[BugFix] Fix metrics error for --num-scheduler-steps > 1 (#8234)

vllm/engine/llm_engine.py

@@ -1718,6 +1718,15 @@ class LLMEngine:
                     # TPOTs.
                     latency = seq_group.get_last_latency(now)
                     time_per_output_tokens_iter.append(latency)
+                    if seq_group.state.current_step == 0:
+                        # For async_output_proc, the do_log_stats()
+                        # is called following init_multi_step(), which
+                        # sets the current_step to zero.
+                        actual_num_batched_tokens +=\
+                            seq_group.state.num_steps - 1
+                    else:
+                        actual_num_batched_tokens +=\
+                            seq_group.state.current_step - 1
 
                 # Because of chunked prefill, we can have a single sequence
                 # group that does multiple prompt_runs. To prevent logging