Flash-MSA: Accelerating Million-Token Training with Sparse Attention Kernelsnanduruganesh.github.io34 points·by rawsh·11 hours ago·3 comments
Zyphra ZAYA1-base: First large-scale model trained on AMDzyphra.com6 points·by rawsh·8 months ago·1 comments
Debugging divergence between engine and transformers logprobs for RLgist.github.com2 points·by rawsh·10 months ago·0 comments