add single request test case for aclgraph

[lilinsiman]

What this PR does / why we need it?

This pr adds online single request DP2 test case for aclgraph

Does this PR introduce any user-facing change?

no

How was this patch tested?

ut

• vLLM version: v0.11.0rc3
• vLLM main: https://github.com/vllm-project/vllm/commit/v0.11.0

[github-actions]

👋 Hi! Thank you for contributing to the vLLM Ascend project. The following points will speed up your PR merge:‌‌

• A PR should do only one thing, smaller PRs enable faster reviews.
• Every PR should include unit tests and end-to-end tests ‌to ensure it works and is not broken by other future PRs.
• Write the commit message by fulfilling the PR description to help reviewer and future developers understand.

If CI fails, you can run linting and testing checks locally according Contributing and Testing.

[gemini-code-assist]

Code Review

This pull request adds a new end-to-end test for aclgraph with data parallelism. The test is well-structured, but there's a critical issue in how the test server is initialized that will prevent the test from running. I've also identified a potential source of test flakiness due to a hardcoded port and provided a comprehensive suggestion to fix both issues, making the test more robust.

[gemini-code-assist]

There are a couple of issues with the RemoteOpenAIServer setup:

1. Incorrect RemoteOpenAIServer call (Critical): The current call to RemoteOpenAIServer is incorrect. The __init__ signature in tests/e2e/conftest.py is (self, model: str, server_host: str, server_port: int, vllm_serve_args: list[str], ...). Your call provides server_args (a list) as the second positional argument for server_host (a string), and misses the required vllm_serve_args argument. This will cause a TypeError at runtime.

2. Hardcoded port (High): The test uses a hardcoded port 20002. This can cause flaky tests due to port conflicts when running tests in parallel or if the port is already occupied. It's better to use a dynamically allocated port.

Here is a suggested fix that addresses both issues by using vllm.utils.get_open_port to dynamically allocate a port and corrects the RemoteOpenAIServer call. The import is added inside the function for simplicity of the suggestion.

    from vllm.utils import get_open_port

    port = get_open_port()
    env_dict = {
        "TASK_QUEUE_ENABLE": "1",
        "HCCL_OP_EXPANSION_MODE": "AIV",
    }
    server_args = [
        "--no-enable-prefix-caching", "--tensor-parallel-size",
        "1", "--data-parallel-size", str(dp_size), "--port", str(port),
        "--trust-remote-code", "--gpu-memory-utilization", "0.9"
    ]
    request_keyword_args: dict[str, Any] = {
        **api_keyword_args,
    }
    with RemoteOpenAIServer(model=model,
                            server_host="localhost",
                            server_port=port,
                            vllm_serve_args=server_args,
                            env_dict=env_dict,
                            auto_port=False) as server:

[yiz-liu]

@lilinsiman Please check these suggested changes.

[lilinsiman]

@lilinsiman Please check these suggested changes.

After checking, it was found that the suggested modification was indeed better. Calling the get_open_port function can automatically return an available port, ensuring that the port is not occupied and thus guaranteeing its availability.

[wangxiyuan]

please enable this test in .github workflow

[lilinsiman]

change the case name

[lilinsiman]

delete port code