The Anatomy of an Instruction Pipeline Hazardhiraditya.github.io5 points·by somnial·3 godziny temu·0 comments
Width vs. Depth: Speculating on the Marginblog.doubleword.ai17 points·by somnial·12 dni temu·1 comments
Pushing memory bound CUDA kernels past the speed of light with data compressionfergusfinn.com2 points·by somnial·2 miesiące temu·0 comments
Speculative KV coding: ~4× losslessly compressed KV cache using a small modelfergusfinn.com2 points·by somnial·2 miesiące temu·0 comments
LLM powered data structures: A lock-free binary search treefergusfinn.com1 points·by somnial·6 miesięcy temu·0 comments
Parallel Primitives for Multi-Agent Workflowsfergusfinn.com1 points·by somnial·6 miesięcy temu·0 comments
somnial·17 dni temu·discussthis is a blog post from a company that hosts open weights LLMs (https://www.doubleword.ai/). I think its possible it might have been tongue in cheek
somnial·w zeszłym miesiącu·discusshttps://fergusfinn.com/blog/economics-of-speculative-decodin...good point tho - plus for Deepseek the shared expert increases the overlap slightly
somnial·w zeszłym miesiącu·discusstrue, but no reason the predictor model couldn't use linear attention (i.e. mamba, GDN etc) to predict KV caches