Satoro·1d agoBuild LLM from Scratch: Part 2Hi folks! 👋 Welcome to Part 2 of our Build an LLM from Scratch series!
Himanshu Singh·1d agoEvery Word Asks a Question: Understanding Attention with Gemma (Part 1)Understanding attention with a simple walk-through of Q, K, V, attention heads and the matrices involved, using real code from Google’s…
Ray Ng·1d agoSame PC, Same GPU: A Small CUDA Sync Took 68 µs on Windows and 15 µs on LinuxWhy our speech model ran slower than real time on a capable PC, and the short script that showed where the time goes
Vortex 404·1d agoAWS Open-Sourced a PyTorch Backend for Trainium — One Line of Code, No CUDATorchNeuron lets existing PyTorch code target Amazon’s own chips with .to(‘neuron’) instead of a rewrite — and Amazon published its own gap…
Anandha L Ranganathan·1d agotorch.compile() in PyTorchWhat Dynamo, AOTAutograd and Inductor each do, and how they work together — A beginners guide.
The Red Pill·2d agoThe Day I Found NaN Loss and Had No Idea WhyMy loss looked healthy until step 118. Then it jumped to inf, then nan, and every metric after that was useless. The run failed the same…
guillain bisimwa·3d agoWhen AI becomes an artist: Teaching a Neural Network to repaint your photosA network trained to recognise cats and cars can also paint. You give it two pictures: one whose content you want to keep, and one whose…
Anandha L Ranganathan·4d agoRegularization with Gradient descentPlease read here the gradient descent — introduction
Esteban Gómez·4d agoTracking machine learning experiments with NuBoardWhen I train a model, I usually run many variations of it: a few learning rates, a couple of seeds, an ablation or two. TensorBoard shows…
