Files
ComfyUI/main.py

594 lines
24 KiB
Python
Raw Permalink Normal View History

import comfy.options
comfy.options.enable_args_parsing()
from comfy.cli_args import args
from comfy.cli_args import get_console_log_level, get_file_log_outputs
if args.list_feature_flags:
import json
from comfy_api.feature_flags import CLI_FEATURE_FLAG_REGISTRY
print(json.dumps(CLI_FEATURE_FLAG_REGISTRY, indent=2)) # noqa: T201
raise SystemExit(0)
import os
import importlib.util
import shutil
import importlib.metadata
import folder_paths
import time
from comfy.cli_args import enables_dynamic_vram
from app.logger import setup_logger
console_log_level = get_console_log_level(args.verbose)
file_log_outputs = get_file_log_outputs(args.verbose)
setup_logger(log_level=console_log_level, file_outputs=file_log_outputs, use_stdout=args.log_stdout)
from app.assets.seeder import asset_seeder
from app.assets.services import register_output_files
import itertools
import utils.extra_config
from utils.mime_types import init_mime_types
import faulthandler
import logging
import signal
2025-04-12 18:58:20 -04:00
import sys
Support for async node functions (#8830) * Support for async execution functions This commit adds support for node execution functions defined as async. When a node's execution function is defined as async, we can continue executing other nodes while it is processing. Standard uses of `await` should "just work", but people will still have to be careful if they spawn actual threads. Because torch doesn't really have async/await versions of functions, this won't particularly help with most locally-executing nodes, but it does work for e.g. web requests to other machines. In addition to the execute function, the `VALIDATE_INPUTS` and `check_lazy_status` functions can also be defined as async, though we'll only resolve one node at a time right now for those. * Add the execution model tests to CI * Add a missing file It looks like this got caught by .gitignore? There's probably a better place to put it, but I'm not sure what that is. * Add the websocket library for automated tests * Add additional tests for async error cases Also fixes one bug that was found when an async function throws an error after being scheduled on a task. * Add a feature flags message to reduce bandwidth We now only send 1 preview message of the latest type the client can support. We'll add a console warning when the client fails to send a feature flags message at some point in the future. * Add async tests to CI * Don't actually add new tests in this PR Will do it in a separate PR * Resolve unit test in GPU-less runner * Just remove the tests that GHA can't handle * Change line endings to UNIX-style * Avoid loading model_management.py so early Because model_management.py has a top-level `logging.info`, we have to be careful not to import that file before we call `setup_logging`. If we do, we end up having the default logging handler registered in addition to our custom one.
2025-07-10 11:46:19 -07:00
from comfy_execution.progress import get_progress_state
from comfy_execution.utils import get_executing_context
from comfy_api import feature_flags
from app.database.db import init_db, dependencies_available
if __name__ == "__main__":
#NOTE: These do not do anything on core ComfyUI, they are for custom nodes.
os.environ['HF_HUB_DISABLE_TELEMETRY'] = '1'
os.environ['DO_NOT_TRACK'] = '1'
faulthandler.enable(file=sys.stderr, all_threads=args.debug_hang)
if __name__ == "__main__" and args.debug_hang:
dumping_traceback = False
def dump_traceback_on_sigint(signum, frame):
global dumping_traceback
if dumping_traceback:
raise KeyboardInterrupt
dumping_traceback = True
faulthandler.dump_traceback(file=sys.stderr, all_threads=True)
raise KeyboardInterrupt
signal.signal(signal.SIGINT, dump_traceback_on_sigint)
import comfy_aimdo.control
if enables_dynamic_vram():
simple_vram_headroom = None if args.reserve_vram is None else int(args.reserve_vram * 1024 ** 3)
try:
comfy_aimdo.control.init(simple_vram_headroom=simple_vram_headroom, nvml_pressure=not args.disable_nvml_pressure)
except TypeError:
# comfy-aimdo 0.4.10 protocol.
try:
comfy_aimdo.control.init(simple_vram_headroom=simple_vram_headroom)
except TypeError:
# comfy-aimdo 0.4.9 protocol.
comfy_aimdo.control.init()
if os.name == "nt":
os.environ['MIMALLOC_PURGE_DELAY'] = '0'
if __name__ == "__main__":
os.environ['TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL'] = '1'
if args.default_device is not None:
default_dev = args.default_device
devices = list(range(32))
devices.remove(default_dev)
devices.insert(0, default_dev)
devices = ','.join(map(str, devices))
os.environ['CUDA_VISIBLE_DEVICES'] = str(devices)
os.environ['HIP_VISIBLE_DEVICES'] = str(devices)
if args.cuda_device is not None:
os.environ['CUDA_VISIBLE_DEVICES'] = str(args.cuda_device)
os.environ['HIP_VISIBLE_DEVICES'] = str(args.cuda_device)
os.environ["ASCEND_RT_VISIBLE_DEVICES"] = str(args.cuda_device)
logging.info("Set cuda device to: {}".format(args.cuda_device))
if args.oneapi_device_selector is not None:
os.environ['ONEAPI_DEVICE_SELECTOR'] = args.oneapi_device_selector
logging.info("Set oneapi device selector to: {}".format(args.oneapi_device_selector))
if args.deterministic:
if 'CUBLAS_WORKSPACE_CONFIG' not in os.environ:
os.environ['CUBLAS_WORKSPACE_CONFIG'] = ":4096:8"
import cuda_malloc
if "rocm" in cuda_malloc.get_torch_version_noimport():
os.environ['OCL_SET_SVM_SIZE'] = '262144' # set at the request of AMD
def handle_comfyui_manager_unavailable():
manager_req_path = os.path.join(os.path.dirname(os.path.abspath(folder_paths.__file__)), "manager_requirements.txt")
uv_available = shutil.which("uv") is not None
pip_cmd = f"{sys.executable} -m pip install -r {manager_req_path}"
msg = f"\n\nTo use the `--enable-manager` feature, the `comfyui-manager` package must be installed first.\ncommand:\n\t{pip_cmd}"
if uv_available:
msg += f"\nor using uv:\n\tuv pip install -r {manager_req_path}"
msg += "\n"
logging.warning(msg)
args.enable_manager = False
if args.enable_manager:
if importlib.util.find_spec("comfyui_manager"):
import comfyui_manager
if not comfyui_manager.__file__ or not comfyui_manager.__file__.endswith('__init__.py'):
handle_comfyui_manager_unavailable()
else:
handle_comfyui_manager_unavailable()
def apply_custom_paths():
# extra model paths
extra_model_paths_config_path = os.path.join(os.path.dirname(os.path.realpath(__file__)), "extra_model_paths.yaml")
if os.path.isfile(extra_model_paths_config_path):
utils.extra_config.load_extra_path_config(extra_model_paths_config_path)
if args.extra_model_paths_config:
for config_path in itertools.chain(*args.extra_model_paths_config):
utils.extra_config.load_extra_path_config(config_path)
# --base-directory
if args.base_directory:
logging.info(f"Setting base directory to: {folder_paths.base_path}")
# --models-directory
if args.models_directory:
logging.info(f"Setting models directory to: {folder_paths.models_dir}")
# --output-directory, --input-directory, --user-directory
if args.output_directory:
output_dir = os.path.abspath(args.output_directory)
logging.info(f"Setting output directory to: {output_dir}")
folder_paths.set_output_directory(output_dir)
# These are the default folders that checkpoints, clip and vae models will be saved to when using CheckpointSave, etc.. nodes
folder_paths.add_model_folder_path("checkpoints", os.path.join(folder_paths.get_output_directory(), "checkpoints"))
folder_paths.add_model_folder_path("clip", os.path.join(folder_paths.get_output_directory(), "clip"))
folder_paths.add_model_folder_path("vae", os.path.join(folder_paths.get_output_directory(), "vae"))
folder_paths.add_model_folder_path("diffusion_models",
os.path.join(folder_paths.get_output_directory(), "diffusion_models"))
folder_paths.add_model_folder_path("loras", os.path.join(folder_paths.get_output_directory(), "loras"))
if args.input_directory:
input_dir = os.path.abspath(args.input_directory)
logging.info(f"Setting input directory to: {input_dir}")
folder_paths.set_input_directory(input_dir)
if args.user_directory:
user_dir = os.path.abspath(args.user_directory)
logging.info(f"Setting user directory to: {user_dir}")
folder_paths.set_user_directory(user_dir)
def execute_prestartup_script():
if args.disable_all_custom_nodes and len(args.whitelist_custom_nodes) == 0:
return
def execute_script(script_path):
module_name = os.path.splitext(script_path)[0]
try:
spec = importlib.util.spec_from_file_location(module_name, script_path)
module = importlib.util.module_from_spec(spec)
spec.loader.exec_module(module)
return True
except Exception as e:
logging.error(f"Failed to execute startup-script: {script_path} / {e}")
return False
node_paths = folder_paths.get_folder_paths("custom_nodes")
for custom_node_path in node_paths:
possible_modules = os.listdir(custom_node_path)
node_prestartup_times = []
for possible_module in possible_modules:
module_path = os.path.join(custom_node_path, possible_module)
if args.enable_manager:
if comfyui_manager.should_be_disabled(module_path):
continue
if os.path.isfile(module_path) or module_path.endswith(".disabled") or module_path == "__pycache__":
continue
script_path = os.path.join(module_path, "prestartup_script.py")
if os.path.exists(script_path):
if args.disable_all_custom_nodes and possible_module not in args.whitelist_custom_nodes:
logging.info(f"Prestartup Skipping {possible_module} due to disable_all_custom_nodes and whitelist_custom_nodes")
continue
time_before = time.perf_counter()
success = execute_script(script_path)
node_prestartup_times.append((time.perf_counter() - time_before, module_path, success))
if len(node_prestartup_times) > 0:
logging.info("\nPrestartup times for custom nodes:")
for n in sorted(node_prestartup_times):
if n[2]:
import_message = ""
else:
import_message = " (PRESTARTUP FAILED)"
logging.info("{:6.1f} seconds{}: {}".format(n[0], import_message, n[1]))
logging.info("")
apply_custom_paths()
init_mime_types()
if args.enable_manager:
comfyui_manager.prestartup()
execute_prestartup_script()
# Main code
2023-04-05 20:32:59 -04:00
import asyncio
2023-01-03 01:53:32 -05:00
import threading
import gc
if 'torch' in sys.modules:
logging.warning("WARNING: Potential Error in code: Torch already imported, torch should never be imported before this point.")
import comfy.utils
import execution
2023-04-05 20:32:59 -04:00
import server
Support for async node functions (#8830) * Support for async execution functions This commit adds support for node execution functions defined as async. When a node's execution function is defined as async, we can continue executing other nodes while it is processing. Standard uses of `await` should "just work", but people will still have to be careful if they spawn actual threads. Because torch doesn't really have async/await versions of functions, this won't particularly help with most locally-executing nodes, but it does work for e.g. web requests to other machines. In addition to the execute function, the `VALIDATE_INPUTS` and `check_lazy_status` functions can also be defined as async, though we'll only resolve one node at a time right now for those. * Add the execution model tests to CI * Add a missing file It looks like this got caught by .gitignore? There's probably a better place to put it, but I'm not sure what that is. * Add the websocket library for automated tests * Add additional tests for async error cases Also fixes one bug that was found when an async function throws an error after being scheduled on a task. * Add a feature flags message to reduce bandwidth We now only send 1 preview message of the latest type the client can support. We'll add a console warning when the client fails to send a feature flags message at some point in the future. * Add async tests to CI * Don't actually add new tests in this PR Will do it in a separate PR * Resolve unit test in GPU-less runner * Just remove the tests that GHA can't handle * Change line endings to UNIX-style * Avoid loading model_management.py so early Because model_management.py has a top-level `logging.info`, we have to be careful not to import that file before we call `setup_logging`. If we do, we end up having the default logging handler registered in addition to our custom one.
