thu-ml / SageAttention
Quantized Attention achieves speedup of 2-3x and 3-5x compared to FlashAttention and xformers, without lossing end-to-end metrics across language, image, and video models.
See what the GitHub community is most excited about today.
Quantized Attention achieves speedup of 2-3x and 3-5x compared to FlashAttention and xformers, without lossing end-to-end metrics across language, image, and video models.
FlashMLA: Efficient MLA decoding kernels
FlashInfer: Kernel Library for LLM Serving
Instant neural graphics primitives: lightning fast NeRF and more
DeepEP: an efficient expert-parallel communication library
CUDA accelerated rasterization of gaussian splatting
Tile primitives for speedy kernels
NCCL Tests
cuGraph - RAPIDS Graph Analytics Library
cuVS - a library for vector search and clustering on the GPU
[MICRO'23, MLSys'22] TorchSparse: Efficient Training and Inference Framework for Sparse Convolution on GPUs.
CUDA Library Samples
RAFT contains fundamental widely-used algorithms and primitives for machine learning and information retrieval. The algorithms are CUDA-accelerated and form building blocks for more easily writing high performance applications.
This package contains the original 2012 AlexNet code.
CUDA Kernel Benchmarking Library
PyTorch bindings for CUTLASS grouped GEMM.
[ARCHIVED] Cooperative primitives for CUDA C++. See https://github.com/NVIDIA/cccl
LLM training in simple, raw C/CUDA