Projects sharing features with Trlx

EasyR1 is a distributed model training system and reinforcement learning framework for large language and vision-language models. It functions as a multimodal trainer and an implementation of a Proximal Policy Optimization pipeline designed to refine the reasoning and perception capabilities of models that process both text and images. The system specializes in distributing reinforcement learning workloads across multiple compute nodes to manage high memory requirements. It optimizes hardware utilization through padding-free training and fine-tuning to fit large models onto available graphics

OpenRLHF is a training framework and alignment library designed for reinforcement learning from human feedback across distributed GPU clusters. It provides tools for aligning large language models and multimodal vision-language models using algorithms such as PPO, GRPO, and DPO. The framework distinguishes itself through a distributed inference engine that overlaps sample rollout with training to increase throughput. It supports scaling to models exceeding 70 billion parameters via parameter sharding and handles long-context sequences through ring-attention sequence parallelism. The project

This is a PyTorch implementation of reinforcement learning from human feedback designed to align large language models with human values and preferences. It provides a framework for the PaLM architecture and incorporates parameter-efficient fine-tuning to adapt models while minimizing the number of updated weights. The system enables the development of reward models that act as scoring mechanisms built from human preference data. These models evaluate generative outputs to guide the alignment process. The workflow covers policy optimization using a clipped objective, reward modeling based on

This project is a framework for aligning large language models with human preferences. It provides a library for optimizing model behavior by mapping preference data directly to a policy objective, bypassing the need for a separate reward model. The framework utilizes a closed-form mathematical objective to adjust model weights by maximizing the likelihood of preferred responses relative to dispreferred ones. It includes a toolkit for supervised fine-tuning, allowing users to establish a stable baseline on demonstration data before applying preference-based optimization techniques. The libra

This project provides a comprehensive technical guide and framework for engineering large-scale machine learning systems. It covers the full lifecycle of model development, focusing on the infrastructure and computational principles required to build, train, and serve generative AI models across distributed GPU clusters. The repository distinguishes itself by offering deep-dive tutorials and implementation strategies for complex system challenges. It emphasizes high-performance architectural primitives, such as collective communication orchestration, distributed tensor sharding, and static gr

This library provides a comprehensive framework for fine-tuning, aligning, and distilling transformer-based language models. It serves as a toolkit for adapting models to specialized domains through supervised learning, while offering advanced methodologies to improve output quality and reasoning capabilities. The project distinguishes itself through specialized alignment and optimization techniques, including direct preference optimization and reinforcement learning, which allow models to be tuned against human preferences without complex reward modeling. It further supports training efficie

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

This project is an alignment framework and suite of pipelines for training language models using supervised fine-tuning and preference optimization. It provides tools for executing large-scale distributed training across multiple GPUs and compute nodes, alongside a system for measuring model helpfulness and dialogue quality through single-turn and multi-turn benchmarks. The framework includes specialized tools for direct preference optimization to refine model behavior using paired data without a separate reward model. It also supports constitutional AI alignment and the training of reward mo

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

Tinker Cookbook is an open-source framework for fine-tuning large language models, supporting supervised learning, reinforcement learning, and parameter-efficient techniques like LoRA adapters. It provides a complete pipeline for aligning models with human preferences through multi-stage RLHF workflows, from supervised fine-tuning through preference optimization to reinforcement learning. The framework distinguishes itself through recipe-based training orchestration, where fine-tuning workflows are defined as composable recipe files that chain data loading, model configuration, and training l

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

Pythia is a multimodal research framework and distributed training system designed for building, training, and evaluating large models that combine visual and linguistic data. It provides a modular environment for developing vision-language models, focusing on the integration of image and text inputs into shared feature representations. The framework utilizes a modular architecture that decouples model building blocks into interchangeable components, allowing for flexible configuration of vision and language modules. It includes a benchmark suite for executing reference models against standar

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 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

This project is a comprehensive toolkit designed for the full lifecycle management of large language and multimodal models. It functions as a unified orchestrator that handles the entire development process, ranging from dataset preparation and supervised fine-tuning to advanced reinforcement learning alignment and production-ready inference deployment. The platform distinguishes itself through a specialized reinforcement learning library that supports complex optimization algorithms, including group relative policy optimization and leave-one-out techniques, to improve model instruction-follo

