diff --git a/tests/integration/defs/conftest.py b/tests/integration/defs/conftest.py index 561d9bf3272..e5f4a4b0f64 100644 --- a/tests/integration/defs/conftest.py +++ b/tests/integration/defs/conftest.py @@ -2362,6 +2362,7 @@ def pytest_configure(config): tqdm.tqdm.monitor_interval = 0 if config.getoption("--run-ray"): os.environ["TLLM_DISABLE_MPI"] = "1" + os.environ["TLLM_RAY_FORCE_LOCAL_CLUSTER"] = "1" # Initialize PeriodicJUnitXML reporter if enabled periodic = config.getoption("--periodic-junit", default=False) diff --git a/tests/integration/defs/disaggregated/test_disaggregated_single_gpu.py b/tests/integration/defs/disaggregated/test_disaggregated_single_gpu.py index d6b63d3ab3c..434b1ef88a2 100644 --- a/tests/integration/defs/disaggregated/test_disaggregated_single_gpu.py +++ b/tests/integration/defs/disaggregated/test_disaggregated_single_gpu.py @@ -5,7 +5,7 @@ import cloudpickle import pytest -from defs.conftest import skip_no_hopper +from defs.conftest import skip_no_hopper, skip_ray from mpi4py import MPI from mpi4py.futures import MPIPoolExecutor @@ -244,6 +244,7 @@ def verify_disaggregated(model, generation_overlap, enable_cuda_graph, prompt, print("All workers terminated.") +@skip_ray @pytest.mark.parametrize("model", ["TinyLlama-1.1B-Chat-v1.0"]) @pytest.mark.parametrize("generation_overlap", [False, True]) @pytest.mark.parametrize("enable_cuda_graph", [False, True]) @@ -259,6 +260,7 @@ def test_disaggregated_simple_llama(model, generation_overlap, ]) +@skip_ray @skip_no_hopper @pytest.mark.parametrize("model", ["DeepSeek-V3-Lite-fp8/fp8"]) @pytest.mark.parametrize("generation_overlap", [False, True]) @@ -276,6 +278,7 @@ def test_disaggregated_simple_deepseek(model, generation_overlap, ]) +@skip_ray @skip_no_hopper @pytest.mark.parametrize("model", ["Qwen3-8B-FP8"]) @pytest.mark.parametrize("generation_overlap", [False, True]) diff --git a/tests/integration/defs/test_unittests.py b/tests/integration/defs/test_unittests.py index 190ea5111e2..d2cd80692cf 100644 --- a/tests/integration/defs/test_unittests.py +++ b/tests/integration/defs/test_unittests.py @@ -77,6 +77,7 @@ def test_unittests_v2(llm_root, llm_venv, case: str, output_dir, request): test_prefix = "unittest" waives_file = request.config.getoption("--waives-file") + run_ray = request.config.getoption("--run-ray") num_workers = 1 @@ -138,6 +139,9 @@ def test_unittests_v2(llm_root, llm_venv, case: str, output_dir, request): waives_file = os.path.abspath(waives_file) command += [f"--waives-file={waives_file}"] + if run_ray: + command += ["--run-ray"] + command += arg_list print(f"Running unit test:\"python {' '.join(command)}\"")