[Bugfix] Forward session headers to consistent-hash routing in PD discovery mode

[simondanielsson]

Purpose

This preserves session affinity for both completions and chat completions when running in --vllm-pd-disaggration mode with service discovery.

Previously, the X-Session-Id header was ignored in that case.

Test Plan

CI passes.

Test Result

OK

---

Essential Elements of an Effective PR Description Checklist

• The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
• The test plan, such as providing test command.
• The test results, such as pasting the results comparison before and after, or e2e results

[hsliuustc0106]

fix rust checks please

[simondanielsson]

Environment issue in NV PD tests causing CI to fail: https://buildkite.com/vllm/router/builds/786/canvas?sid=01a0812d-1087-4bb6-92ac-2546f27be662&tab=output

[simondanielsson]

@hsliuustc0106 Who can I contact to get a review on this PR? Thanks

[Copilot]

🟡 Changes recommended

The new per-request header normalization introduces avoidable allocations and should be gated on whether the active policies actually need headers, using ASCII-only lowercasing for header names.

Once you've addressed the issues Copilot identified, you can request another Copilot review.

Pull request overview

This PR fixes session affinity in vLLM prefill/decode (PD) discovery mode by ensuring request headers (notably X-Session-Id) are forwarded into policy-based worker selection, so consistent-hash routing can remain stable across requests.

Changes:

• Normalize incoming HTTP headers and pass them into policy-based worker selection for PD discovery routing.
• Update discovery-mode worker selection to use the header-aware policy selection API.
• Add tests verifying consistent-hash routing uses session headers (and that cache-aware routing ignores them).

File summaries

File	Description
src/routers/http/vllm_pd_router.rs	Normalizes headers for discovery-mode selection, threads headers into policy-based routing, and adds regression tests for session-affinity behavior.

Review details

Suppressed comments (1)

src/routers/http/vllm_pd_router.rs:608

• process_vllm_request normalizes all headers into a new HashMap for every request, even when the active policies don’t consult headers. This adds avoidable per-request allocations; you can gate the normalization on needs_headers() for the prefill/decode policies and pass None otherwise.

        let request_text = serde_json::to_string(&request_json).ok();
        let request_str = request_text.as_deref();
        let request_headers = headers.map(normalize_request_headers);

• Files reviewed: 1/1 changed files
• Comments generated: 1
• Review effort level: Lite

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.