calibration benchmark for zero overhead schedule

vllm/benchmarks/benchmark_throughput.py

@@ -3,6 +3,7 @@
 import argparse
 import dataclasses
 import json
+import os
 import random
 import time
 from pathlib import Path
@@ -214,8 +215,9 @@ def run_vllm(
                         use_tqdm=False)
 
     use_beam_search = False
-    print("sleep 1")
-    time.sleep(1)
+    if os.environ.get('VLLM_ZERO_OVERHEAD') == '1':
+        print("sleep 1")
+        time.sleep(1) # ZERO_OVERHEAD : sleep and wait the last step in warmup
     if not use_beam_search:
         if args.profile:
             profile_dir = args.profile_result_dir