perf(client): read prime-rl's packed completion_logprobs

[faresobeid]

Lets generate() read prime-rl's packed per-token logprobs.

prime-rl's /inference/v1/generate server now returns the sampled-token logprobs as one base64 float32 array, choice.completion_logprobs = {data, shape, dtype}, instead of choice.logprobs.content (one JSON object per token). The per-token objects were a large part of the vLLM API server's GC load under RL traffic (see the prime-rl PR).

• _parse_completion_logprobs reads completion_logprobs when present and applies the same checks as the per-token path: count matches completion_ids, values finite, no -9999.0 sentinel. Returns the same list[float].
• Without the field (stock vLLM), nothing changes.
• sampling_mask is passed through as before. On prime-rl's server it is now a packed {ids, counts} dict; verifiers decodes it.

Companions: prime-rl PrimeIntellect-ai/prime-rl#3900 / verifiers PrimeIntellect-ai/verifiers#2778. Land renderers -> verifiers -> prime-rl; the prime-rl submodule pins must move to the merged commits.

Validation

• tests/test_client.py: 32 passed (one new test for the packed path, incl. the sentinel check).
• ruff check / ruff format --check (0.15.12) on the changed files: clean.

---

Note

Low Risk
Localized response parsing with fallback to stock vLLM; same validation semantics and output shape for downstream RL/verifier consumers.

Overview
Adds support for prime-rl’s compact per-token logprob payload on /inference/v1/generate responses.

When choice.completion_logprobs is present ({data, shape, dtype} with base64 float32), generate() decodes it via new _parse_packed_completion_logprobs and still returns the same list[float] as before. Validation mirrors the legacy path: length must match completion_ids, values must be finite, and vLLM’s -9999 missing-logprob sentinel is rejected. If the packed field is absent, behavior is unchanged and choice.logprobs.content is parsed as today.

A comment on sampling_mask notes that prime-rl may return a packed CSR {ids, counts} form (still passed through unchanged). Tests cover the happy path plus sentinel and non-finite failures.

^{Reviewed by Cursor Bugbot for commit e161362. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Add packed completion_logprobs decoding to client renderer

• Adds _parse_packed_completion_logprobs in renderers/client.py, which decodes base64 packed logprob payloads into a NumPy array using the supplied dtype and returns per-token floats
• generate now prefers choice.completion_logprobs when present, falling back to the existing choice.logprobs.content parsing
• Length mismatches, non-finite values, and values at or below the -9999 missing-logprob sentinel raise MalformedGenerateResponseError
• Risk: importing renderers.client now requires NumPy as a dependency

^{Macroscope summarized 19c558f.}

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 8a9619d. Configure here.}

[macroscopeapp]

Approvability

Verdict: Approved at e134d56

Macroscope's review found this PR approvable — This is a localized, backward-compatible performance path for decoding packed completion logprobs; existing response handling remains unchanged and NumPy was already a required dependency. The new behavior is covered by focused tests for valid and invalid payloads.

Notes:

• Macroscope's correctness review did not run, so approvability was decided on eligibility alone.

No code changes detected at 19c558f. Prior analysis still applies.

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