2025-07-10 11:46:19 -07:00
from protocol import BinaryEventTypes
import nodes
import comfy.model_management
2025-01-25 15:03:57 -05:00
import comfyui_version
import app.logger
import hook_breaker_ac10a0
import comfy.memory_management
import comfy.model_patcher
if args.enable_dynamic_vram or (enables_dynamic_vram() and comfy.model_management.is_nvidia() and not comfy.model_management.is_wsl()):
if (not args.enable_dynamic_vram) and (comfy.model_management.torch_version_numeric < (2, 8)):
logging.warning("Unsupported Pytorch detected. DynamicVRAM support requires Pytorch version 2.8 or later. Falling back to legacy ModelPatcher. VRAM estimates may be unreliable especially on Windows")
else:
try:
aimdo_initialized = comfy_aimdo.control.init_devices((d.index, int(args.vram_headroom * 1024 ** 3)) for d in comfy.model_management.get_all_torch_devices())
except TypeError:
# comfy-aimdo 0.4.9 protocol.
aimdo_initialized = comfy_aimdo.control.init_devices(d.index for d in comfy.model_management.get_all_torch_devices())
if aimdo_initialized:
if console_log_level == 'DEBUG':
comfy_aimdo.control.set_log_debug()
elif console_log_level == 'DETAIL':
try:
comfy_aimdo.control.set_log_detail()
except AttributeError:
comfy_aimdo.control.set_log_info()
elif console_log_level == 'CRITICAL':
comfy_aimdo.control.set_log_critical()
elif console_log_level == 'ERROR':
comfy_aimdo.control.set_log_error()
elif console_log_level == 'WARNING':
comfy_aimdo.control.set_log_warning()
else: #INFO
comfy_aimdo.control.set_log_info()
comfy.model_patcher.CoreModelPatcher = comfy.model_patcher.ModelPatcherDynamic
comfy.memory_management.aimdo_enabled = True
logging.info("DynamicVRAM support detected and enabled")
else:
logging.warning("No working comfy-aimdo install detected. DynamicVRAM support disabled. Falling back to legacy ModelPatcher. VRAM estimates may be unreliable especially on Windows")
def cuda_malloc_warning():
device = comfy.model_management.get_torch_device()
device_name = comfy.model_management.get_torch_device_name(device)
cuda_malloc_warning = False
if "cudaMallocAsync" in device_name:
for b in cuda_malloc.blacklist:
if b in device_name:
cuda_malloc_warning = True
if cuda_malloc_warning:
2024-03-11 16:24:47 -04:00
logging.warning("\nWARNING: this card most likely does not support cuda-malloc, if you get \"CUDA error\" please run ComfyUI with: --disable-cuda-malloc\n")
def _collect_output_absolute_paths(history_result: dict) -> list[str]:
"""Extract absolute file paths for output items from a history result."""
paths: list[str] = []
seen: set[str] = set()
for node_output in history_result.get("outputs", {}).values():
for items in node_output.values():
if not isinstance(items, list):
continue
for item in items:
if not isinstance(item, dict):
continue
item_type = item.get("type")
if item_type not in ("output", "temp"):
continue
base_dir = folder_paths.get_directory_by_type(item_type)
if base_dir is None:
continue
base_dir = os.path.abspath(base_dir)
filename = item.get("filename")
if not filename:
continue
abs_path = os.path.abspath(
os.path.join(base_dir, item.get("subfolder", ""), filename)
)
if not abs_path.startswith(base_dir + os.sep) and abs_path != base_dir:
continue
if abs_path not in seen:
seen.add(abs_path)
paths.append(abs_path)
return paths
def prompt_worker(q, server_instance):
current_time: float = 0.0
Multi-threaded load of models from disk (big load time speedups & Offload to disk) (CORE-43,CORE-152,CORE-164,CORE-165,CORE-117) (#13802) * model_management: disable non-dynamic smart memory Disable smart memory outright for non dynamic models. This is a minor step towards deprecation of --disable-dynamic-vram and the legacy ModelPatcher. This is needed for estimate-free model development, where new models can opt-out of supplying a memory estimate and not have to worry about hard VRAM allocations due to legacy non-dynamic model patchers This is also a general stability increase for a lot of stray use cases where estimates may still be off and going forward we are not going to accurately maintain such estimates. * pinned_memory: implement with aimdo growable buffer Use a single growable buffer so we can do threaded pre-warming on pinned memory. * mm: use aimdo to do transfer from disk to pin Aimdo implements a faster threaded loader. * Add stream host pin buffer for AIMDO casts Introduce per-offload-stream HostBuffer reuse for pinned staging, include it in cast buffer reset synchronization. Defer actual casts that go via this pin path to a separate pass such that the buffer can be allocated monolithically (to avoid cudaHostRegister thrash). * remove old pin path * Implement JIT pinned memory pressure Replace the predictive pin pressure mechanism with JIT PIN memory pressure. * LowVRAMPatch: change to two-phase visit * lora: re-implement as inplace swiss-army-knife operation * prepare for multiple pin sets * implement pinned loras * requirements: comfy-aimdo 0.4.0 * ops: remove unused arg This was defeatured in aimdo iteration * ops: sync the CPU with only the offload stream activity This was syncing with the offload stream which itself is synced with the compute stream, so this was syncing CPU with compute transitively. Define the event to sync it more gently. * pins: implement freeing intermediate for pinned memory Pinning is more important than inactive intermediates and the stream pin buffer is more important than even active intermediates. * execution: implement pin eviction on RAM presure Add back proper pin freeing on RAM pressure * implement pin registration swaps Uncap the windows pins from 50% by extending the pool and have a pressure mechanism to move the pin reservations om demand. This unfortunately implies a GPU sync to do the freeing so significant hysterisis needs to be added to consolidate these pressure events. * cli_args/execution: Implement lower background cache-ram threshold Limit the amount of RAM background intermediates can use, so that switching workflows doesn't degrade performance too much. * make default * bump aimdo * model-patcher: force-cast tiny weights Flux 2 gets crazy stalls due to a mix of tiny and giant weights creating lopsided steam buffer rotations which creates stalls. * ops: refactor in prep for chunking * mm: delegate pin-on-the-way to aimdo Aimdo is able to chunk and slice this on the way for better CPU->GPU overlap. The main advantage is the ability to shorten the bus contention window between previous weight transfer and the next weights vbar fault. * bump aimdo * pinning updates * specify hostbuf max allocation size There a signs of virtual memory exhaustion on some linux systems when throwing 128GB for every little piece. Pass the actual to save aimdo from over-estimates * tests: update execution tests for caching The default caching changed to ram-cache so update these tests accordingly. Remove the LRU 0 test as this also falls through to RAM cache.
