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

Keep NVFP4 blockscale swizzle padding on the input device (#39141)

Browse Source
This commit is contained in:
Ziang Li
2026-09-11 22:18:02 -07:00
committed by GitHub
parent c389dd0863
commit dbd4302bbb
1 changed files with 4 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/layers/quantization/utils.py
+4 -1
View File
@@ -585,7 +585,10 @@ def swizzle_blockscale(scale: torch.Tensor):
round_up_multiple = lambda x, m: (x + m - 1) // m * m round_up_multiple = lambda x, m: (x + m - 1) // m * m
M_padded = round_up_multiple(M, 128) M_padded = round_up_multiple(M, 128)
K_padded = round_up_multiple(K, 4) K_padded = round_up_multiple(K, 4)
padded_scale = torch.zeros((B, M_padded, K_padded), dtype=scale.dtype) # Online weight updates do not necessarily run under a CUDA device context.
padded_scale = torch.zeros(
(B, M_padded, K_padded), dtype=scale.dtype, device=scale.device
)
padded_scale[:B, :M, :K] = scale padded_scale[:B, :M, :K] = scale
batches, rows, cols = padded_scale.shape batches, rows, cols = padded_scale.shape
assert rows % 128 == 0 assert rows % 128 == 0