From 44a92e54b9efdd56be68a915ed717f4964f2e6ab Mon Sep 17 00:00:00 2001 From: billishyahao Date: Tue, 1 Sep 2026 23:35:31 +0800 Subject: [PATCH] [AMD] fix aiter cannot get heuristic kernel regression (#37438) --- python/sglang/srt/layers/attention/aiter_backend.py | 13 ++++++++++--- 1 file changed, 10 insertions(+), 3 deletions(-) diff --git a/python/sglang/srt/layers/attention/aiter_backend.py b/python/sglang/srt/layers/attention/aiter_backend.py index 37672d57f..fb7981fbb 100755 --- a/python/sglang/srt/layers/attention/aiter_backend.py +++ b/python/sglang/srt/layers/attention/aiter_backend.py @@ -466,9 +466,16 @@ class AiterAttnBackend(AttentionBackend): intra_batch_mode = False # Zero-pad topology (h12->qh16): prefer Gluon decode over PS kernel. if self.head_pad_mode == "zero" and self.kv_cache_dtype == fp8_dtype: - _use_mla_ps_kernel = False - fast_mode = False - intra_batch_mode = False + # Disable ps only when gluon kernel is selected to avoid falling + # back to incorrect aiter kernel + if prefer_mla_gluon_decode( + head_pad_mode=self.head_pad_mode, + num_head=self.num_head, + kv_cache_dtype=self.kv_cache_dtype, + ): + _use_mla_ps_kernel = False + fast_mode = False + intra_batch_mode = False log_mla_gluon_capability(logger) self.max_split_per_batch = 32 if _use_mla_ps_kernel else None