CVE-2026-54234
Description
vLLM is a high-throughput and memory-efficient inference and serving engine for LLMs. Prior to 0.24.0, a frontend-legal multi-request speculative decoding workload can cause the rejection sampler to produce a recovered token equal to the model vocabulary size boundary value, which is then converted to negative one when the engine selects the next live token for a request and is written back into the drafter's input ids; that out-of-vocabulary value is later consumed by the model's embedding and attention path and crashes the engine worker with a GPU device-side assertion. The same triggering request sequence is reachable through the public gRPC Generate and Abort endpoints, so a remote client that can send generation requests can crash the shared engine worker, aborting concurrent requests and causing a service-wide denial of service for other clients of the deployment until the worker is restarted. This issue is fixed in version 0.24.0.
Predictions
Heuristic predictions, AS-IS, for prioritization only.
Mitigations
No mitigations published for this CVE yet.
The vendor-content worker queues fetches as references arrive (check back in a few minutes). Or — if you've already worked around this in production — publish your fix to the community-verified tier.
✚ Propose a mitigation on Community → Mitigations published via the community go through AI scoring + 2 human reviewers + 7-day silent objection window before landing here withsource_tier=community-verified.
References
CWEs
CWE-20 CWE-1284
Community-verified mitigations for this CVE will appear above when contributors publish them.
Verify integrity in audit chain (admin only). AS-IS.