layer(img, vec) # (N, T, patch_size ** 2 * out_channels) return img - def process_img(self, x, index=0, h_offset=0, w_offset=0): + def process_img(self, x, index=0, h_offset=0, w_offset=0, transformer_options={}): bs, c, h, w = x.shape patch_size = self.patch_size x = comfy.ldm.common_dit.pad_to_patch_size(x, (patch_size, patch_size)) @@ -222,10 +222,22 @@ def process_img(self, x, index=0, h_offset=0, w_offset=0): h_offset = ((h_offset + (patch_size // 2)) // patch_size) w_offset = ((w_offset + (patch_size // 2)) // patch_size) - img_ids = torch.zeros((h_len, w_len, 3), device=x.device, dtype=x.dtype) + steps_h = h_len + steps_w = w_len + + rope_options = transformer_options.get("rope_options", None) + if rope_options is not None: + h_len = (h_len - 1.0) * rope_options.get("scale_y", 1.0) + 1.0 + w_len = (w_len - 1.0) * rope_options.get("scale_x", 1.0) + 1.0 + + index += rope_options.get("shift_t", 0.0) + h_offset += rope_options.get("shift_y", 0.0) + w_offset += rope_options.get("shift_x", 0.0) + + img_ids = torch.zeros((steps_h, steps_w, 3), device=x.device, dtype=x.dtype) img_ids[:, :, 0] = img_ids[:, :, 1] + index - img_ids[:, :, 1] = img_ids[:, :, 1] + torch.linspace(h_offset, h_len - 1 + h_offset, steps=h_len, device=x.device, dtype=x.dtype).unsqueeze(1) - img_ids[:, :, 2] = img_ids[:, :, 2] + torch.linspace(w_offset, w_len - 1 + w_offset, steps=w_len, device=x.device, dtype=x.dtype).unsqueeze(0) + img_ids[:, :, 1] = img_ids[:, :, 1] + torch.linspace(h_offset, h_len - 1 + h_offset, steps=steps_h, device=x.device, dtype=x.dtype).unsqueeze(1) + img_ids[:, :, 2] = img_ids[:, :, 2] + torch.linspace(w_offset, w_len - 1 + w_offset, steps=steps_w, device=x.device, dtype=x.dtype).unsqueeze(0) return img, repeat(img_ids, "h w c -> b (h w) c", b=bs) def forward(self, x, timestep, context, y=None, guidance=None, ref_latents=None, control=None, transformer_options={}, **kwargs): @@ -241,7 +253,7 @@ def _forward(self, x, timestep, context, y=None, guidance=None, ref_latents=None h_len = ((h_orig + (patch_size // 2)) // patch_size) w_len = ((w_orig + (patch_size // 2)) // patch_size) - img, img_ids = self.process_img(x) + img, img_ids = self.process_img(x, transformer_options=transformer_options) img_tokens = img.shape[1] if ref_latents is not None: h = 0 diff --git a/comfy/ldm/qwen_image/controlnet.py b/comfy/ldm/qwen_image/controlnet.py index 92ac3cf0afd6..a6d408104a00 100644 --- a/comfy/ldm/qwen_image/controlnet.py +++ b/comfy/ldm/qwen_image/controlnet.py @@ -44,7 +44,7 @@ def forward( txt_start = round(max(((x.shape[-1] + (self.patch_size // 2)) // self.patch_size) // 2, ((x.shape[-2] + (self.patch_size // 2)) // self.patch_size) // 2)) txt_ids = torch.arange(txt_start, txt_start + context.shape[1], device=x.device).reshape(1, -1, 1).repeat(x.shape[0], 1, 3) ids = torch.cat((txt_ids, img_ids), dim=1) - image_rotary_emb = self.pe_embedder(ids).squeeze(1).unsqueeze(2).to(x.dtype) + image_rotary_emb = self.pe_embedder(ids).to(x.dtype).contiguous() del ids, txt_ids, img_ids hidden_states = self.img_in(hidden_states) + self.controlnet_x_embedder(hint) diff --git a/comfy/ldm/qwen_image/model.py b/comfy/ldm/qwen_image/model.py index e5d0d17c1d6e..427ea19c1179 100644 --- a/comfy/ldm/qwen_image/model.py +++ b/comfy/ldm/qwen_image/model.py @@ -236,10 +236,10 @@ def forward( img_mod1, img_mod2 = img_mod_params.chunk(2, dim=-1) txt_mod1, txt_mod2 = txt_mod_params.chunk(2, dim=-1) - img_normed = self.img_norm1(hidden_states) - img_modulated, img_gate1 = self._modulate(img_normed, img_mod1) - txt_normed = self.txt_norm1(encoder_hidden_states) - txt_modulated, txt_gate1 = self._modulate(txt_normed, txt_mod1) + img_modulated, img_gate1 = self._modulate(self.img_norm1(hidden_states), img_mod1) + del img_mod1 + txt_modulated, txt_gate1 = self._modulate(self.txt_norm1(encoder_hidden_states), txt_mod1) + del txt_mod1 img_attn_output, txt_attn_output = self.attn( hidden_states=img_modulated, @@ -248,16 +248,20 @@ def forward( image_rotary_emb=image_rotary_emb, transformer_options=transformer_options, ) + del img_modulated + del txt_modulated hidden_states = hidden_states + img_gate1 * img_attn_output encoder_hidden_states = encoder_hidden_states + txt_gate1 * txt_attn_output + del img_attn_output + del txt_attn_output + del img_gate1 + del txt_gate1 - img_normed2 = self.img_norm2(hidden_states) - img_modulated2, img_gate2 = self._modulate(img_normed2, img_mod2) + img_modulated2, img_gate2 = self._modulate(self.img_norm2(hidden_states), img_mod2) hidden_states = torch.addcmul(hidden_states, img_gate2, self.img_mlp(img_modulated2)) - txt_normed2 = self.txt_norm2(encoder_hidden_states) - txt_modulated2, txt_gate2 = self._modulate(txt_normed2, txt_mod2) + txt_modulated2, txt_gate2 = self._modulate(self.txt_norm2(encoder_hidden_states), txt_mod2) encoder_hidden_states = torch.addcmul(encoder_hidden_states, txt_gate2, self.txt_mlp(txt_modulated2)) return encoder_hidden_states, hidden_states diff --git a/comfy/model_management.py b/comfy/model_management.py index 0d040e55ea8c..a21df54b3f38 100644 --- a/comfy/model_management.py +++ b/comfy/model_management.py @@ -504,6 +504,7 @@ def model_load(self, lowvram_model_memory=0, force_patch_weights=False): if use_more_vram == 0: use_more_vram = 1e32 self.model_use_more_vram(use_more_vram, force_patch_weights=force_patch_weights) + real_model = self.model.model if is_intel_xpu() and not args.disable_ipex_optimize and 'ipex' in globals() and real_model is not None: @@ -689,7 +690,10 @@ def load_models_gpu(models, memory_required=0, force_patch_weights=False, minimu current_free_mem = get_free_memory(torch_dev) + loaded_memory lowvram_model_memory = max(128 * 1024 * 1024, (current_free_mem - minimum_memory_required), min(current_free_mem * MIN_WEIGHT_MEMORY_RATIO, current_free_mem - minimum_inference_memory())) - lowvram_model_memory = max(0.1, lowvram_model_memory - loaded_memory) + lowvram_model_memory = lowvram_model_memory - loaded_memory + + if lowvram_model_memory == 0: + lowvram_model_memory = 0.1 if vram_set_state == VRAMState.NO_VRAM: lowvram_model_memory = 0.1 @@ -1085,25 +1089,36 @@ def cast_to_device(tensor, device, dtype, copy=False): PINNED_MEMORY = {} TOTAL_PINNED_MEMORY = 0 -if PerformanceFeature.PinnedMem in args.fast: - if WINDOWS: - MAX_PINNED_MEMORY = get_total_memory(torch.device("cpu")) * 0.45 # Windows limit is apparently 50% - else: - MAX_PINNED_MEMORY = get_total_memory(torch.device("cpu")) * 0.95 -else: - MAX_PINNED_MEMORY = -1 +MAX_PINNED_MEMORY = -1 +if not args.disable_pinned_memory: + if is_nvidia() or is_amd(): + if WINDOWS: + MAX_PINNED_MEMORY = get_total_memory(torch.device("cpu")) * 0.45 # Windows limit is apparently 50% + else: + MAX_PINNED_MEMORY = get_total_memory(torch.device("cpu")) * 0.95 + logging.info("Enabled pinned memory {}".format(MAX_PINNED_MEMORY // (1024 * 1024))) + def pin_memory(tensor): global TOTAL_PINNED_MEMORY if MAX_PINNED_MEMORY <= 0: return False - if not is_nvidia(): + if type(tensor) is not torch.nn.parameter.Parameter: return False if not is_device_cpu(tensor.device): return False + if tensor.is_pinned(): + #NOTE: Cuda does detect when a tensor is already pinned and would + #error below, but there are proven cases where this also queues an error + #on the GPU async. So dont trust the CUDA API and guard here + return False + + if not tensor.is_contiguous(): + return False + size = tensor.numel() * tensor.element_size() if (TOTAL_PINNED_MEMORY + size) > MAX_PINNED_MEMORY: return False @@ -1121,13 +1136,21 @@ def unpin_memory(tensor): if MAX_PINNED_MEMORY <= 0: return False - if not is_nvidia(): - return False - if not is_device_cpu(tensor.device): return False ptr = tensor.data_ptr() + size = tensor.numel() * tensor.element_size() + + size_stored = PINNED_MEMORY.get(ptr, None) + if size_stored is None: + logging.warning("Tried to unpin tensor not pinned by ComfyUI") + return False + + if size != size_stored: + logging.warning("Size of pinned tensor changed") + return False + if torch.cuda.cudart().cudaHostUnregister(ptr) == 0: TOTAL_PINNED_MEMORY -= PINNED_MEMORY.pop(ptr) if len(PINNED_MEMORY) == 0: diff --git a/comfy/model_patcher.py b/comfy/model_patcher.py index 5a31a8734b07..cf1b0d4412bc 100644 --- a/comfy/model_patcher.py +++ b/comfy/model_patcher.py @@ -843,7 +843,7 @@ def unpatch_model(self, device_to=None, unpatch_weights=True): self.object_patches_backup.clear() - def partially_unload(self, device_to, memory_to_free=0): + def partially_unload(self, device_to, memory_to_free=0, force_patch_weights=False): with self.use_ejected(): hooks_unpatched = False memory_freed = 0 @@ -887,13 +887,19 @@ def partially_unload(self, device_to, memory_to_free=0): module_mem += move_weight_functions(m, device_to) if lowvram_possible: if weight_key in self.patches: - _, set_func, convert_func = get_key_weight(self.model, weight_key) - m.weight_function.append(LowVramPatch(weight_key, self.patches, convert_func, set_func)) - patch_counter += 1 + if force_patch_weights: + self.patch_weight_to_device(weight_key) + else: + _, set_func, convert_func = get_key_weight(self.model, weight_key) + m.weight_function.append(LowVramPatch(weight_key, self.patches, convert_func, set_func)) + patch_counter += 1 if bias_key in self.patches: - _, set_func, convert_func = get_key_weight(self.model, bias_key) - m.bias_function.append(LowVramPatch(bias_key, self.patches, convert_func, set_func)) - patch_counter += 1 + if force_patch_weights: + self.patch_weight_to_device(bias_key) + else: + _, set_func, convert_func = get_key_weight(self.model, bias_key) + m.bias_function.append(LowVramPatch(bias_key, self.patches, convert_func, set_func)) + patch_counter += 1 cast_weight = True if cast_weight: @@ -909,6 +915,7 @@ def partially_unload(self, device_to, memory_to_free=0): self.model.model_lowvram = True self.model.lowvram_patch_counter += patch_counter self.model.model_loaded_weight_memory -= memory_freed + logging.info("loaded partially: {:.2f} MB loaded, lowvram patches: {}".format(self.model.model_loaded_weight_memory / (1024 * 1024), self.model.lowvram_patch_counter)) return memory_freed def partially_load(self, device_to, extra_memory=0, force_patch_weights=False): @@ -921,6 +928,9 @@ def partially_load(self, device_to, extra_memory=0, force_patch_weights=False): extra_memory += (used - self.model.model_loaded_weight_memory) self.patch_model(load_weights=False) + if extra_memory < 0 and not unpatch_weights: + self.partially_unload(self.offload_device, -extra_memory, force_patch_weights=force_patch_weights) + return 0 full_load = False if self.model.model_lowvram == False and self.model.model_loaded_weight_memory > 0: self.apply_hooks(self.forced_hooks, force_apply=True) diff --git a/comfy/ops.py b/comfy/ops.py index 733bff99d190..2a90a5ba210b 100644 --- a/comfy/ops.py +++ b/comfy/ops.py @@ -77,7 +77,10 @@ def cast_bias_weight(s, input=None, dtype=None, device=None, bias_dtype=None, of # will add async-offload support to your cast and improve performance. if input is not None: if dtype is None: - dtype = input.dtype + if isinstance(input, QuantizedTensor): + dtype = input._layout_params["orig_dtype"] + else: + dtype = input.dtype if bias_dtype is None: bias_dtype = dtype if device is None: @@ -110,9 +113,9 @@ def cast_bias_weight(s, input=None, dtype=None, device=None, bias_dtype=None, of for f in s.bias_function: bias = f(bias) - weight = weight.to(dtype=dtype) - if weight_has_function: + if weight_has_function or weight.dtype != dtype: with wf_context: + weight = weight.to(dtype=dtype) for f in s.weight_function: weight = f(weight) @@ -534,18 +537,7 @@ def forward(self, *args, **kwargs): # ============================================================================== # Mixed Precision Operations # ============================================================================== -from .quant_ops import QuantizedTensor - -QUANT_FORMAT_MIXINS = { - "float8_e4m3fn": { - "dtype": torch.float8_e4m3fn, - "layout_type": "TensorCoreFP8Layout", - "parameters": { - "weight_scale": torch.nn.Parameter(torch.zeros((), dtype=torch.float32), requires_grad=False), - "input_scale": torch.nn.Parameter(torch.zeros((), dtype=torch.float32), requires_grad=False), - } - } -} +from .quant_ops import QuantizedTensor, QUANT_ALGOS class MixedPrecisionOps(disable_weight_init): _layer_quant_config = {} @@ -596,23 +588,24 @@ def _load_from_state_dict(self, state_dict, prefix, local_metadata, if quant_format is None: raise ValueError(f"Unknown quantization format for layer {layer_name}") - mixin = QUANT_FORMAT_MIXINS[quant_format] - self.layout_type = mixin["layout_type"] + qconfig = QUANT_ALGOS[quant_format] + self.layout_type = qconfig["comfy_tensor_layout"] - scale_key = f"{prefix}weight_scale" + weight_scale_key = f"{prefix}weight_scale" layout_params = { - 'scale': state_dict.pop(scale_key, None), - 'orig_dtype': MixedPrecisionOps._compute_dtype + 'scale': state_dict.pop(weight_scale_key, None), + 'orig_dtype': MixedPrecisionOps._compute_dtype, + 'block_size': qconfig.get("group_size", None), } if layout_params['scale'] is not None: - manually_loaded_keys.append(scale_key) + manually_loaded_keys.append(weight_scale_key) self.weight = torch.nn.Parameter( - QuantizedTensor(weight.to(device=device, dtype=mixin["dtype"]), self.layout_type, layout_params), + QuantizedTensor(weight.to(device=device), self.layout_type, layout_params), requires_grad=False ) - for param_name, param_value in mixin["parameters"].items(): + for param_name in qconfig["parameters"]: param_key = f"{prefix}{param_name}" _v = state_dict.pop(param_key, None) if _v is None: @@ -643,7 +636,7 @@ def forward(self, input, *args, **kwargs): if (getattr(self, 'layout_type', None) is not None and getattr(self, 'input_scale', None) is not None and not isinstance(input, QuantizedTensor)): - input = QuantizedTensor.from_float(input, self.layout_type, scale=self.input_scale, fp8_dtype=self.weight.dtype) + input = QuantizedTensor.from_float(input, self.layout_type, scale=self.input_scale, dtype=self.weight.dtype) return self._forward(input, self.weight, self.bias) diff --git a/comfy/quant_ops.py b/comfy/quant_ops.py index c56e32a73ad1..1d058bece5a7 100644 --- a/comfy/quant_ops.py +++ b/comfy/quant_ops.py @@ -74,6 +74,12 @@ def _copy_layout_params(params): new_params[k] = v return new_params +def _copy_layout_params_inplace(src, dst, non_blocking=False): + for k, v in src.items(): + if isinstance(v, torch.Tensor): + dst[k].copy_(v, non_blocking=non_blocking) + else: + dst[k] = v class QuantizedLayout: """ @@ -318,13 +324,13 @@ def generic_to_dtype_layout(func, args, kwargs): def generic_copy_(func, args, kwargs): qt_dest = args[0] src = args[1] - + non_blocking = args[2] if len(args) > 2 else False if isinstance(qt_dest, QuantizedTensor): if isinstance(src, QuantizedTensor): # Copy from another quantized tensor - qt_dest._qdata.copy_(src._qdata) + qt_dest._qdata.copy_(src._qdata, non_blocking=non_blocking) qt_dest._layout_type = src._layout_type - qt_dest._layout_params = _copy_layout_params(src._layout_params) + _copy_layout_params_inplace(src._layout_params, qt_dest._layout_params, non_blocking=non_blocking) else: # Copy from regular tensor - just copy raw data qt_dest._qdata.copy_(src) @@ -336,6 +342,26 @@ def generic_copy_(func, args, kwargs): def generic_has_compatible_shallow_copy_type(func, args, kwargs): return True + +@register_generic_util(torch.ops.aten.empty_like.default) +def generic_empty_like(func, args, kwargs): + """Empty_like operation - creates an empty tensor with the same quantized structure.""" + qt = args[0] + if isinstance(qt, QuantizedTensor): + # Create empty tensor with same shape and dtype as the quantized data + hp_dtype = kwargs.pop('dtype', qt._layout_params["orig_dtype"]) + new_qdata = torch.empty_like(qt._qdata, **kwargs) + + # Handle device transfer for layout params + target_device = kwargs.get('device', new_qdata.device) + new_params = _move_layout_params_to_device(qt._layout_params, target_device) + + # Update orig_dtype if dtype is specified + new_params['orig_dtype'] = hp_dtype + + return QuantizedTensor(new_qdata, qt._layout_type, new_params) + return func(*args, **kwargs) + # ============================================================================== # FP8 Layout + Operation Handlers # ============================================================================== @@ -378,6 +404,13 @@ def dequantize(qdata, scale, orig_dtype, **kwargs): def get_plain_tensors(cls, qtensor): return qtensor._qdata, qtensor._layout_params['scale'] +QUANT_ALGOS = { + "float8_e4m3fn": { + "storage_t": torch.float8_e4m3fn, + "parameters": {"weight_scale", "input_scale"}, + "comfy_tensor_layout": "TensorCoreFP8Layout", + }, +} LAYOUTS = { "TensorCoreFP8Layout": TensorCoreFP8Layout, diff --git a/comfy/sd1_clip.py b/comfy/sd1_clip.py index f8a7c2a1b43c..3066de2d7a50 100644 --- a/comfy/sd1_clip.py +++ b/comfy/sd1_clip.py @@ -460,7 +460,7 @@ def load_embed(embedding_name, embedding_directory, embedding_size, embed_key=No return embed_out class SDTokenizer: - def __init__(self, tokenizer_path=None, max_length=77, pad_with_end=True, embedding_directory=None, embedding_size=768, embedding_key='clip_l', tokenizer_class=CLIPTokenizer, has_start_token=True, has_end_token=True, pad_to_max_length=True, min_length=None, pad_token=None, end_token=None, min_padding=None, tokenizer_data={}, tokenizer_args={}): + def __init__(self, tokenizer_path=None, max_length=77, pad_with_end=True, embedding_directory=None, embedding_size=768, embedding_key='clip_l', tokenizer_class=CLIPTokenizer, has_start_token=True, has_end_token=True, pad_to_max_length=True, min_length=None, pad_token=None, end_token=None, min_padding=None, pad_left=False, tokenizer_data={}, tokenizer_args={}): if tokenizer_path