[Bugfix] Fix embedding model hangs with --enable-metrics (#2822)

python/sglang/srt/configs/model_config.py

@@ -128,7 +128,7 @@ class ModelConfig:
         self.num_hidden_layers = self.hf_text_config.num_hidden_layers
         self.vocab_size = self.hf_text_config.vocab_size
 
-        # Veirfy quantization
+        # Verify quantization
         self._verify_quantization()
 
         # Cache attributes

python/sglang/srt/managers/tokenizer_manager.py

@@ -688,7 +688,7 @@ class TokenizerManager:
                     if self.enable_metrics:
                         completion_tokens = (
                             recv_obj.completion_tokens[i]
-                            if recv_obj.completion_tokens
+                            if getattr(recv_obj, "completion_tokens", None)
                             else 0
                         )
 
@@ -716,7 +716,11 @@ class TokenizerManager:
                                 time.time() - state.created_time
                             )
                             # Compute time_per_output_token for the non-streaming case
-                            if not state.obj.stream and completion_tokens >= 1:
+                            if (
+                                hasattr(state.obj, "stream")
+                                and not state.obj.stream
+                                and completion_tokens >= 1
+                            ):
                                 self.metrics_collector.observe_time_per_output_token(
                                     (time.time() - state.created_time)
                                     / completion_tokens

python/sglang/srt/model_executor/model_runner.py

@@ -724,7 +724,7 @@ class ModelRunner:
         elif forward_batch.forward_mode.is_idle():
             return self.forward_idle(forward_batch)
         else:
-            raise ValueError(f"Invaid forward mode: {forward_batch.forward_mode}")
+            raise ValueError(f"Invalid forward mode: {forward_batch.forward_mode}")
 
     def sample(
         self, logits_output: LogitsProcessorOutput, forward_batch: ForwardBatch