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

revert BLOCK and num_warps on HIP (#3722)

Browse Source
This commit is contained in:
HAI
2025-02-20 07:30:18 -08:00
committed by GitHub
parent 1eb8eade2b
commit 6252ade985
1 changed files with 2 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

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

@@ -329,8 +329,8 @@ def extend_attention_fwd(
BLOCK_DV = triton.next_power_of_2(Lv) BLOCK_DV = triton.next_power_of_2(Lv)
if is_hip_: if is_hip_:
BLOCK_M, BLOCK_N = (32, 32) BLOCK_M, BLOCK_N = (64, 64)
num_warps = 2 num_warps = 4
else: else:
if is_cuda_available and CUDA_CAPABILITY[0] >= 9: if is_cuda_available and CUDA_CAPABILITY[0] >= 9: