重构 Qwen2LM 输出处理逻辑，以改进token处理逻辑

Author: hexisyztem

cosyvoice/llm/llm.py

@@ -496,22 +496,24 @@ def inference_wrapper(self, lm_input, sampling, min_len, max_len, uuid):
                 self.vllm_output_queue[uuid] = queue.Queue()
             out_tokens = []
             while True:
+                top_ids = None
                 with self.lock:
                     if self.vllm_output_queue[uuid].empty() is True:
                         request_outputs: List[RequestOutput] = self.vllm.step()
                         for request_output in request_outputs:
-                            top_ids = list(request_output.outputs[0].token_ids)[-1]
-                            self.vllm_output_queue[request_output.request_id].put(top_ids)
-                if self.vllm_output_queue[uuid].empty() is False:
-                    top_ids = self.vllm_output_queue[uuid].get()
+                            self.vllm_output_queue[request_output.request_id].put(list(request_output.outputs[0].token_ids)[-1])
+                    if self.vllm_output_queue[uuid].empty() is False:
+                        top_ids = self.vllm_output_queue[uuid].get()
+                if top_ids is not None:
                     if top_ids in self.stop_token_ids:
                         break
                     # in stream mode, yield token one by one
                     yield top_ids
                     out_tokens.append(top_ids)
                     if len(out_tokens) == max_len:
                         break
-                time.sleep(0.001)
+                else:
+                    time.sleep(0.001)
             with self.lock:
                 self.vllm_output_queue.pop(uuid)
         else:

cosyvoice/vllm/cosyvoice2.py

@@ -23,7 +23,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 """Inference-only Qwen2 model compatible with HuggingFace weights."""
-from typing import Optional
+from typing import Optional, Union, Iterable
 from packaging.version import parse as vparse
 import vllm