Pushing memory bound CUDA kernels past the speed of light with data compressionfergusfinn.com·2 pts·somnial·0
Speculative KV coding: ~4× losslessly compressed KV cache using a small modelfergusfinn.com·2 pts·somnial·0