2026-05-21 10:03:58 +10:00
cache_ram = 0
cache_ram_inactive = 0
if not args.cache_classic and not args.cache_none and args.cache_lru <= 0:
cache_ram = min(10.0, max(2.0, comfy.model_management.total_ram * 0.10 / 1024.0))
cache_ram_inactive = min(128.0, comfy.model_management.total_ram / 1024.0)
Multi-threaded load of models from disk (big load time speedups & Offload to disk) (CORE-43,CORE-152,CORE-164,CORE-165,CORE-117) (#13802) * model_management: disable non-dynamic smart memory Disable smart memory outright for non dynamic models. This is a minor step towards deprecation of --disable-dynamic-vram and the legacy ModelPatcher. This is needed for estimate-free model development, where new models can opt-out of supplying a memory estimate and not have to worry about hard VRAM allocations due to legacy non-dynamic model patchers This is also a general stability increase for a lot of stray use cases where estimates may still be off and going forward we are not going to accurately maintain such estimates. * pinned_memory: implement with aimdo growable buffer Use a single growable buffer so we can do threaded pre-warming on pinned memory. * mm: use aimdo to do transfer from disk to pin Aimdo implements a faster threaded loader. * Add stream host pin buffer for AIMDO casts Introduce per-offload-stream HostBuffer reuse for pinned staging, include it in cast buffer reset synchronization. Defer actual casts that go via this pin path to a separate pass such that the buffer can be allocated monolithically (to avoid cudaHostRegister thrash). * remove old pin path * Implement JIT pinned memory pressure Replace the predictive pin pressure mechanism with JIT PIN memory pressure. * LowVRAMPatch: change to two-phase visit * lora: re-implement as inplace swiss-army-knife operation * prepare for multiple pin sets * implement pinned loras * requirements: comfy-aimdo 0.4.0 * ops: remove unused arg This was defeatured in aimdo iteration * ops: sync the CPU with only the offload stream activity This was syncing with the offload stream which itself is synced with the compute stream, so this was syncing CPU with compute transitively. Define the event to sync it more gently. * pins: implement freeing intermediate for pinned memory Pinning is more important than inactive intermediates and the stream pin buffer is more important than even active intermediates. * execution: implement pin eviction on RAM presure Add back proper pin freeing on RAM pressure * implement pin registration swaps Uncap the windows pins from 50% by extending the pool and have a pressure mechanism to move the pin reservations om demand. This unfortunately implies a GPU sync to do the freeing so significant hysterisis needs to be added to consolidate these pressure events. * cli_args/execution: Implement lower background cache-ram threshold Limit the amount of RAM background intermediates can use, so that switching workflows doesn't degrade performance too much. * make default * bump aimdo * model-patcher: force-cast tiny weights Flux 2 gets crazy stalls due to a mix of tiny and giant weights creating lopsided steam buffer rotations which creates stalls. * ops: refactor in prep for chunking * mm: delegate pin-on-the-way to aimdo Aimdo is able to chunk and slice this on the way for better CPU->GPU overlap. The main advantage is the ability to shorten the bus contention window between previous weight transfer and the next weights vbar fault. * bump aimdo * pinning updates * specify hostbuf max allocation size There a signs of virtual memory exhaustion on some linux systems when throwing 128GB for every little piece. Pass the actual to save aimdo from over-estimates * tests: update execution tests for caching The default caching changed to ram-cache so update these tests accordingly. Remove the LRU 0 test as this also falls through to RAM cache.
2026-05-21 10:03:58 +10:00
if len(args.cache_ram) > 0:
cache_ram = args.cache_ram[0]
if len(args.cache_ram) > 1:
cache_ram_inactive = args.cache_ram[1]
cache_type = execution.CacheType.RAM_PRESSURE
if args.cache_classic:
cache_type = execution.CacheType.CLASSIC
elif args.cache_lru > 0:
cache_type = execution.CacheType.LRU
elif args.cache_none:
execution: fold in dependency aware caching / Fix --cache-none with loops/lazy etc (Resubmit) (#10440) * execution: fold in dependency aware caching This makes --cache-none compatiable with lazy and expanded subgraphs. Currently the --cache-none option is powered by the DependencyAwareCache. The cache attempts to maintain a parallel copy of the execution list data structure, however it is only setup once at the start of execution and does not get meaninigful updates to the execution list. This causes multiple problems when --cache-none is used with lazy and expanded subgraphs as the DAC does not accurately update its copy of the execution data structure. DAC has an attempt to handle subgraphs ensure_subcache however this does not accurately connect to nodes outside the subgraph. The current semantics of DAC are to free a node ASAP after the dependent nodes are executed. This means that if a subgraph refs such a node it will be requed and re-executed by the execution_list but DAC wont see it in its to-free lists anymore and leak memory. Rather than try and cover all the cases where the execution list changes from inside the cache, move the while problem to the executor which maintains an always up-to-date copy of the wanted data-structure. The executor now has a fast-moving run-local cache of its own. Each _to node has its own mini cache, and the cache is unconditionally primed at the time of add_strong_link. add_strong_link is called for all of static workflows, lazy links and expanded subgraphs so its the singular source of truth for output dependendencies. In the case of a cache-hit, the executor cache will hold the non-none value (it will respect updates if they happen somehow as well). In the case of a cache-miss, the executor caches a None and will wait for a notification to update the value when the node completes. When a node completes execution, it simply releases its mini-cache and in turn its strong refs on its direct anscestor outputs, allowing for ASAP freeing (same as the DependencyAwareCache but a little more automatic). This now allows for re-implementation of --cache-none with no cache at all. The dependency aware cache was also observing the dependency sematics for the objects and UI cache which is not accurate (this entire logic was always outputs specific). This also prepares for more complex caching strategies (such as RAM pressure based caching), where a cache can implement any freeing strategy completely independently of the DepedancyAwareness requirement. * main: re-implement --cache-none as no cache at all The execution list now tracks the dependency aware caching more correctly that the DependancyAwareCache. Change it to a cache that does nothing. * test_execution: add --cache-none to the test suite --cache-none is now expected to work universally. Run it through the full unit test suite. Propagate the server parameterization for whether or not the server is capabale of caching, so that the minority of tests that specifically check for cache hits can if else. Hard assert NOT caching in the else to give some coverage of --cache-none expected behaviour to not acutally cache.