is None: tokenizer_path = os.path.join(os.path.dirname(os.path.realpath(__file__)), "sd1_tokenizer") self.tokenizer = tokenizer_class.from_pretrained(tokenizer_path, **tokenizer_args) @@ -468,6 +468,7 @@ def __init__(self, tokenizer_path=None, max_length=77, pad_with_end=True, embedd self.min_length = tokenizer_data.get("{}_min_length".format(embedding_key), min_length) self.end_token = None self.min_padding = min_padding + self.pad_left = pad_left empty = self.tokenizer('')["input_ids"] self.tokenizer_adds_end_token = has_end_token @@ -522,6 +523,12 @@ def _try_get_embedding(self, embedding_name:str): return (embed, "{} {}".format(embedding_name[len(stripped):], leftover)) return (embed, leftover) + def pad_tokens(self, tokens, amount): + if self.pad_left: + for i in range(amount): + tokens.insert(0, (self.pad_token, 1.0, 0)) + else: + tokens.extend([(self.pad_token, 1.0, 0)] * amount) def tokenize_with_weights(self, text:str, return_word_ids=False, tokenizer_options={}, **kwargs): ''' @@ -600,7 +607,7 @@ def tokenize_with_weights(self, text:str, return_word_ids=False, tokenizer_optio if self.end_token is not None: batch.append((self.end_token, 1.0, 0)) if self.pad_to_max_length: - batch.extend([(self.pad_token, 1.0, 0)] * (remaining_length)) + self.pad_tokens(batch, remaining_length) #start new batch batch = [] if self.start_token is not None: @@ -614,11 +621,11 @@ def tokenize_with_weights(self, text:str, return_word_ids=False, tokenizer_optio if self.end_token is not None: batch.append((self.end_token, 1.0, 0)) if min_padding is not None: - batch.extend([(self.pad_token, 1.0, 0)] * min_padding) + self.pad_tokens(batch, min_padding) if self.pad_to_max_length and len(batch) < self.max_length: - batch.extend([(self.pad_token, 1.0, 0)] * (self.max_length - len(batch))) + self.pad_tokens(batch, self.max_length - len(batch)) if min_length is not None and len(batch) < min_length: - batch.extend([(self.pad_token, 1.0, 0)] * (min_length - len(batch))) + self.pad_tokens(batch, min_length - len(batch)) if not return_word_ids: batched_tokens = [[(t, w) for t, w,_ in x] for x in batched_tokens] diff --git a/comfy_api_nodes/apis/PixverseController.py b/comfy_api_nodes/apis/PixverseController.py deleted file mode 100644 index 310c0f54655d..000000000000 --- a/comfy_api_nodes/apis/PixverseController.py +++ /dev/null @@ -1,17 +0,0 @@ -# generated by datamodel-codegen: -# filename: filtered-openapi.yaml -# timestamp: 2025-04-29T23:44:54+00:00 - -from __future__ import annotations - -from typing import Optional - -from pydantic import BaseModel - -from . import PixverseDto - - -class ResponseData(BaseModel): - ErrCode: Optional[int] = None - ErrMsg: Optional[str] = None - Resp: Optional[PixverseDto.V2OpenAPII2VResp] = None diff --git a/comfy_api_nodes/apis/PixverseDto.py b/comfy_api_nodes/apis/PixverseDto.py deleted file mode 100644 index 323c38e96390..000000000000 --- a/comfy_api_nodes/apis/PixverseDto.py +++ /dev/null @@ -1,57 +0,0 @@ -# generated by datamodel-codegen: -# filename: filtered-openapi.yaml -# timestamp: 2025-04-29T23:44:54+00:00 - -from __future__ import annotations - -from typing import Optional - -from pydantic import BaseModel, Field - - -class V2OpenAPII2VResp(BaseModel): - video_id: Optional[int] = Field(None, description='Video_id') - - -class V2OpenAPIT2VReq(BaseModel): - aspect_ratio: str = Field( - ..., description='Aspect ratio (16:9, 4:3, 1:1, 3:4, 9:16)', examples=['16:9'] - ) - duration: int = Field( - ..., - description='Video duration (5, 8 seconds, --model=v3.5 only allows 5,8; --quality=1080p does not support 8s)', - examples=[5], - ) - model: str = Field( - ..., description='Model version (only supports v3.5)', examples=['v3.5'] - ) - motion_mode: Optional[str] = Field( - 'normal', - description='Motion mode (normal, fast, --fast only available when duration=5; --quality=1080p does not support fast)', - examples=['normal'], - ) - negative_prompt: Optional[str] = Field( - None, description='Negative prompt\n', max_length=2048 - ) - prompt: str = Field(..., description='Prompt', max_length=2048) - quality: str = Field( - ..., - description='Video quality ("360p"(Turbo model), "540p", "720p", "1080p")', - examples=['540p'], - ) - seed: Optional[int] = Field(None, description='Random seed, range: 0 - 2147483647') - style: Optional[str] = Field( - None, - description='Style (effective when model=v3.5, "anime", "3d_animation", "clay", "comic", "cyberpunk") Do not include style parameter unless needed', - examples=['anime'], - ) - template_id: Optional[int] = Field( - None, - description='Template ID (template_id must be activated before use)', - examples=[302325299692608], - ) - water_mark: Optional[bool] = Field( - False, - description='Watermark (true: add watermark, false: no watermark)', - examples=[False], - ) diff --git a/comfy_api_nodes/apis/client.py b/comfy_api_nodes/apis/client.py deleted file mode 100644 index bdaddcc8812b..000000000000 --- a/comfy_api_nodes/apis/client.py +++ /dev/null @@ -1,981 +0,0 @@ -""" -API Client Framework for api.comfy.org. - -This module provides a flexible framework for making API requests from ComfyUI nodes. -It supports both synchronous and asynchronous API operations with proper type validation. - -Key Components: --------------- -1. ApiClient - Handles HTTP requests with authentication and error handling -2. ApiEndpoint - Defines a single HTTP endpoint with its request/response models -3. ApiOperation - Executes a single synchronous API operation - -Usage Examples: --------------- - -# Example 1: Synchronous API Operation -# ------------------------------------ -# For a simple API call that returns the result immediately: - -# 1. Create the API client -api_client = ApiClient( - base_url="https://api.example.com", - auth_token="your_auth_token_here", - comfy_api_key="your_comfy_api_key_here", - timeout=30.0, - verify_ssl=True -) - -# 2. Define the endpoint -user_info_endpoint = ApiEndpoint( - path="/v1/users/me", - method=HttpMethod.GET, - request_model=EmptyRequest, # No request body needed - response_model=UserProfile, # Pydantic model for the response - query_params=None -) - -# 3. Create the request object -request = EmptyRequest() - -# 4. Create and execute the operation -operation = ApiOperation( - endpoint=user_info_endpoint, - request=request -) -user_profile = await operation.execute(client=api_client) # Returns immediately with the result - - -# Example 2: Asynchronous API Operation with Polling -# ------------------------------------------------- -# For an API that starts a task and requires polling for completion: - -# 1. Define the endpoints (initial request and polling) -generate_image_endpoint = ApiEndpoint( - path="/v1/images/generate", - method=HttpMethod.POST, - request_model=ImageGenerationRequest, - response_model=TaskCreatedResponse, - query_params=None -) - -check_task_endpoint = ApiEndpoint( - path="/v1/tasks/{task_id}", - method=HttpMethod.GET, - request_model=EmptyRequest, - response_model=ImageGenerationResult, - query_params=None -) - -# 2. Create the request object -request = ImageGenerationRequest( - prompt="a beautiful sunset over mountains", - width=1024, - height=1024, - num_images=1 -) - -# 3. Create and execute the polling operation -operation = PollingOperation( - initial_endpoint=generate_image_endpoint, - initial_request=request, - poll_endpoint=check_task_endpoint, - task_id_field="task_id", - status_field="status", - completed_statuses=["completed"], - failed_statuses=["failed", "error"] -) - -# This will make the initial request and then poll until completion -result = await operation.execute(client=api_client) # Returns the final ImageGenerationResult when done -""" - -from __future__ import annotations -import aiohttp -import asyncio -import logging -import io -import os -import socket -from aiohttp.client_exceptions import ClientError, ClientResponseError -from typing import Type, Optional, Any, TypeVar, Generic, Callable -from enum import Enum -import json -from urllib.parse import urljoin, urlparse -from pydantic import BaseModel, Field -import uuid # For generating unique operation IDs - -from server import PromptServer -from comfy.cli_args import args -from comfy import utils -from . import request_logger - -T = TypeVar("T", bound=BaseModel) -R = TypeVar("R", bound=BaseModel) -P = TypeVar("P", bound=BaseModel) # For poll response - -PROGRESS_BAR_MAX = 100 - - -class NetworkError(Exception): - """Base exception for network-related errors with diagnostic information.""" - pass - - -class LocalNetworkError(NetworkError): - """Exception raised when local network connectivity issues are detected.""" - pass - - -class ApiServerError(NetworkError): - """Exception raised when the API server is unreachable but internet is working.""" - pass - - -class EmptyRequest(BaseModel): - """Base class for empty request bodies. - For GET requests, fields