A NCCL extension library, designed to efficiently offload GPU memory allocated by the NCCL communication library.

A construction kit for reinforcement learning environment management.

Checkpoint-engine is a simple middleware to update model weights in LLM inference engines

PyTorch Single Controller

A modular, primitive-first, python-first PyTorch library for Reinforcement Learning.

PyTorch-native post-training at scale

Contexts Optical Compression

The best ChatGPT that $100 can buy.

A lightweight, powerful framework for multi-agent workflows

Official Repo For "Sa2VA: Marrying SAM2 with LLaVA for Dense Grounded Understanding of Images and Videos"

Fully Open Framework for Democratized Multimodal Training

PyTorch bindings for CUTLASS and CUBLAS Grouped GEMM, Permute and Unpermute.

AdaptiveGEMM: FP8 GEMM with Adaptation to Various Lengths of Group M

A debugging and profiling tool that can trace and visualize python code execution

how to optimize some algorithm in cuda.

Reference PyTorch implementation and models for DINOv3

Implementation for FP8/INT8 Rollout for RL training without performence drop.

VeOmni: Scaling Any Modality Model Training with Model-Centric Distributed Recipe Zoo

gpt-oss-120b and gpt-oss-20b are two open-weight language models by OpenAI

Scalable toolkit for efficient model reinforcement

(best/better) practices of megatron on veRL and tuning guide

SkyRL: A Modular Full-stack RL Library for LLMs

siiRL: Shanghai Innovation Institute RL Framework for Advanced LLMs and Multi-Agent Systems

A Scientific Multimodal Foundation Model

Train speculative decoding models effortlessly and port them smoothly to SGLang serving.