minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 13 Packages Projects Releases Wiki Activity

[nvidia] Gemma4 nvfp4 fix (#22079)

Browse Source
This commit is contained in:
Shu Wang
2026-04-10 08:44:34 +08:00
committed by GitHub
parent cebd9c2a1e
commit 5638d40f3a
1 changed files with 8 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/layers/attention/triton_ops/extend_attention.py
+8
View File
@@ -72,6 +72,14 @@ def _get_block_sizes_for_extend_attention(Lq: int, Lv: int):
BLOCK_M, BLOCK_N = (64, 64) BLOCK_M, BLOCK_N = (64, 64)
else: else:
BLOCK_M, BLOCK_N = (32, 32) BLOCK_M, BLOCK_N = (32, 32)
elif _is_cuda and CUDA_CAPABILITY[0] == 10:
# Blackwell data-center architecture (GB200, B200, sm_100a)
# sm_100a has different register constraints from Hopper; Hopper block sizes
# cause PTX register exhaustion (>255 regs) for large head dims (Lq=512).
if Lq <= 256:
BLOCK_M, BLOCK_N = (64, 64)
else:
BLOCK_M, BLOCK_N = (16, 64)
elif _is_cuda and CUDA_CAPABILITY[0] >= 9: elif _is_cuda and CUDA_CAPABILITY[0] >= 9:
# Hopper architecture (H100, etc.) # Hopper architecture (H100, etc.)
if Lq <= 256: if Lq <= 256: