Disable mla persistent kernel when not using fp8 kv_cache (#17327)

2026-01-20 12:29:54 +08:00
parent 1b192cf198
commit 6988a0f570
1 changed files with 3 additions and 1 deletions
--- a/python/sglang/srt/layers/attention/aiter_backend.py
+++ b/python/sglang/srt/layers/attention/aiter_backend.py
@@ -187,7 +187,9 @@ class AiterAttnBackend(AttentionBackend):

            # current persist a16w16 mla_decode kernel does not support head_num = 128
            # need to fall back to non-persist
-            if self.kv_cache_dtype is not fp8_dtype and self.enable_dp_attention:
+            # only use mla_ps_kernel when fp8 kv_cache
+            # for non-fp8 kv_cache, use non-persist kernel to avoid performance degradation
+            if self.kv_cache_dtype is not fp8_dtype:
                _use_mla_ps_kernel = False
                fast_mode = False
                intra_batch_mode = False