Ring attention implementation with flash attention

High-performance FlashAttention Implementation for Ascend NPU

PTO instruction set architecture

🚀 Efficient implementations for emerging model architectures

A pure-Python implementation of the Nvidia CuTe layout algebra intended to be approachable and easy to learn.

Custom kernel collections using https://github.com/PTO-ISA/pto-isa

Pythonic interface and JIT compiler for https://gitcode.com/cann/pto-isa

Open-source transpiler for CUDA Tile (13.1) migration

FlyDSL is the Python front‑end of the project: a Flexible Layout Python DSL for expressing tiling, partitioning, data movement, and kernel structure at a high level.

CUDA Tile IR is an MLIR-based intermediate representation and compiler infrastructure for CUDA kernel optimization, focusing on tile-based computation patterns and optimizations targeting NVIDIA te…

Tile-based language built for AI computation across all scales

⚡️Write HGEMM from scratch using Tensor Cores with WMMA, MMA and CuTe API, Achieve Peak⚡️ Performance.

A small OpenCL benchmark program to measure peak GPU/CPU performance.

USP: Unified (a.k.a. Hybrid, 2D) Sequence Parallel Attention for Long Context Transformers Model Training and Inference

A tool for bandwidth measurements on NVIDIA GPUs.

FlashAttention (Metal Port)

Puzzles for learning Triton

WALDSAC (RRANSAC with SPRT)

Exploring the scalable matrix extension of the Apple M4 processor

SGLang is a high-performance serving framework for large language models and multimodal models.

Apple GPU microarchitecture

Scientific computing with Metal in C++: Matrix multiplication example

Interactive architecture diagrams for codebases

TTS support with GGML

Everything we actually know about the Apple Neural Engine (ANE)