merlo
Inference on SmolLM2-135M on 2 Raspberry Pi Zero W to hit 3 tokens/s.
blog, paper
Model for trained with multi-turn reinforcement learning to write fast CUDA kernels.
civetta, passero
civetta implements everything you need for supervised training (autograd, dataloaders, optimizers, …).
passero implements an env template and various classical RL algorithms.
blog, hn
Trained model with reinforcement learning to do fast (sub 5s) codebase-wide context agentic context retrieval.
code
Library for working with Series Parallel Structures. For more info on what series parallel structures are, see here.