will be sent as query parameters.""" - - pass - - -class UploadRequest(BaseModel): - file_name: str = Field(..., description="Filename to upload") - content_type: Optional[str] = Field( - None, - description="Mime type of the file. For example: image/png, image/jpeg, video/mp4, etc.", - ) - - -class UploadResponse(BaseModel): - download_url: str = Field(..., description="URL to GET uploaded file") - upload_url: str = Field(..., description="URL to PUT file to upload") - - -class HttpMethod(str, Enum): - GET = "GET" - POST = "POST" - PUT = "PUT" - DELETE = "DELETE" - PATCH = "PATCH" - - -class ApiClient: - """ - Client for making HTTP requests to an API with authentication, error handling, and retry logic. - """ - - def __init__( - self, - base_url: str, - auth_token: Optional[str] = None, - comfy_api_key: Optional[str] = None, - timeout: float = 3600.0, - verify_ssl: bool = True, - max_retries: int = 3, - retry_delay: float = 1.0, - retry_backoff_factor: float = 2.0, - retry_status_codes: Optional[tuple[int, ...]] = None, - session: Optional[aiohttp.ClientSession] = None, - ): - self.base_url = base_url - self.auth_token = auth_token - self.comfy_api_key = comfy_api_key - self.timeout = timeout - self.verify_ssl = verify_ssl - self.max_retries = max_retries - self.retry_delay = retry_delay - self.retry_backoff_factor = retry_backoff_factor - # Default retry status codes: 408 (Request Timeout), 429 (Too Many Requests), - # 500, 502, 503, 504 (Server Errors) - self.retry_status_codes = retry_status_codes or (408, 429, 500, 502, 503, 504) - self._session: Optional[aiohttp.ClientSession] = session - self._owns_session = session is None # Track if we have to close it - - @staticmethod - def _generate_operation_id(path: str) -> str: - """Generates a unique operation ID for logging.""" - return f"{path.strip('/').replace('/', '_')}_{uuid.uuid4().hex[:8]}" - - @staticmethod - def _create_json_payload_args( - data: Optional[dict[str, Any]] = None, - headers: Optional[dict[str, str]] = None, - ) -> dict[str, Any]: - return { - "json": data, - "headers": headers, - } - - def _create_form_data_args( - self, - data: dict[str, Any] | None, - files: dict[str, Any] | None, - headers: Optional[dict[str, str]] = None, - multipart_parser: Callable | None = None, - ) -> dict[str, Any]: - if headers and "Content-Type" in headers: - del headers["Content-Type"] - - if multipart_parser and data: - data = multipart_parser(data) - - if isinstance(data, aiohttp.FormData): - form = data # If the parser already returned a FormData, pass it through - else: - form = aiohttp.FormData(default_to_multipart=True) - if data: # regular text fields - for k, v in data.items(): - if v is None: - continue # aiohttp fails to serialize "None" values - # aiohttp expects strings or bytes; convert enums etc. - form.add_field(k, str(v) if not isinstance(v, (bytes, bytearray)) else v) - - if files: - file_iter = files if isinstance(files, list) else files.items() - for field_name, file_obj in file_iter: - if file_obj is None: - continue # aiohttp fails to serialize "None" values - # file_obj can be (filename, bytes/io.BytesIO, content_type) tuple - if isinstance(file_obj, tuple): - filename, file_value, content_type = self._unpack_tuple(file_obj) - else: - file_value = file_obj - filename = getattr(file_obj, "name", field_name) - content_type = "application/octet-stream" - - form.add_field( - name=field_name, - value=file_value, - filename=filename, - content_type=content_type, - ) - return {"data": form, "headers": headers or {}} - - @staticmethod - def _create_urlencoded_form_data_args( - data: dict[str, Any], - headers: Optional[dict[str, str]] = None, - ) -> dict[str, Any]: - headers = headers or {} - headers["Content-Type"] = "application/x-www-form-urlencoded" - return { - "data": data, - "headers": headers, - } - - def get_headers(self) -> dict[str, str]: - """Get headers for API requests, including authentication if available""" - headers = {"Content-Type": "application/json", "Accept": "application/json"} - - if self.auth_token: - headers["Authorization"] = f"Bearer {self.auth_token}" - elif self.comfy_api_key: - headers["X-API-KEY"] = self.comfy_api_key - - return headers - - async def _check_connectivity(self, target_url: str) -> dict[str, bool]: - """ - Check connectivity to determine if network issues are local or server-related. - - Args: - target_url: URL to check connectivity to - - Returns: - Dictionary with connectivity status details - """ - results = { - "internet_accessible": False, - "api_accessible": False, - "is_local_issue": False, - "is_api_issue": False, - } - timeout = aiohttp.ClientTimeout(total=5.0) - async with aiohttp.ClientSession(timeout=timeout) as session: - try: - async with session.get("https://www.google.com", ssl=self.verify_ssl) as resp: - results["internet_accessible"] = resp.status < 500 - except (ClientError, asyncio.TimeoutError, socket.gaierror): - results["is_local_issue"] = True - return results # cannot reach the internet – early exit - - # Now check API health endpoint - parsed = urlparse(target_url) - health_url = f"{parsed.scheme}://{parsed.netloc}/health" - try: - async with session.get(health_url, ssl=self.verify_ssl) as resp: - results["api_accessible"] = resp.status < 500 - except ClientError: - pass # leave as False - - results["is_api_issue"] = results["internet_accessible"] and not results["api_accessible"] - return results - - async def request( - self, - method: str, - path: str, - params: Optional[dict[str, Any]] = None, - data: Optional[dict[str, Any]] = None, - files: Optional[dict[str, Any] | list[tuple[str, Any]]] = None, - headers: Optional[dict[str, str]] = None, - content_type: str = "application/json", - multipart_parser: Callable | None = None, - retry_count: int = 0, # Used internally for tracking retries - ) -> dict[str, Any]: - """ - Make an HTTP request to the API with automatic retries for transient errors. - - Args: - method: HTTP method (GET, POST, etc.) - path: API endpoint path (will be joined with base_url) - params: Query parameters - data: body data - files: Files to upload - headers: Additional headers - content_type: Content type of the request. Defaults to application/json. - retry_count: Internal parameter for tracking retries, do not set manually - - Returns: - Parsed JSON response - - Raises: - LocalNetworkError: If local network connectivity issues are detected - ApiServerError: If the API server is unreachable but internet is working - Exception: For other request failures - """ - - # Build full URL and merge headers - relative_path = path.lstrip("/") - url = urljoin(self.base_url, relative_path) - self._check_auth(self.auth_token, self.comfy_api_key) - - request_headers = self.get_headers() - if headers: - request_headers.update(headers) - if files: - request_headers.pop("Content-Type", None) - if params: - params = {k: v for k, v in params.items() if v is not None} # aiohttp fails to serialize None values - - logging.debug("[DEBUG] Request Headers: %s", request_headers) - logging.debug("[DEBUG] Files: %s", files) - logging.debug("[DEBUG] Params: %s", params) - logging.debug("[DEBUG] Data: %s", data) - - if content_type == "application/x-www-form-urlencoded": - payload_args = self._create_urlencoded_form_data_args(data or {}, request_headers) - elif content_type == "multipart/form-data": - payload_args = self._create_form_data_args(data, files, request_headers, multipart_parser) - else: - payload_args = self._create_json_payload_args(data, request_headers) - - operation_id = self._generate_operation_id(path) - request_logger.log_request_response( - operation_id=operation_id, - request_method=method, - request_url=url, - request_headers=request_headers, - request_params=params, - request_data=data if content_type == "application/json" else "[form-data or other]", - ) - - session = await self._get_session() - try: - async with session.request( - method, - url, - params=params, - ssl=self.verify_ssl, - **payload_args, - ) as resp: - if resp.status >= 400: - try: - error_data = await resp.json() - except (aiohttp.ContentTypeError, json.JSONDecodeError): - error_data = await resp.text() - - return await self._handle_http_error( - ClientResponseError(resp.request_info, resp.history, status=resp.status, message=error_data), - operation_id, - method, - url, - params, - data, - files, - headers, - content_type, - multipart_parser, - retry_count=retry_count, - response_content=error_data, - ) - - # Success – parse JSON (safely) and log - try: - payload = await resp.json() - response_content_to_log = payload - except (aiohttp.ContentTypeError, json.JSONDecodeError): - payload = {} - response_content_to_log = await resp.text() - - request_logger.log_request_response( - operation_id=operation_id, - request_method=method, - request_url=url, - response_status_code=resp.status, - response_headers=dict(resp.headers), - response_content=response_content_to_log, - ) - return