[NPU] Qwen3-VL-30B use split_qkv_rmsnorm_rope for extend (#29505)

This commit is contained in:
silencejade
2026-06-29 14:35:19 +08:00
committed by GitHub
parent 38d4ffcd86
commit f85cc94d82
+1 -4
View File
@@ -661,10 +661,7 @@ class Qwen3MoeAttention(nn.Module):
):
if hidden_states.shape[0] == 0:
return hidden_states, forward_batch, None
if (
not _is_npu
or forward_batch.forward_mode.is_extend_or_draft_extend_or_mixed()
):
if not _is_npu:
return self.forward_prepare_native(
positions=positions,
hidden_states=hidden_states,