2025-10-23 05:49:05 +10:00
cache_type = execution.CacheType.NONE
Multi-threaded load of models from disk (big load time speedups & Offload to disk) (CORE-43,CORE-152,CORE-164,CORE-165,CORE-117) (#13802) * model_management: disable non-dynamic smart memory Disable smart memory outright for non dynamic models. This is a minor step towards deprecation of --disable-dynamic-vram and the legacy ModelPatcher. This is needed for estimate-free model development, where new models can opt-out of supplying a memory estimate and not have to worry about hard VRAM allocations due to legacy non-dynamic model patchers This is also a general stability increase for a lot of stray use cases where estimates may still be off and going forward we are not going to accurately maintain such estimates. * pinned_memory: implement with aimdo growable buffer Use a single growable buffer so we can do threaded pre-warming on pinned memory. * mm: use aimdo to do transfer from disk to pin Aimdo implements a faster threaded loader. * Add stream host pin buffer for AIMDO casts Introduce per-offload-stream HostBuffer reuse for pinned staging, include it in cast buffer reset synchronization. Defer actual casts that go via this pin path to a separate pass such that the buffer can be allocated monolithically (to avoid cudaHostRegister thrash). * remove old pin path * Implement JIT pinned memory pressure Replace the predictive pin pressure mechanism with JIT PIN memory pressure. * LowVRAMPatch: change to two-phase visit * lora: re-implement as inplace swiss-army-knife operation * prepare for multiple pin sets * implement pinned loras * requirements: comfy-aimdo 0.4.0 * ops: remove unused arg This was defeatured in aimdo iteration * ops: sync the CPU with only the offload stream activity This was syncing with the offload stream which itself is synced with the compute stream, so this was syncing CPU with compute transitively. Define the event to sync it more gently. * pins: implement freeing intermediate for pinned memory Pinning is more important than inactive intermediates and the stream pin buffer is more important than even active intermediates. * execution: implement pin eviction on RAM presure Add back proper pin freeing on RAM pressure * implement pin registration swaps Uncap the windows pins from 50% by extending the pool and have a pressure mechanism to move the pin reservations om demand. This unfortunately implies a GPU sync to do the freeing so significant hysterisis needs to be added to consolidate these pressure events. * cli_args/execution: Implement lower background cache-ram threshold Limit the amount of RAM background intermediates can use, so that switching workflows doesn't degrade performance too much. * make default * bump aimdo * model-patcher: force-cast tiny weights Flux 2 gets crazy stalls due to a mix of tiny and giant weights creating lopsided steam buffer rotations which creates stalls. * ops: refactor in prep for chunking * mm: delegate pin-on-the-way to aimdo Aimdo is able to chunk and slice this on the way for better CPU->GPU overlap. The main advantage is the ability to shorten the bus contention window between previous weight transfer and the next weights vbar fault. * bump aimdo * pinning updates * specify hostbuf max allocation size There a signs of virtual memory exhaustion on some linux systems when throwing 128GB for every little piece. Pass the actual to save aimdo from over-estimates * tests: update execution tests for caching The default caching changed to ram-cache so update these tests accordingly. Remove the LRU 0 test as this also falls through to RAM cache.
