Tag

Provides research-grade implementations and pretrained models for sequence tasks (translation, LM, speech). Offers multi-GPU training, fast generation (beam/sampling/lexical constraints), mixed-precision, and state sharding — aimed at researchers reproducing or extending papers.

ONNX (Open Neural Network Exchange) is an open ecosystem that provides an open source format for AI models, including deep learning and traditional ML. It defines an extensible computation graph model, built-in operators, and standard data types, focusing on inferencing capabilities. Widely supported across frameworks and hardware, it enables interoperability and accelerates AI innovation.

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.

xFormers is a modular, hackable, and efficiency-oriented Transformer building-blocks library from Facebook Research (Meta). It provides multiple attention implementations (memory-efficient exact attention, sparse and block-sparse attention), fused operators, and custom CUDA kernels to speed up transformer research and prototyping within the PyTorch ecosystem.

Enables on-device inference for PyTorch models—export, optimize, and run LLMs, vision, and speech models across mobile, embedded, and edge devices with a tiny runtime and multiple hardware backends.

LLM Transparency Tool (LLM-TT) is an open-source interactive toolkit from Facebook Research for analyzing the internal workings of Transformer-based language models. It lets users run inferences, explore contribution graphs tied to selected tokens, inspect representations after any block, and drill down to attention heads, FFN blocks and individual neurons to see how they promote or suppress output tokens. The project provides Docker and local installation instructions and a live demo hosted on Hugging Face.

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.

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.

Promptable image/video segmentation that finds, segments, and tracks all instances of an open‑vocabulary concept from short text or exemplars. Key differences: an automated 4M‑concept data engine, a presence token for finer text discrimination, and a detector–tracker design. Checkpoints on Hugging Face; GPU required.

DINOv3 is Meta AI Research's reference PyTorch implementation and model collection for a family of self-supervised vision foundation models. It provides high-resolution dense patch features, multiple pretrained backbones (ViT and ConvNeXt variants), pretrained heads for classification/detection/segmentation/depth, and integration examples for PyTorch Hub and Hugging Face. The repo contains training and evaluation scripts, notebooks, and instructions to obtain model weights.

SAM-Audio (Segment Anything in Audio) is a Meta / Facebook Research foundation model and GitHub project for isolating target sounds from complex audio mixtures using text, visual, or temporal prompts. The repository provides inference code, example notebooks, links to pretrained checkpoints on Hugging Face, and evaluation/reranking tools (Judge, CLAP). Models are released in multiple sizes (small, base, large) and variants optimized for visual prompting.

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.