feat(client): pass through sampling-mask logprobs and by-handle payload segments

[faresobeid]

On top of #174's packed completion logprobs, renderers.client.generate() passes two more optional prime-rl /inference/v1/generate fields through. Both are pure pass-through: no request-side change, no parsing or validation (same as sampling_mask). Stock vLLM sends neither, and then the keys are None.

• sampling_mask_logprobs = choice.get("sampling_mask_logprobs"): the sampler's logprobs at the sampling-mask ids, packed base64 float32 {data, shape, dtype}, parallel to the packed sampling_mask ids. verifiers decodes it (feat(v1): optional sampler logprobs on SamplingMask verifiers#2756). Consumer: prime-rl score centering, feat: exact score centering over the sampling mask (arXiv:2609.20807) prime-rl#3794.
• payload = choice.get("payload"): by-handle segments of a shared file that prime-rl's inference server writes instead of returning routed_experts / sampling_mask inline. verifiers puts them on traces (feat(v1): by-handle payload segments on traces verifiers#2765). Consumer: feat: pass router-replay ids and sampling masks by handle on multi-node runs prime-rl#3902.

Stack

Base: #174 (perf/packed-completion-logprobs). Order: #174 -> PrimeIntellect-ai/verifiers#2778 -> prime-rl#3900; then this PR with PrimeIntellect-ai/verifiers#2756 (-> prime-rl#3794) and PrimeIntellect-ai/verifiers#2765 (-> prime-rl#3902).

Validation

• pytest tests/test_client.py --noconftest (32 passed), ruff check, ruff format --check on renderers/client.py.

Note

Add sampling_mask_logprobs and payload to generate client result

The async generation client in client.py now reads two new fields from the generation choice and returns them. sampling_mask_logprobs carries the server's renormalized sampling-mask logprobs (prime-rl may send these with packed sampling-mask data). payload carries by-handle side-array data. Both keys are None when the choice does not provide the field.

^{Macroscope summarized d3a1af0.}

---

Note

Low Risk
Optional response-field pass-through only; no request changes, validation, or behavior when fields are absent.

Overview
generate() now surfaces optional inference metadata from the generation choice without parsing or validation.

It returns sampling_mask_logprobs alongside the existing sampling_mask: when prime-rl’s server runs with score centering, this is the packed float32 logprob per kept vocab id (parallel to the mask ids); stock vLLM omits it and the value is None. The docstring and inline comments describe that behavior.

The result dict also passes through payload (by-handle side arrays for large routed_experts / sampling_mask segments written by the inference server instead of inline JSON).

^{Reviewed by Cursor Bugbot for commit d3a1af0. Bugbot is set up for automated code reviews on this repo. Configure here.}

[macroscopeapp]

Approvability

Verdict: Approved at d3a1af0

Macroscope's review found this PR approvable — This one-file change adds two optional server-response fields to generate() without changing requests, existing outputs, or control flow. The pass-through is small, backwards-compatible, and isolated, with no sensitive, deployment, or major runtime impact.

Notes:

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

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

Dismissing prior approval to re-evaluate d3a1af0