The most powerful and modular diffusion model GUI, api and backend with a graph/nodes interface. The fastest local inference engine in the world.
Limit Windows multi-GPU visibility (#15737)
rattus5ab2f7a2d676
diff --git a/comfy/cli_args.py b/comfy/cli_args.pyindex c6660846d..659e772ed 100644--- a/comfy/cli_args.py+++ b/comfy/cli_args.py@@ -74,7 +74,7 @@ parser.add_argument("--temp-directory", type=str, default=None, help="Set the Coparser.add_argument("--input-directory", type=str, default=None, help="Set the ComfyUI input directory. Overrides --base-directory.")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=str, default=None, metavar="DEVICE_ID", help="Set the ids of cuda devices this instance will use, as a comma-separated list (e.g. '0' or '0,1'). All other devices will not be visible.")+parser.add_argument("--cuda-device", type=str, default=None, metavar="DEVICE_ID", help="Set the ids of cuda devices this instance will use, as a comma-separated list (e.g. '0' or '0,1'), or 'all' to leave all currently visible devices available. All other devices will not be visible.")parser.add_argument("--default-device", type=int, default=None, metavar="DEFAULT_DEVICE_ID", help="Set the id of the default device, all other devices will stay visible.")cm_group = parser.add_mutually_exclusive_group()cm_group.add_argument("--cuda-malloc", action="store_true", help="Enable cudaMallocAsync (enabled by default for torch 2.0 and up).")diff --git a/cuda_malloc.py b/cuda_malloc.pyindex 8c4422db8..022c9ae88 100644--- a/cuda_malloc.py+++ b/cuda_malloc.py@@ -28,19 +28,19 @@ def get_gpu_names():device_info = DISPLAY_DEVICEA()device_info.cb = ctypes.sizeof(device_info)device_index = 0- gpu_names = set()+ gpu_names = []while user32.EnumDisplayDevicesA(None, device_index, ctypes.byref(device_info), 0):device_index += 1- gpu_names.add(device_info.DeviceString.decode('utf-8'))+ gpu_names.append(device_info.DeviceString.decode('utf-8'))return gpu_namesreturn enum_display_devices()else:- gpu_names = set()+ gpu_names = []out = subprocess.check_output(['nvidia-smi', '-L'])for l in out.split(b'\n'):if len(l) > 0:- gpu_names.add(l.decode('utf-8').split(' (UUID')[0])+ gpu_names.append(l.decode('utf-8').split(' (UUID')[0])return gpu_namesblacklist = {"GeForce GTX TITAN X", "GeForce GTX 980", "GeForce GTX 970", "GeForce GTX 960", "GeForce GTX 950", "GeForce 945M",diff --git a/main.py b/main.pyindex b6f7d6cef..9fadc25b9 100644--- a/main.py+++ b/main.py@@ -17,7 +17,7 @@ import importlib.metadataimport folder_pathsimport timefrom comfy.cli_args import enables_dynamic_vram-from app.logger import setup_logger+from app.logger import setup_logger, log_startup_warningconsole_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)@@ -41,6 +41,44 @@ if __name__ == "__main__":os.environ['HF_HUB_DISABLE_TELEMETRY'] = '1'os.environ['DO_NOT_TRACK'] = '1'+ import cuda_malloc++ if os.name == "nt":+ cuda_visibility = os.environ.get("CUDA_VISIBLE_DEVICES")+ device_selection = args.cuda_device++ try:+ gpu_count = sum("NVIDIA" in name.upper() for name in cuda_malloc.get_gpu_names())+ except OSError:+ gpu_count = 0++ if gpu_count > 1:+ warning = None+ multiple_visible = False+ if device_selection is None and args.default_device is None and cuda_visibility is None:+ os.environ["CUDA_VISIBLE_DEVICES"] = "0"+ warning = "Multiple NVIDIA GPUs detected. ComfyUI will use GPU 0 only on Windows by default. To restore all GPUs, pass --cuda-device all --disable-pinned-memory."+ elif device_selection == "all":+ multiple_visible = cuda_visibility is None or "," in cuda_visibility+ elif device_selection is not None:+ multiple_visible = "," in device_selection+ elif args.default_device is not None:+ multiple_visible = True+ else:+ multiple_visible = "," in cuda_visibility++ if multiple_visible and not args.disable_pinned_memory:+ warning = "Multiple NVIDIA GPUs are visible on Windows with pinned memory enabled. Restart with --disable-pinned-memory to avoid CUDA host-transfer failures."++ if warning:+ log_startup_warning(f"""+________________________________________________________________________+WARNING WARNING WARNING WARNING WARNING++{warning}+________________________________________________________________________+""".strip())+faulthandler.enable(file=sys.stderr, all_threads=args.debug_hang)if __name__ == "__main__" and args.debug_hang:dumping_traceback = False@@ -74,7 +112,7 @@ if os.name == "nt":if __name__ == "__main__":os.environ['TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL'] = '1'- if args.default_device is not None:+ if args.default_device is not None and args.cuda_device != "all":default_dev = args.default_devicedevices = list(range(32))devices.remove(default_dev)@@ -83,7 +121,9 @@ if __name__ == "__main__":os.environ['CUDA_VISIBLE_DEVICES'] = str(devices)os.environ['HIP_VISIBLE_DEVICES'] = str(devices)- if args.cuda_device is not None:+ if args.cuda_device == "all":+ logging.info("Set cuda devices to all")+ elif 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)@@ -97,7 +137,6 @@ if __name__ == "__main__":if 'CUBLAS_WORKSPACE_CONFIG' not in os.environ:os.environ['CUBLAS_WORKSPACE_CONFIG'] = ":4096:8"- import cuda_mallocif "rocm" in cuda_malloc.get_torch_version_noimport():os.environ['OCL_SET_SVM_SIZE'] = '262144' # set at the request of AMD