xinyun/ComfyUI
1
0
Fork 0
mirror of https://git.datalinker.icu/comfyanonymous/ComfyUI synced 2026-08-02 23:07:04 +08:00
Code Issues Packages Projects Releases Wiki Activity

quant ops: Dequantize weight in-place

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 2025-11-27 20:46:37 +10:00
parent f17251bec6
commit d114ea0d44
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
def dequantize(qdata, scale, orig_dtype, **kwargs):
plain_tensor = torch.ops.aten._to_copy.default(qdata, dtype=orig_dtype)
return plain_tensor * scale
plain_tensor.mul_(scale)
return plain_tensor
@classmethod
def get_plain_tensors(cls, qtensor):