Add option to warm up

benchmarks/benchmark_serving.py

@@ -374,6 +374,24 @@ def mock_requests(total_mock_requests: int):
   return data
 
 
+def sample_warmup_requests(requests):
+  interesting_buckets = [
+        0,
+        16,
+        32,
+        64,
+        128,
+        256,
+        512,
+        1024,]
+
+  for start, end in zip(interesting_buckets[:-1], interesting_buckets[1:]):
+    for request in requests:
+      if start < request.prompt_len <= end:
+        yield request
+        break
+
+
 def main(args: argparse.Namespace):
   print(args)
   random.seed(args.seed)
@@ -390,6 +408,23 @@ def main(args: argparse.Namespace):
   else:
     input_requests = sample_requests(args.dataset, args.num_prompts, tokenizer, args.max_output_length)
 
+  if args.warmup_first:
+    print('Warm up start:' )
+    warmup_requests = list(sample_warmup_requests(input_requests)) * 2
+    benchmark_result, request_outputs = asyncio.run(
+        benchmark(
+            api_url=api_url,
+            tokenizer=tokenizer,
+            input_requests=warmup_requests,
+            request_rate=args.request_rate,
+            disable_tqdm=args.disable_tqdm,
+            session_cache=args.session_cache,
+            priority=args.priority,
+            threads=args.threads,
+        )
+    )
+    print('Warm up done')
+
   benchmark_result, request_outputs = asyncio.run(
       benchmark(
           api_url=api_url,
@@ -551,6 +586,14 @@ def main(args: argparse.Namespace):
           "File path to store request outputs"
       ),
   )
+  parser.add_argument(
+      "--warmup-first",
+      type=bool,
+      default=False,
+      help=(
+          "Whether to send warmup req first"
+      ),
+  )
 
   args = parser.parse_args()
   main(args)