[misc][ci] fix cpu test with plugins (vllm-project#7489)

.buildkite/run-cpu-test.sh

@@ -23,7 +23,7 @@ docker exec cpu-test-avx2 bash -c "python3 examples/offline_inference.py"
 # Run basic model test
 docker exec cpu-test bash -c "
   pip install pytest Pillow protobuf
-  pytest -v -s tests/models -m \"not vlm\" --ignore=tests/models/test_embedding.py --ignore=tests/models/test_registry.py --ignore=tests/models/test_jamba.py --ignore=tests/models/test_danube3_4b.py" # Mamba and Danube3-4B on CPU is not supported
+  pytest -v -s tests/models -m \"not vlm\" --ignore=tests/models/test_embedding.py --ignore=tests/models/test_oot_registration.py --ignore=tests/models/test_registry.py --ignore=tests/models/test_jamba.py --ignore=tests/models/test_danube3_4b.py" # Mamba and Danube3-4B on CPU is not supported
 
 # online inference
 docker exec cpu-test bash -c "

.buildkite/test-pipeline.yaml

@@ -292,6 +292,7 @@ steps:
   - pytest -v -s distributed/test_chunked_prefill_distributed.py
   - pytest -v -s distributed/test_multimodal_broadcast.py
   - pytest -v -s spec_decode/e2e/test_integration_dist_tp2.py
+  - pip install -e ./plugins/vllm_add_dummy_model
   - pytest -v -s distributed/test_distributed_oot.py
   - CUDA_VISIBLE_DEVICES=0,1 pytest -v -s test_sharded_state_loader.py
   - CUDA_VISIBLE_DEVICES=0,1 pytest -v -s distributed/test_utils.py

tests/models/test_oot_registration.py

@@ -4,19 +4,18 @@
 
 from vllm import LLM, SamplingParams
 
-# NOTE: the order of the tests is important
-# the first test does not load any plugins
-# the second test loads the plugin
-# they share the same process, so the plugin is loaded for the second test
+from ..utils import fork_new_process_for_each_test
 
 
+@fork_new_process_for_each_test
 def test_plugin(dummy_opt_path):
     os.environ["VLLM_PLUGINS"] = ""
     with pytest.raises(Exception) as excinfo:
         LLM(model=dummy_opt_path, load_format="dummy")
     assert "are not supported for now" in str(excinfo.value)
 
 
+@fork_new_process_for_each_test
 def test_oot_registration(dummy_opt_path):
     os.environ["VLLM_PLUGINS"] = "register_dummy_model"
     prompts = ["Hello, my name is", "The text does not matter"]

tests/utils.py

@@ -10,17 +10,17 @@
 from typing import Any, Callable, Dict, List, Optional
 
 import openai
-import ray
 import requests
 from transformers import AutoTokenizer
 from typing_extensions import ParamSpec
 
 from vllm.distributed import (ensure_model_parallel_initialized,
                               init_distributed_environment)
 from vllm.entrypoints.openai.cli_args import make_arg_parser
+from vllm.platforms import current_platform
 from vllm.utils import FlexibleArgumentParser, get_open_port, is_hip
 
-if is_hip():
+if current_platform.is_rocm():
     from amdsmi import (amdsmi_get_gpu_vram_usage,
                         amdsmi_get_processor_handles, amdsmi_init,
                         amdsmi_shut_down)
@@ -32,7 +32,7 @@ def _nvml():
             yield
         finally:
             amdsmi_shut_down()
-else:
+elif current_platform.is_cuda():
     from pynvml import (nvmlDeviceGetHandleByIndex, nvmlDeviceGetMemoryInfo,
                         nvmlInit, nvmlShutdown)
 
@@ -43,6 +43,11 @@ def _nvml():
             yield
         finally:
             nvmlShutdown()
+else:
+
+    @contextmanager
+    def _nvml():
+        yield
 
 
 VLLM_PATH = Path(__file__).parent.parent
@@ -293,6 +298,8 @@ def multi_process_parallel(
     pp_size: int,
     test_target: Any,
 ) -> None:
+    import ray
+
     # Using ray helps debugging the error when it failed
     # as compared to multiprocessing.
     # NOTE: We need to set working_dir for distributed tests,