Open-source alternatives to DeepEP

NCCL is a high-performance communication library and distributed GPU computing framework designed for executing collective and point-to-point data exchanges across multiple GPUs in single or multi-node systems. It serves as an RDMA GPU transport layer and memory orchestrator, facilitating high-bandwidth synchronization of data and model gradients for distributed GPU training and inference. The library is distinguished by its ability to execute communication primitives directly from GPU kernels, removing the host CPU from the critical path. It utilizes topology-aware path selection to optimize

Accelerate is a PyTorch distributed training library that abstracts the boilerplate required to run models across multiple GPUs, TPUs, and CPUs. It functions as a deep learning model scaler and distributed hardware orchestrator, allowing the same training script to run on different hardware backends without modifying the core logic. The project provides a distributed training command line interface for configuring compute environments and launching jobs across single or multi-node clusters. It includes a mixed precision training framework to implement FP16 and BF16 precision, reducing memory

JAX is a hardware-accelerated array library and automatic differentiation system for numerical computing. It provides a framework compatible with NumPy that extends array operations with a just-in-time compiler to transform Python functions into optimized kernels for execution on GPU and TPU accelerators. The system differentiates itself through the use of an XLA-based compiler and a single program multiple data sharding model. These capabilities allow the library to distribute large-scale computations across multiple hardware accelerators using both automatic parallelization and manual shard

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

Megatron-LM is a distributed transformer training library and large language model training framework designed to scale models across thousands of GPUs. It functions as a GPU-optimized deep learning toolkit and a scaling engine for mixture-of-experts architectures, enabling the training of models with hundreds of billions of parameters. The project implements multi-dimensional model parallelism, combining tensor, pipeline, data, expert, and context-based workload distribution. It specifically optimizes mixture-of-experts architectures through integrated memory and communication improvements t

PyTorch Lightning is a deep learning research framework that provides a structured environment for organizing machine learning code. It functions as a unified trainer orchestrator, centralizing the execution flow by managing the interaction between hardware resources, data loaders, and model components. By decoupling model architecture from training logic, the framework enables researchers to maintain clean, modular codebases that remain portable across different environments. The framework distinguishes itself through a hardware-agnostic abstraction layer that scales deep learning workloads

ColossalAI is a distributed deep learning framework designed for training and deploying massive artificial intelligence models across clusters of hardware accelerators. It functions as a parallel computing engine that partitions model workloads and data across multiple processors to maximize memory efficiency and throughput. The platform distinguishes itself through a comprehensive suite of parallelization strategies, including multi-dimensional tensor parallelism and pipeline-based model parallelism, which segment neural network layers and stages across devices. To support large-scale genera

DGL is a Python library for building and training graph neural networks. It functions as a graph message passing framework and a geometric deep learning tool, enabling the development of models that analyze graph-structured data. The library is designed for large-scale graph processing, utilizing distributed training and neighbor sampling to handle datasets with billions of edges. It provides specialized support for heterogeneous graph modeling, allowing for the representation of complex real-world entities with multiple node and edge types. Its capabilities cover a wide range of graph tasks

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

Dask is a parallel computing framework and distributed task scheduler designed to scale Python data science workflows from single machines to large clusters. It functions as a cluster resource manager that orchestrates computational logic by representing tasks and their dependencies as directed acyclic graphs. This architecture allows the system to automate the distribution of workloads across available hardware while managing complex execution requirements. The project distinguishes itself through a lazy evaluation engine that defers data operations until they are explicitly requested, enabl

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

A bidirectional pipeline parallelism algorithm for computation-communication overlap in DeepSeek V3/R1 training.

Expert Parallelism Load Balancer

This repository is the official documentation for TensorFlow, a machine learning framework. It provides comprehensive guides, tutorials, and API references for building, training, and deploying machine learning models. The documentation covers the full lifecycle of machine learning projects, from constructing data pipelines and building neural networks with high-level APIs to customizing training loops and deploying trained models in production, on edge devices, or in browsers. The documentation includes step-by-step tutorials for a range of tasks, including reinforcement learning, ranking mo

verl is a distributed training system designed for large language model alignment and reinforcement learning. It provides a framework for executing post-training pipelines, including supervised fine-tuning and reinforcement learning from human feedback, to refine model behavior and agentic capabilities. The system utilizes a hybrid training and inference engine that optimizes memory and communication when switching between model generation and gradient updates. It supports multi-modal reinforcement learning for models processing both image and text data, and implements algorithms such as PPO