This project is a language model finetuning framework designed to adapt large language models to specific datasets using supervised fine-tuning and low-rank adaptation. It serves as a distributed training manager that coordinates workloads and synchronizes gradients across multiple processing units to scale performance. The framework includes a specialized toolkit for low-rank adaptation to update a subset of model weights, reducing memory and hardware requirements. It provides capabilities for instruction fine-tuning, domain adaptation, and the optimization of function calling to improve how

MedicalGPT is an open-source framework for fine-tuning large language models, with a dedicated focus on adapting general models to the medical domain. It provides a complete pipeline that covers continued pretraining on domain-specific corpora, supervised instruction tuning, tokenizer vocabulary extension with medical terminology, and alignment to clinician preferences through direct preference optimization, reinforcement learning, or knowledge distillation. The framework also supports training models to invoke external tools and functions in multi-turn clinical conversations. The platform di

Chainer is an open-source deep learning framework built around define-by-run automatic differentiation, where computation graphs are constructed dynamically during forward execution. This imperative approach allows networks to be built using standard Python control flow, with gradients computed automatically through reverse-mode differentiation on the dynamically recorded graph. The framework supports GPU acceleration through a NumPy-compatible array backend with CUDA and cuDNN support, and provides a pluggable device abstraction that lets users switch between CPU and GPU computation without c

This project is a training pipeline and framework for developing Chinese language models based on the Llama 2 architecture. It functions as a distributed GPU trainer and dataset preprocessing toolkit designed for both the initial pre-training of baseline models and subsequent supervised fine-tuning. The system distinguishes itself through a specialized workflow for Chinese text, incorporating a data curation pipeline that uses similarity hashing for deduplication and a tokenization process that converts raw text into memory-mapped binary files for efficient disk access. It implements a superv

This project provides a comprehensive collection of educational resources and technical guides for training, fine-tuning, and deploying machine learning models using PyTorch and Hugging Face. It serves as a practical reference for scaling deep learning workflows, offering structured instructions for managing large-scale architectures across distributed hardware accelerators. The repository distinguishes itself by focusing on the end-to-end lifecycle of large language models, specifically emphasizing containerized deployment and performance optimization. It details workflows for parameter-effi

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

Sana is a framework for high-resolution image and video synthesis based on a linear diffusion transformer. It provides a toolkit for the training, fine-tuning, and execution of text-to-image and text-to-video models, as well as a video generative world model capable of simulating physical environments with precise spatial control. The project is distinguished by its use of linear complexity layers to handle high resolutions and its support for long-form, minute-length video generation in real time. It implements a two-stage inference paradigm that separates structural generation from visual t

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

LMFlow is a comprehensive suite for large language model fine-tuning, context extension, multimodal processing, and inference execution. It provides a toolkit for updating model parameters through full tuning or memory-efficient adapter algorithms, alongside an inference engine for executing tuned models via command-line or web-based interfaces. The framework includes a dedicated alignment suite for supervised tuning and reward model training to refine model behavior. It features a context window extender to increase maximum input lengths and a multimodal framework for building chatbots that

This project is a distributed machine learning platform and sparse deep learning framework designed for training and serving models with high-dimensional sparse data. It functions as an online model serving infrastructure and recommendation system engine, enabling real-time item retrieval and scoring using deep tree matching and neural networks. The system distinguishes itself through a multi-task learning framework that optimizes multiple objective functions within a shared representation space. It features a specialized online serving infrastructure that supports dynamic model hot-loading a

This project is a PyTorch implementation of 3D residual networks designed for video action recognition. It provides a spatiotemporal architecture that analyzes both spatial frames and temporal motion to classify human activities within video clips. The system includes a distributed model training framework to accelerate learning across multiple compute nodes. It supports the deployment and fine-tuning of pre-trained model weights, allowing the adaptation of existing networks to specific new datasets. The codebase covers the full pipeline for spatiotemporal learning, including video dataset p