Jsgforever

A remote client accessing the OpenAI-compatible API can cause a denial of service by submitting a specific legal image where normalized pixels sum to zero when processed alone in a per-step encoder batch. This sequence leads the ` parse and validate image input` function to return no embedding, which is then propagated by `embed multimodal`. Consequently, the `sanity check mm encoder outputs` function triggers a fatal assertion, shutting down the engine for all users. This affects the DeepSeek-OCR, DeepSeek-OCR-2, and Unlimited-OCR image paths.

A denial of service issue allows remote attackers to crash the engine by sending inconsistent multimodal metadata to the disaggregated '/inference/v1/generate' endpoint. This occurs when a valid multimodal item is replayed with repeated or overlapping `mm placeholders` ranges, or when `kwargs data` and `mm hashes` lists are mismatched. These actions result in a negative text span during Qwen2.5-VL position construction, which terminates the engine.

The software uses a client-supplied multimodal `uuid` verbatim as a cache identity without performing consistency checks on cache hits. This allows unauthenticated attackers to desynchronize the P0/P1 processor caches, triggering a fatal token-placeholder mismatch that crashes the entire serving instance, requiring a restart to recover.