Enable torch.compile for triton backend (#1422)

python/sglang/bench_latency.py

@@ -479,7 +479,8 @@ def main(server_args, bench_args):
 
 
 if __name__ == "__main__":
-    # TODO(kevin85421): Make the parser setup unit testable.
+    multiprocessing.set_start_method("spawn", force=True)
+
     parser = argparse.ArgumentParser()
     ServerArgs.add_cli_args(parser)
     BenchArgs.add_cli_args(parser)