RomanNum3ral
/
ComfyUI
mirror of https://github.com/Comfy-Org/ComfyUI.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

fix: preserve linear bias during INT4 reconstruction

This commit is contained in:
JWLHS 2026-07-14 04:13:00 +08:00
parent 335eb29ca5
commit 4ee3d16d84
1 changed files with 5 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
comfy/quantization/torchao/quantize.py
View File

@ -37,7 +37,6 @@ def quantize_model(
return model return model
# torchao plain_int32 safetensors suffixes
_QUANT_SUFFIXES = ( _QUANT_SUFFIXES = (
".weight_scale", ".weight_zp", ".weight_scale", ".weight_zp",
".weight_b0", ".weight_b1", ".weight_b0", ".weight_b1",
@ -89,6 +88,8 @@ def reconstruct_int4_state_dict(
for suffix in _QUANT_SUFFIXES: for suffix in _QUANT_SUFFIXES:
sd.pop(f"{base}{suffix}", None) sd.pop(f"{base}{suffix}", None)
bias = sd.pop(f"{base}.bias", None)
in_f = qdata.shape[1] * 8 in_f = qdata.shape[1] * 8
out_f = qdata.shape[0] out_f = qdata.shape[0]
@ -99,12 +100,15 @@ def reconstruct_int4_state_dict(
scale=scale.to(torch.float16), scale=scale.to(torch.float16),
zp=zp.to(torch.int8), zp=zp.to(torch.int8),
block_size=(b0.item(), b1.item()), block_size=(b0.item(), b1.item()),
bias=bias,
) )
if device.type != "cpu": if device.type != "cpu":
layer._qdata = layer._qdata.to(device) layer._qdata = layer._qdata.to(device)
layer._scale = layer._scale.to(device) layer._scale = layer._scale.to(device)
layer._zp = layer._zp.to(device) layer._zp = layer._zp.to(device)
if bias is not None:
layer.bias.data = layer.bias.data.to(device)
replaced[base] = layer replaced[base] = layer