From 2d6c12e2fd666d5159891249226d6d3a9a07f4e4 Mon Sep 17 00:00:00 2001 From: sushil Dubey Date: Tue, 25 Aug 2026 10:57:42 +0530 Subject: [PATCH] [AMD] Don't request the unused softmax LSE in the AITER diffusion backend (#32926) Co-authored-by: Claude Opus 5 (1M context) Co-authored-by: YC Yen-Ching Tseng --- .../multimodal_gen/runtime/layers/attention/backends/aiter.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/python/sglang/multimodal_gen/runtime/layers/attention/backends/aiter.py b/python/sglang/multimodal_gen/runtime/layers/attention/backends/aiter.py index f897214b7..9606c4849 100755 --- a/python/sglang/multimodal_gen/runtime/layers/attention/backends/aiter.py +++ b/python/sglang/multimodal_gen/runtime/layers/attention/backends/aiter.py @@ -204,14 +204,14 @@ class AITerImpl(AttentionImpl): ) # BF16 path - output, _ = aiter.flash_attn_func( + output = aiter.flash_attn_func( query, key, value, dropout_p=self.dropout_p, causal=self.causal, return_attn_probs=False, - return_lse=True, + return_lse=False, ) return output