vLLM is an inference and serving engine for large language models. From 0.20.2rc0 until 0.26.0, safe_load_prompt_embeds in vllm/renderers/embed_utils.py uses torch.sparse.check_sparse_tensor_invariants, whose process-global save, enable, and restore state can be raced by concurrent prompt_embeds parts submitted to POST /v1/chat/completions through AsyncMultiModalItemTracker.resolve_items, asyncio.gather, and the default executor, allowing an invalid sparse tensor to reach tensor.to_dense despite the CVE-2025-62164 guard when enable_prompt_embeds is enabled. This issue is fixed in version 0.26.0.
The product contains a concurrent code sequence that requires temporary, exclusive access to a shared resource, but a timing window exists in which the shared resource can be modified by another code sequence operating concurrently.
A race condition occurs within concurrent environments, and it is effectively a property of a code sequence. Depending on the context, a code sequence may be in the form of a function call, a small number of instructions, a series of program invocations, etc. A race condition violates these properties, which are closely related:
A race condition exists when an “interfering code sequence” can still access the shared resource, violating exclusivity. The interfering code sequence could be “trusted” or “untrusted.” A trusted interfering code sequence occurs within the product; it cannot be modified by the attacker, and it can only be invoked indirectly. An untrusted interfering code sequence can be authored directly by the attacker, and typically it is external to the vulnerable product.