vLLM versions up to and including 0.29.0 contain a memory exhaustion flaw. The software fails to validate the tp_size parameter supplied in kv_transfer_params on OpenAI‑compatible completion endpoints, allowing an attacker to request an unbounded value. When the decoder worker processes such a request, it allocates that amount of memory, quickly consuming all available RAM and causing the kernel to invoke the OOM killer to terminate the worker process. This results in a denial‑of‑service condition for the affected deployment. The weakness is a classic example of uncontrolled memory allocation, aligned with CWE‑789.

The vulnerability affects the vllm-project vllm product. All releases through 0.29.0 are impacted; deployments using any of these versions are at risk. The flaw is triggered by the OpenAI‑compatible completion endpoints exposed by the service.

The CVSS score of 8.7 labels this issue as high severity. No EPSS score is available to indicate current exploitation likelihood, and the vulnerability is not listed in CISA's KEV catalog. Based on the description, it is inferred that attackers could exploit it remotely via the exposed completion API without requiring privileged access, assuming they can supply arbitrary tp_size values. Because the flaw leads directly to memory allocation, the attacker’s race condition is almost deterministic once the request reaches the server. The kernel OOM mechanism will terminate the offending process, potentially disrupting all concurrent decoding activities and halting service until a restart or cleanup occurs.