[NPU] Qwen3-VL-8B use split_qkv_rmsnorm_rope for extend (#29627)

This commit is contained in:
silencejade
2026-06-30 15:55:50 +08:00
committed by GitHub
parent bae78a44da
commit ff51acd67b
+1 -4
View File
@@ -287,10 +287,7 @@ class Qwen3Attention(nn.Module):
positions, hidden_states, forward_batch
)
save_kv_cache = False
elif (
not _is_npu
or forward_batch.forward_mode.is_extend_or_draft_extend_or_mixed()
):
elif not _is_npu:
q, k, v = self.forward_prepare_native(
positions=positions,
hidden_states=hidden_states,