andreas/ComfyUI
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

ops: Fix vanilla-fp8 loaded lora quality (#12390)

Browse Source 
This was missing the stochastic rounding required for fp8 downcast
to be consistent with model_patcher.patch_weight_to_device.

Missed in testing as I spend too much time with quantized tensors
and overlooked the simpler ones.
This commit is contained in:
rattus
2026-02-10 10:38:28 -08:00
committed by GitHub
parent f719f9c062
commit 123a7874a9
1 changed files with 2 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files
comfy/ops.py
+2 -2
View File
@@ -169,8 +169,8 @@ def cast_bias_weight_with_vbar(s, dtype, device, bias_dtype, non_blocking, compu
if orig.dtype == dtype and len(fns) == 0: if orig.dtype == dtype and len(fns) == 0:
#The layer actually wants our freshly saved QT #The layer actually wants our freshly saved QT
x = y x = y
else: elif update_weight:
y = x y = comfy.float.stochastic_rounding(x, orig.dtype, seed = comfy.utils.string_to_seed(s.seed_key))
if update_weight: if update_weight:
orig.copy_(y) orig.copy_(y)
for f in fns: for f in fns: