[sglang-miles] Cherry-pick #24462: fix /pause_generation for --tokenizer-worker-num > 1

[ByronHsu]

Motivation

Cherry-pick of #24462 onto sglang-miles.

Without this fix, /pause_generation and /continue_generation only mutate the is_pause flag on the single TokenizerWorker whose uvicorn process happens to receive the HTTP request, leaving the other tokenizer workers free to keep dispatching new generation requests during weight transfer. This silently breaks RL weight updates whenever --tokenizer-worker-num > 1.

Modifications

Applies 7deed98e1 cleanly on top of sglang-miles:

• TokenizerWorkerRegistration and PauseContinueBroadcast IPC messages in io_struct.py.
• Worker-registration step in MultiTokenizerRouter that tracks every tokenizer worker's IPC name on startup.
• Pause/continue handling in the router that broadcasts PauseContinueBroadcast to all registered workers (and still forwards the original request to the scheduler for non-abort modes).
• New TokenizerWorker.pause_generation / continue_generation methods that send the request to the router and await an asyncio.Future resolved by the broadcast handler — so the HTTP response only returns after the pause state has propagated.

Checklist

• Format your code according to the Code Formatting with Pre-Commit.
• For reviewers: if you intend to acknowledge my contribution, please do so by including Co-authored-by: bingyuhsu <byronhsu1230@gmail.com> in the commit message after the PR is merged.

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!