[refactor] refactor model runner capture model

[weiguihua2]

What this PR does / why we need it?

Refactor the capture_model method in model_runner to directly reuse the method from vLLM.

Currently, most of the logic in the capture_model method is similar to that in the vllm code. Directly using the vllm method can reduce the maintenance cost of the vllm-ascend code. Modify as follows:
1、refactor capture_model function, directly inheriting community methods
2、refactor initialize_aclgraph_capture function, move to initialize_attn_backend

Does this PR introduce any user-facing change?

No

How was this patch tested?

• vLLM version: release/v0.13.0
• vLLM main: vllm-project/vllm@ad32e3e

[gemini-code-assist]

Code Review

This pull request refactors the model capturing logic to inherit from the community GPUModelRunner, which simplifies the code and improves maintainability. It also refactors how CUDA graph support is determined for attention backends, moving from a static attribute to a dynamic method. These changes align the Ascend-specific codebase more closely with the upstream vLLM project. My review has identified a critical bug in a new context manager that fails to restore a patched function, and a performance issue in initialization logic where a function is called redundantly in a loop.

I am having trouble creating individual review comments. Click here to see my feedback.

vllm_ascend/worker/model_runner_v1.py (3201-3207)

The context manager _replace_gpu_model_runner_function_wrapper does not correctly restore the original graph_capture function. The finally block re-assigns the patched function instead of restoring the original one. This will lead to a permanent patch, which can cause unexpected behavior in other parts of the codebase that might rely on the original graph_capture from the parent module.

def _replace_gpu_model_runner_function_wrapper(target_module_name):
    target_module = sys.modules[target_module_name]
    original_graph_capture = target_module.graph_capture
    try:
        target_module.graph_capture = graph_capture
        yield
    finally:
        target_module.graph_capture = original_graph_capture

vllm_ascend/worker/model_runner_v1.py (2709-2713)

The function get_attn_backends_for_group is called twice for each kv_cache_group_spec. It's called once in the first loop to populate attention_backend_maps and attention_backend_list, and then again in this second loop. This is inefficient. The results from the first loop should be reused here.

        for i, attn_backends_map in enumerate(attention_backend_maps):
            self.attn_groups.append(create_attn_groups(attn_backends_map, i))

[github-actions]

This pull request has conflicts, please resolve those before we can evaluate the pull request.

[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.

[github-actions]

This pull request has conflicts, please resolve those before we can evaluate the pull request.

[yiz-liu]

We should iterate through attention_backend_maps now, please fix this in another PR.

[weiguihua2]

OK

[github-actions]

This pull request has conflicts, please resolve those before we can evaluate the pull request.

[github-actions]

This pull request has conflicts, please resolve those before we can evaluate the pull request.