标签：cuda

A massively parallel, optimal functional runtime in Rust

Tile primitives for speedy kernels

LLM training in simple, raw C/CUDA

Code and data for paper “Deep Painterly Harmonization”: https://arxiv.org/abs/1804.03189