OpenPi is a vision-language-action robot control framework designed to generate physical control actions for robotic systems. It functions as a distributed robot model trainer, a model format converter, and a robot action streaming server. The framework provides tools for transforming model checkpoints between different framework formats to ensure interoperability across various development environments. It also includes a server that uses websocket connections to stream model-generated control actions from remote inference servers to physical robot hardware in real-time. The system supports

PocketFlow is an integrated toolkit for deep learning model compression, distributed training, and mobile format optimization. It provides a system for reducing the size and complexity of neural networks to improve inference efficiency, featuring a dedicated engine for knowledge distillation and a mobile model optimizer. The framework differentiates itself through an automated hyperparameter tuning system that uses reinforcement learning and statistical models to determine optimal compression ratios and layer-wise bit allocation. It also includes a distributed training system that utilizes mu

Open-Instruct is a distributed training and instruction tuning framework for large language models. It functions as a coordinator for supervised fine-tuning, reinforcement learning from human feedback pipelines, and tool-use training, providing specialized roles for dataset curation and model alignment. The project distinguishes itself through a high-performance training architecture that utilizes actor-based distributed coordination and hybrid sharding to manage large GPU clusters. It implements advanced alignment techniques including direct preference optimization, group relative policy opt

DeepSpeed is a distributed deep learning optimization library and framework designed for the training and inference of massive AI models. It serves as a model parallelism orchestrator and a toolkit for scaling large language models across multiple GPUs and compute nodes. The project distinguishes itself through 3D parallelism orchestration, which combines data, pipeline, and tensor parallelism. It utilizes ZeRO-based memory partitioning to eliminate redundant storage and employs CPU-offload memory management to move weights and optimizer states to system RAM. Additionally, it provides special

This project is a JAX-based transformer framework and large language model trainer designed for building and training distributed models on TPU hardware accelerators. It provides a system for pretraining and fine-tuning autoregressive models by splitting weights and computations across a mesh of devices to reduce memory overhead and increase processing speed. The framework includes a TPU compute orchestrator for provisioning resources and automating dependency installation across remote distributed nodes. It also features a model weight converter capable of transforming and resharding checkpo

Swift is a toolkit for the full-parameter and parameter-efficient fine-tuning of large language and multimodal models. It functions as a multimodal model trainer for text, image, video, and audio data, and includes specialized tools for model compression and reinforcement learning from human feedback. The framework provides an alignment toolkit for optimizing model behavior using preference learning algorithms and reinforcement learning. It integrates parameter-efficient fine-tuning methods to adapt models with minimal memory and compute requirements, alongside utilities for reducing hardware

AReaL is a system for agent orchestration, distributed model training, and parameter-efficient tuning. It provides a framework for developing multi-turn reasoning agents and training large models using reinforcement learning from human feedback. The project implements a toolkit for improving the visual reasoning and geometry problem solving capabilities of vision-language models. It utilizes a memory-efficient tuning system to optimize mathematical and reasoning models across different inference backends. The infrastructure supports large-scale training through tensor, pipeline, and expert p

PyText is an extensible PyTorch-based framework for building, training, and deploying custom natural language processing models, including text classifiers, sequence taggers, and intent-slot predictors. It provides a modular toolkit that allows developers to assemble these models using pluggable registries for model architectures, data formats, and tensorizers, all configurable through YAML files without requiring code changes. The framework distinguishes itself through its comprehensive support for the full NLP model lifecycle, from training to production inference. It includes pre-built neu

NeMo is a comprehensive framework designed for the development, training, and deployment of large-scale conversational and generative artificial intelligence models. It provides an integrated platform for building multimodal systems, encompassing speech processing, language modeling, and reinforcement learning alignment. The framework is built to handle the entire lifecycle of AI development, from data curation and model pretraining to production-ready service deployment. The platform distinguishes itself through advanced distributed training capabilities, including tensor and pipeline parall

Metaseq is a transformer sequence modeling toolkit designed for training, fine-tuning, and deploying sequence-to-sequence models using open pre-trained weights. It provides a comprehensive framework for large language model training, including dedicated tools for sequence dataset processing and a standalone inference server for generating text via API requests. The project features specialized utilities for model quantization to reduce parameter precision to eight bits, which lowers memory usage and increases inference speed. It also includes a checkpoint conversion pipeline to transform mode

Scenic is a research framework designed for the development and training of deep learning models, with a specific focus on computer vision and multimodal transformer architectures. It provides a comprehensive toolkit for defining neural network structures, managing large-scale data pipelines, and executing training workflows across distributed hardware environments. The framework is built upon a functional programming paradigm that utilizes hardware-agnostic tensor abstractions and just-in-time compilation to maximize computational efficiency. By employing modular layer composition, it allows