wangbo/ComfyUI
1
0
Fork 0
mirror of https://github.com/comfyanonymous/ComfyUI.git synced 2026-01-24 21:30:15 +08:00
Code Issues Actions 12 Packages Projects Releases Wiki Activity

Merge branch 'comfyanonymous:master' into master

Browse Source
This commit is contained in:
patientx 2025-02-23 14:51:43 +03:00 committed by GitHub
commit 8142770e5f
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 7 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

7
comfy/model_management.py
View File

@ -257,9 +257,12 @@ except:
# torch.backends.cuda.enable_flash_sdp(True) # torch.backends.cuda.enable_flash_sdp(True)
# torch.backends.cuda.enable_mem_efficient_sdp(True) # torch.backends.cuda.enable_mem_efficient_sdp(True)
PRIORITIZE_FP16 = False # TODO: remove and replace with something that shows exactly which dtype is faster than the other
try: try:
if is_nvidia() and args.fast: if is_nvidia() and args.fast:
torch.backends.cuda.matmul.allow_fp16_accumulation = True torch.backends.cuda.matmul.allow_fp16_accumulation = True
PRIORITIZE_FP16 = True # TODO: limit to cards where it actually boosts performance
except: except:
pass pass
@ -682,6 +685,10 @@ def unet_dtype(device=None, model_params=0, supported_dtypes=[torch.float16, tor
if model_params * 2 > free_model_memory: if model_params * 2 > free_model_memory:
return fp8_dtype return fp8_dtype
if PRIORITIZE_FP16:
if torch.float16 in supported_dtypes and should_use_fp16(device=device, model_params=model_params):
return torch.float16
for dt in supported_dtypes: for dt in supported_dtypes:
if dt == torch.float16 and should_use_fp16(device=device, model_params=model_params): if dt == torch.float16 and should_use_fp16(device=device, model_params=model_params):
if torch.float16 in supported_dtypes: if torch.float16 in supported_dtypes: