[nxphi47]

Bug

With --intra-node-data-parallel-size greater than 1, these endpoints return a 500:

• GET /v1/models
• GET /get_model_info
• GET /get_server_info
• GET /health_generate

The error message is:

Request failed: error sending request for url (http://0.0.0.1/v1/models)

Inference requests (/v1/chat/completions and so on) are not affected.

Cause

With intra-node DP, each worker's URL is tagged with its DP rank, e.g. http://localhost:8100@1.

select_first_worker() returned this tagged URL, and proxy_get_request() appended the endpoint to it. The router therefore requested http://localhost:8100@1/v1/models. That URL parses as user localhost:8100 at host 1, which resolves to 0.0.0.1.

Inference requests already strip the tag via base_url(), and health checks via extract_dp_rank(). Only the proxied GETs used the tagged URL.

Fix

select_first_worker() and select_first_worker_for_model() now return base_url(). For a DP-aware worker that is the URL without @rank; for a regular worker it is unchanged.

Tests

• New unit test: test_select_first_worker_dp_aware registers four DP-aware workers and checks that the selected URL has no @rank.
• CI: the Test workflow passed: nightly build, cargo test, clippy with -D warnings, and rustfmt.
• Unrelated failure: Trivy vulnerability scanner fails because it cannot download its vulnerability database. It fails the same way on the other open PRs (fix/raise-nofile-limit, feat/pd-merge-payload-fields).
• Reproduction: run the v0.2.3 release binary in front of a stub worker that serves /health and /v1/models:

--intra-node-data-parallel-size	GET /v1/models through the router
1	returns the stub's model list
2	Request failed: error sending request for url (http://0.0.0.0/v1/models)

Why it matters

prime-rl starts this router with --intra-node-data-parallel-size set to the vLLM data-parallel size. Every prime-rl inference server with DP > 1 therefore has a broken /v1/models. Three effects:

• prime-rl's eval and orchestrator model check fails at startup unless skip_model_check is set.
• The orchestrator cannot read max_model_len. This fails silently.
• Agent harnesses cannot read the context window, so they never compact automatically.

[macroscopeapp]

Approvability

Verdict: Approved at a530aed

Macroscope's review found this PR approvable — This is a focused bug fix that removes the DP rank from URLs used by the router's proxied GET endpoints, while leaving regular-worker URLs unchanged. The accompanying unit test covers the DP-aware case and the diff introduces no broader API, schema, or infrastructure changes.

^{You can add or adjust custom eligibility rules. Learn more.}