xinyun/ComfyUI
1
0
Fork 0
mirror of https://git.datalinker.icu/comfyanonymous/ComfyUI synced 2025-12-08 21:44:33 +08:00
Code Issues Packages Projects Releases Wiki Activity

Speed up lora compute and lower memory usage by doing it in fp16. (#11161)

Browse Source
This commit is contained in:
comfyanonymous 2025-12-06 15:36:20 -08:00 committed by GitHub
parent 7ac7d69d94
commit 50ca97e776
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
2 changed files with 17 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

14
comfy/model_management.py
View File

@ -1492,6 +1492,20 @@ def extended_fp16_support():
return True
LORA_COMPUTE_DTYPES = {}
def lora_compute_dtype(device):
dtype = LORA_COMPUTE_DTYPES.get(device, None)
if dtype is not None:
return dtype
if should_use_fp16(device):
dtype = torch.float16
else:
dtype = torch.float32
LORA_COMPUTE_DTYPES[device] = dtype
return dtype
def soft_empty_cache(force=False):
global cpu_state
if cpu_state == CPUState.MPS:

5
comfy/model_patcher.py
View File

@ -614,10 +614,11 @@ class ModelPatcher:
if key not in self.backup:
self.backup[key] = collections.namedtuple('Dimension', ['weight', 'inplace_update'])(weight.to(device=self.offload_device, copy=inplace_update), inplace_update)
temp_dtype = comfy.model_management.lora_compute_dtype(device_to)
if device_to is not None:
temp_weight = comfy.model_management.cast_to_device(weight, device_to, torch.float32, copy=True)
temp_weight = comfy.model_management.cast_to_device(weight, device_to, temp_dtype, copy=True)
else:
temp_weight = weight.to(torch.float32, copy=True)
temp_weight = weight.to(temp_dtype, copy=True)
if convert_func is not None:
temp_weight = convert_func(temp_weight, inplace=True)