AI Train2019
Trains transformer models from 2B to 462B parameters across thousands of GPUs by combining tensor, pipeline, context, and expert parallelism. Ships composable building blocks (Megatron Core) plus reference scripts, with FP8/FP4 and ~47% MFU on H100s.
AI Train2020
thu-ml (Tsinghua University Machine Learning Group)Tsinghua University, appliedAI Institute for Europe
Deep reinforcement learning library on pure PyTorch and Gymnasium, with 30+ algorithms across on-policy, off-policy, and offline RL. Exposes both a one-call high-level interface and a procedural API, plus vectorized envs and reproducible MuJoCo benchmarks.
AI Infra2020
Optimizes distributed PyTorch training and inference for very large models with ZeRO memory partitioning, parallelism, MoE, offload, and compression. Best when GPU memory, training cost, or cluster throughput is the bottleneck.
AI Train2020
PyTorch library for operator learning: neural networks that map between whole function spaces, not fixed grids, so a model trained at one resolution runs at any other. Bundles FNO, Tensorized FNO and related architectures, mainly for solving PDEs.
AI Infra2020
Sits between PyTorch and micrograd: eager tensors with autograd plus a small, fully hackable compiler that fuses operations into kernels. Adding a new accelerator backend takes about 25 low-level ops, so it runs on CUDA, Metal, AMD, and WebGPU.
AI Infra2021
Runs, manages, and scales AI workloads across 20+ clouds, Kubernetes, Slurm, and on-prem from one YAML or Python spec. Auto-provisions GPUs/TPUs, fails over across regions and providers when capacity is short, and routes jobs to the cheapest option.
AI Train2021
HPC-AI Technology Inc. (Colossal-AI team), Shenggui Li +1HPC-AI Technology Inc., National University of Singapore
Scales a single-GPU training script to thousands of GPUs through a unified interface, combining data, pipeline, tensor, and sequence parallelism. Its Gemini memory manager offloads tensors across GPU, CPU, and NVMe so models far larger than VRAM still fit.
MLOps2021
Orchestrates ML training pipelines and production agent workflows from one Python codebase that runs unchanged from a laptop to Kubernetes or any cloud. Auto-versions artifacts, models, and agent checkpoints, with no orchestrator or framework lock-in.
AI Infra2022
Fused CUDA kernels that compute exact attention without ever writing the full N×N score matrix to GPU memory, cutting memory from quadratic to linear and speeding up training and inference on A100/H100. Ships FlashAttention-2/3 plus KV-cache decode paths.
AI Image2022
Unifies successive YOLO generations — YOLOv8, YOLO11, YOLOv3 and newer — under one package and a single `YOLO` API spanning detection, segmentation, classification, pose, oriented boxes and tracking, plus one-line export to ONNX, TensorRT and CoreML.
AI Train2022
Reproduces GPT-2 (124M) from scratch on OpenWebText in ~4 days on an 8xA100 node, with the whole stack kept to two ~300-line files: train.py for the loop and model.py for the architecture. A char-level Shakespeare run finishes in ~3 minutes on one GPU.
AI Train2023
Modular PyTorch-based framework for building, training, and deploying physics-informed ML models (neural operators, PINNs, GNNs, diffusion). Provides GPU‑optimized training, domain-specific datapipes for meshes/point clouds, distributed scaling and a model zoo.