feat(renderers): switch client to vLLM 0.20 /inference/v1/generate

[hallerite]

Summary

vLLM 0.20 ships a unified tokens-in / tokens-out endpoint at /inference/v1/generate that supersedes the bespoke /v1/generate handler prime-rl shipped on top of vllm 0.19. Migrate verifiers' RendererClient onto the new endpoint and pin the renderers package to its lean rewrite.

Companion PRs:

• feat: switch client to vLLM 0.20 /inference/v1/generate renderers#1 — lean generate() rewrite (formerly this PR's packages/renderers/; now lives in its own repo)
• feat: unify renderer + teacher endpoints onto vLLM 0.20 /inference/v1/generate prime-rl#2408 — consumer migration onto the new endpoint

What changed

Renderers pin (pyproject.toml + uv.lock)

• [tool.uv.sources] pins renderers to PrimeIntellect-ai/renderers@40bc2a6 (the head of the companion PR — the lean generate() rewrite).
• packages/renderers/ deleted from the verifiers tree; the package is no longer vendored.
• Version constraint bumped to renderers>=0.1.6 in [project] deps and the renderers extra. Once renderers-v0.1.6 publishes to PyPI, drop [tool.uv.sources] and let it resolve from PyPI directly.
• Drops the uv pip install -e packages/renderers CI hack — no longer needed once renderers resolves through [tool.uv.sources].
• Deletes .github/workflows/publish-renderers.yml — the publish flow now lives in the renderers repo (ci: add PyPI publish workflow renderers#2).

RendererClient adapter (verifiers/clients/renderer_client.py)

• get_native_response builds a sampling_params dict from the caller's flat sampling_args / extra_body and calls the new generate(...) with named args. This is the right place for the OpenAI-shaped → lean adaptation; the renderers package itself no longer carries OpenAI-SDK conventions.
• from_native_response reads request_id instead of id; Usage is reconstructed from token-list lengths (the new endpoint doesn't return a usage block).

Test plan

• tests/test_renderer_client.py + tests/test_renderer_e2e.py updated for the new request_id / sampling_params shapes — 42/42 pass against the external pin.
• e2e renderer rollout against a live vllm 0.20 server (prime-rl#2408 + this PR's client): 20-step multi_reverse_text RL run, 2688 calls to /inference/v1/generate, eval Avg@4 = 0.83 — identical numbers to the fat-API version.
• ruff format --check / ruff check clean.
• ty check verifiers passes (0 errors).
• Rebased onto current main (picks up Make renderers optional and add PyPI publish workflow #1279 — renderers as optional / PyPI publish workflow).

Notes

• VLMs are still blocked by prime-rl's validate_renderer_vs_vlm config validator. The new endpoint already supports MM features end-to-end; lifting the ban needs the renderer client to build features client-side (HF processor → MultiModalKwargsItem → base64 msgpack). That's a separate PR — easier to review on its own once this lands.

---

Note

Medium Risk
Moderate risk because it changes the inference request/response contract (generate call shape, request_id, and usage reconstruction) and removes the vendored packages/renderers implementation in favor of a pinned external dependency.

Overview
Switches RendererClient over to renderers.client.generate, adapting OpenAI-style sampling args into a sampling_params dict, passing through cache_salt/priority/headers, and updating response handling to use request_id and reconstruct usage from token lengths.

Removes the in-repo packages/renderers implementation (and its publish-renderers GitHub workflow) and pins renderers via pyproject.toml (renderers>=0.1.6 plus a git tool.uv.sources override) so the renderer code is consumed as an external package.

^{Reviewed by Cursor Bugbot for commit b494fb7. Bugbot is set up for automated code reviews on this repo. Configure here.}