From e138f7960ac9a8bddc4eeead09f0545b860ded8b Mon Sep 17 00:00:00 2001 From: "jacky.cheng" Date: Wed, 25 Feb 2026 12:50:28 +0800 Subject: [PATCH] [AMD] Fix accuracy while using --enable-dp-attention (#19247) Co-authored-by: yichiche@amd.com --- python/sglang/srt/layers/attention/aiter_backend.py | 5 ++++- 1 file changed, 4 insertions(+), 1 deletion(-) mode change 100644 => 100755 python/sglang/srt/layers/attention/aiter_backend.py diff --git a/python/sglang/srt/layers/attention/aiter_backend.py b/python/sglang/srt/layers/attention/aiter_backend.py old mode 100644 new mode 100755 index 97d6ee573..fd8661234 --- a/python/sglang/srt/layers/attention/aiter_backend.py +++ b/python/sglang/srt/layers/attention/aiter_backend.py @@ -203,7 +203,10 @@ class AiterAttnBackend(AttentionBackend): # need to fall back to non-persist # only use mla_ps_kernel when fp8 kv_cache # for non-fp8 kv_cache on tp8, use non-persist kernel to avoid performance degradation - if self.num_head == 16 and self.kv_cache_dtype is not fp8_dtype: + # head_num=16 (tp8 perf issue), head_num=128 (unsupported, like tp1 or --enable-dp-attention with tp8-dp8) + if ( + self.num_head == 16 or self.num_head == 128 + ) and self.kv_cache_dtype is not fp8_dtype: _use_mla_ps_kernel = False fast_mode = False intra_batch_mode = False