payload - - except (ClientError, asyncio.TimeoutError, socket.gaierror) as e: - # Treat as *connection* problem – optionally retry, else escalate - if retry_count < self.max_retries: - delay = self.retry_delay * (self.retry_backoff_factor ** retry_count) - logging.warning("Connection error. Retrying in %.2fs (%s/%s): %s", delay, retry_count + 1, - self.max_retries, str(e)) - await asyncio.sleep(delay) - return await self.request( - method, - path, - params=params, - data=data, - files=files, - headers=headers, - content_type=content_type, - multipart_parser=multipart_parser, - retry_count=retry_count + 1, - ) - # One final connectivity check for diagnostics - connectivity = await self._check_connectivity(self.base_url) - if connectivity["is_local_issue"]: - raise LocalNetworkError( - "Unable to connect to the API server due to local network issues. " - "Please check your internet connection and try again." - ) from e - raise ApiServerError( - f"The API server at {self.base_url} is currently unreachable. " - f"The service may be experiencing issues. Please try again later." - ) from e - - @staticmethod - def _check_auth(auth_token, comfy_api_key): - """Verify that an auth token is present or comfy_api_key is present""" - if auth_token is None and comfy_api_key is None: - raise Exception("Unauthorized: Please login first to use this node.") - return auth_token or comfy_api_key - - @staticmethod - async def upload_file( - upload_url: str, - file: io.BytesIO | str, - content_type: str | None = None, - max_retries: int = 3, - retry_delay: float = 1.0, - retry_backoff_factor: float = 2.0, - ) -> aiohttp.ClientResponse: - """Upload a file to the API with retry logic. - - Args: - upload_url: The URL to upload to - file: Either a file path string, BytesIO object, or tuple of (file_path, filename) - content_type: Optional mime type to set for the upload - max_retries: Maximum number of retry attempts - retry_delay: Initial delay between retries in seconds - retry_backoff_factor: Multiplier for the delay after each retry - """ - headers: dict[str, str] = {} - skip_auto_headers: set[str] = set() - if content_type: - headers["Content-Type"] = content_type - else: - # tell aiohttp not to add Content-Type that will break the request signature and result in a 403 status. - skip_auto_headers.add("Content-Type") - - # Extract file bytes - if isinstance(file, io.BytesIO): - file.seek(0) - data = file.read() - elif isinstance(file, str): - with open(file, "rb") as f: - data = f.read() - else: - raise ValueError("File must be BytesIO or str path") - - parsed = urlparse(upload_url) - basename = os.path.basename(parsed.path) or parsed.netloc or "upload" - operation_id = f"upload_{basename}_{uuid.uuid4().hex[:8]}" - request_logger.log_request_response( - operation_id=operation_id, - request_method="PUT", - request_url=upload_url, - request_headers=headers, - request_data=f"[File data {len(data)} bytes]", - ) - - delay = retry_delay - for attempt in range(max_retries + 1): - try: - timeout = aiohttp.ClientTimeout(total=None) # honour server side timeouts - async with aiohttp.ClientSession(timeout=timeout) as session: - async with session.put( - upload_url, data=data, headers=headers, skip_auto_headers=skip_auto_headers, - ) as resp: - resp.raise_for_status() - request_logger.log_request_response( - operation_id=operation_id, - request_method="PUT", - request_url=upload_url, - response_status_code=resp.status, - response_headers=dict(resp.headers), - response_content="File uploaded successfully.", - ) - return resp - except (ClientError, asyncio.TimeoutError) as e: - request_logger.log_request_response( - operation_id=operation_id, - request_method="PUT", - request_url=upload_url, - response_status_code=e.status if hasattr(e, "status") else None, - response_headers=dict(e.headers) if hasattr(e, "headers") else None, - response_content=None, - error_message=f"{type(e).__name__}: {str(e)}", - ) - if attempt < max_retries: - logging.warning( - "Upload failed (%s/%s). Retrying in %.2fs. %s", attempt + 1, max_retries, delay, str(e) - ) - await asyncio.sleep(delay) - delay *= retry_backoff_factor - else: - raise NetworkError(f"Failed to upload file after {max_retries + 1} attempts: {e}") from e - - async def _handle_http_error( - self, - exc: ClientResponseError, - operation_id: str, - *req_meta, - retry_count: int, - response_content: dict | str = "", - ) -> dict[str, Any]: - status_code = exc.status - if status_code == 401: - user_friendly = "Unauthorized: Please login first to use this node." - elif status_code == 402: - user_friendly = "Payment Required: Please add credits to your account to use this node." - elif status_code == 409: - user_friendly = "There is a problem with your account. Please contact support@comfy.org." - elif status_code == 429: - user_friendly = "Rate Limit Exceeded: Please try again later." - else: - if isinstance(response_content, dict): - if "error" in response_content and "message" in response_content["error"]: - user_friendly = f"API Error: {response_content['error']['message']}" - if "type" in response_content["error"]: - user_friendly += f" (Type: {response_content['error']['type']})" - else: # Handle cases where error is just a JSON dict with unknown format - user_friendly = f"API Error: {json.dumps(response_content)}" - else: - if len(response_content) < 200: # Arbitrary limit for display - user_friendly = f"API Error (raw): {response_content}" - else: - user_friendly = f"API Error (raw, status {response_content})" - - request_logger.log_request_response( - operation_id=operation_id, - request_method=req_meta[0], - request_url=req_meta[1], - response_status_code=exc.status, - response_headers=dict(req_meta[5]) if req_meta[5] else None, - response_content=response_content, - error_message=f"HTTP Error {exc.status}", - ) - - logging.debug("[DEBUG] API Error: %s (Status: %s)", user_friendly, status_code) - if response_content: - logging.debug("[DEBUG] Response content: %s", response_content) - - # Retry if eligible - if status_code in self.retry_status_codes and retry_count < self.max_retries: - delay = self.retry_delay * (self.retry_backoff_factor ** retry_count) - logging.warning( - "HTTP error %s. Retrying in %.2fs (%s/%s)", - status_code, - delay, - retry_count + 1, - self.max_retries, - ) - await asyncio.sleep(delay) - return await self.request( - req_meta[0], # method - req_meta[1].replace(self.base_url, ""), # path - params=req_meta[2], - data=req_meta[3], - files=req_meta[4], - headers=req_meta[5], - content_type=req_meta[6], - multipart_parser=req_meta[7], - retry_count=retry_count + 1, - ) - - raise Exception(user_friendly) from exc - - @staticmethod - def _unpack_tuple(t): - """Helper to normalise (filename, file, content_type) tuples.""" - if len(t) == 3: - return t - elif len(t) == 2: - return t[0], t[1], "application/octet-stream" - else: - raise ValueError("files tuple must be (filename, file[, content_type])") - - async def _get_session(self) -> aiohttp.ClientSession: - if self._session is None or self._session.closed: - timeout = aiohttp.ClientTimeout(total=self.timeout) - self._session = aiohttp.ClientSession(timeout=timeout) - self._owns_session = True - return self._session - - async def close(self) -> None: - if self._owns_session and self._session and not self._session.closed: - await self._session.close() - - async def __aenter__(self) -> "ApiClient": - """Allow usage as async‑context‑manager – ensures clean teardown""" - return self - - async def __aexit__(self, exc_type, exc, tb): - await self.close() - - -class ApiEndpoint(Generic[T, R]): - """Defines an API endpoint with its request and response types""" - - def __init__( - self, - path: str, - method: HttpMethod, - request_model: Type[T], - response_model: Type[R], - query_params: Optional[dict[str, Any]] = None, - ): - """Initialize an API endpoint definition. - - Args: - path: The URL path for this endpoint, can include placeholders like {id} - method: The HTTP method to use (GET, POST, etc.) - request_model: Pydantic model class that defines the structure and validation rules for API requests to this endpoint - response_model: Pydantic model class that defines the structure and validation rules for API responses from this endpoint - query_params: Optional dictionary of query parameters to include in the request - """ - self.path = path - self.method = method - self.request_model = request_model - self.response_model = response_model - self.query_params = query_params or {} - - -class SynchronousOperation(Generic[T, R]): - """Represents a single synchronous API operation.""" - - def __init__( - self, - endpoint: ApiEndpoint[T, R], - request: T, - files: Optional[dict[str, Any] | list[tuple[str, Any]]] = None, - api_base: str | None = None, - auth_token: