[fix]修复llm_engine.py 越界报错

vllm/engine/llm_engine.py

@@ -989,15 +989,16 @@ class LLMEngine:
                 output = [outputs_by_sequence_group[0][i]]
 
             # tree style speculative decoding may generate empty output in first step
-            if outputs and isinstance(output[0], CompletionSequenceGroupOutput):
-                samples = [o.samples[0] for o in output]
-                valid_samples = [
-                    sample for sample in samples
-                    if sample.output_token != VLLM_INVALID_TOKEN_ID
-                ]
-                if len(valid_samples) == 0:
-                    empty_seq_indices.append(i)
-                    continue
+            if self.speculative_config and self.speculative_config.tree_style_spec_decoding:
+                if outputs and isinstance(output[0], CompletionSequenceGroupOutput):
+                    samples = [o.samples[0] for o in output]
+                    valid_samples = [
+                        sample for sample in samples
+                        if sample.output_token != VLLM_INVALID_TOKEN_ID
+                    ]
+                    if len(valid_samples) == 0:
+                        empty_seq_indices.append(i)
+                        continue
 
             if not is_async:
                 seq_group.update_num_computed_tokens(