WombatKV
An object-storage-native key-value cache for LLM inference that lets any engine pointed at the same S3 bucket reuse prefilled context across restarts and machines.
WombatKV stores whatever an inference engine prefills once, such as system prompts, RAG context, codebase context or conversation history, as content-addressed blocks in an S3 bucket, following the pattern of durable truth on object storage with a free-to-die compute layer on top.
Because the storage key is derived from the model, tokenizer, dtype, LoRA and prefix chain, any engine that computes the same chain finds the same blocks, with no engine ID, session ID or coordination needed. This lets a restarted process skip re-prefilling, lets several conversations over the same shared document share one prefill, lets concurrent reviewers of the same pull request share one prefill, and lets separate engines on separate machines pointed at the same bucket prefix-cache between each other.
The 0.1.0-alpha release ships a ds4 (DeepSeek-V4-Flash) adapter, validated on an Apple M3 Max with local MinIO and on Linux with daemon-mode transport benchmarks.