perf(dpa): add session-aware routing and post-prefill decode protection

[yhl-amd]

Summary

• propagate X-Dynamo-Session-ID / X-Correlation-ID through the OpenAI path and load-place new DPA sessions using estimated uncached prefill tokens plus request-equivalent decode pressure
• keep established sessions on an immutable cache owner so later agent turns preserve prefix-cache locality; child correlation IDs are placed independently
• add an opt-in post-prefill decode interval, synchronized across DPA ranks through the existing delayer reduction
• arm the interval only after a prefill forward actually completes, rather than when the delayer merely permits admission
• expose routing counters/gauges and document the new controls and production defaults

Enablement

export ATOM_DP_SESSION_AFFINITY=1
export ATOM_DP_LB_REQ_EQUIV=512
export ATOM_ENABLE_PREFILL_DELAYER=1
export ATOM_PREFILL_DECODE_INTERVAL=10

# server
--dp-load-balance least_tokens

All new behavior is opt-in: session affinity defaults to off and the post-prefill decode interval defaults to 0. Existing sessions never migrate.

Benchmark evidence

On the 3600-second DPA8 agentic comparison, the combined routing + PDI configuration measured:

• input throughput: +0.05% (neutral)
• mean TTFT: -4.15%
• p50 TTFT: -3.10%
• mean ITL: -1.16%
• p50 ITL: -0.23%

That run also included the dense low-batch-size CUDA Graph change, so the result is evidence for the combined configuration and is not presented as isolated PDI attribution.

A separate 600-second replay-cost migration experiment reduced input throughput by 0.71%, cache read by 0.94 pp, and actual prefix hits by roughly 2.04 pp. This PR therefore keeps established session ownership strict and excludes migration.

The prior +4.47% DPA result depended on a separate device-synchronization change and is intentionally not claimed here.

Validation

• 69 passed, 1 skipped: DP routing, API header propagation, prefill delayer, and partial-prefill-tail tests
• 33 passed: sequence and prefill scheduler tests
• Black check passed on changed Python files
• Ruff passed on changed Python files
• compileall and git diff --check passed

Scope notes

This PR intentionally excludes device synchronization, session migration, LMCache/offload, MegaMoE fixed-slot transport, and independent-replica experiments.

InferenceX currently reports Unsupported agentic server metrics backend for ATOM metrics during final aggregation. Raw AIPerf JSON/CSV export completes successfully; aggregator support is outside this ATOM change.

[github-actions]

🏷️ CI Guide

Runs automatically on every eligible PR before approval:

• ✅ Pre Checkin: Black, Ruff, catalog schema validation, non-GPU unit tests

Heavy model tests:

• ✅ Run after the PR is approved and Pre Checkin passes
• ✅ Run immediately when an approval review is submitted
• ✅ Can be requested before approval with labels

Label	Tests
ci:full	Run all heavy PR model tests: native ATOM, vLLM, and SGLang
ci:atom	Run native ATOM model accuracy tests
ci:vllm	Run ATOM vLLM OOT model accuracy tests
ci:sglang	Run ATOM SGLang model accuracy tests

Heavy jobs are skipped when the PR is not approved and no matching ci:* label is present.
Add labels via the sidebar or gh pr edit 2016 --add-label <label>