HackerTrans
TopNewTrendsCommentsPastAskShowJobs

somnial

9 karmajoined 9 mesi fa

Submissions

The Anatomy of an Instruction Pipeline Hazard

hiraditya.github.io
7 points·by somnial·8 ore fa·0 comments

Width vs. Depth: Speculating on the Margin

blog.doubleword.ai
17 points·by somnial·12 giorni fa·1 comments

Pushing memory bound CUDA kernels past the speed of light with data compression

fergusfinn.com
2 points·by somnial·2 mesi fa·0 comments

Speculative KV coding: ~4× losslessly compressed KV cache using a small model

fergusfinn.com
2 points·by somnial·2 mesi fa·0 comments

[untitled]

1 points·by somnial·2 mesi fa·0 comments

70x faster cold(ish) starts for SGLang

fergusfinn.com
1 points·by somnial·3 mesi fa·0 comments

LLM powered data structures: A lock-free binary search tree

fergusfinn.com
1 points·by somnial·6 mesi fa·0 comments

Parallel Primitives for Multi-Agent Workflows

fergusfinn.com
1 points·by somnial·6 mesi fa·0 comments

Scheduling in LLM Inference

fergusfinn.com
1 points·by somnial·8 mesi fa·0 comments

How fast can an LLM go?

fergusfinn.com
2 points·by somnial·9 mesi fa·0 comments

comments

somnial
·17 giorni fa·discuss
this 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
·mese scorso·discuss
https://fergusfinn.com/blog/economics-of-speculative-decodin...

good point tho - plus for Deepseek the shared expert increases the overlap slightly
somnial
·mese scorso·discuss
true, but no reason the predictor model couldn't use linear attention (i.e. mamba, GDN etc) to predict KV caches