diff --git a/benchmarks/benchmark/tools/profile-generator/container/benchmark_serving.py b/benchmarks/benchmark/tools/profile-generator/container/benchmark_serving.py index 6cefd2e56..8ae1109ee 100644 --- a/benchmarks/benchmark/tools/profile-generator/container/benchmark_serving.py +++ b/benchmarks/benchmark/tools/profile-generator/container/benchmark_serving.py @@ -181,7 +181,7 @@ async def send_request( elif backend == "jetstream": pload = { "prompt": prompt, - "max_tokens": 1, + "max_tokens": output_len, } else: raise ValueError(f"Unknown backend: {backend}")