Suppport qwen model and solve some problems (#75)

python/sglang/srt/managers/detokenizer_manager.py

@@ -55,6 +55,7 @@ class DetokenizerManager:
                         first_token = self.tokenizer.convert_ids_to_tokens(
                             int(output_tokens[i][0])
                         )
+                        first_token = first_token.decode("utf-8")
                         if first_token.startswith("▁"):
                             output_strs[i] = " " + output_strs[i]
 

python/sglang/srt/managers/router/model_runner.py

@@ -240,6 +240,7 @@ class ModelRunner:
         from sglang.srt.models.llama2 import LlamaForCausalLM
         from sglang.srt.models.llava import LlavaLlamaForCausalLM
         from sglang.srt.models.mixtral import MixtralForCausalLM
+        from sglang.srt.models.qwen import QWenLMHeadModel
 
         # Select model class
         architectures = getattr(self.model_config.hf_config, "architectures", [])
@@ -258,6 +259,9 @@ class ModelRunner:
             if arch == "MixtralForCausalLM":
                 model_class = MixtralForCausalLM
                 break
+            if arch == "QWenLMHeadModel":
+                model_class = QWenLMHeadModel
+                break
         if model_class is None:
             raise ValueError(f"Unsupported architectures: {architectures}")