diff --git a/tests/kernels/moe/conftest.py b/tests/kernels/moe/conftest.py index a217fe684e..0a210edc4a 100644 --- a/tests/kernels/moe/conftest.py +++ b/tests/kernels/moe/conftest.py @@ -2,6 +2,32 @@ # SPDX-FileCopyrightText: Copyright contributors to the vLLM project import pytest +from vllm.distributed import cleanup_dist_env_and_memory + +NEEDS_CLEAN_ENTRY = frozenset( + { + # Fail unless they start from a clean distributed and memory state, + # which they used to get only as a side effect of the preceding test + # cleaning up after itself. + "test_moe_layer.py", + "test_ocp_mx_moe.py", + "test_zero_expert_moe.py", + # Execute no test on ROCm, where this was measured, so nothing is known + # about how they behave without the per-test cleanup. Cleaned on entry + # as a precaution; drop once they have been seen green without it. + "test_batched_deepgemm.py", + "test_deepep_deepgemm_moe.py", + "test_deepep_moe.py", + "test_deepep_v2_moe.py", + "test_deepgemm.py", + "test_grouped_topk.py", + "test_mxfp4_moe.py", + "test_shared_fused_moe_routed_transform.py", + "test_silu_mul_per_token_group_quant_fp8_colmajor.py", + "test_situ_mul_fp8_quant.py", + } +) + def pytest_addoption(parser): parser.addoption( @@ -12,3 +38,23 @@ def pytest_addoption(parser): @pytest.fixture def subtests(request): return request.config.getoption("--subtests") + + +@pytest.fixture() +def should_do_global_cleanup_after_test() -> bool: + """Drop the per-test global cleanup for this directory. + + A full ``gc.collect()`` plus an emptied caching allocator after each of the + ~1450 small kernel tests per shard costs more than the tests themselves. + The tests that depend on isolation get a clean state on entry instead, in + the hook below. + """ + return False + + +def pytest_runtest_setup(item): + # Not tryfirst: the skipping plugin evaluates skip marks in a tryfirst + # hook, so tests about to be skipped never reach this, while fixture setup + # still runs afterwards. + if item.path.name in NEEDS_CLEAN_ENTRY: + cleanup_dist_env_and_memory() diff --git a/tests/kernels/moe/test_fused_topk.py b/tests/kernels/moe/test_fused_topk.py index 825cd20263..9c196a884d 100644 --- a/tests/kernels/moe/test_fused_topk.py +++ b/tests/kernels/moe/test_fused_topk.py @@ -45,7 +45,8 @@ def torch_topk( @pytest.mark.skipif( - not current_platform.is_cuda(), reason="This test is skipped on non-CUDA platform." + not current_platform.is_cuda_alike(), + reason="This test is skipped on non-CUDA platform.", ) @pytest.mark.parametrize("num_tokens", [1, 33, 56]) @pytest.mark.parametrize("hidden_size", [1024, 2048]) @@ -89,7 +90,8 @@ def test_fused_topk( @pytest.mark.skipif( - not current_platform.is_cuda(), reason="This test is skipped on non-CUDA platform." + not current_platform.is_cuda_alike(), + reason="This test is skipped on non-CUDA platform.", ) @pytest.mark.parametrize("num_tokens", [1, 33, 56]) @pytest.mark.parametrize("hidden_size", [1024, 2048]) @@ -138,7 +140,8 @@ def test_fused_topk_bias( @pytest.mark.skipif( - not current_platform.is_cuda(), reason="This test is skipped on non-CUDA platform." + not current_platform.is_cuda_alike(), + reason="This test is skipped on non-CUDA platform.", ) @pytest.mark.parametrize("num_experts", [6, 8, 16]) @pytest.mark.parametrize("topk", [3, 4]) @@ -205,7 +208,8 @@ def test_fused_topk_nan_inf_clamp( @pytest.mark.skipif( - not current_platform.is_cuda(), reason="This test is skipped on non-CUDA platform." + not current_platform.is_cuda_alike(), + reason="This test is skipped on non-CUDA platform.", ) @pytest.mark.parametrize("num_experts", [6, 8, 16]) @pytest.mark.parametrize("topk", [3, 4])