feat(router): opt-in parent session affinity for initial placement

[YAMY1234]

What

CLOSES: DYN-3521 DGH-1094

Opt-in parent session affinity for initial placement (proof-of-concept for #11343): when DYN_ROUTER_PARENT_AFFINITY=1 and a request's nvext.session_control carries parent_session_id, an unbound session's first request is placed on the worker (and DP rank) currently bound to its parent session. The session then binds its own id on route, so the hint only influences initial placement; parent lookup reuses the existing sticky machinery, including the ineligible-worker fallback. Default off; the new field is optional and wire-compatible.

Why

Agentic workloads spawn subagent sessions that share a large prefix with their parent. On deployments without KV-event prefix indexing, a child session's first request is placed blindly and re-prefills the shared prefix once per child. Details and design discussion in #11343.

Status / scope

• Draft on the pre-header session_control protocol (base 3169427d) — this is the protocol window we could validate end-to-end with an agentic trace-replay client emitting parent_session_id.
• Validated live on a 2-node disaggregated SGLang deployment: the policy routes (debug-logged) child first-requests onto the parent's worker/DP-rank as designed.
• Known follow-ups before this is mergeable:
  1. Port to current main — the sticky coordinator was reworked by feat(router): coordinate session affinity across replicas [DYN-3249] #11079 (session claims); on main the parent id is already parsed from the agentic identity headers, so the policy should consume that instead of a new body field.
  2. Config surface: promote the env toggle to a router config option.
  3. Load-aware guard: adopt the parent placement only under a load threshold, so wide sibling fan-outs don't over-concentrate a single DP rank.
  4. Tests.

Happy to do the main port if the direction looks right — opening early for feedback on the policy shape.

[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.

[dynamo-ops]

Unsigned Commits Detected

The following commits are not GPG-signed and must be signed before CI can run:

Commit	Reason
9f0e778	unsigned
4bea22b	unsigned

Please sign your commits and push again. See the GitHub docs on commit signature verification for help.

[JasperChan929]

Hi @YAMY1234, thanks for working on this.

I see that commit 9f0e778 already reworked the implementation onto the header-based session-affinity architecture, but the PR is still a draft, has not received reviews, and CI is currently blocked by the unsigned commits.

I have completed additional end-to-end validation for #11343:

• on 2 physical RTX 4090 GPUs, all 35/35 tested B2 event/predicted trials at fan-out 2/4 and an exact 16k prefix placed siblings across both GPU workers and computed one full prefix per selected worker;
• on 8 physical RTX 4090D GPUs, all 80/80 B2 event/predicted trials at fan-out 4/8 and exact 16k/32k prefixes selected 4/8 workers and computed 4/8 full prefix copies;
• the corresponding B1 affinity controls selected one worker and computed one shared prefix;
• a strict single-worker C2 matrix observed no duplicate full-prefix prefill in 120/120 trials.

I also opened the benchmark-only contribution request #14045 so this behavior can be tested independently of the runtime implementation.

Are you planning to resume and finish #11347? If so, I would be happy to help with benchmark validation, tests, rebasing, or other follow-up work. If not, would you be comfortable with me coordinating with the maintainers on a current-main successor PR, while preserving attribution to your implementation?

I will wait for your and the maintainers' direction before starting a separate runtime implementation.