AI Train2017
Sequence modeling toolkit for training custom models for translation, summarization, and language modeling. Reference implementation behind RoBERTa, BART, mBART, XLM-R, and wav2vec 2.0, with multi-GPU and mixed-precision training.
AI Infra2017
Defines a portable model format and operator set for moving trained machine learning models across frameworks, runtimes, and hardware targets without locking the model to one toolchain.
AI Image2019
Modular implementations of object detection, instance/semantic/panoptic segmentation and related vision models for research and deployment. Offers a large model zoo, export to TorchScript/Caffe2, and PyTorch-native optimizations for faster training and extensibility.
AI Infra2021
Drop-in transformer building blocks with custom CUDA kernels: memory-efficient exact attention (up to ~10x faster), block-sparse attention, fused softmax/layernorm/SwiGLU ops. Cuts VRAM and speeds up diffusion and LLM training on Nvidia GPUs.
AI Deploy2022
Deploys PyTorch models directly on phones, microcontrollers, and embedded hardware via ahead-of-time compilation to a ~50KB C++ runtime. Delegates subgraphs to 12+ backends (XNNPACK, CoreML, Qualcomm, ARM Ethos-U) with torchao quantization.
AI Infra2023
Traces how Transformer LLMs route information from input to output, attributing each block's effect to individual attention heads and feed-forward neurons. Click any edge to see what a head promotes or suppresses in vocabulary space.
AI Coding2025
Performs fast static type checking and provides a language server with code navigation, semantic highlighting, and completions for Python. Processes ~1.85M lines/sec and completes IDE rechecks typically under 10ms — intended for responsive editor workflows and large codebases.
AI Video2025
Provides PyTorch code, pretrained checkpoints, and evaluation tooling for V-JEPA 2 — a Meta FAIR family of self-supervised video encoders and an action-conditioned world model. Includes training recipes, HuggingFace checkpoints, evaluation probes, and robot post‑training artifacts.
AI Image2025
Detects, segments, and tracks every instance of an open-vocabulary concept in images and video from a text phrase or visual exemplar, not just one object per prompt. An 848M-param model reaching ~75-80% of human accuracy across 270K concepts.
AI Image2025
Self-supervised vision foundation model producing dense, patch-level features that transfer to classification, segmentation, depth, and detection with a frozen backbone. Spans ViT-S (21M) to ViT-7B (6.7B params), plus ConvNeXt and satellite variants.
AI Audio2025
Isolates any single sound from a complex audio mixture using a text description, a visual cue from a video frame, or a time span, returning both the isolated target and the residual. Released in small, base, and large sizes plus visual-prompt variants.
AI Infra2025
Provides a Gymnasium-style API and tooling to create, deploy, and interact with isolated execution environments for agentic RL training. Includes async/sync clients, a web interface, CLI, Docker-based deployment, and Hugging Face Spaces integration.