A fast communication-overlapping library for tensor/expert parallelism on GPUs.

Our first fully AI generated deep learning system

🚀🚀🚀 This repository lists some awesome public CUDA, cuda-python, cuBLAS, cuDNN, CUTLASS, TensorRT, TensorRT-LLM, Triton, TVM, MLIR, PTX and High Performance Computing (HPC) projects.

Examples of CUDA implementations by Cutlass CuTe

CUTLASS and CuTe Examples

Suckless no dependencies CUDA/C++ port of NVIDIA's DVLT, feed it images, get a 3D point cloud + camera poses. NO python. One fast 5MB binary.

Performance of the C++ interface of flash attention and flash attention v2 in large language model (LLM) inference scenarios.

This is my 🔥 100 Days of GPU — a wild, hands-on journey through CUDA/CUTLASS kernels, Triton spells, and PTX sorcery.

GEMM and Winograd based convolutions using CUTLASS

Minimal FlashAttention in CUDA C++/CuTe: readable WMMA/CuTe kernels, no NxN workspace, up to 4.5x faster than naive PyTorch

NCU-driven iterative optimization workflow for CUDA/CUTLASS/Triton/CuTe DSL kernels.

SGLang Kernel Wheel Index

study of cutlass

NVFP4 inference on Blackwell GeForce (RTX 5090/5080/5070 Ti/RTX PRO 6000) — SM120 patches for vLLM + FlashInfer + CUTLASS. 175 tok/s on Qwen3.6-35B MoE.

Multiple GEMM operators are constructed with cutlass to support LLM inference.

A cutlass cute implementation of headdim-64 flashattentionv2 TensorRT plugin for LightGlue. Run on Jetson Orin NX 8GB with TensorRT 8.5.2.

CUDA/Triton/CuTe DSL/CUTLASS kernel optimization agent

Native Blackwell (sm_100) tcgen05 training backward for the gated-linear-recurrence family (GDN-2/GLA/KDA/SSD), plus a contract-grade verifier that falsifies published GPU kernels. Six open Mamba-3 Triton kernels including the first open MIMO backward; GRPO trainer graded on contract correctness. state-spaces/mamba#904

Reading notes on the open source code of AI infrastructure (cutlass, hpc, sglang, llm, compiler, etc.)

Faster attention kernels for serving TML's Inkling model on vLLM. 2.7x over the shipping path on H100, and the only implementation that runs on A100.