Merge branch 'v0.5.4_dev_qwennext' into 'v0.5.4_dev'

适配qwen3-next

See merge request OpenDAS/sglang!4

python/sglang/srt/models/qwen3_next.py

@@ -396,7 +396,7 @@ class Qwen3GatedDeltaNet(nn.Module):
     def _forward_input_proj(self, hidden_states: torch.Tensor):
         DUAL_STREAM_TOKEN_THRESHOLD = 1024 if not _is_npu else 0
         seq_len, _ = hidden_states.shape
-        if seq_len < DUAL_STREAM_TOKEN_THRESHOLD:
+        if seq_len < DUAL_STREAM_TOKEN_THRESHOLD and self.alt_stream is not None:
             current_stream = torch.cuda.current_stream()
             self.alt_stream.wait_stream(current_stream)
             projected_states_qkvz, _ = self.in_proj_qkvz(hidden_states)