Optional[str] = None, - comfy_api_key: Optional[str] = None, - auth_kwargs: Optional[dict[str, str]] = None, - timeout: float = 7200.0, - verify_ssl: bool = True, - content_type: str = "application/json", - multipart_parser: Callable | None = None, - max_retries: int = 3, - retry_delay: float = 1.0, - retry_backoff_factor: float = 2.0, - ) -> None: - self.endpoint = endpoint - self.request = request - self.files = files - self.api_base: str = api_base or args.comfy_api_base - self.auth_token = auth_token - self.comfy_api_key = comfy_api_key - if auth_kwargs is not None: - self.auth_token = auth_kwargs.get("auth_token", self.auth_token) - self.comfy_api_key = auth_kwargs.get("comfy_api_key", self.comfy_api_key) - self.timeout = timeout - self.verify_ssl = verify_ssl - self.content_type = content_type - self.multipart_parser = multipart_parser - self.max_retries = max_retries - self.retry_delay = retry_delay - self.retry_backoff_factor = retry_backoff_factor - - async def execute(self, client: Optional[ApiClient] = None) -> R: - owns_client = client is None - if owns_client: - client = ApiClient( - base_url=self.api_base, - auth_token=self.auth_token, - comfy_api_key=self.comfy_api_key, - timeout=self.timeout, - verify_ssl=self.verify_ssl, - max_retries=self.max_retries, - retry_delay=self.retry_delay, - retry_backoff_factor=self.retry_backoff_factor, - ) - - try: - request_dict: Optional[dict[str, Any]] - if isinstance(self.request, EmptyRequest): - request_dict = None - else: - request_dict = self.request.model_dump(exclude_none=True) - for k, v in list(request_dict.items()): - if isinstance(v, Enum): - request_dict[k] = v.value - - logging.debug("[DEBUG] API Request: %s %s", self.endpoint.method.value, self.endpoint.path) - logging.debug("[DEBUG] Request Data: %s", json.dumps(request_dict, indent=2)) - logging.debug("[DEBUG] Query Params: %s", self.endpoint.query_params) - - response_json = await client.request( - self.endpoint.method.value, - self.endpoint.path, - params=self.endpoint.query_params, - data=request_dict, - files=self.files, - content_type=self.content_type, - multipart_parser=self.multipart_parser, - ) - - logging.debug("=" * 50) - logging.debug("[DEBUG] RESPONSE DETAILS:") - logging.debug("[DEBUG] Status Code: 200 (Success)") - logging.debug("[DEBUG] Response Body: %s", json.dumps(response_json, indent=2)) - logging.debug("=" * 50) - - parsed_response = self.endpoint.response_model.model_validate(response_json) - logging.debug("[DEBUG] Parsed Response: %s", parsed_response) - return parsed_response - finally: - if owns_client: - await client.close() - - -class TaskStatus(str, Enum): - """Enum for task status values""" - - COMPLETED = "completed" - FAILED = "failed" - PENDING = "pending" - - -class PollingOperation(Generic[T, R]): - """Represents an asynchronous API operation that requires polling for completion.""" - - def __init__( - self, - poll_endpoint: ApiEndpoint[EmptyRequest, R], - completed_statuses: list[str], - failed_statuses: list[str], - *, - status_extractor: Callable[[R], Optional[str]], - progress_extractor: Callable[[R], Optional[float]] | None = None, - result_url_extractor: Callable[[R], Optional[str]] | None = None, - price_extractor: Callable[[R], Optional[float]] | None = None, - request: Optional[T] = None, - api_base: str | None = None, - auth_token: Optional[str] = None, - comfy_api_key: Optional[str] = None, - auth_kwargs: Optional[dict[str, str]] = None, - poll_interval: float = 5.0, - max_poll_attempts: int = 120, # Default max polling attempts (10 minutes with 5s interval) - max_retries: int = 3, # Max retries per individual API call - retry_delay: float = 1.0, - retry_backoff_factor: float = 2.0, - estimated_duration: Optional[float] = None, - node_id: Optional[str] = None, - ) -> None: - self.poll_endpoint = poll_endpoint - self.request = request - self.api_base: str = api_base or args.comfy_api_base - self.auth_token = auth_token - self.comfy_api_key = comfy_api_key - if auth_kwargs is not None: - self.auth_token = auth_kwargs.get("auth_token", self.auth_token) - self.comfy_api_key = auth_kwargs.get("comfy_api_key", self.comfy_api_key) - self.poll_interval = poll_interval - self.max_poll_attempts = max_poll_attempts - self.max_retries = max_retries - self.retry_delay = retry_delay - self.retry_backoff_factor = retry_backoff_factor - self.estimated_duration = estimated_duration - self.status_extractor = status_extractor or (lambda x: getattr(x, "status", None)) - self.progress_extractor = progress_extractor - self.result_url_extractor = result_url_extractor - self.price_extractor = price_extractor - self.node_id = node_id - self.completed_statuses = completed_statuses - self.failed_statuses = failed_statuses - self.final_response: Optional[R] = None - self.extracted_price: Optional[float] = None - - async def execute(self, client: Optional[ApiClient] = None) -> R: - owns_client = client is None - if owns_client: - client = ApiClient( - base_url=self.api_base, - auth_token=self.auth_token, - comfy_api_key=self.comfy_api_key, - max_retries=self.max_retries, - retry_delay=self.retry_delay, - retry_backoff_factor=self.retry_backoff_factor, - ) - try: - return await self._poll_until_complete(client) - finally: - if owns_client: - await client.close() - - def _display_text_on_node(self, text: str): - if not self.node_id: - return - if self.extracted_price is not None: - text = f"Price: ${self.extracted_price}\n{text}" - PromptServer.instance.send_progress_text(text, self.node_id) - - def _display_time_progress_on_node(self, time_completed: int | float): - if not self.node_id: - return - if self.estimated_duration is not None: - remaining = max(0, int(self.estimated_duration) - time_completed) - message = f"Task in progress: {time_completed}s (~{remaining}s remaining)" - else: - message = f"Task in progress: {time_completed}s" - self._display_text_on_node(message) - - def _check_task_status(self, response: R) -> TaskStatus: - try: - status = self.status_extractor(response) - if status in self.completed_statuses: - return TaskStatus.COMPLETED - if status in self.failed_statuses: - return TaskStatus.FAILED - return TaskStatus.PENDING - except Exception as e: - logging.error("Error extracting status: %s", e) - return TaskStatus.PENDING - - async def _poll_until_complete(self, client: ApiClient) -> R: - """Poll until the task is complete""" - consecutive_errors = 0 - max_consecutive_errors = min(5, self.max_retries * 2) # Limit consecutive errors - - if self.progress_extractor: - progress = utils.ProgressBar(PROGRESS_BAR_MAX) - - status = TaskStatus.PENDING - for poll_count in range(1, self.max_poll_attempts + 1): - try: - logging.debug("[DEBUG] Polling attempt #%s", poll_count) - - request_dict = None if self.request is None else self.request.model_dump(exclude_none=True) - - if poll_count == 1: - logging.debug( - "[DEBUG] Poll Request: %s %s", - self.poll_endpoint.method.value, - self.poll_endpoint.path, - ) - logging.debug( - "[DEBUG] Poll Request Data: %s", - json.dumps(request_dict, indent=2) if request_dict else "None", - ) - - # Query task status - resp = await client.request( - self.poll_endpoint.method.value, - self.poll_endpoint.path, - params=self.poll_endpoint.query_params, - data=request_dict, - ) - consecutive_errors = 0 # reset on success - response_obj: R = self.poll_endpoint.response_model.model_validate(resp) - - # Check if task is complete - status = self._check_task_status(response_obj) - logging.debug("[DEBUG] Task Status: %s", status) - - # If progress extractor is provided, extract progress - if self.progress_extractor: - new_progress = self.progress_extractor(response_obj) - if new_progress is not None: - progress.update_absolute(new_progress, total=PROGRESS_BAR_MAX) - - if self.price_extractor: - price = self.price_extractor(response_obj) - if price is not None: - self.extracted_price = price - - if status == TaskStatus.COMPLETED: - message = "Task completed successfully" - if self.result_url_extractor: - result_url = self.result_url_extractor(response_obj) - if result_url: - message = f"Result URL: {result_url}" - logging.debug("[DEBUG] %s", message) - self._display_text_on_node(message) - self.final_response = response_obj - if self.progress_extractor: - progress.update(100) - return self.final_response - if status == TaskStatus.FAILED: - message = f"Task failed: {json.dumps(resp)}" - logging.error("[DEBUG] %s", message) - raise Exception(message) - logging.debug("[DEBUG] Task still pending, continuing to poll...") - # Task pending – wait - for i in range(int(self.poll_interval)): - self._display_time_progress_on_node((poll_count - 1) * self.poll_interval + i) - await asyncio.sleep(1) - - except (LocalNetworkError, ApiServerError, NetworkError) as e: - consecutive_errors += 1 - if consecutive_errors >= max_consecutive_errors: - raise Exception( - f"Polling aborted after {consecutive_errors} network errors: {str(e)}" - ) from e - logging.warning( - "Network error (%s/%s): %s", - consecutive_errors, - max_consecutive_errors, - str(e), - ) - await asyncio.sleep(self.poll_interval) - except Exception as e: - # For other errors, increment count and potentially abort - consecutive_errors += 1 - if consecutive_errors >= max_consecutive_errors or status == TaskStatus.FAILED: - raise Exception( - f"Polling aborted after {consecutive_errors} consecutive errors: {str(e)}" - ) from e - - logging.error("[DEBUG] Polling error: %s", str(e)) - logging.warning( - "Error during polling (attempt %s/%s): %s. Will retry in %s seconds.", - poll_count, - self.max_poll_attempts, - str(e), - self.poll_interval, - ) - await asyncio.sleep(self.poll_interval) - - # If we've exhausted all polling attempts - raise Exception( - f"Polling timed out after {self.max_poll_attempts} attempts (" f"{self.max_poll_attempts * self.poll_interval} seconds). " - "The operation may still be running on the server but is taking longer than expected." - ) diff --git a/comfy_api_nodes/nodes_rodin.py b/comfy_api_nodes/nodes_rodin.py index ad402923645c..e60e7a6d67ed 100644 --- a/comfy_api_nodes/nodes_rodin.py +++ b/comfy_api_nodes/nodes_rodin.py @@ -5,12 +5,9 @@ """ -from __future__ import annotations from inspect import cleandoc import folder_paths as comfy_paths -import aiohttp import os -import asyncio import logging import math from typing import Optional @@ -26,11 +23,11 @@ Rodin3DDownloadResponse, JobStatus, ) -from comfy_api_nodes.apis.client import ( +from comfy_api_nodes.util import ( + sync_op, + poll_op, ApiEndpoint, - HttpMethod, - SynchronousOperation, - PollingOperation, + download_url_to_bytesio, ) from comfy_api.latest import ComfyExtension, IO @@ -121,35 +118,31 @@ def tensor_to_filelike(tensor, max_pixels: int = 2048*2048): async def create_generate_task( + cls: type[IO.ComfyNode], images=None, seed=1, material="PBR", quality_override=18000, tier="Regular", mesh_mode="Quad", - TAPose = False, - auth_kwargs: Optional[dict[str, str]] = None, + ta_pose: bool = False, ): if images is None: raise Exception("Rodin 3D generate requires at least 1 image.") if len(images) > 5: raise Exception("Rodin 3D generate requires up to 5 image.") - path = "/proxy/rodin/api/v2/rodin" - operation = SynchronousOperation( - endpoint=ApiEndpoint( - path=path, - method=HttpMethod.POST, - request_model=Rodin3DGenerateRequest, - response_model=Rodin3DGenerateResponse, - ), - request=Rodin3DGenerateRequest( + response = await sync_op( + cls, + ApiEndpoint(path="/proxy/rodin/api/v2/rodin", method="POST"), + response_model=Rodin3DGenerateResponse, + data=Rodin3DGenerateRequest( seed=seed, tier=tier, material=material, quality_override=quality_override, mesh_mode=mesh_mode, - TAPose=TAPose, + TAPose=ta_pose, ), files=[ ( @@ -159,11 +152,8 @@ async def create_generate_task( for image in images if image is not None ], content_type="multipart/form-data", - auth_kwargs=auth_kwargs, ) - response = await operation.execute() - if hasattr(response, "error"): error_message = f"Rodin3D Create 3D generate Task Failed. Message: {response.message}, error: {response.error}" logging.error(error_message) @@ -187,74 +177,46 @@ def check_rodin_status(response: Rodin3DCheckStatusResponse) -> str: return "DONE" return "Generating" +def extract_progress(response: Rodin3DCheckStatusResponse) -> Optional[int]: + if not response.jobs: + return None + completed_count = sum(1 for job in response.jobs if job.status == JobStatus.Done) + return int((completed_count / len(response.jobs)) * 100) -async def poll_for_task_status( - subscription_key, auth_kwargs: Optional[dict[str, str]] = None, -) -> Rodin3DCheckStatusResponse: - poll_operation = PollingOperation( - poll_endpoint=ApiEndpoint( - path="/proxy/rodin/api/v2/status", - method=HttpMethod.POST, - request_model=Rodin3DCheckStatusRequest, - response_model=Rodin3DCheckStatusResponse, - ), - request=Rodin3DCheckStatusRequest(subscription_key=subscription_key), - completed_statuses=["DONE"], - failed_statuses=["FAILED"], + +async def poll_for_task_status(subscription_key: str, cls: type[IO.ComfyNode]) -> Rodin3DCheckStatusResponse: + logging.info("[ Rodin3D API - CheckStatus ] Generate Start!") + return await poll_op( + cls, + ApiEndpoint(path="/proxy/rodin/api/v2/status", method="POST"), + response_model=Rodin3DCheckStatusResponse, + data=Rodin3DCheckStatusRequest(subscription_key=subscription_key), status_extractor=check_rodin_status, - poll_interval=3.0, - auth_kwargs=auth_kwargs, + progress_extractor=extract_progress, ) - logging.info("[ Rodin3D API - CheckStatus ] Generate Start!") - return await poll_operation.execute() -async def get_rodin_download_list(uuid, auth_kwargs: Optional[dict[str, str]] = None) -> Rodin3DDownloadResponse: +async def get_rodin_download_list(uuid: str, cls: type[IO.ComfyNode]) -> Rodin3DDownloadResponse: logging.info("[ Rodin3D API - Downloading ] Generate Successfully!") - operation = SynchronousOperation( - endpoint=ApiEndpoint( - path="/proxy/rodin/api/v2/download", - method=HttpMethod.POST, - request_model=Rodin3DDownloadRequest, - response_model=Rodin3DDownloadResponse, - ), - request=Rodin3DDownloadRequest(task_uuid=uuid), - auth_kwargs=auth_kwargs, + return await sync_op( + cls, + ApiEndpoint(path="/proxy/rodin/api/v2/download", method="POST"), + response_model=Rodin3DDownloadResponse, + data=Rodin3DDownloadRequest(task_uuid=uuid), + monitor_progress=False, ) - return await operation.execute() -async def download_files(url_list, task_uuid): +async def download_files(url_list, task_uuid: str): result_folder_name = f"Rodin3D_{task_uuid}" save_path = os.path.join(comfy_paths.get_output_directory(), result_folder_name) os.makedirs(save_path, exist_ok=True) model_file_path = None - async with aiohttp.ClientSession() as session: - for i in url_list.list: - file_path = os.path.join(save_path, i.name) - if file_path.endswith(".glb"): - model_file_path = os.path.join(result_folder_name, i.name) - logging.info("[ Rodin3D API - download_files ] Downloading file: %s", file_path) - max_retries = 5 - for attempt in range(max_retries): - try: - async with session.get(i.url) as resp: - resp.raise_for_status() - with open(file_path, "wb") as f: - async for chunk in resp.content.iter_chunked(32 * 1024): - f.write(chunk) - break - except Exception as e: - logging.info("[ Rodin3D API - download_files ] Error downloading %s:%s", file_path, str(e)) - if attempt < max_retries - 1: - logging.info("Retrying...") - await asyncio.sleep(2) - else: - logging.info( - "[ Rodin3D API - download_files ] Failed to download %s after %s attempts.", - file_path, - max_retries, - ) + for i in url_list.list: + file_path = os.path.join(save_path, i.name) + if file_path.endswith(".glb"): + model_file_path = os.path.join(result_folder_name, i.name) + await download_url_to_bytesio(i.url, file_path) return model_file_path @@ -276,6 +238,7 @@ def define_schema(cls) -> IO.Schema: hidden=[ IO.Hidden.auth_token_comfy_org, IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, ], is_api_node=True, ) @@ -294,21 +257,17 @@ async def execute( for i in range(num_images): m_images.append(Images[i]) mesh_mode, quality_override = get_quality_mode(Polygon_count) - auth = { - "auth_token": cls.hidden.auth_token_comfy_org, - "comfy_api_key": cls.hidden.api_key_comfy_org, - } task_uuid, subscription_key = await create_generate_task( + cls, images=m_images, seed=Seed, material=Material_Type, quality_override=quality_override, tier=tier, mesh_mode=mesh_mode, - auth_kwargs=auth, ) - await poll_for_task_status(subscription_key, auth_kwargs=auth) - download_list = await get_rodin_download_list(task_uuid, auth_kwargs=auth) + await poll_for_task_status(subscription_key, cls) + download_list = await get_rodin_download_list(task_uuid, cls) model = await download_files(download_list, task_uuid) return IO.NodeOutput(model) @@ -332,6 +291,7 @@ def define_schema(cls) -> IO.Schema: hidden=[ IO.Hidden.auth_token_comfy_org, IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, ], is_api_node=True, ) @@ -350,21 +310,17 @@ async def execute( for i in range(num_images): m_images.append(Images[i]) mesh_mode, quality_override = get_quality_mode(Polygon_count) - auth = { - "auth_token": cls.hidden.auth_token_comfy_org, - "comfy_api_key": cls.hidden.api_key_comfy_org, - } task_uuid, subscription_key = await create_generate_task( + cls, images=m_images, seed=Seed, material=Material_Type, quality_override=quality_override, tier=tier, mesh_mode=mesh_mode, - auth_kwargs=auth, ) - await poll_for_task_status(subscription_key, auth_kwargs=auth) - download_list = await get_rodin_download_list(task_uuid, auth_kwargs=auth) + await poll_for_task_status(subscription_key, cls) + download_list = await get_rodin_download_list(task_uuid, cls) model = await download_files(download_list, task_uuid) return IO.NodeOutput(model) @@ -388,6 +344,7 @@ def define_schema(cls) -> IO.Schema: hidden=[ IO.Hidden.auth_token_comfy_org, IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, ], is_api_node=True, ) @@ -400,27 +357,22 @@ async def execute( Material_Type, Polygon_count, ) -> IO.NodeOutput: - tier = "Smooth" num_images = Images.shape[0] m_images = [] for i in range(num_images): m_images.append(Images[i]) mesh_mode, quality_override = get_quality_mode(Polygon_count) - auth = { - "auth_token": cls.hidden.auth_token_comfy_org, - "comfy_api_key": cls.hidden.api_key_comfy_org, - } task_uuid, subscription_key = await create_generate_task( + cls, images=m_images, seed=Seed, material=Material_Type, quality_override=quality_override, - tier=tier, + tier="Smooth", mesh_mode=mesh_mode, - auth_kwargs=auth, ) - await poll_for_task_status(subscription_key, auth_kwargs=auth) - download_list = await get_rodin_download_list(task_uuid, auth_kwargs=auth) + await poll_for_task_status(subscription_key, cls) + download_list = await get_rodin_download_list(task_uuid, cls) model = await download_files(download_list, task_uuid) return IO.NodeOutput(model) @@ -451,6 +403,7 @@ def define_schema(cls) -> IO.Schema: hidden=[ IO.Hidden.auth_token_comfy_org, IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, ], is_api_node=True, ) @@ -461,29 +414,21 @@ async def execute( Images, Seed, ) -> IO.NodeOutput: - tier = "Sketch" num_images = Images.shape[0] m_images = [] for i in range(num_images): m_images.append(Images[i]) - material_type = "PBR" - quality_override = 18000 - mesh_mode = "Quad" - auth = { - "auth_token": cls.hidden.auth_token_comfy_org, - "comfy_api_key": cls.hidden.api_key_comfy_org, - } task_uuid, subscription_key = await create_generate_task( + cls, images=m_images, seed=Seed, - material=material_type, - quality_override=quality_override, - tier=tier, - mesh_mode=mesh_mode, - auth_kwargs=auth, + material="PBR", + quality_override=18000, + tier="Sketch", + mesh_mode="Quad", ) - await poll_for_task_status(subscription_key, auth_kwargs=auth) - download_list = await get_rodin_download_list(task_uuid, auth_kwargs=auth) + await poll_for_task_status(subscription_key, cls) + download_list = await get_rodin_download_list(task_uuid, cls) model = await download_files(download_list, task_uuid) return IO.NodeOutput(model) @@ -522,6 +467,7 @@ def define_schema(cls) -> IO.Schema: hidden=[ IO.Hidden.auth_token_comfy_org, IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, ], is_api_node=True, ) @@ -541,22 +487,18 @@ async def execute( for i in range(num_images): m_images.append(Images[i]) mesh_mode, quality_override = get_quality_mode(Polygon_count) - auth = { - "auth_token": cls.hidden.auth_token_comfy_org, - "comfy_api_key": cls.hidden.api_key_comfy_org, - } task_uuid, subscription_key = await create_generate_task( + cls, images=m_images, seed=Seed, material=Material_Type, quality_override=quality_override, tier=tier, mesh_mode=mesh_mode, - TAPose=TAPose, - auth_kwargs=auth, + ta_pose=TAPose, ) - await poll_for_task_status(subscription_key, auth_kwargs=auth) - download_list = await get_rodin_download_list(task_uuid, auth_kwargs=auth) + await poll_for_task_status(subscription_key, cls) + download_list = await get_rodin_download_list(task_uuid, cls) model = await download_files(download_list, task_uuid) return IO.NodeOutput(model) diff --git a/comfy_api_nodes/util/client.py b/comfy_api_nodes/util/client.py index 65bb35f0f092..2d5dcd648486 100644 --- a/comfy_api_nodes/util/client.py +++ b/comfy_api_nodes/util/client.py @@ -16,9 +16,9 @@ from comfy import utils from comfy_api.latest import IO -from comfy_api_nodes.apis import request_logger from server import PromptServer +from . import request_logger from ._helpers import ( default_base_url, get_auth_header, @@ -77,7 +77,7 @@ class _PollUIState: _RETRY_STATUS = {408, 429, 500, 502, 503, 504} -COMPLETED_STATUSES = ["succeeded", "succeed", "success", "completed", "finished"] +COMPLETED_STATUSES = ["succeeded", "succeed", "success", "completed", "finished", "done"] FAILED_STATUSES = ["cancelled", "canceled", "fail", "failed", "error"] QUEUED_STATUSES = ["created", "queued", "queueing", "submitted"] diff --git a/comfy_api_nodes/util/download_helpers.py b/comfy_api_nodes/util/download_helpers.py index 364874bedb9d..14207dc6889c 100644 --- a/comfy_api_nodes/util/download_helpers.py +++ b/comfy_api_nodes/util/download_helpers.py @@ -12,8 +12,8 @@ from comfy_api.input_impl import VideoFromFile from comfy_api.latest import IO as COMFY_IO -from comfy_api_nodes.apis import request_logger +from . import request_logger from ._helpers import ( default_base_url, get_auth_header, diff --git a/comfy_api_nodes/apis/request_logger.py b/comfy_api_nodes/util/request_logger.py similarity index 100% rename from comfy_api_nodes/apis/request_logger.py rename to comfy_api_nodes/util/request_logger.py index c6974d35cce4..ac52e2eabbdd 100644 --- a/comfy_api_nodes/apis/request_logger.py +++ b/comfy_api_nodes/util/request_logger.py @@ -1,11 +1,11 @@ from __future__ import annotations -import os import datetime +import hashlib import json import logging +import os import re -import hashlib from typing import Any import folder_paths diff --git a/comfy_api_nodes/util/upload_helpers.py b/comfy_api_nodes/util/upload_helpers.py index 7bfc61704a2a..632450d9be86 100644 --- a/comfy_api_nodes/util/upload_helpers.py +++ b/comfy_api_nodes/util/upload_helpers.py @@ -13,8 +13,8 @@ from comfy_api.latest import IO, Input from comfy_api.util import VideoCodec, VideoContainer -from comfy_api_nodes.apis import request_logger +from . import request_logger from ._helpers import is_processing_interrupted, sleep_with_interrupt from .client import ( ApiEndpoint, diff --git a/folder_paths.py b/folder_paths.py index f110d832bb23..98b7d340c609 100644 --- a/folder_paths.py +++ b/folder_paths.py @@ -4,6 +4,7 @@ import time import mimetypes import logging +import sys from typing import Literal, List from collections.abc import Collection @@ -115,6 +116,16 @@ def set_input_directory(input_dir: str) -> None: global input_directory input_directory = input_dir +def set_custom_nodes_directory(custom_nodes_dir: str) -> None: + global folder_names_and_paths + + delimiter = ":" + if sys.platform in ("win32", "cygwin"): + delimiter = ";" + + directories = custom_nodes_dir.split(delimiter) + folder_names_and_paths["custom_nodes"] = (directories, set()) + def get_output_directory() -> str: global output_directory return output_directory diff --git a/main.py b/main.py index e1b0f1620bbc..a6b241da48f7 100644 --- a/main.py +++ b/main.py @@ -32,7 +32,7 @@ def apply_custom_paths(): for config_path in itertools.chain(*args.extra_model_paths_config): utils.extra_config.load_extra_path_config(config_path) - # --output-directory, --input-directory, --user-directory + # --output-directory, --input-directory, --user-directory, --custom-nodes-directory if args.output_directory: output_dir = os.path.abspath(args.output_directory) logging.info(f"Setting output directory to: {output_dir}") @@ -56,6 +56,11 @@ def apply_custom_paths(): logging.info(f"Setting user directory to: {user_dir}") folder_paths.set_user_directory(user_dir) + if args.custom_nodes_directory: + custom_nodes_dir = os.path.abspath(args.custom_nodes_directory) + logging.info(f"Setting custom nodes directory to: {custom_nodes_dir}") + folder_paths.set_custom_nodes_directory(custom_nodes_dir) + def execute_prestartup_script(): if args.disable_all_custom_nodes and len(args.whitelist_custom_nodes) == 0: @@ -110,7 +115,6 @@ def execute_script(script_path): import threading import gc - if os.name == "nt": os.environ['MIMALLOC_PURGE_DELAY'] = '0' @@ -155,6 +159,7 @@ def execute_script(script_path): import app.logger import hook_breaker_ac10a0 + def cuda_malloc_warning(): device = comfy.model_management.get_torch_device() device_name = comfy.model_management.get_torch_device_name(device) @@ -293,7 +298,7 @@ def cleanup_temp(): def setup_database(): try: - from app.database.db import init_db, dependencies_available + from app.database.db import dependencies_available, init_db if dependencies_available(): init_db() except Exception as e: diff --git a/server.py b/server.py index 5d773b10a57f..d059d3dc9085 100644 --- a/server.py +++ b/server.py @@ -2,6 +2,7 @@ import sys import asyncio import traceback +import time import nodes import folder_paths @@ -733,6 +734,7 @@ async def post_prompt(request): for sensitive_val in execution.SENSITIVE_EXTRA_DATA_KEYS: if sensitive_val in extra_data: sensitive[sensitive_val] = extra_data.pop(sensitive_val) + extra_data["create_time"] = int(time.time() * 1000) # timestamp in milliseconds self.prompt_queue.put((number, prompt_id, prompt, extra_data, outputs_to_execute, sensitive)) response = {"prompt_id": prompt_id, "number": number, "node_errors": valid[3]} return web.json_response(response)