fix pending error in zero overhead

benchmarks/benchmark_serving.py

@@ -570,6 +570,8 @@ async def benchmark(
     else:
         print("Initial test run completed. Starting main benchmark run...")
 
+    time.sleep(0.1) # ZERO_OVERHEAD : sleep and wait the last step in warmup
+    
     if profile:
         print("Starting profiler...")
         profile_input = RequestFuncInput(model=model_id,

vllm/sequence.py

@@ -315,7 +315,8 @@ class SequenceData(msgspec.Struct,
         effect_offset = self._effective_length - len(self.output_token_ids)
         if effect_offset < 0:
             self._output_token_ids[effect_offset] = token_id
-            self._new_appended_tokens[effect_offset] = token_id
+            if len(self._new_appended_tokens) >= effect_offset * -1:
+                self._new_appended_tokens[effect_offset] = token_id
             self._cached_all_token_ids[effect_offset] = token_id
             self._effective_length += 1
 
@@ -848,17 +849,19 @@ class SequenceGroup:
     def set_last_token_time(self, now: float) -> None:
         """Sets the last token time for Request level timings."""
         # If still in prefill phase, assertion fails.
-        assert not self.is_prefill(), (
-            "seq_group.set_last_token_time() should not be called "
-            "if the seq_group is in prefill phase.")
+        if not self.seqs[0].zero_overhead:
+            assert not self.is_prefill(), (
+                "seq_group.set_last_token_time() should not be called "
+                "if the seq_group is in prefill phase.")
         self.last_token_latency = now - self.metrics.last_token_time
         self.metrics.last_token_time = now
 
     def get_last_token_latency(self) -> float:
         """Returns the latency of the last token."""
-        assert not self.is_prefill(), (
-            "seq_group.get_last_token_latency() should not be called "
-            "if the seq_group is in prefill phase.")
+        if not self.seqs[0].zero_overhead:
+            assert not self.is_prefill(), (
+                "seq_group.get_last_token_latency() should not be called "
+                "if the seq_group is in prefill phase.")
         return self.last_token_latency
 
     def maybe_set_first_token_time(self, time: float) -> None: