[CI] Add new runner and enable QwQ multinpu test (#417)

### What this PR does / why we need it?

- Add a new runner to the continuous integration system and keep the
original CI runner until the new runner runs stably
- Add distributed test cases

### Does this PR introduce _any_ user-facing change?
No

### How was this patch tested?
CI passed

---------

Signed-off-by: wangli <wangli858794774@gmail.com>

tests/conftest.py

@@ -17,14 +17,17 @@
 # limitations under the License.
 #
 
+import gc
 from typing import List, Optional, Tuple, TypeVar, Union
 
 import numpy as np
 import pytest
+import torch
 from PIL import Image
 from vllm import LLM, SamplingParams
 from vllm.config import TaskOption
-from vllm.distributed import cleanup_dist_env_and_memory
+from vllm.distributed.parallel_state import (destroy_distributed_environment,
+                                             destroy_model_parallel)
 from vllm.inputs import ExplicitEncoderDecoderPrompt, TextPrompt, TokensPrompt
 from vllm.logger import init_logger
 from vllm.outputs import RequestOutput
@@ -37,6 +40,7 @@ from tests.model_utils import (TokensTextLogprobs,
 logger = init_logger(__name__)
 
 _M = TypeVar("_M")
+
 _PromptMultiModalInput = Union[List[_M], List[List[_M]]]
 
 PromptImageInput = _PromptMultiModalInput[Image.Image]
@@ -44,6 +48,13 @@ PromptAudioInput = _PromptMultiModalInput[Tuple[np.ndarray, int]]
 PromptVideoInput = _PromptMultiModalInput[np.ndarray]
 
 
+def cleanup_dist_env_and_memory():
+    destroy_model_parallel()
+    destroy_distributed_environment()
+    gc.collect()
+    torch.npu.empty_cache()
+
+
 class VllmRunner:
 
     def __init__(

tests/test_offline_inference.py

@@ -31,20 +31,13 @@ import vllm_ascend  # noqa: F401
 MODELS = [
     "Qwen/Qwen2.5-0.5B-Instruct",
 ]
-os.environ["VLLM_USE_MODELSCOPE"] = "True"
 os.environ["PYTORCH_NPU_ALLOC_CONF"] = "max_split_size_mb:256"
 
-TARGET_TEST_SUITE = os.environ.get("TARGET_TEST_SUITE", "L4")
-
 
 @pytest.mark.parametrize("model", MODELS)
 @pytest.mark.parametrize("dtype", ["half", "float16"])
 @pytest.mark.parametrize("max_tokens", [5])
-def test_models(
-    model: str,
-    dtype: str,
-    max_tokens: int,
-) -> None:
+def test_models(model: str, dtype: str, max_tokens: int) -> None:
     # 5042 tokens for gemma2
     # gemma2 has alternating sliding window size of 4096
     # we need a prompt with more than 4096 tokens to test the sliding window
@@ -60,6 +53,28 @@ def test_models(
         vllm_model.generate_greedy(example_prompts, max_tokens)
 
 
+@pytest.mark.multinpu
+@pytest.mark.parametrize("model, distributed_executor_backend", [
+    ("Qwen/QwQ-32B", "mp"),
+])
+def test_models_distributed(vllm_runner, model: str,
+                            distributed_executor_backend: str) -> None:
+    example_prompts = [
+        "vLLM is a high-throughput and memory-efficient inference and serving engine for LLMs.",
+        "Briefly describe the major milestones in the development of artificial intelligence from 1950 to 2020.",
+        "Compare and contrast artificial intelligence with human intelligence in terms of processing information.",
+    ]
+    dtype = "half"
+    max_tokens = 5
+    with vllm_runner(
+            model,
+            dtype=dtype,
+            tensor_parallel_size=4,
+            distributed_executor_backend=distributed_executor_backend,
+    ) as vllm_model:
+        vllm_model.generate_greedy(example_prompts, max_tokens)
+
+
 if __name__ == "__main__":
     import pytest
     pytest.main([__file__])