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

fix: temporary solution for DeepSeek V2 H100 layout conversion issue (#1060)

Browse Source 
Co-authored-by: ispobock <ISPObaoke@163.com>
This commit is contained in:
Yineng Zhang
2024-08-13 13:48:54 +08:00
committed by GitHub
parent 162f3ccb01
commit 65915f9f3e
1 changed files with 3 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
python/sglang/srt/layers/extend_attention.py
View File

@@ -275,7 +275,9 @@ def extend_attention_fwd(
BLOCK_DPE = 0 BLOCK_DPE = 0
BLOCK_DV = Lv BLOCK_DV = Lv
if CUDA_CAPABILITY[0] >= 8: if CUDA_CAPABILITY[0] >= 9:
BLOCK_M, BLOCK_N = (128, 64)
elif CUDA_CAPABILITY[0] >= 8:
BLOCK_M, BLOCK_N = (128, 128) if Lq <= 128 else (64, 64) BLOCK_M, BLOCK_N = (128, 128) if Lq <= 128 else (64, 64)
else: else:
BLOCK_M, BLOCK_N = (64, 64) if Lq <= 128 else (32, 32) BLOCK_M, BLOCK_N = (64, 64) if Lq <= 128 else (32, 32)