2026-05-21 10:03:58 +10:00
e = execution.PromptExecutor(server_instance, cache_type=cache_type, cache_args={ "lru" : args.cache_lru, "ram" : cache_ram, "ram_inactive" : cache_ram_inactive } )
last_gc_collect = 0
need_gc = False
gc_collect_interval = 10.0
2023-01-03 01:53:32 -05:00
while True:
timeout = 1000.0
if need_gc:
timeout = max(gc_collect_interval - (current_time - last_gc_collect), 0.0)
queue_item = q.get(timeout=timeout)
if queue_item is not None:
item, item_id = queue_item
execution_start_time = time.perf_counter()
prompt_id = item[1]
server_instance.last_prompt_id = prompt_id
sensitive = item[5]
extra_data = item[3].copy()
for k in sensitive:
extra_data[k] = sensitive[k]
asset_seeder.pause()
e.execute(item[2], prompt_id, extra_data, item[4])
need_gc = True
remove_sensitive = lambda prompt: prompt[:5] + prompt[6:]
2024-01-11 08:38:18 -05:00
q.task_done(item_id,
Execution Model Inversion (#2666) * Execution Model Inversion This PR inverts the execution model -- from recursively calling nodes to using a topological sort of the nodes. This change allows for modification of the node graph during execution. This allows for two major advantages: 1. The implementation of lazy evaluation in nodes. For example, if a "Mix Images" node has a mix factor of exactly 0.0, the second image input doesn't even need to be evaluated (and visa-versa if the mix factor is 1.0). 2. Dynamic expansion of nodes. This allows for the creation of dynamic "node groups". Specifically, custom nodes can return subgraphs that replace the original node in the graph. This is an incredibly powerful concept. Using this functionality, it was easy to implement: a. Components (a.k.a. node groups) b. Flow control (i.e. while loops) via tail recursion c. All-in-one nodes that replicate the WebUI functionality d. and more All of those were able to be implemented entirely via custom nodes, so those features are *not* a part of this PR. (There are some front-end changes that should occur before that functionality is made widely available, particularly around variant sockets.) The custom nodes associated with this PR can be found at: https://github.com/BadCafeCode/execution-inversion-demo-comfyui Note that some of them require that variant socket types ("*") be enabled. * Allow `input_info` to be of type `None` * Handle errors (like OOM) more gracefully * Add a command-line argument to enable variants This allows the use of nodes that have sockets of type '*' without applying a patch to the code. * Fix an overly aggressive assertion. This could happen when attempting to evaluate `IS_CHANGED` for a node during the creation of the cache (in order to create the cache key). * Fix Pyright warnings * Add execution model unit tests * Fix issue with unused literals Behavior should now match the master branch with regard to undeclared inputs. Undeclared inputs that are socket connections will be used while undeclared inputs that are literals will be ignored. * Make custom VALIDATE_INPUTS skip normal validation Additionally, if `VALIDATE_INPUTS` takes an argument named `input_types`, that variable will be a dictionary of the socket type of all incoming connections. If that argument exists, normal socket type validation will not occur. This removes the last hurdle for enabling variant types entirely from custom nodes, so I've removed that command-line option. I've added appropriate unit tests for these changes. * Fix example in unit test This wouldn't have caused any issues in the unit test, but it would have bugged the UI if someone copy+pasted it into their own node pack. * Use fstrings instead of '%' formatting syntax * Use custom exception types. * Display an error for dependency cycles Previously, dependency cycles that were created during node expansion would cause the application to quit (due to an uncaught exception). Now, we'll throw a proper error to the UI. We also make an attempt to 'blame' the most relevant node in the UI. * Add docs on when ExecutionBlocker should be used * Remove unused functionality * Rename ExecutionResult.SLEEPING to PENDING * Remove superfluous function parameter * Pass None for uneval inputs instead of default This applies to `VALIDATE_INPUTS`, `check_lazy_status`, and lazy values in evaluation functions. * Add a test for mixed node expansion This test ensures that a node that returns a combination of expanded subgraphs and literal values functions correctly. * Raise exception for bad get_node calls. * Minor refactor of IsChangedCache.get * Refactor `map_node_over_list` function * Fix ui output for duplicated nodes * Add documentation on `check_lazy_status` * Add file for execution model unit tests * Clean up Javascript code as per review * Improve documentation Converted some comments to docstrings as per review * Add a new unit test for mixed lazy results This test validates that when an output list is fed to a lazy node, the node will properly evaluate previous nodes that are needed by any inputs to the lazy node. No code in the execution model has been changed. The test already passes. * Allow kwargs in VALIDATE_INPUTS functions When kwargs are used, validation is skipped for all inputs as if they had been mentioned explicitly. * List cached nodes in `execution_cached` message This was previously just bugged in this PR.
2024-08-15 08:21:11 -07:00
e.history_result,
2024-01-11 08:38:18 -05:00
status=execution.PromptQueue.ExecutionStatus(
status_str='success' if e.success else 'error',
completed=e.success,
messages=e.status_messages), process_item=remove_sensitive)
if server_instance.client_id is not None:
server_instance.send_sync("executing", {"node": None, "prompt_id": prompt_id}, server_instance.client_id)
current_time = time.perf_counter()
execution_time = current_time - execution_start_time
# Log Time in a more readable way after 10 minutes
if execution_time > 600:
execution_time = time.strftime("%H:%M:%S", time.gmtime(execution_time))
2026-05-25 04:00:55 +02:00
logging.info(f"Prompt executed in {execution_time}", extra={'color': 'green'})
else:
2026-05-25 04:00:55 +02:00
logging.info("Prompt executed in {:.2f} seconds".format(execution_time), extra={'color': 'green'})
if not asset_seeder.is_disabled():
paths = _collect_output_absolute_paths(e.history_result)
register_output_files(paths, job_id=prompt_id)
flags = q.get_flags()
free_memory = flags.get("free_memory", False)
if flags.get("unload_models", free_memory):
comfy.model_management.unload_all_models()
need_gc = True
last_gc_collect = 0
if free_memory:
e.reset()
need_gc = True
last_gc_collect = 0
if need_gc:
current_time = time.perf_counter()
if (current_time - last_gc_collect) > gc_collect_interval:
gc.collect()
comfy.model_management.soft_empty_cache()
last_gc_collect = current_time
need_gc = False
hook_breaker_ac10a0.restore_functions()
if not asset_seeder.is_disabled():
asset_seeder.enqueue_enrich(roots=("output",), compute_hashes=args.enable_asset_hashing)
asset_seeder.resume()
2023-06-07 15:15:38 +02:00
async def run(server_instance, address='', port=8188, verbose=True, call_on_start=None):
addresses = []
for addr in address.split(","):
addresses.append((addr, port))
await asyncio.gather(
server_instance.start_multi_address(addresses, call_on_start, verbose), server_instance.publish_loop()
)
2023-01-03 01:53:32 -05:00
def hijack_progress(server_instance):
Support for async node functions (#8830) * Support for async execution functions This commit adds support for node execution functions defined as async. When a node's execution function is defined as async, we can continue executing other nodes while it is processing. Standard uses of `await` should "just work", but people will still have to be careful if they spawn actual threads. Because torch doesn't really have async/await versions of functions, this won't particularly help with most locally-executing nodes, but it does work for e.g. web requests to other machines. In addition to the execute function, the `VALIDATE_INPUTS` and `check_lazy_status` functions can also be defined as async, though we'll only resolve one node at a time right now for those. * Add the execution model tests to CI * Add a missing file It looks like this got caught by .gitignore? There's probably a better place to put it, but I'm not sure what that is. * Add the websocket library for automated tests * Add additional tests for async error cases Also fixes one bug that was found when an async function throws an error after being scheduled on a task. * Add a feature flags message to reduce bandwidth We now only send 1 preview message of the latest type the client can support. We'll add a console warning when the client fails to send a feature flags message at some point in the future. * Add async tests to CI * Don't actually add new tests in this PR Will do it in a separate PR * Resolve unit test in GPU-less runner * Just remove the tests that GHA can't handle * Change line endings to UNIX-style * Avoid loading model_management.py so early Because model_management.py has a top-level `logging.info`, we have to be careful not to import that file before we call `setup_logging`. If we do, we end up having the default logging handler registered in addition to our custom one.
2025-07-10 11:46:19 -07:00
def hook(value, total, preview_image, prompt_id=None, node_id=None):
executing_context = get_executing_context()
if prompt_id is None and executing_context is not None:
prompt_id = executing_context.prompt_id
if node_id is None and executing_context is not None:
node_id = executing_context.node_id
comfy.model_management.throw_exception_if_processing_interrupted()
Support for async node functions (#8830) * Support for async execution functions This commit adds support for node execution functions defined as async. When a node's execution function is defined as async, we can continue executing other nodes while it is processing. Standard uses of `await` should "just work", but people will still have to be careful if they spawn actual threads. Because torch doesn't really have async/await versions of functions, this won't particularly help with most locally-executing nodes, but it does work for e.g. web requests to other machines. In addition to the execute function, the `VALIDATE_INPUTS` and `check_lazy_status` functions can also be defined as async, though we'll only resolve one node at a time right now for those. * Add the execution model tests to CI * Add a missing file It looks like this got caught by .gitignore? There's probably a better place to put it, but I'm not sure what that is. * Add the websocket library for automated tests * Add additional tests for async error cases Also fixes one bug that was found when an async function throws an error after being scheduled on a task. * Add a feature flags message to reduce bandwidth We now only send 1 preview message of the latest type the client can support. We'll add a console warning when the client fails to send a feature flags message at some point in the future. * Add async tests to CI * Don't actually add new tests in this PR Will do it in a separate PR * Resolve unit test in GPU-less runner * Just remove the tests that GHA can't handle * Change line endings to UNIX-style * Avoid loading model_management.py so early Because model_management.py has a top-level `logging.info`, we have to be careful not to import that file before we call `setup_logging`. If we do, we end up having the default logging handler registered in addition to our custom one.
2025-07-10 11:46:19 -07:00
if prompt_id is None:
prompt_id = server_instance.last_prompt_id
if node_id is None:
node_id = server_instance.last_node_id
progress = {"value": value, "max": total, "prompt_id": prompt_id, "node": node_id}
Support for async node functions (#8830) * Support for async execution functions This commit adds support for node execution functions defined as async. When a node's execution function is defined as async, we can continue executing other nodes while it is processing. Standard uses of `await` should "just work", but people will still have to be careful if they spawn actual threads. Because torch doesn't really have async/await versions of functions, this won't particularly help with most locally-executing nodes, but it does work for e.g. web requests to other machines. In addition to the execute function, the `VALIDATE_INPUTS` and `check_lazy_status` functions can also be defined as async, though we'll only resolve one node at a time right now for those. * Add the execution model tests to CI * Add a missing file It looks like this got caught by .gitignore? There's probably a better place to put it, but I'm not sure what that is. * Add the websocket library for automated tests * Add additional tests for async error cases Also fixes one bug that was found when an async function throws an error after being scheduled on a task. * Add a feature flags message to reduce bandwidth We now only send 1 preview message of the latest type the client can support. We'll add a console warning when the client fails to send a feature flags message at some point in the future. * Add async tests to CI * Don't actually add new tests in this PR Will do it in a separate PR * Resolve unit test in GPU-less runner * Just remove the tests that GHA can't handle * Change line endings to UNIX-style * Avoid loading model_management.py so early Because model_management.py has a top-level `logging.info`, we have to be careful not to import that file before we call `setup_logging`. If we do, we end up having the default logging handler registered in addition to our custom one.
2025-07-10 11:46:19 -07:00
get_progress_state().update_progress(node_id, value, total, preview_image)
server_instance.send_sync("progress", progress, server_instance.client_id)
if preview_image is not None:
Support for async node functions (#8830) * Support for async execution functions This commit adds support for node execution functions defined as async. When a node's execution function is defined as async, we can continue executing other nodes while it is processing. Standard uses of `await` should "just work", but people will still have to be careful if they spawn actual threads. Because torch doesn't really have async/await versions of functions, this won't particularly help with most locally-executing nodes, but it does work for e.g. web requests to other machines. In addition to the execute function, the `VALIDATE_INPUTS` and `check_lazy_status` functions can also be defined as async, though we'll only resolve one node at a time right now for those. * Add the execution model tests to CI * Add a missing file It looks like this got caught by .gitignore? There's probably a better place to put it, but I'm not sure what that is. * Add the websocket library for automated tests * Add additional tests for async error cases Also fixes one bug that was found when an async function throws an error after being scheduled on a task. * Add a feature flags message to reduce bandwidth We now only send 1 preview message of the latest type the client can support. We'll add a console warning when the client fails to send a feature flags message at some point in the future. * Add async tests to CI * Don't actually add new tests in this PR Will do it in a separate PR * Resolve unit test in GPU-less runner * Just remove the tests that GHA can't handle * Change line endings to UNIX-style * Avoid loading model_management.py so early Because model_management.py has a top-level `logging.info`, we have to be careful not to import that file before we call `setup_logging`. If we do, we end up having the default logging handler registered in addition to our custom one.
