seq.data._effective_length+=1

benchmarks/benchmark_throughput.py

@@ -179,7 +179,7 @@ def run_vllm(
     sampling_params: List[SamplingParams] = []
     for request in requests:
         prompts.append(
-            TextPrompt(prompt="helloworld",
+            TextPrompt(prompt="helloword",
                        multi_modal_data=request.multi_modal_data))
         sampling_params.append(
             SamplingParams(

vllm/engine/llm_engine.py

@@ -1514,9 +1514,9 @@ class LLMEngine:
                 for token_id, seq_id in zip(sample_out_list, sample_out_ids):
                     if seq.seq_id == seq_id:
                         sample.output_token = token_id[0]
+                        seq.data._effective_length+=1
                         seq.fix_last_token_id(sample.output_token)
-                        self.fix_process_model_output(ctx_output_queue,ctx_request_outputs,
-            ctx_multi_step_stream_outputs)
+                        self.fix_process_model_output(ctx_output_queue,ctx_request_outputs,ctx_multi_step_stream_outputs)
                         break
 
     def _advance_to_next_step(
@@ -1613,8 +1613,8 @@ class LLMEngine:
                 last_outputs_ids = last_outputs_ids, 
                 last_outputs_sample = last_outputs_tensor)
             if allow_async_output_proc:
-                execute_model_req.async_callback = self.async_callbacks[
-                    virtual_engine]
+                    execute_model_req.async_callback = self.async_callbacks[
+                        virtual_engine]
 
             #profile.ProfRangeAutoPush('model_executor')
             outputs = self.model_executor.execute_model(