[RL] add --skip-warmup

python/sglang/bench_offline_throughput.py

@@ -57,7 +57,7 @@ class BenchArgs:
     extra_request_body: Optional[str] = None
     apply_chat_template: bool = False
     profile: bool = False
-    skip_warmup: bool = False
+    skip_bench_warmup: bool = False
     do_not_exit: bool = False
     prompt_suffix: str = ""
 
@@ -172,7 +172,7 @@ def add_cli_args(parser: argparse.ArgumentParser):
             "SGLANG_TORCH_PROFILER_DIR to enable profiler.",
         )
         parser.add_argument(
-            "--skip-warmup",
+            "--skip-bench-warmup",
             action="store_true",
             help="Skip the warmup batches.",
         )
@@ -346,7 +346,7 @@ def throughput_test(
     )
 
     # Warm up
-    if not bench_args.skip_warmup:
+    if not bench_args.skip_bench_warmup:
         logging.info("\nWarmup...")
         throughput_test_once(
             backend_name=bench_args.backend,

python/sglang/bench_one_batch_server.py

@@ -42,7 +42,7 @@ class BenchArgs:
     input_len_step_percentage: float = 0.0
     result_filename: str = "result.jsonl"
     base_url: str = ""
-    skip_warmup: bool = False
+    skip_bench_warmup: bool = False
     show_report: bool = False
     profile: bool = False
     profile_by_stage: bool = False
@@ -75,7 +75,7 @@ def add_cli_args(parser: argparse.ArgumentParser):
             "--result-filename", type=str, default=BenchArgs.result_filename
         )
         parser.add_argument("--base-url", type=str, default=BenchArgs.base_url)
-        parser.add_argument("--skip-warmup", action="store_true")
+        parser.add_argument("--skip-bench-warmup", action="store_true")
         parser.add_argument("--show-report", action="store_true")
         parser.add_argument("--profile", action="store_true")
         parser.add_argument("--profile-by-stage", action="store_true")
@@ -258,7 +258,7 @@ def run_benchmark(server_args: ServerArgs, bench_args: BenchArgs):
     tokenizer = get_tokenizer(tokenizer_path)
 
     # warmup
-    if not bench_args.skip_warmup:
+    if not bench_args.skip_bench_warmup:
         print("=" * 8 + " Warmup Begin " + "=" * 8)
         run_one_case(
             base_url,

python/sglang/srt/entrypoints/http_server.py

@@ -932,11 +932,9 @@ def launch_server(
         warmup_thread.join()
 
 
-def _wait_and_warmup(
+def _execute_server_warmup(
     server_args: ServerArgs,
     pipe_finish_writer: Optional[multiprocessing.connection.Connection],
-    image_token_text: str,
-    launch_callback: Optional[Callable[[], None]] = None,
 ):
     headers = {}
     url = server_args.url()
@@ -961,7 +959,7 @@ def _wait_and_warmup(
             pipe_finish_writer.send(last_traceback)
         logger.error(f"Initialization failed. warmup error: {last_traceback}")
         kill_process_tree(os.getpid())
-        return
+        return success
 
     model_info = res.json()
 
@@ -1035,10 +1033,25 @@ def _wait_and_warmup(
             pipe_finish_writer.send(last_traceback)
         logger.error(f"Initialization failed. warmup error: {last_traceback}")
         kill_process_tree(os.getpid())
-        return
+        return False
 
     # Debug print
     # logger.info(f"warmup request returns: {res.json()=}")
+    return success
+
+
+def _wait_and_warmup(
+    server_args: ServerArgs,
+    pipe_finish_writer: Optional[multiprocessing.connection.Connection],
+    image_token_text: str,
+    launch_callback: Optional[Callable[[], None]] = None,
+):
+    if not server_args.skip_warmup:
+        if not _execute_server_warmup(
+            server_args,
+            pipe_finish_writer,
+        ):
+            return
 
     logger.info("The server is fired up and ready to roll!")
 

python/sglang/srt/server_args.py

@@ -46,6 +46,7 @@ class ServerArgs:
     tokenizer_path: Optional[str] = None
     tokenizer_mode: str = "auto"
     skip_tokenizer_init: bool = False
+    skip_warmup: bool = False
     load_format: str = "auto"
     model_loader_extra_config: str = "{}"
     trust_remote_code: bool = False
@@ -597,6 +598,11 @@ def add_cli_args(parser: argparse.ArgumentParser):
             action="store_true",
             help="If set, skip init tokenizer and pass input_ids in generate request.",
         )
+        parser.add_argument(
+            "--skip-warmup",
+            action="store_true",
+            help="If set, skip warmup.",
+        )
         parser.add_argument(
             "--load-format",
             type=str,