xinyun/ComfyUI
1
0
Fork 0
mirror of https://git.datalinker.icu/comfyanonymous/ComfyUI synced 2025-12-09 22:14:34 +08:00
Code Issues Packages Projects Releases Wiki Activity

quant ops: Dequantize weight in-place (#10935)

Browse Source 
In flux2 these weights are huge (200MB). As plain_tensor is a throw-away
deep copy, do this multiplication in-place to save VRAM.
This commit is contained in:
rattus 2025-11-28 02:06:30 +10:00 committed by GitHub
parent f17251bec6
commit 3f382a4f98
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 2 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

3
comfy/quant_ops.py
View File

@ -425,7 +425,8 @@ class TensorCoreFP8Layout(QuantizedLayout):
@staticmethod @staticmethod
def dequantize(qdata, scale, orig_dtype, **kwargs): def dequantize(qdata, scale, orig_dtype, **kwargs):
plain_tensor = torch.ops.aten._to_copy.default(qdata, dtype=orig_dtype) plain_tensor = torch.ops.aten._to_copy.default(qdata, dtype=orig_dtype)
return plain_tensor * scale plain_tensor.mul_(scale)
return plain_tensor
@classmethod @classmethod
def get_plain_tensors(cls, qtensor): def get_plain_tensors(cls, qtensor):