support basic long_seq feature st

[LookAround0301]

What this PR does / why we need it?

support basic long_seq feature st

Does this PR introduce any user-facing change?

How was this patch tested?

• vLLM version: v0.12.0
• vLLM main: vllm-project/vllm@ad32e3e

[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 introduces end-to-end smoke tests for a long sequence feature using PCP/DCP on multi-card setups. The tests cover eager mode, full graph compilation, and piecewise execution. My review focuses on improving code quality and maintainability. I've identified a misleading docstring and significant code duplication across the test functions. I've provided suggestions to correct the docstring and refactor the tests to eliminate redundancy, making them easier to maintain in the future.

[gemini-code-assist]

The file's docstring contains an incorrect command to run the tests. It refers to test_qwen3_moe.py instead of the current file, test_long_sequence_basic.py. This is likely a copy-paste error and can be confusing for other developers.

Suggested change

	"""Compare the short outputs of HF and vLLM when using greedy sampling.
	Run `pytest tests/e2e/multicard/test_qwen3_moe.py`.
	"""
	"""Compare the short outputs of HF and vLLM when using greedy sampling.
	Run `pytest tests/e2e/multicard/long_sequence/test_long_sequence_basic.py`.
	"""

[gemini-code-assist]

The three test functions (test_pcp_dcp_basic, test_pcp_dcp_full_graph, test_pcp_dcp_piece_wise) are highly repetitive, making the code difficult to maintain. Key components like prompts, sampling_params, and the VllmRunner configuration are duplicated in each function.

A better approach is to refactor this using pytest.mark.parametrize. This will create a single, parameterized test, eliminating redundancy and making the test configurations explicit and easier to manage. The suggested code implements this refactoring. Please note that import pytest is included in the suggestion and should be moved to the top of the file.

import pytest


PROMPTS = [
    "The capital of France is",
    "Hello, my name is Tom, I am",
    "The president of United States is",
    "AI future is"
]
SAMPLING_PARAMS = SamplingParams(max_tokens=32, temperature=0.0)
DEEPSEEK_MODEL = "deepseek-ai/DeepSeek-V2-Lite-Chat"
QWEN_MODEL = "vllm-ascend/Qwen3-30B-A3B-W8A8"

BASE_DEEPSEEK_ARGS = {
    "max_model_len": 1024,
    "tensor_parallel_size": 2,
    "prefill_context_parallel_size": 2,
    "decode_context_parallel_size": 2,
    "max_num_batched_tokens": 1024,
    "enable_expert_parallel": True,
    "block_size": 128
}
BASE_QWEN_ARGS = {
    "max_model_len": 1024,
    "tensor_parallel_size": 8,
    "prefill_context_parallel_size": 2,
    "decode_context_parallel_size": 2,
    "enable_expert_parallel": True,
    "block_size": 128,
    "quantization": "ascend",
}

def _run_models(deepseek_vllm_runner_args, qwen_vllm_runner_args):
    with VllmRunner(DEEPSEEK_MODEL, **deepseek_vllm_runner_args) as runner:
        runner.model.generate(PROMPTS, SAMPLING_PARAMS)
    
    with VllmRunner(QWEN_MODEL, **qwen_vllm_runner_args) as runner:
        runner.model.generate(PROMPTS, SAMPLING_PARAMS)

@pytest.mark.parametrize("extra_args", [
    {"enforce_eager": True},
    {
        "enforce_eager": False,
        "compilation_config": {
            "cudagraph_mode": "FULL_DECODE_ONLY",
            "cudagraph_capture_sizes": [4, 8, 24, 48, 60]
        },
    },
    {"enforce_eager": False},
], ids=["basic", "full_graph", "piece_wise"])
def test_pcp_dcp(extra_args):
    deepseek_args = {**BASE_DEEPSEEK_ARGS, **extra_args}
    qwen_args = {**BASE_QWEN_ARGS, **extra_args}
    _run_models(deepseek_args, qwen_args)