vLLM Security Advisory: Incomplete CVE-2025-62164 Fix Bypassed

According to the GitHub Security Advisory database, vLLM revision `26587f9519e22a5c4549ead7595ad9ca3229c4fd` contains an incomplete remediation for CVE-2025-62164. The issue arises from concurrent prompt-embedding parts in `/v1/chat/completions` requests bypassing the intended protection mechanism. Specifically, the safeguard relies on PyTorch's `torch.sparse.check_sparse_tensor_invariants()` context, which operates on process-global state. When two prompt parts are processed concurrently, one context can exit and restore the global flag to `False` before the second part loads its tensor, allowing invalid sparse payloads to pass through.

This vulnerability is triggered when the `--enable-prompt-embeds` flag is set, which is default-off. However, it does not depend on other configurations like `renderer_num_workers > 1` or multimodal model settings. Additionally, API authentication is optional unless explicitly configured via CLI or environment variables.

Operators should carefully evaluate whether the `--enable-prompt-embeds` flag is necessary for their deployment and ensure that API authentication is enforced to mitigate risks. Testing for concurrency-related issues in prompt handling is also advisable, as this pattern of incomplete remediation highlights the challenges of securing global state in multithreaded environments.