Skip to content

Add portable vLLM token artifacts for Responses continuations #70

Description

@franciscojavierarceo

Parent tracker: #69
Design and measured evidence: #65

Summary

Add portable, validated token artifacts for Responses continuations after exact llm-d routing and reliable checkpoint persistence are established.

The artifact path is a transport/rendering optimization. It must preserve the KV-cache hit ratio achieved by exact routing and must not depend on a process-local handle that is valid only before llm-d selects a replica.

Scope

  • Accept a private portable artifact reference or certified prompt tokens in vLLM and feed the resolved complete token sequence into the existing input/APC path.
  • Replace unbounded process-local prompt dictionaries with a bounded, tenant-scoped, expiring local cache backed by an authenticated portable artifact service or inline reseed path.
  • Bind artifacts and append certificates to model, tokenizer, renderer/template/parser, instructions, tools, cache_salt, LoRA, multimodal identity, and expiration.
  • Define typed artifact miss, expiry, validation failure, and reseed behavior with safe fallback to ordinary full Responses rendering.
  • Accumulate all generated streaming token IDs and preserve tool-round boundaries before optionally advancing an artifact.
  • Coordinate llm-d placement with authenticated router receipts or server-owned artifact resolution; endpoint-local shortcuts may only be applied after endpoint selection.
  • Avoid returning full prompt-token arrays to public clients or creating a second KV-cache namespace outside vLLM APC.

Acceptance criteria

  • Any selected compatible replica can resolve the artifact or safely reseed from authenticated material.
  • Artifact miss, expiry, restart, wrong-pod selection, and validation failure deterministically fall back without changing model-visible token IDs.
  • Portable artifact execution produces the same vLLM APC lookup identity as ordinary full rendering.
  • Cache-hit ratio remains equal to the exact-routing baseline while long-context request transfer, reconstruction, rendering/tokenization work, or TTFT improves.
  • Local artifact storage is bounded, expiring, tenant-isolated, observable, and safe under process rotation.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions