From 9305d100999b7ccaf5f03e4c85f9dd278ae9e1cd Mon Sep 17 00:00:00 2001 From: cen121212 Date: Thu, 25 Jun 2026 08:55:52 +0800 Subject: [PATCH] [NPU] adapt_fused_rope_qk_mqa_optimize (#28872) --- python/sglang/srt/layers/rotary_embedding/base.py | 5 +---- 1 file changed, 1 insertion(+), 4 deletions(-) diff --git a/python/sglang/srt/layers/rotary_embedding/base.py b/python/sglang/srt/layers/rotary_embedding/base.py index 5c20732e0..39cb8f7e2 100644 --- a/python/sglang/srt/layers/rotary_embedding/base.py +++ b/python/sglang/srt/layers/rotary_embedding/base.py @@ -294,10 +294,7 @@ class RotaryEmbedding(MultiPlatformOp): else: cos_sin = self.cos_sin_cache.index_select(0, positions) - if ( - fused_rope_qk_mqa is not None - and query.shape[0] * query.shape[1] < 65535 - ): + if fused_rope_qk_mqa is not None and query.shape[0] < 65535: return fused_rope_qk_mqa( query, key,