From a2956a0f8e8f44cc79c14a9d3b45167631b7c249 Mon Sep 17 00:00:00 2001 From: Andreas Karatzas Date: Mon, 16 Mar 2026 03:08:51 -0500 Subject: [PATCH] [ROCm][CI] Retrying in case of batch variance effects and reducing flakiness (#36442) Signed-off-by: Andreas Karatzas --- tests/v1/e2e/general/test_async_scheduling.py | 8 ++++++-- 1 file changed, 6 insertions(+), 2 deletions(-) diff --git a/tests/v1/e2e/general/test_async_scheduling.py b/tests/v1/e2e/general/test_async_scheduling.py index acb08997c7..8e1eddb0f6 100644 --- a/tests/v1/e2e/general/test_async_scheduling.py +++ b/tests/v1/e2e/general/test_async_scheduling.py @@ -7,7 +7,10 @@ from typing import Any import pytest import torch._dynamo.config as dynamo_config -from tests.utils import large_gpu_mark, single_gpu_only +from tests.utils import ( + large_gpu_mark, + single_gpu_only, +) from vllm import SamplingParams from vllm.logprobs import Logprob from vllm.platforms import current_platform @@ -150,6 +153,7 @@ def test_with_eagle3_spec_decoding(sample_json_schema, monkeypatch: pytest.Monke run_tests(monkeypatch, MTP_MODEL, test_configs, test_sampling_params) +@pytest.mark.flaky(reruns=2, only_on=current_platform.is_rocm()) def test_with_ngram_gpu_spec_decoding(monkeypatch: pytest.MonkeyPatch): """Test ngram_gpu speculative decoding with different configurations. @@ -202,7 +206,6 @@ def run_tests( with monkeypatch.context() as m: # lock matmul precision to full FP32 (IEEE) m.setenv("VLLM_FLOAT32_MATMUL_PRECISION", "highest") - # m.setenv("VLLM_BATCH_INVARIANT", "1") outputs: list[tuple[str, list, list]] = [] for n, ( test_preemption, @@ -351,6 +354,7 @@ def run_test( speculative_config=spec_config, disable_log_stats=False, attention_config=attention_config, + enable_prefix_caching=False if current_platform.is_rocm() else None, **cache_arg, ) as vllm_model: results = []