jinx/ComfyUI
1
Fork 0
mirror of https://github.com/comfyanonymous/ComfyUI.git synced 2025-08-02 23:14:49 +08:00
Code Issues Activity

Add a weight_dtype fp8_e4m3fn_fast to the Diffusion Model Loader node.

Browse Source 
This is used to load weights in fp8 and use fp8 matrix multiplication.
This commit is contained in:
comfyanonymous
2024-10-09 19:43:17 -04:00
parent 203942c8b2
commit e38c94228b
6 changed files with 27 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files

13
comfy/model_management.py
View File

@@ -145,7 +145,7 @@ total_ram = psutil.virtual_memory().total / (1024 * 1024)
logging.info("Total VRAM {:0.0f} MB, total RAM {:0.0f} MB".format(total_vram, total_ram))
try:
logging.info("pytorch version: {}".format(torch.version.__version__))
logging.info("pytorch version: {}".format(torch_version))
except:
pass
@@ -1065,6 +1065,9 @@ def should_use_bf16(device=None, model_params=0, prioritize_performance=True, ma
return False
def supports_fp8_compute(device=None):
if not is_nvidia():
return False
props = torch.cuda.get_device_properties(device)
if props.major >= 9:
return True
@@ -1072,6 +1075,14 @@ def supports_fp8_compute(device=None):
return False
if props.minor < 9:
return False
if int(torch_version[0]) < 2 or (int(torch_version[0]) == 2 and int(torch_version[2]) < 3):
return False
if WINDOWS:
if (int(torch_version[0]) == 2 and int(torch_version[2]) < 4):
return False
return True
def soft_empty_cache(force=False):