From c7a78b382a8d521fb19403a8dd9c65c32c13ce08 Mon Sep 17 00:00:00 2001 From: Sy03 <24841857+Sy0307@users.noreply.github.com> Date: Wed, 19 Aug 2026 16:25:54 +0800 Subject: [PATCH] [CI/Build] Match perf warmups to benchmark concurrency Use enough warmup requests to cover a full concurrent batch while preserving the existing two-request minimum. Signed-off-by: Sy03 <24841857+Sy0307@users.noreply.github.com> --- tests/dfx/conftest.py | 3 ++- tests/dfx/perf/scripts/run_benchmark.py | 1 + 2 files changed, 3 insertions(+), 1 deletion(-) diff --git a/tests/dfx/conftest.py b/tests/dfx/conftest.py index 7e3ab77c7..2e22da531 100644 --- a/tests/dfx/conftest.py +++ b/tests/dfx/conftest.py @@ -653,6 +653,7 @@ def run_benchmark( random_input_len: Any | None = None, random_output_len: Any | None = None, resource_label: str | None = None, + num_warmups: int = 2, ) -> dict[str, Any]: """Run one ``vllm bench serve --omni`` iteration and return parsed metrics. @@ -677,7 +678,7 @@ def run_benchmark( + args + [ "--num-warmups", - "2", + str(num_warmups), "--save-result", "--result-dir", os.environ.get("BENCHMARK_DIR", "tests"), diff --git a/tests/dfx/perf/scripts/run_benchmark.py b/tests/dfx/perf/scripts/run_benchmark.py index 03dfac1a0..1eee88ac7 100644 --- a/tests/dfx/perf/scripts/run_benchmark.py +++ b/tests/dfx/perf/scripts/run_benchmark.py @@ -279,5 +279,6 @@ def test_performance_benchmark(omni_server, benchmark_params): random_input_len=params.get("random_input_len"), random_output_len=params.get("random_output_len"), resource_label=resource_label, + num_warmups=max(2, int(concurrency)), ) assert_result(result, params, num_prompt)