LLM training in simple, raw C/CUDA

A massively parallel, optimal functional runtime in Rust

DeepEP: an efficient expert-parallel communication library

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

DeepGEMM: clean and efficient FP8 GEMM kernels with fine-grained scaling

CUDA accelerated rasterization of gaussian splatting

This package contains the original 2012 AlexNet code.

🤖FFPA: Extend FlashAttention-2 with Split-D, ~O(1) SRAM complexity for large headdim, 1.8x~3x↑🎉 vs SDPA EA.