vLLM — low-severity GPU logits state corruption via allowed_token_ids validation gap

A validation-bound mismatch between tokenizer vocabulary length and model output logits width in vLLM allows crafted token IDs to corrupt internal GPU sampling state.

Though rated low severity, this affects the widely-used vLLM engine's core sampling logic; the precision of the catalogued CVE is worth tracking even at low individual blast radius.

SamplingParams._validate_allowed_token_ids() validates supplied token IDs against tokenizer length rather than the model's actual output logits width, allowing out-of-range token IDs to pass validation and corrupt LogitBiasState / GPU logits state.

vLLM < 0.29.0

Upgrade to vLLM 0.29.0 or later.