[NPU] Qwen3-VL-8B use split_qkv_rmsnorm_rope for extend (#29627)
This commit is contained in:
@@ -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,
|
||||
|
||||
Reference in New Issue
Block a user