revert BLOCK and num_warps on HIP (#3722)

2025-02-20 07:30:18 -08:00
parent 1eb8eade2b
commit 6252ade985
1 changed files with 2 additions and 2 deletions
--- a/python/sglang/srt/layers/attention/triton_ops/extend_attention.py
+++ b/python/sglang/srt/layers/attention/triton_ops/extend_attention.py
@@ -329,8 +329,8 @@ def extend_attention_fwd(
    BLOCK_DV = triton.next_power_of_2(Lv)

    if is_hip_:
-        BLOCK_M, BLOCK_N = (32, 32)
-        num_warps = 2
+        BLOCK_M, BLOCK_N = (64, 64)
+        num_warps = 4

    else:
        if is_cuda_available and CUDA_CAPABILITY[0] >= 9: