Technology

In modern LLM applications like RAG and Agents, the model is constantly fed new context. For example, in RAG, we retrieve relevant documents and stuff them into the prompt.

The issue is that this dynamically retrieved context doesn't always appear at the beginning of the input sequence. Traditional KV caching only reuses a "common prefix," so if the new information isn't at the very start, the cache hit rate plummets, and your GPU ends up recomputing the same things over and over.

CacheBlend changes the game by allowing for the reuse of pre-computed KV caches regardless of their position in the input sequence.

This makes it possible to achieve a 100% KV Cache hit rate in applications like RAG. The performance gains are significant:

• Faster Time-To-First-Token (TTFT): Get your initial response much quicker.
• More Throughput: Serve significantly more users with the same hardware.
• Almost lossless Output Quality: All of this is achieved with little degradation in the model's generation quality.

CacheBlend works by intelligently handling the two main challenges of reusing non-prefix caches:

• Positional Encoding Update: It efficiently updates positional encodings to ensure the model always knows the correct position of each token, even when we're stitching together cached and new data.
• Selective Attention Recalculation: Instead of recomputing everything, it strategically recalculates only the minimal cross-attention needed between the new and cached chunks to maintain perfect generation quality.

An interactive CacheBlend demo is available at: https://github.com/LMCache/LMCache-Examples/tree/main/demo-rag-blending

https://archive.ph/lC5R0

Thoughts on this? I hate LLM but I think this article does a disservice by depicting the victims as perfectly normal before their mental health collapsed. They must have had some kind of preexisting problems that got exacerbated by their use of ChatGPT.

The point about these chatbots being sycophantic is extremely true though. I am not sure whether they are designed to be this way--whether it is because it sells more or if LLMs are too stupid to be argumentative. I have felt its effects personally when using Deepseek. I have noticed that often in its reasoning section it will say something like "the user is very astute" and it feels good to read that as someone who is socially isolated and is never complimented because of that.

I guess the lesson here is to use these chatbots as tools rather than friends because they are not capable of being the latter. I have tried a few times having discussions about politics with Deepseek but it is a terrible experience because of the aforementioned predisposition to being sycophantic. It always devolves into being a yes man.

https://archive.ph/xbx3M

https://archive.ph/Q9FEn

https://github.com/VectorSpaceLab/OmniGen2

https://archive.ph/MnjB5

https://archive.vn/Vj4U8

https://archive.ph/MrmJt

Jan-nano-128k is model fine-tuned to improve performance when enable YaRN scaling (instead of having degraded performance). This model will require YaRN Scaling supported from inference engine.

• It can uses tools continuously, repeatedly.
• It can perform deep research
• Extremely persistent

gguf can be found at: https://huggingface.co/Menlo/Jan-nano-128k-gguf