Open-source alternatives to RLinf

Apex is a high-performance toolkit for PyTorch designed to coordinate distributed training, execute fused GPU kernels, manage mixed precision, and implement optimized distributed optimizers. It provides specialized tools for scaling model training across multiple GPUs and nodes to increase processing speed and throughput. The library features high-performance implementations of Adam and LAMB optimizers to reduce synchronization overhead and memory bottlenecks. It utilizes fused CUDA kernels to combine neural network operations, reducing memory overhead and increasing execution speed. The too

NNI is an AutoML toolkit designed to automate machine learning lifecycles. It functions as a hyperparameter optimization framework, a neural architecture search tool, and a model compression suite. The project provides a distributed training orchestrator to manage machine learning workloads across local machines, remote servers, and cloud platforms. It enables the discovery of efficient model structures through reinforcement learning and one-shot optimization methods, while utilizing Bayesian and evolutionary algorithms to automate hyperparameter tuning. Additional capabilities include tools

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

Higgsfield is a distributed machine learning training framework and GPU cluster orchestrator designed for scaling neural networks with billions of parameters. It functions as a large model sharding system and a containerized deployment tool to manage computational workflows across heterogeneous compute resources. The platform provides a centralized interface for experiment management, enabling the monitoring of real-time telemetry, performance metrics, and logs. It ensures reproducible results by using container isolation to standardize dependencies across different computing environments. T

IsaacGymEnvs is a GPU-accelerated physics sandbox and robotics policy training suite designed for reinforcement learning. It serves as a vectorized robotic simulator that runs thousands of parallel environments on GPUs to accelerate the training of neural networks. The project provides a sim-to-real transfer framework that utilizes domain randomization and physics variations to ensure policies trained in simulation are robust enough for deployment on real hardware. It distinguishes itself through a high-performance architecture that uses tensor-based state management to handle observations an

This project is a research framework and toolkit designed for training large-scale vision transformers and multimodal language models. It provides a comprehensive suite for vision-language pretraining, enabling the development of models that map images and text into shared latent spaces. The framework is distinguished by its capabilities in high-fidelity image generation and multimodal research, utilizing normalizing flows and variational autoencoders to produce images from text prompts or class labels. It supports the development of both generative and contrastive models, allowing for a wide

Search-R1 is a distributed training system and reinforcement learning framework designed to create search-augmented language models. It provides an architecture for scaling model workloads across head and worker nodes while optimizing how models interleave internal reasoning with external tool calls. The system focuses on refining model behavior through custom reward signals and reinforcement learning to improve tool-use formatting and information retrieval. It implements an interleaved reasoning-search loop that allows models to alternate between internal thought generation and external data

Torchtune is a PyTorch-native library for fine-tuning, aligning, and quantizing large language models. It provides a config-driven system for instantiating components, orchestrating distributed training, and managing parameter-efficient fine-tuning with quantization support, all through YAML-based configurations and command-line overrides. The library distinguishes itself through its comprehensive post-training workflow orchestration, combining supervised fine-tuning, preference optimization (DPO, PPO, GRPO), knowledge distillation, and quantization-aware training in a single configurable pip

rllm is an asynchronous reinforcement learning framework for training language agents. It provides a unified pipeline that runs the same agent code for both evaluation and training, automatically capturing traces for gradient computation. The framework supports distributed reinforcement learning across multiple GPUs and nodes using pluggable backends, and executes agents in isolated sandboxes—either locally or in the cloud—for safe and scalable rollout collection. It trains agents built with LangGraph, SmolAgents, OpenAI Agents SDK, or custom frameworks without requiring core logic changes. T

llm-foundry is a training framework for large language models, providing a system for foundation model pre-training and supervised fine-tuning. It includes a distributed trainer for scaling workloads across multiple nodes and GPUs, a dataset streaming pipeline for loading data from cloud storage, and a parameter-efficient fine-tuning implementation. The framework distinguishes itself through its use of parameter sharding and high-throughput data streaming to maintain stability during large-scale training. It incorporates low-rank adaptation to reduce computational costs and uses eight-bit flo

OpenManus-RL is a reinforcement learning framework and distributed training pipeline designed to train large language models as agents. It serves as an agentic reasoning optimizer and reward model trainer, providing the infrastructure to improve model decision-making through reward-based policy optimization. The project distinguishes itself through a distributed architecture that supports parameter sharding across multiple compute nodes and a coordinated rollout system for collecting interaction trajectories. It incorporates advanced reasoning strategies, such as Tree-of-Thoughts and Monte Ca

Axolotl is a configuration-driven framework designed for the fine-tuning, evaluation, and quantization of large language models. It functions as a comprehensive orchestrator for distributed training, enabling users to manage complex workflows across multi-node and multi-GPU environments. By utilizing structured configuration files, the platform streamlines the setup of training parameters, dataset paths, and hardware distribution strategies. The project distinguishes itself through its support for diverse training methodologies, including full-parameter tuning, parameter-efficient adaptation,

Baselines is a comprehensive suite of frameworks for reinforcement learning algorithm implementation, imitation learning, and training orchestration. It provides a library of standardized learning algorithms used to benchmark and replicate research results, alongside a deep learning policy framework for constructing neural network architectures such as multi-layer perceptrons, convolutional networks, and long short-term memory networks. The project includes a specialized imitation learning toolkit that enables agents to mimic expert behavior through behavior cloning and generative adversarial

