298977887/ComfyUI
1
0
Fork 0
mirror of https://github.com/comfyanonymous/ComfyUI.git synced 2025-04-18 18:33:30 +00:00
Code Issues Packages Projects Releases Wiki Activity

Speed up SDXL on 16xx series with fp16 weights and manual cast.

Browse Source
This commit is contained in:
comfyanonymous 2024-02-04 13:23:43 -05:00
parent 98b80ad1f5
commit 24129d78e6
1 changed files with 3 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
comfy/model_management.py
View File

@ -496,7 +496,7 @@ def unet_dtype(device=None, model_params=0):
return torch.float8_e4m3fn
if args.fp8_e5m2_unet:
return torch.float8_e5m2
if should_use_fp16(device=device, model_params=model_params):
if should_use_fp16(device=device, model_params=model_params, manual_cast=True):
return torch.float16
return torch.float32
@ -696,7 +696,7 @@ def is_device_mps(device):
return True
return False
def should_use_fp16(device=None, model_params=0, prioritize_performance=True):
def should_use_fp16(device=None, model_params=0, prioritize_performance=True, manual_cast=False):
global directml_enabled
if device is not None:
@ -738,7 +738,7 @@ def should_use_fp16(device=None, model_params=0, prioritize_performance=True):
if x in props.name.lower():
fp16_works = True
if fp16_works:
if fp16_works or manual_cast:
free_model_memory = (get_free_memory() * 0.9 - minimum_inference_memory())
if (not prioritize_performance) or model_params * 4 > free_model_memory:
return True