Awesome GitHub RepositoriesTraining Throughput Optimization

Axolotl is a distributed training orchestrator and fine-tuning framework for large language models, multimodal systems, and quantized models. It provides a structured environment for specializing pre-trained models through full parameter updates or low-rank adaptation, as well as aligning model outputs with human expectations via preference tuning pipelines and reward modeling. The system distinguishes itself through a configuration-driven pipeline that manages preprocessing and training workflows via a single file for reproducibility. It implements high-throughput optimizations such as multi

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

ERNIE is a development toolkit for training, fine-tuning, and deploying large language models built on the PaddlePaddle deep learning platform. It provides a comprehensive suite of core components, including an inference server for vision and language models, a training and fine-tuning toolkit, and a framework for building retrieval-augmented generation systems using private knowledge bases. The project features multimodal AI models capable of reasoning across text, images, and video to perform complex visual understanding and information extraction. It distinguishes itself through specialize

LARK is a development toolkit for training, fine-tuning, and deploying large language models and multimodal models based on PaddlePaddle. It functions as a comprehensive framework that includes an LLM training orchestrator, an inference server, and a multimodal model framework for processing text, image, and video inputs. The project features a retrieval-augmented generation system for building conversational applications that integrate web search and private knowledge bases. It provides specific capabilities for multimodal reasoning and complex logic, enabling the extraction of structured da

gpt-neox is a distributed training system and framework for building large-scale autoregressive language models. It implements the transformer architecture and provides a toolkit for training models with billions of parameters by distributing weights across compute clusters. The framework distinguishes itself through extensive support for distributed model parallelism, including pipeline and sequence parallelism, to overcome single-device memory limits. It further supports sparse model architectures using a mixture of experts system with Sinkhorn-based routing. The project covers a broad ran

xtuner este un motor de antrenare cuprinzător pentru modele de limbaj mari (LLM), oferind un toolkit pentru pre-antrenare, fine-tuning supervizat și optimizarea modelelor multimodale vision-language. Servește ca un accelerator de antrenare distribuită și un framework specializat pentru scalarea modelelor Mixture-of-Experts și alinierea comportamentului modelului prin învățare prin consolidare din feedback uman (RLHF). Proiectul se distinge prin optimizări avansate de memorie și calcul, cum ar fi paralelismul de secvență pentru ferestre de context ultra-lungi și paralelismul de pipeline intercalat pentru a reduce timpul de inactivitate al GPU-ului. Oferă o suită dedicată pentru optimizarea preferințelor, implementând tehnici precum Group Relative Policy Optimization și Direct Preference Optimization pentru a rafina politicile modelului și sistemele de recompensă. Zonele largi de capabilități acoperă antrenarea distribuită a modelelor pe mai multe noduri, pregătirea seturilor de date multimodale și gestionarea fine-tuning-ului bazat pe adaptoare. Motorul include, de asemenea, instrumente pentru evaluarea modelului, fuziunea ponderilor (weight merging) și exportul parametrilor antrenați către motoarele de inferență. Antrenarea este gestionată prin fișiere de configurare standardizate și launchere distribuite pentru a asigura rezultate consistente pe clusterele de calcul.

warp-ctc este o bibliotecă de înaltă performanță pentru calcularea pierderii de clasificare temporală conexionistă (CTC) pentru antrenarea modelelor de deep learning secvență-la-secvență. Oferă un strat de stabilitate numerică folosind calculul în spațiu logaritmic pentru a preveni subdepășirea (underflow) și erorile de precizie în timpul calculelor de probabilitate pentru secvențe lungi. Biblioteca utilizează nuclee accelerate hardware pentru a calcula pierderea în paralel pe arhitecturi CPU și GPU. Se concentrează pe creșterea throughput-ului de antrenare prin optimizarea pașilor de programare dinamică ai algoritmului CTC. Aceste capabilități susțin antrenarea modelelor pentru recunoașterea vorbirii, recunoașterea optică a caracterelor scrise de mână și maparea generală secvență-la-secvență. Proiectul include integrare pentru calcularea pierderii și antrenarea fără aliniere în TensorFlow.

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

RLinf is a distributed reinforcement learning orchestrator and embodied AI training framework. It provides the infrastructure to train vision-language-action models and robotic policies using a combination of reinforcement learning and supervised fine-tuning. The system is designed for scaling workloads across GPU clusters, managing the placement of actors, rollout workers, and environment components. It features a specialized robotics data collection pipeline for gathering teleoperated demonstrations and simulation trajectories into standardized replay buffers, alongside a hardware interface