Open-source alternatives to Cutlass

This repository is a collection of reference implementations and programming examples for the CUDA Toolkit. It serves as a GPGPU implementation guide and a parallel computing reference, providing code for using graphics hardware to perform general-purpose calculations and high-performance parallel processing. The project provides specific samples for GPU kernel development and resource management. These include demonstrations of multi-GPU communication, peer-to-peer memory access, and system hardware inspection to coordinate distributed GPU resources. The codebase covers a wide range of capa

AISystem is a comprehensive AI full-stack infrastructure project covering the entire pipeline from AI chip architecture to high-level training frameworks. It encompasses the development of AI compiler frameworks, inference engines, and distributed training orchestrators designed to coordinate workloads across a heterogeneous compute stack of CPUs, GPUs, and NPUs. The project focuses on the deep integration of software and hardware, employing software-hardware co-design to align tensor layouts with physical memory structures. It provides specialized capabilities for accelerating Transformer mo

Flashlight is a C++ machine learning library and deep learning framework designed for building and training neural networks. It functions as a tensor manipulation library and an automatic differentiation engine that tracks operations to calculate gradients via backpropagation for model optimization. The project is distinguished by its role as a distributed training framework, utilizing all-reduce gradient synchronization and distributed environments to scale machine learning workloads across multiple nodes and devices. It features a backend-agnostic memory interface and RAII-based management

Triton is a parallel computing framework and high-level programming language designed for writing custom compute kernels. It functions as a deep learning compiler, translating complex mathematical operations into high-throughput instructions that maximize hardware utilization and memory efficiency on graphics processing units. The framework distinguishes itself through a hardware-agnostic compute abstraction that allows developers to define kernels without manual low-level tuning. It employs just-in-time compilation to generate optimized binary instructions at runtime, utilizing static data f

FlashMLA is an LLM attention kernel library and inference acceleration library providing a collection of high-performance CUDA kernels. It implements multi-head latent attention mechanisms designed to reduce memory overhead and increase throughput during the forward and backward passes of large language model inference. The library utilizes quantized cache attention kernels to improve computation efficiency across both sparse and dense token processing. It specifically optimizes the prefill and decoding phases of model inference through these latent attention implementations. The project cov

Flashlight is a standalone C++ machine learning library and tensor library used for building and training neural networks. It functions as a comprehensive neural network framework and automatic differentiation engine, providing the tools to construct computation graphs and calculate gradients via backpropagation. The project serves as a distributed training framework, utilizing all-reduce operations to synchronize gradients and parameters across multiple compute nodes and devices. It distinguishes itself through deep integration of high-performance tensor manipulation, native device memory in

This project is a comprehensive educational resource and programming course covering C++ language semantics and features from C++03 through C++26. It provides structured tutorials and technical guides focused on modern C++ development. The material offers specialized instruction on template metaprogramming, including the use of type traits and compile-time computations. It features detailed guides on concurrency and parallelism for multi-core execution, as well as a reference for software design applying SOLID principles and RAII. Additionally, it covers build performance optimization to redu

Sglang is a high-performance inference engine and serving system designed for large language and multimodal models. It provides a programmable interface for orchestrating complex generation workflows, enabling developers to coordinate multi-turn dialogues, tool invocations, and reasoning chains through a domain-specific language. The platform is built to support production-scale deployments, offering an OpenAI-compatible API that allows for integration with existing application ecosystems. The system distinguishes itself through a disaggregated architecture that separates compute-intensive pr

xtensor is a C++ multidimensional array library for numerical computing that provides N-dimensional containers with an interface mirroring the NumPy API. It utilizes a lazy evaluation expression engine to defer numerical computations until assignment, which minimizes memory allocations and intermediate copies. The library features a foreign memory array adaptor that allows it to wrap external buffers, such as NumPy arrays, to perform numerical operations in-place without duplicating data. It further optimizes performance through lazy broadcasting and a system that manages the lifetime of temp

Warp is a Python framework that JIT-compiles Python functions into CUDA kernels for GPU-accelerated parallel computation, with built-in automatic differentiation and multi-framework array interoperability. At its core, it provides a GPU kernel compilation system that enables writing and executing custom GPU kernels directly from Python, while supporting automatic gradient computation through those kernels for integration with machine learning pipelines. The framework also includes tile-based cooperative computing, where thread blocks partition into tiles for shared-memory and tensor-core opera

Zen-C is a multi-target systems language and source-to-source compiler that translates high-level logic into human-readable GNU C or C11 code. It functions as a JIT-enabled programming language with an in-process compiler for real-time interactive code evaluation and testing. The project serves as a CUDA GPU kernel generator, mapping specialized syntax to CUDA C++ using device attributes to target graphics hardware. It acts as an interoperability layer capable of emitting compatible code for C++, Objective-C, and Lisp to bridge native system frameworks and libraries. The language includes an

TileLang is a Python-embedded domain-specific language compiler that JIT-compiles and autotunes GPU kernels. It uses a tile-based DSL, automatic software pipelining, and parallel autotuning to generate optimized GPU kernels at runtime. It supports tensor core operations with Pythonic syntax, automatic memory management, and thread mapping. The compiler searches over tile sizes, thread counts, and scheduling policies, compiling and benchmarking candidates in parallel to find the fastest kernel. It also caches compiled binaries and tuning results to disk for reuse across sessions. TileLang inc

bitsandbytes is a quantization library for large language models that reduces memory footprints using k-bit quantization. It provides a framework for 4-bit low-rank adaptation, tools for 8-bit model compression, and memory-efficient optimizer extensions for PyTorch. The project enables the training of large models on limited hardware through 4-bit quantization and low-rank adaptation weights. It also facilitates faster inference by compressing models to 8-bit precision using vector-wise quantization. The library covers a range of memory optimization capabilities, including optimizer memory r

