diff --git a/python/sglang/srt/layers/attention/triton_ops/extend_attention.py b/python/sglang/srt/layers/attention/triton_ops/extend_attention.py
index 74ad872bc..608f9bab0 100644
--- 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: