EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity

Fix mismatch between padded_scales shape and reshape dimensions in modelopt quantization (#8766)

Browse Source
This commit is contained in:
Jianwei Dong
2025-08-12 14:44:40 +08:00
committed by GitHub
parent c46c75f8c0
commit 83262dcb29
1 changed files with 4 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
python/sglang/srt/layers/quantization/modelopt_quant.py
View File

@@ -677,9 +677,9 @@ class ModelOptFp4LinearMethod(LinearMethodBase):
padded_scales = padded_scales.permute((0, 1, 4, 3, 2, 5)) padded_scales = padded_scales.permute((0, 1, 4, 3, 2, 5))
padded_scales = padded_scales.contiguous().cuda() padded_scales = padded_scales.contiguous().cuda()
padded_scales = ( padded_scales = (
padded_scales.reshape(M, K) padded_scales.reshape(M_padded, K_padded)
if scale_ndim == 2 if scale_ndim == 2
else padded_scales.reshape(B, M, K) else padded_scales.reshape(B, M_padded, K_padded)
) )
layer.weight_scale_interleaved = Parameter(padded_scales, requires_grad=False) layer.weight_scale_interleaved = Parameter(padded_scales, requires_grad=False)
@@ -878,9 +878,9 @@ class ModelOptNvFp4FusedMoEMethod(FusedMoEMethodBase):
swizzled_scale = padded_scale.permute((0, 1, 4, 3, 2, 5)) swizzled_scale = padded_scale.permute((0, 1, 4, 3, 2, 5))
swizzled_scale = swizzled_scale.contiguous().cuda() swizzled_scale = swizzled_scale.contiguous().cuda()
return ( return (
swizzled_scale.reshape(M, K) swizzled_scale.reshape(M_padded, K_padded)
if scale_ndim == 2 if scale_ndim == 2
else swizzled_scale.reshape(B, M, K) else swizzled_scale.reshape(B, M_padded, K_padded)
) )
def prepare_static_weights_for_kernel( def prepare_static_weights_for_kernel(