From cf0b549d4828b6b9f6d98277c36784ab3c79ff6d Mon Sep 17 00:00:00 2001 From: comfyanonymous Date: Fri, 28 Feb 2025 02:48:20 -0500 Subject: [PATCH] --fast now takes a number as argument to indicate how fast you want it. The idea is that you can indicate how much quality vs speed you want. At the moment: --fast 2 enables fp16 accumulation if your pytorch supports it. --fast 5 enables fp8 matrix mult on fp8 models and the optimization above. --fast without a number enables all optimizations. --- comfy/cli_args.py | 2 +- comfy/model_management.py | 3 ++- comfy/ops.py | 2 +- 3 files changed, 4 insertions(+), 3 deletions(-) diff --git a/comfy/cli_args.py b/comfy/cli_args.py index a906ff1c..10c142e6 100644 --- a/comfy/cli_args.py +++ b/comfy/cli_args.py @@ -130,7 +130,7 @@ parser.add_argument("--default-hashing-function", type=str, choices=['md5', 'sha parser.add_argument("--disable-smart-memory", action="store_true", help="Force ComfyUI to agressively offload to regular ram instead of keeping models in vram when it can.") parser.add_argument("--deterministic", action="store_true", help="Make pytorch use slower deterministic algorithms when it can. Note that this might not make images deterministic in all cases.") -parser.add_argument("--fast", action="store_true", help="Enable some untested and potentially quality deteriorating optimizations.") +parser.add_argument("--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.") 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.") diff --git a/comfy/model_management.py b/comfy/model_management.py index afbb133d..5eb2e5ad 100644 --- a/comfy/model_management.py +++ b/comfy/model_management.py @@ -280,9 +280,10 @@ if ENABLE_PYTORCH_ATTENTION: PRIORITIZE_FP16 = False # TODO: remove and replace with something that shows exactly which dtype is faster than the other try: - if is_nvidia() and args.fast: + if is_nvidia() and args.fast >= 2: torch.backends.cuda.matmul.allow_fp16_accumulation = True PRIORITIZE_FP16 = True # TODO: limit to cards where it actually boosts performance + logging.info("Enabled fp16 accumulation.") except: pass diff --git a/comfy/ops.py b/comfy/ops.py index 30014477..905ea90f 100644 --- a/comfy/ops.py +++ b/comfy/ops.py @@ -360,7 +360,7 @@ 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) and not disable_fast_fp8: + if fp8_compute and (fp8_optimizations or args.fast >= 5) and not disable_fast_fp8: return fp8_ops if compute_dtype is None or weight_dtype == compute_dtype: