karpathy / llm.c
LLM training in simple, raw C/CUDA
See what the GitHub community is most excited about this week.
LLM training in simple, raw C/CUDA
Tile primitives for speedy kernels
GPU accelerated decision optimization
[ICLR2025, ICML2025, NeurIPS2025 Spotlight] Quantized Attention achieves speedup of 2-5x compared to FlashAttention, without lossing end-to-end metrics across language, image, and video models.
DeepEP: an efficient expert-parallel communication library
cuVS - a library for vector search and clustering on the GPU
Fast CUDA matrix multiplication from scratch
FlashInfer: Kernel Library for LLM Serving
CUDA Kernel Benchmarking Library
[ARCHIVED] Cooperative primitives for CUDA C++. See https://github.com/NVIDIA/cccl
DeepGEMM: clean and efficient FP8 GEMM kernels with fine-grained scaling
Instant neural graphics primitives: lightning fast NeRF and more