From f2d8fbf4590daed7245800840a48298f4c961a9f Mon Sep 17 00:00:00 2001 From: stefankoncarevic Date: Mon, 5 Oct 2026 17:32:33 +0200 Subject: [PATCH] [CI/Build] Run the prefill token scoring test under batch invariance (#60056) Signed-off-by: Stefan Koncarevic --- tests/v1/sample/test_logprobs.py | 4 ++++ 1 file changed, 4 insertions(+) diff --git a/tests/v1/sample/test_logprobs.py b/tests/v1/sample/test_logprobs.py index 317ac10d41..4152662fbd 100644 --- a/tests/v1/sample/test_logprobs.py +++ b/tests/v1/sample/test_logprobs.py @@ -1407,6 +1407,10 @@ def test_prompt_logprobs_with_chunking_and_preemption(): def test_prompt_logprob_token_ids_with_chunking_and_preemption(monkeypatch): """Per-row scores stay row-aligned across chunked prefill and preemption.""" monkeypatch.setenv("VLLM_USE_V2_MODEL_RUNNER", "1") + # The ragged check below compares a second generate call against the first + # one. Default kernel selection varies between processes, which moves bf16 + # logprob tails by up to 0.2; under this the two runs are bit-identical. + monkeypatch.setenv("VLLM_BATCH_INVARIANT", "1") prompts = [ "The following numbers of the sequence "