From 6988a0f5706c05ebc40ba9ef5c09243bcbf9886c Mon Sep 17 00:00:00 2001 From: Thomas Wang Date: Tue, 20 Jan 2026 12:29:54 +0800 Subject: [PATCH] Disable mla persistent kernel when not using fp8 kv_cache (#17327) --- python/sglang/srt/layers/attention/aiter_backend.py | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-) diff --git a/python/sglang/srt/layers/attention/aiter_backend.py b/python/sglang/srt/layers/attention/aiter_backend.py index e71c051a4..a99e230e2 100644 --- 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