2025-07-10 11:46:19 -07:00
# Only send old method if client doesn't support preview metadata
if not feature_flags.supports_feature(
server_instance.sockets_metadata,
server_instance.client_id,
"supports_preview_metadata",
):
server_instance.send_sync(
BinaryEventTypes.UNENCODED_PREVIEW_IMAGE,
preview_image,
server_instance.client_id,
)
comfy.utils.set_progress_bar_global_hook(hook)
2023-01-03 01:53:32 -05:00
2023-06-07 15:15:38 +02:00
def cleanup_temp():
temp_dir = folder_paths.get_temp_directory()
if os.path.exists(temp_dir):
shutil.rmtree(temp_dir, ignore_errors=True)
2023-06-07 15:15:38 +02:00
def setup_database():
try:
if dependencies_available():
init_db()
if args.enable_assets:
if asset_seeder.start(roots=("models", "input", "output"), prune_first=True, compute_hashes=args.enable_asset_hashing):
logging.info("Background asset scan initiated for models, input, output")
except Exception as e:
if "database is locked" in str(e):
logging.error(
"Database is locked. Another ComfyUI process is already using this database.\n"
"To resolve this, specify a separate database file for this instance:\n"
" --database-url sqlite:///path/to/another.db"
)
sys.exit(1)
if args.enable_assets:
logging.error(
f"Failed to initialize database: {e}\n"
"The --enable-assets flag requires a working database connection.\n"
"To resolve this, try one of the following:\n"
" 1. Install the latest requirements: pip install -r requirements.txt\n"
" 2. Specify an alternative database URL: --database-url sqlite:///path/to/your.db\n"
" 3. Use an in-memory database: --database-url sqlite:///:memory:"
)
sys.exit(1)
logging.error(f"Failed to initialize database. Please ensure you have installed the latest requirements. If the error persists, please report this as in future the database will be required: {e}")
def start_comfyui(asyncio_loop=None):
"""
Starts the ComfyUI server using the provided asyncio event loop or creates a new one.
Returns the event loop, server instance, and a function to start the server asynchronously.
"""
if args.temp_directory:
temp_dir = os.path.join(os.path.abspath(args.temp_directory), "temp")
2024-03-11 16:24:47 -04:00
logging.info(f"Setting temp directory to: {temp_dir}")
folder_paths.set_temp_directory(temp_dir)
cleanup_temp()
if not asyncio_loop:
asyncio_loop = asyncio.new_event_loop()
asyncio.set_event_loop(asyncio_loop)
prompt_server = server.PromptServer(asyncio_loop)
if args.enable_manager and not args.disable_manager_ui:
comfyui_manager.start()
hook_breaker_ac10a0.save_functions()
asyncio_loop.run_until_complete(nodes.init_extra_nodes(
init_custom_nodes=(not args.disable_all_custom_nodes) or len(args.whitelist_custom_nodes) > 0,
init_api_nodes=not args.disable_api_nodes
))
# Re-apply Comfy's cuDNN benchmark policy after custom-node imports. Benchmark
# mode can request near-card-sized autotune workspaces, and some custom nodes set it at import time.
comfy.model_management.set_cudnn_benchmark()
hook_breaker_ac10a0.restore_functions()
cuda_malloc_warning()
setup_database()
prompt_server.add_routes()
hijack_progress(prompt_server)
threading.Thread(target=prompt_worker, daemon=True, args=(prompt_server.prompt_queue, prompt_server,)).start()
2023-02-07 21:57:17 -05:00
2023-04-05 20:32:59 -04:00
if args.quick_test_for_ci:
exit(0)
os.makedirs(folder_paths.get_temp_directory(), exist_ok=True)
call_on_start = None
2023-05-06 16:59:40 -04:00
if args.auto_launch:
def startup_server(scheme, address, port):
import webbrowser
2023-08-01 01:14:17 -04:00
if os.name == 'nt' and address == '0.0.0.0':
address = '127.0.0.1'
if ':' in address:
address = "[{}]".format(address)
webbrowser.open(f"{scheme}://{address}:{port}")
call_on_start = startup_server
async def start_all():
await prompt_server.setup()
await run(prompt_server, address=args.listen, port=args.port, verbose=not args.dont_print_server, call_on_start=call_on_start)
# Returning these so that other code can integrate with the ComfyUI loop and server
return asyncio_loop, prompt_server, start_all
if __name__ == "__main__":
# Running directly, just start ComfyUI.
2025-04-12 18:58:20 -04:00
logging.info("Python version: {}".format(sys.version))
2025-01-25 15:03:57 -05:00
logging.info("ComfyUI version: {}".format(comfyui_version.__version__))
for package in ("comfy-aimdo", "comfy-kitchen"):
try:
logging.info("{} version: {}".format(package, importlib.metadata.version(package)))
except:
pass
2025-03-07 19:53:07 -05:00
if sys.version_info.major == 3 and sys.version_info.minor < 10:
logging.warning("WARNING: You are using a python version older than 3.10, please upgrade to a newer one. 3.12 and above is recommended.")
if args.disable_dynamic_vram:
logging.warning(
"Dynamic vram disabled with argument. If you have any issues with "
"dynamic vram enabled please give us a detailed reports as this "
"argument will be removed soon. If you use gguf we recommend keeping "
"dynamic vram enabled and using native ComfyUI model formats instead. "
"ComfyUI native formats like fp8 will be faster even if they are larger than your memory."
)
event_loop, _, start_all_func = start_comfyui()
2023-06-07 15:15:38 +02:00
try:
x = start_all_func()
app.logger.print_startup_warnings()
event_loop.run_until_complete(x)
2023-06-07 15:15:38 +02:00
except KeyboardInterrupt:
2024-03-11 16:24:47 -04:00
logging.info("\nStopped server")
finally:
asset_seeder.shutdown()
cleanup_temp()