minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

Fix block-scale swizzling device placement (#37849)

Browse Source
This commit is contained in:
Lianmin Zheng
2026-09-03 10:47:04 -07:00
committed by GitHub
parent 23ab10a63e
commit 619ab2bcce
1 changed files with 1 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/layers/quantization/utils.py
+1 -1
View File
@@ -615,7 +615,7 @@ def swizzle_blockscale(scale: torch.Tensor):
assert cols % 4 == 0 assert cols % 4 == 0
padded_scale = padded_scale.reshape(batches, rows // 128, 4, 32, cols // 4, 4) padded_scale = padded_scale.reshape(batches, rows // 128, 4, 32, cols // 4, 4)
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().to(scale.device)
return ( return (
swizzled_scale.reshape(M_padded, K_padded) swizzled_scale.reshape(M_padded, K_padded)
if scale_ndim == 2 if scale_ndim == 2