[Misc]Fix AttributeError in RecomputeScheduler due to missing routed experts extraction logic

[SOMEONEUNSEEN]

What this PR does / why we need it?

part of #10215

Summary
Fixes a crash in RecomputeScheduler / AsyncRecomputeScheduler when a request finishes, caused by calling a non-existent _get_routed_experts() method:

AttributeError: 'AsyncRecomputeScheduler' object has no attribute '_get_routed_experts'
This is triggered when all of the following are true:

Async scheduling is enabled (uses AsyncRecomputeScheduler)
A request stops in the current step (stopped=True)
Reproducible on both vLLM v0.21.0 and main commit 9090368.

Does this PR introduce any user-facing change?

Yes. After this fix, in async scheduling + recompute scheduler scenarios:

The service no longer crashes due to the missing _get_routed_experts method
When enable_return_routed_experts=True, routed experts are correctly returned via EngineCoreOutput
No API or interface changes.

How was this patch tested?

• vLLM version: v0.21.0
• vLLM main: vllm-project/vllm@9090368

[gemini-code-assist]

Summary of Changes

Hello, I'm Gemini Code Assist¹! I'm currently reviewing this pull request and will post my feedback shortly. In the meantime, here's a summary to help you and other reviewers quickly get up to speed!

This pull request addresses an AttributeError in the RecomputeScheduler by introducing a more comprehensive mechanism for extracting and managing routed experts. It ensures that routing data is correctly persisted and retrieved, accounting for different vLLM version requirements and specific request states such as prompt processing versus token generation.

Highlights

• Routed Experts Logic: Implemented robust extraction logic for routed experts in the RecomputeScheduler to resolve potential AttributeError issues.
• Version Compatibility: Added conditional logic to handle routed experts differently based on the vLLM version, ensuring compatibility with version 0.21.0 and newer.
• Data Persistence: Added functionality to persist per-step routed experts into the scheduler-side slot buffer to support correct retrieval during request processing.

New Features

🧠 You can now enable Memory (public preview) to help Gemini Code Assist learn from your team's feedback. This makes future code reviews more consistent and personalized to your project's style. Click here to enable Memory in your admin console.

Using Gemini Code Assist

The full guide for Gemini Code Assist can be found on our documentation page, here are some quick tips.

Invoking Gemini

You can request assistance from Gemini at any point by creating a comment using either /gemini <command> or @gemini-code-assist <command>. Below is a summary of the supported commands on the current page.

Feature	Command	Description
Code Review	/gemini review	Performs a code review for the current pull request in its current state.
Pull Request Summary	/gemini summary	Provides a summary of the current pull request in its current state.
Comment	@gemini-code-assist	Responds in comments when explicitly tagged, both in pull request comments and review comments.
Help	/gemini help	Displays a list of available commands.

Customization

To customize the Gemini Code Assist for GitHub experience, repository maintainers can create a configuration file and/or provide a custom code review style guide (such as PEP-8 for Python) by creating and adding files to a .gemini/ folder in the base of the repository. Detailed instructions can be found here.

Limitations & Feedback

Gemini Code Assist may make mistakes. Please leave feedback on any instances where its feedback is incorrect or counterproductive. You can react with 👍 and 👎 on @gemini-code-assist comments. If you're interested in giving your feedback about your experience with Gemini Code Assist for GitHub and other Google products, sign up here.

Footnotes

1. Review the Privacy Notices, Generative AI Prohibited Use Policy, Terms of Service, and learn how to configure Gemini Code Assist in GitHub here. Gemini can make mistakes, so double check it and use code with caution. ↩

[gemini-code-assist]

Code Review

Suggested PR Title:

[Ops][Feature] Support returning routed experts in RecomputeScheduler

Suggested PR Summary:

### What this PR does / why we need it?
This pull request adds support for retrieving and persisting routed experts (MoE routing data) in the `RecomputeScheduler` for Ascend. It introduces version-specific handling for vLLM (supporting both version "0.21.0" and others) to store batch routing data and map it back to individual requests.

Feedback:
- A critical issue was identified where `self._re_block_ids` is never populated in `RecomputeScheduler`, causing `self._re_block_ids.pop(req_id, [])` to always return an empty list. This should be replaced with `self.kv_cache_manager.get_block_ids(req_id)`.
- A potential `KeyError` could occur when accessing `num_scheduled_tokens[rid]`. It is safer to use `num_scheduled_tokens.get(rid, 0)`.

### Does this PR introduce _any_ user-facing change?
No, this is an internal update to the scheduler.

### How was this patch tested?
No tests were provided in this PR.

[gemini-code-assist]

The self._re_block_ids dictionary is never populated in RecomputeScheduler.schedule (which completely overrides the parent Scheduler.schedule method). As a result, self._re_block_ids.pop(req_id, []) will always return an empty list [], causing self.routed_experts_mgr.get to be called with no block IDs and fail to retrieve the routed experts. To fix this, retrieve the block IDs directly from the kv_cache_manager using self.kv_cache_manager.get_block_ids(req_id).

Suggested change

	block_ids = self._re_block_ids.pop(req_id, [])
	block_ids = self.kv_cache_manager.get_block_ids(req_id)

[gemini-code-assist]

If a request ID in model_runner_output.req_ids is missing from num_scheduled_tokens due to any discrepancy, accessing num_scheduled_tokens[rid] directly will raise a KeyError and crash the engine. Using .get(rid, 0) is safer and prevents potential crashes.

Suggested change

	offset += num_scheduled_tokens[rid]
	offset += num_scheduled_tokens.get(rid, 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.