Neilalfred

**Name of the Vulnerable Software and Affected Versions** vLLM versions prior to 0.24.0 **Description** A flaw in the rejection sampler during multi-request speculative decoding workloads allows the production of a recovered token equal to the model vocabulary size boundary value. This value is converted to negative one when the engine selects the next live token and is written into the drafter's input ids. This out-of-vocabulary value is subsequently processed by the model's embedding and attention path, resulting in a GPU device-side assertion that crashes the engine worker. This condition can be triggered remotely via the 'Generate' and 'Abort' gRPC endpoints, allowing a client to cause a service-wide denial of service by aborting concurrent requests and crashing the shared engine worker. **Recommendations** Update vLLM to version 0.24.0.