The most powerful and modular diffusion model GUI, api and backend with a graph/nodes interface. The fastest local inference engine in the world.
Use enum list for --fast options (#7024)
Chenlei Hu4d55f16ae872
diff --git a/comfy/cli_args.py b/comfy/cli_args.pyindex 10c142e67..6aee14b3e 100644--- a/comfy/cli_args.py+++ b/comfy/cli_args.py@@ -130,7 +130,12 @@ parser.add_argument("--default-hashing-function", type=str, choices=['md5', 'shaparser.add_argument("--disable-smart-memory", action="store_true", help="Force ComfyUI to agressively offload to regular ram instead of keeping models in vram when it can.")parser.add_argument("--deterministic", action="store_true", help="Make pytorch use slower deterministic algorithms when it can. Note that this might not make images deterministic in all cases.")-parser.add_argument("--fast", metavar="number", type=int, const=99, default=0, nargs="?", help="Enable some untested and potentially quality deteriorating optimizations. You can pass a number from 0 to 10 for a bigger speed vs quality tradeoff. Using --fast with no number means maximum speed. 2 or larger enables fp16 accumulation, 5 or larger enables fp8 matrix multiplication.")++class PerformanceFeature(enum.Enum):+ Fp16Accumulation = "fp16_accumulation"+ Fp8Optimization = "fp8_optimization"++parser.add_argument("--fast", nargs="*", type=PerformanceFeature, help="Enable some untested and potentially quality deteriorating optimizations.")parser.add_argument("--dont-print-server", action="store_true", help="Don't print server output.")parser.add_argument("--quick-test-for-ci", action="store_true", help="Quick test for CI.")@@ -194,3 +199,14 @@ if args.disable_auto_launch:if args.force_fp16:args.fp16_unet = True+++# '--fast' is not provided, use an empty set+if args.fast is None:+ args.fast = set()+# '--fast' is provided with an empty list, enable all optimizations+elif args.fast == []:+ args.fast = set(PerformanceFeature)+# '--fast' is provided with a list of performance features, use that list+else:+ args.fast = set(args.fast)diff --git a/comfy/model_management.py b/comfy/model_management.pyindex 5eb2e5ad6..bc90e3dff 100644--- a/comfy/model_management.py+++ b/comfy/model_management.py@@ -19,7 +19,7 @@import psutilimport loggingfrom enum import Enum-from comfy.cli_args import args+from comfy.cli_args import args, PerformanceFeatureimport torchimport sysimport platform@@ -280,7 +280,7 @@ if ENABLE_PYTORCH_ATTENTION:PRIORITIZE_FP16 = False # TODO: remove and replace with something that shows exactly which dtype is faster than the othertry:- if is_nvidia() and args.fast >= 2:+ if is_nvidia() and PerformanceFeature.Fp16Accumulation in args.fast:torch.backends.cuda.matmul.allow_fp16_accumulation = TruePRIORITIZE_FP16 = True # TODO: limit to cards where it actually boosts performancelogging.info("Enabled fp16 accumulation.")diff --git a/comfy/ops.py b/comfy/ops.pyindex 905ea90f6..d98b2b0e7 100644--- a/comfy/ops.py+++ b/comfy/ops.py@@ -18,7 +18,7 @@import torchimport comfy.model_management-from comfy.cli_args import args+from comfy.cli_args import args, PerformanceFeatureimport comfy.floatcast_to = comfy.model_management.cast_to #TODO: remove once no more references@@ -360,7 +360,11 @@ def pick_operations(weight_dtype, compute_dtype, load_device=None, disable_fast_if scaled_fp8 is not None:return scaled_fp8_ops(fp8_matrix_mult=fp8_compute, scale_input=True, override_dtype=scaled_fp8)- if fp8_compute and (fp8_optimizations or args.fast >= 5) and not disable_fast_fp8:+ if (+ fp8_compute and+ (fp8_optimizations or PerformanceFeature.Fp8Optimization in args.fast) and+ not disable_fast_fp8+ ):return fp8_opsif compute_dtype is None or weight_dtype == compute_dtype: