[bugfix][refactor] fix recompute_scheduler break with vllm 0.12.0 & support async scheduling & refactor recompute_scheduler.py

[linfeng-yuan]

What this PR does / why we need it?

Currently, the initialization and fundamental functions of RecomputeScheduler are broken with vLLM v0.12.0. This PR fixes the conflicts of RecomputeScheduler and refactor its implementations by inheriting original Scheduler of vLLM. Meanwhile, this PR also supports async cheduling with recompute scheduler by implementing AsyncRecomputeScheduler which is simply inherited AsncyScheduler of vLLM and RecomputeScheduler of vLLM-Ascend with python MRO.

Does this PR introduce any user-facing change?

No. The switch naming is the same as v0.11.0 : recompute_scheduler_enable

How was this patch tested?

E2E serving with 2P1D dsv3.1 passed. The performance was the same as original vllm scheduler with async_scheduling and preempted requests in D Nodes are successfully transfered to Proxy and further to P Node. This significantly improves the performance and robustness of PD disaggregation deployments.

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

[gemini-code-assist]

Code Review

The pull request introduces significant refactoring to the scheduler components, aiming to fix recompute issues, support async scheduling, and align with vLLM 0.12.0. A new AsyncRecomputeScheduler is added, and the core RecomputeScheduler is refactored to inherit from the base vLLM Scheduler class. However, there are critical issues related to class definitions and naming consistency that need to be addressed to ensure the code is functional.

[gemini-code-assist]

The @dataclass decorator and the class definition for RecomputeSchedulerConfig appear to be missing. This will lead to a syntax error or NameError when the initialize_from_config method is called, as it's currently defined as a standalone method without a class context. The class definition, likely inheriting from SchedulerConfig, needs to be restored.

@dataclass
class RecomputeSchedulerConfig(SchedulerConfig):
    pass

    @classmethod

[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.

[linfeng-yuan]

/gemini review

[gemini-code-assist]

Code Review

This pull request introduces an AsyncRecomputeScheduler by inheriting from AsyncScheduler and RecomputeScheduler, and refactors the RecomputeScheduler to align with upstream vLLM changes. The RecomputeScheduler now inherits from vllm.v1.core.sched.scheduler.Scheduler, significantly reducing its __init__ method and moving much of its core logic to the base class. Key changes include updating RecomputeSchedulerConfig to accept VllmConfig and dynamically select between synchronous and asynchronous recompute schedulers, optimizing scheduling for async operations by avoiding extra steps, refining preemption logic to restore token budgets and related data, and adjusting speculative decoding token calculations. Additionally, the handling of encoder inputs and KV connector metadata is updated, and KV cache event collection and publishing are moved to the update_from_output method. The RecomputeSchedulerOutput dataclass now inherits from SchedulerOutput, simplifying its structure.