Galahad KV cache: Corbenic AI beta stores encrypted attention states on NVMe for server reuse
The beta product stores encrypted attention states to disk, allowing servers to reuse processing work without re-reading identical prompts.
Key points
- Galahad beta launched on 1 October 2026 for vLLM, SGLang and Llama.cpp
- The system uses Blaise retrieval and Taliesin caching components
- Founder states no special hardware is required beyond existing NVMe drives
Corbenic AI released the beta version of Galahad on 1 October 2026, a library that stores encrypted key-value cache data on local NVMe drives. The tool allows servers to reload attention states when identical text reappears, avoiding redundant processing. According to TechRadar Pro, founder Sietse Schelpe described the mechanism as turning the server's existing SSD into part of the AI's working memory.
Root cause
The performance gain stems from avoiding repeated computation of attention states. In standard serving, models recompute the prefill stage from the first token for every query about a document. Corbenic's paper notes that while engines like vLLM reuse this state in GPU memory, it is lost under memory pressure or restarts. Galahad addresses this by persisting the state to disk using the Taliesin component, which fingerprints input bytes, model and tenant.
Attack path
The system relies on exact matching of input parameters. The Taliesin component writes cache blocks to disk based on a fingerprint of the exact input bytes. When those bytes reappear, the system loads the stored state. Schelpe told TechRadar Pro that this stored copy only activates once GPU memory has been cleared or the server has restarted, meaning it does not replace RAM usage during active sessions.
Affected versions
The beta supports specific inference engines. According to TechRadar Pro, Galahad integrates with vLLM, SGLang and Llama.cpp. The solution does not require special hardware, utilising the NVMe drive already present in the server. The company positions the tool as a small library offering exact, fail-closed reuse, distinguishing it from broader persistence efforts by Nvidia, Phison and LMCache.
Mitigation
- Ensure NVMe drives have sufficient capacity for cache storage
- Verify encryption keys are managed securely for cached data
- Monitor disk I/O impact during high-concurrency periods
- Test integration with specific vLLM or SGLang configurations
Why it matters
Teams managing large language models often face constraints in GPU memory capacity. Persisting state to local storage allows for more efficient use of limited resources. Operations should evaluate whether their workloads involve repeated prompts. This approach reduces computational overhead for identical requests. It also requires careful management of storage space and data security. Teams usually assess the trade-off between latency and storage costs. Check whether existing infrastructure can handle increased disk I/O.
What to do and how to stay safe: Corbenic AI
- Review current GPU memory utilisation patterns
- Assess frequency of repeated prompt submissions
- Plan storage allocation for encrypted cache data
- Test retrieval latency with production-like workloads
Step-by-step guide: Password Policies Examples That Reduce Password Fatigue
General security guidance from the Vector Bulletin newsroom. It is not confirmed advice from the organisations named in this story.
Frequently asked questions
Does Galahad replace GPU memory?
No, the stored copy only activates once GPU memory is cleared or the server restarts, according to founder Sietse Schelpe.
Which inference engines are supported?
The beta supports vLLM, SGLang and Llama.cpp, as stated by Corbenic AI.
What drives the 14x speedup?
The paper credits most of the speedup to Blaise, a retrieval component, rather than SSD loading alone.
