[Aphoh]

Overview:

Adds the backend-neutral migration hooks for native /generate formats, split out of #16977 so SGLang and vLLM migration can each land on top independently. No format turns migration on in this PR: SGLang and vLLM both bind with migration: Err(..), so runtime behavior is unchanged.

Details:

• GenerateExchange::migrate(delivered_tokens, &mut PreprocessedRequest): the format rebases its request for a retry, or returns a terminal LLMEngineOutput without calling another worker.
• Bound.migration: Result<(), String>: the format's admission decision; Err keeps the request on one worker.
• RequestContinuation on PreprocessedRequest (frontend-only, not serialized) lets the generic Migration operator call back into the format.
• Migration::for_generate and build_generate_pipeline wire the operator into the format pipeline.
• Worker.engine_info + ENGINE_INFO_KEY: worker-advertised engine metadata (e.g. sampling defaults) available at bind time.
• SGLang and vLLM formats get "does not migrate yet" stubs.

Where should the reviewer start?

lib/llm/src/generate_format.rs, then lib/llm/src/migration.rs and lib/llm/src/http/service/format_generate.rs.

Stack: SGLang migration #16977 and vLLM migration #17766 are rebased on this PR.

Related Issues

• Relates to refactor(frontend): unify native generate handlers behind format plugins #16675
• Relates to DEP (light): Pluggable native generate formats with one frontend lifecycle #16674

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[Aphoh]

/ok to test e5f56de

[devin-ai-integration]

✅ Dynamo PR CI passed — run 41409733214 (attempt 1) on e5f56dee4a

Gate checks: ✅ backend-status-check · ✅ deploy-status-check · ✅ dynamo-status-check

_{Posted automatically by Devin for run 41409733214. Updated on every full-CI run of this PR.}