From 5995bfec6368faa48fd8ad159bf9805c15ebc217 Mon Sep 17 00:00:00 2001 From: YAMY <74099316+YAMY1234@users.noreply.github.com> Date: Sat, 21 Feb 2026 21:57:17 -0800 Subject: [PATCH] [Qwen3-Next] Enable fused_qkvzba_split_reshape_cat also for prefill (#18917) --- python/sglang/srt/models/qwen3_next.py | 8 +------- 1 file changed, 1 insertion(+), 7 deletions(-) diff --git a/python/sglang/srt/models/qwen3_next.py b/python/sglang/srt/models/qwen3_next.py index b565cbccc..39a086c98 100644 --- a/python/sglang/srt/models/qwen3_next.py +++ b/python/sglang/srt/models/qwen3_next.py @@ -405,17 +405,11 @@ class Qwen3GatedDeltaNet(nn.Module): hidden_states: torch.Tensor, forward_batch: ForwardBatch, ): - is_cuda_graph = forward_batch.forward_mode.is_cuda_graph() - projected_states_qkvz, projected_states_ba = self._forward_input_proj( hidden_states ) - if ( - self.num_v_heads // self.num_k_heads in [1, 2, 4] - and is_cuda_graph - and not _is_cpu - ): + if self.num_v_heads // self.num_k_heads in [1, 2, 4] and not _is_cpu: mixed_qkv, z, b, a = fused_qkvzba_split_reshape_cat( projected_states_qkvz, projected_states_ba,