CVE-2026-115916

Name of the Vulnerable Software and Affected Versions vLLM versions prior to 0.32.0

Description An off-by-one error in the calculation of max token id within vllm/tokenizers/hf.py allows a pre-tokenized prompt ID equal to the vocabulary size to pass validation. When the tensor-parallel size is one and the vocabulary is an exact multiple of 64, the VocabParallelEmbedding.forward() function fails to mask this value before it is used by F.embedding. This results in a CUDA device-side assertion, which terminates the engine.

Recommendations Update to version 0.32.0.