ManiSkill is a GPU-accelerated robot simulation framework designed for training robotic manipulation skills, benchmarking learning algorithms, and generating synthetic datasets. It serves as a reinforcement learning environment where robot control policies can be developed and evaluated using parallelized physics and rendering on the GPU. The platform is distinguished by its ability to perform sim-to-real transfer, allowing policies trained in virtual environments to be deployed onto physical robotic hardware. It features ray-traced parallel rendering for producing high-frame-rate RGBD and se

ZenML is an extensible machine learning orchestration framework designed to manage the end-to-end lifecycle of data pipelines and AI agent workflows. It functions as a durable orchestrator that executes machine learning tasks as directed acyclic graphs, ensuring that every step is containerized for consistent performance across local, cloud, and hybrid infrastructure. By decoupling pipeline code from underlying compute and storage backends, the platform allows developers to define infrastructure-agnostic stacks that remain portable across diverse environments. The project distinguishes itself

Diffusion Policy is a robot learning framework that uses diffusion models to map visual observations to precise action trajectories. It functions as an imitation learning toolkit and visuomotor policy learner, providing a system to train neural networks that replicate human behavior by generating robotic movements based on image and sensor data. The framework employs a conditional denoising process to sample sequences of robotic movements, allowing it to handle multimodal action distributions where multiple valid trajectories may exist for a single state. It utilizes score-based action modeli

Acme is a reinforcement learning framework and execution engine designed for developing and benchmarking learning algorithms. It provides a library of modular components and reference implementations used to construct agents and establish performance baselines. The system enables the scaling of agent architectures from single-stream execution to large distributed environments. This allows for the transition from initial prototyping to distributed execution for training and evaluation. The framework covers reinforcement learning development and agent architecture prototyping, providing the bu

s1 is a reasoning training framework and GPU cluster orchestrator designed to build and refine large language models. It provides a system for executing supervised fine-tuning on distributed hardware, utilizing gradient checkpointing and hardware optimization to improve model reasoning. The project features a synthetic data generator and dataset builder that produce high-quality training sets. This workflow collects questions, generates model reasoning traces, and applies automated grading loops to filter for correct answers. The framework includes an evaluation suite to compute accuracy and

iii is a distributed service orchestrator and event-driven workflow engine designed to compose and manage cross-language functions and workers through a central execution engine. It functions as a multi-language service mesh and WebSocket service gateway, providing a persistent communication layer for remote service workers. The platform enables dynamic runtime extensions, allowing new workers and capabilities to be deployed and registered into a live environment without requiring system restarts. It distinguishes itself by offering machine-readable skill exposure and agent capability integra

OpenVLA is a vision-language-action model and framework designed for general-purpose robotic manipulation. It provides a robotic policy training framework and a control inference engine that map visual and textual inputs to robotic control actions, enabling zero-shot instruction following on hardware. The project includes a robotics dataset pipeline for standardizing diverse trajectory data and managing dataset mixtures. It supports large-scale model training through distributed GPU compute and sharded data parallelism, alongside parameter-efficient adaptation for fine-tuning models to new ta

This project is a distributed training infrastructure designed for aligning large language models through reinforcement learning. It functions as an end-to-end engine for complex alignment tasks, including proximal policy optimization, direct preference optimization, and iterative self-play. By providing a unified framework for multi-turn interactions and tool-use scenarios, it enables the development of models capable of reasoning and external environment engagement. The framework distinguishes itself through a decoupled architecture that separates model training from sample generation. This

This project is a comprehensive engineering framework and technical reference for managing, scaling, and optimizing distributed machine learning infrastructure. It provides a suite of methodologies and diagnostic tools designed to support large-scale model training and inference on high-performance computing clusters. The project distinguishes itself through a specialized diagnostic toolkit and infrastructure optimization suite that addresses the complexities of multi-node environments. It enables precise control over cluster resources, including hardware maintenance, network topology configu

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

Horizon is a reinforcement learning platform designed for training, evaluating, and deploying agents and contextual bandits using historical data. It serves as an off-policy engine and offline policy evaluation tool, allowing decision-making policies to be optimized and tested without the need for a live simulator. The framework specializes in recommendation system optimization, specifically using slating-based reinforcement learning to optimize the ordering and sequencing of multiple recommendations. It also functions as a contextual bandit framework that manages the balance between explorat

This project is a Python workflow orchestration platform and programmatic data pipeline engine used to author, schedule, and monitor complex data pipelines. It functions as a directed acyclic graph manager and scheduler, allowing users to define data movement and transformation tasks as code to ensure precise execution order and maintainability. The platform distinguishes itself by treating workflows as code, enabling pipelines to be versioned and tested through a standard programming language. It utilizes a system of extensible operators to encapsulate integration logic and employs a templat

AI-System is an educational resource and toolkit designed for learning the hardware and software foundations of deep learning systems. It provides a curriculum and practical exercises for building AI infrastructure, ranging from low-level CUDA kernel development to high-level system management. The project includes a toolkit for developing tensor operations and optimizing GPU performance through direct hardware programming. It also features a framework for distributed training, focusing on resource scheduling and communication protocols to manage large-scale models across multiple computing n