The agent that grows with you

OpenClaw-RL: Train any agent simply by talking

A PyTorch native platform for training generative AI models

Ongoing research training transformer models at scale

A library for accelerating Transformer models on NVIDIA GPUs, including using 8-bit and 4-bit floating point (FP8 and FP4) precision on Hopper, Ada and Blackwell GPUs, to provide better performance…

A safetensors extension to efficiently store sparse quantized tensors on disk

Official PyTorch implementation for "Large Language Diffusion Models"

[FSE'2026] SWE-Factory: Your Automated Factory for Issue Resolution Training Data and Evaluation Benchmarks

verl-agent is an extension of veRL, designed for training LLM/VLM agents via RL. verl-agent is also the official code for paper "Group-in-Group Policy Optimization for LLM Agent Training"

[ICLR 2025] COAT: Compressing Optimizer States and Activation for Memory-Efficient FP8 Training

PyTorch native quantization for training and inference

A collection of research papers on low-precision training methods

FP16xINT4 LLM inference kernel that can achieve near-ideal ~4x speedups up to medium batchsizes of 16-32 tokens.

Efficient Triton Kernels for LLM Training

Bridge Megatron-Core to Hugging Face/Reinforcement Learning

verl/HybridFlow: A Flexible and Efficient RL Post-Training Framework

Community maintained hardware plugin for vLLM on Ascend

A high-throughput and memory-efficient inference and serving engine for LLMs

A real browser preview inside your editor that you can debug.

TinyNeuralNetwork is an efficient and easy-to-use deep learning model compression framework.

TIGER: Time-frequency Interleaved Gain Extraction and Reconstruction for Efficient Speech Separation

Pytorch implementation of "CleanMel: Mel-Spectrogram Enhancement for Improving Both Speech Quality and ASR".

Development repository for the Triton language and compiler

Fast and memory-efficient exact attention

🐳 Efficient Triton implementations for "Native Sparse Attention: Hardware-Aligned and Natively Trainable Sparse Attention"

MoBA: Mixture of Block Attention for Long-Context LLMs

Code for NeurIPS 2024 paper - The GAN is dead; long live the GAN! A Modern Baseline GAN - by Huang et al.

Several simple examples for popular neural network toolkits calling custom CUDA operators.

Homogram is a 3rd-party Telegram client for HarmonyOS 5, driven by ArkTS/ArkUI (UI-layer) and Rust (native-layer).