From abb4ef7e2a0acb299735dbd3b50c7be02db276bf Mon Sep 17 00:00:00 2001 From: mgoin Date: Sat, 28 Jun 2025 16:33:10 +0000 Subject: [PATCH] Try fixing eagle e2e test OOM by reducing block allocation Signed-off-by: mgoin --- tests/spec_decode/e2e/test_eagle_correctness.py | 8 ++++++++ 1 file changed, 8 insertions(+) diff --git a/tests/spec_decode/e2e/test_eagle_correctness.py b/tests/spec_decode/e2e/test_eagle_correctness.py index fd838285aba7..7c369feec415 100644 --- a/tests/spec_decode/e2e/test_eagle_correctness.py +++ b/tests/spec_decode/e2e/test_eagle_correctness.py @@ -370,6 +370,10 @@ def test_llama2_eagle_e2e_greedy_correctness(vllm_runner, common_llm_kwargs, @pytest.mark.parametrize( "common_llm_kwargs", [{ + # 2 for small prompt, 256//16 for generated. + "num_gpu_blocks_override": 2 + 256 // 16, + "max_model_len": (2 + 256 // 16) * 16, + # Skip cuda graph recording for fast test. "enforce_eager": True, @@ -420,6 +424,10 @@ def test_llama3_eagle_e2e_greedy_correctness(vllm_runner, common_llm_kwargs, @pytest.mark.parametrize( "common_llm_kwargs", [{ + # 2 for small prompt, 256//16 for generated. + "num_gpu_blocks_override": 2 + 256 // 16, + "max_model_len": (2 + 256 // 16) * 16, + # Skip cuda graph recording for fast test. "enforce_eager": True,