Additional tuning for grouped paged attention kernel

Changed: - num_stages
sgl-project · Feb 1, 2025 · eb34954 · eb34954
1 parent 9c5980e
commit eb34954
Showing 1 changed file with 1 addition and 1 deletion.
diff --git a/python/sglang/srt/layers/attention/triton_ops/decode_attention.py b/python/sglang/srt/layers/attention/triton_ops/decode_attention.py
@@ -467,7 +467,7 @@ def _decode_grouped_att_m_fwd(
         NUM_KV_SPLITS=NUM_KV_SPLITS,
         logit_cap=logit_cap,
         num_warps=4,
-        num_stages=2,
+        num_stages=1,
         Lk=Lk,
         Lv=Lv,
         **extra_kargs,