minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

[Qwen3-Next] Enable fused_qkvzba_split_reshape_cat also for prefill (#18917)

Browse Source
This commit is contained in:
YAMY
2026-02-22 13:57:17 +08:00
committed by GitHub
parent 8cf003c44b
commit 5995bfec63
1 changed files with 1 additions and 7 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/models/qwen3_next.py
+1 -7
View File
@@ -405,17 +405,11 @@ class Qwen3GatedDeltaNet(nn.Module):
hidden_states: torch.Tensor, hidden_states: torch.Tensor,
forward_batch: ForwardBatch, forward_batch: ForwardBatch,
): ):
is_cuda_graph = forward_batch.forward_mode.is_cuda_graph()
projected_states_qkvz, projected_states_ba = self._forward_input_proj( projected_states_qkvz, projected_states_ba = self._forward_input_proj(
hidden_states hidden_states
) )
if ( if self.num_v_heads // self.num_k_heads in [1, 2, 4] and not _is_cpu:
self.num_v_heads // self.num_k_heads in [1, 2, 4]
and is_cuda_graph
and not _is_cpu
):
mixed_qkv, z, b, a = fused_qkvzba_split_reshape_cat( mixed_qkv, z, b, a = fused_qkvzba_split_reshape_cat(
projected_states_qkvz, projected_states_qkvz,
projected_states_ba, projected_states_ba,