Reject full-context chat prompts before max_tokens underflows

[rasdani]

Summary

• reject rendered standard chat prompts that already fill the model context before vLLM derives max_tokens=0
• preserve /chat/completions/tokens behavior by validating after stitched request.tokens are installed, so TITO checks the actual request tokens
• add focused coverage for the context-limit guard and token-endpoint deferral

Why This Fixes The Bug

Before this change, the standard /v1/chat/completions serving path rendered the prompt and then called vLLM get_max_tokens(...). When prompt_len == max_model_len, that helper can return 0; request.to_sampling_params(max_tokens=0) then surfaces as:

BadRequestError: max_tokens must be at least 1, got 0

That error is a generic model error to the rollout env, so group-scored rollouts can reschedule an otherwise nearly complete group.

After this change, OpenAIServingChatWithTokens.render_chat_request() validates rendered standard chat prompts before vLLM reaches get_max_tokens. If the prompt has no generation room, it raises VLLMValidationError(parameter="input_tokens"), which vLLM serializes as a 400 BadRequestError with a context-length message:

This model's maximum context length is ... However, your request has ... input tokens ... (parameter=input_tokens, value=...)

The verifiers OpenAI client already classifies messages containing maximum context length / context length as OverlongPromptError, and the env handles that as prompt_too_long instead of a generic model failure.

Impact

Long multi-turn rollouts that reach max_model_len now stop through the existing overlong-prompt path instead of sending max_tokens=0 to vLLM. In the observed SLURM run this was not the dominant throughput bottleneck, but each occurrence can still be expensive because group scoring discards partial group progress and reschedules.

Verification

• uv run pytest tests/unit/inference/test_serving_chat_with_tokens.py tests/unit/inference/test_serving_generate.py
• uv run ruff format --check src/prime_rl/inference/vllm/serving_chat_with_tokens.py tests/unit/inference/test_serving_chat_with_tokens.py
• uv run ruff check src/prime_rl/inference/vllm/serving_chat_with_tokens.py tests/unit/inference/test_serving_chat_with_tokens.py

---

Note

Medium Risk
Changes request validation behavior in the OpenAI chat serving path; could alter which requests are rejected and the error shape returned, but is localized and covered by unit tests.

Overview
Adds an explicit context-length guard for standard /v1/chat/completions requests by validating rendered engine_prompts in render_chat_request and returning a VLLMValidationError(parameter="input_tokens") when the prompt already fills the model context (avoiding max_tokens=0 underflow errors).

Refactors the same check into _validate_prompt_has_generation_room() and reuses it in the token-in (/chat/completions/tokens) path, while deferring validation there until after request.tokens are stitched in so the check reflects the actual prompt tokens. Includes new unit tests covering both the guard behavior and the token-endpoint deferral.

^{Reviewed by Cursor Bugbot for commit 82d8849. Bugbot is set up for automated code reviews on this repo. Configure here.}