From 0570f7a06ebcac9d880f18d106e8ceb7c1e94567 Mon Sep 17 00:00:00 2001 From: Yueqian Lin <70319226+linyueqian@users.noreply.github.com> Date: Fri, 18 Sep 2026 11:25:01 -0700 Subject: [PATCH] [CI][Qwen3-Omni] Keep the deploy-YAML literal in the structured-config sampling check (#7803) Signed-off-by: Yueqian Lin --- tests/e2e/offline_inference/test_qwen3_omni.py | 10 ++++++---- 1 file changed, 6 insertions(+), 4 deletions(-) diff --git a/tests/e2e/offline_inference/test_qwen3_omni.py b/tests/e2e/offline_inference/test_qwen3_omni.py index de64c9045..4b744df15 100644 --- a/tests/e2e/offline_inference/test_qwen3_omni.py +++ b/tests/e2e/offline_inference/test_qwen3_omni.py @@ -112,16 +112,18 @@ def test_structured_multistage_config_reaches_runtime(omni_runner, offline_clien assert code2wav_vllm.scheduler_config.async_scheduling is False # Sampling defaults are consumed from StageClient metadata, rather than - # from either the resolver output or VllmConfig. + # from either the resolver output or VllmConfig. These literals mirror + # vllm_omni/deploy/qwen3_omni_moe.yaml exactly. expected_sampling = ( {"temperature": 0.0, "max_tokens": 2048}, {"temperature": 0.9, "top_k": 50, "max_tokens": 4096, "repetition_penalty": 1.05}, { "temperature": 0.0, "top_p": 1.0, - # SamplingParams normalizes the disabled greedy-sampling sentinel - # from the deploy YAML (top_k=-1) to its runtime value (top_k=0). - "top_k": 0, + # Deploy YAML literal. The resolver check below compares it verbatim; + # the runtime check normalizes it through SamplingParams, which turns + # the greedy-sampling sentinel -1 into 0. + "top_k": -1, "max_tokens": 65536, "repetition_penalty": 1.1, },