This is a PyTorch recommendation framework and deep learning recommendation model designed to generate personalized content predictions. It functions as a distributed embedding trainer that processes dense and sparse features through a neural network architecture to predict user preferences. The project implements a CUDA-optimized machine learning system using specialized GPU kernels to accelerate embedding lookup and aggregation. It employs a distributed approach to shard massive sparse feature tables across multiple GPUs, enabling the training of large-scale models. The system utilizes a t

CuPy is a CUDA array computing library that implements a NumPy-compatible interface for executing array operations and numerical computing on NVIDIA GPUs. It serves as a GPU-accelerated numerical library and a CUDA-based SciPy implementation, offloading heavy calculations to graphics hardware to increase processing speed for scientific and engineering workloads. The library enables multi-framework tensor exchange, allowing data buffers to be shared between different deep learning frameworks using standardized memory layouts to avoid memory copies. It also supports custom GPU kernel integratio

ndarray is a multidimensional array library for Rust that serves as a linear algebra framework and scientific computing tool. It provides the core infrastructure for creating and manipulating n-dimensional arrays, functioning as both a parallel array processor and a toolkit for numerical data analysis. The library distinguishes itself by providing efficient slicing and memory views, allowing for data sharing without copying. It leverages optimized backend math libraries for high-speed matrix multiplication and distributes heavy mathematical iterations across multiple CPU threads to accelerate

NumPy is a foundational library for scientific computing in Python, providing a comprehensive framework for managing and manipulating large-scale numerical information. It centers on high-performance multidimensional array objects that serve as the primary data structure for complex mathematical operations and data analysis workflows. The library distinguishes itself through specialized mechanisms for handling multidimensional data, including advanced indexing, slicing, and broadcasting techniques that allow for efficient operations across arrays of varying shapes. It utilizes strided metadat

Numba is a just-in-time compiler that translates high-level Python functions into optimized machine code at runtime. By leveraging the LLVM compiler infrastructure, it provides a framework for accelerating numerical data processing and mathematical computations, enabling performance levels comparable to statically compiled languages. The project distinguishes itself through its ability to perform type-inference-based specialization, which generates machine instructions tailored to the specific data types used during execution. It employs a lazy compilation pipeline that defers translation unt

Darknet is a low-level neural network engine and framework written in C. It is designed for training and deploying deep learning models, with a primary focus on convolutional neural networks. The project serves as a CUDA accelerated deep learning library that offloads heavy mathematical operations to NVIDIA graphics hardware. This acceleration is used to increase processing speed and reduce execution time during the training of large networks. The engine supports a range of activities including deep learning research, image recognition development, and the training of convolutional neural ne

This project is a CUDA programming course and technical guide focused on writing and optimizing GPU kernels for hardware acceleration. It provides structured learning resources for using the CUDA platform to execute operations on silicon architectures. The material covers the optimization of linear algebra kernels and the analysis of machine learning deployment. It includes guidance on identifying acceleration tools, mapping the deep learning ecosystem, and evaluating the frameworks used to move models from research to production environments. The scope extends to GPU performance optimizatio

DeepXDE is a scientific machine learning library and deep learning PDE solver used to compute solutions for forward and inverse ordinary, partial, and integro-differential equations. It functions as a physics-informed neural network library that embeds physical laws and boundary conditions directly into the neural network loss function. The project provides a deep operator network framework for learning operator mappings that approximate relationships between functions in multiphysics problems. It is implemented as a multi-backend tensor library, allowing the system to switch between differen

Xarray is a Python multidimensional array library and labeled dataset framework. It extends the NumPy data structure by adding labels to arrays, allowing for the organization of complex N-dimensional data using named dimensions and coordinates. The library provides a NetCDF data interface for reading and writing scientific data formats such as NetCDF and Zarr. It enables scientific array computing by maintaining the relationship between data and physical coordinates during mathematical operations. The project covers multidimensional data analysis, geospatial data manipulation, and climate da

Gorgonia is a Go library that provides an automatic differentiation engine and a computation graph framework for building and training neural networks. It functions as a CUDA-accelerated tensor library and a SIMD-optimized math library, enabling machine learning workflows entirely within the Go ecosystem. The library distinguishes itself through a dual-backend architecture that dispatches neural network operations to either a GPU or CPU depending on CUDA availability at runtime. It constructs differentiable directed acyclic graphs of tensor operations, supports reverse-mode automatic gradient

Deeplearnjs is a JavaScript deep learning framework and automatic differentiation engine designed for building and training artificial intelligence models within a web browser environment. It functions as a machine learning library that leverages WebGL to provide hardware acceleration for neural networks. The project serves as a high-performance linear algebra library, using the GPU to execute operations on multi-dimensional arrays. This enables the implementation of deep learning models and the execution of client-side machine learning inference. The framework covers the complete automatic

Safetensors is a secure tensor serialization format and library designed for storing and distributing model weights. Its primary purpose is to provide a safe file format for machine learning tensors that prevents the execution of arbitrary or malicious code during the deserialization process. The project is distinguished by its use of zero-copy memory mapping, which reads data from disk directly into memory to minimize overhead. It enables cross-framework compatibility, allowing tensor data to be serialized and deserialized across different machine learning libraries. The system covers high-