-
Notifications
You must be signed in to change notification settings - Fork 907
Home
Compute less. Serve longer.
RedKnot is more than a project.
We are committed to the long-term exploration of KV Cache reuse and KV Cache optimization, and we hope to build an open, objective, and trustworthy community around it.
We believe good systems research should be open to discussion, reproducible in practice, and able to stand the test of the open-source community.
Whether you are working on KV reuse, cache management, attention optimization, long-context serving, or simply curious about this direction — you are welcome here.
We welcome ideas, experiments, discussions, criticism, and contributions from everyone who cares about making KV Cache better.
Open source makes technology accessible. And in an era of technological democratization, the spirit of open source will shine brighter than ever.
RedKnot is an efficient long-context LLM serving framework built on SGLang.
Instead of recomputing everything, RedKnot tries to:
- Reuse what is stable
- Recompute what matters
- Skip what is unnecessary
Our goal is simple:
Compute only what needs to be computed again.
RedKnot explores this idea across KV Cache reuse, attention heads, tokens, MoE experts, and the underlying serving system.
🚀 Quick Start Run RedKnot and reproduce the released results.
🧠 How RedKnot Works Understand KV reuse, sparse execution, and SegPagedAttention.
📊 Benchmarks Performance, quality, and compute-saving results.
🤖 Model Support DeepSeek, Qwen, Mistral, Llama, and more.
🔧 Ascend NPU Ascend adaptation and deployment guides.
📄 Paper Read the technical details behind RedKnot.
KV Cache is becoming one of the most important system abstractions for long-context and agentic LLM serving.
There is still a lot to explore.
If you have a new idea, an unexpected result, a better implementation, or even a disagreement with one of our conclusions, open an issue, start a discussion, or send us a PR.
RedKnot belongs to everyone who helps make it better.
Welcome to RedKnot.