[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>
2026-08-19 16:25:54 +08:00
2 changed files with 3 additions and 1 deletions
@@ -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"),
@@ -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)