Tag
Explore by tags
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.
Converts document images—scans, photos, born-digital PDFs—into structured text in two stages: first map layout and reading order, then parse each element (text, tables, formulas, figures) in parallel, each guided by its own task prompt.
A PyTorch DTensor-native SPMD library for training and fine-tuning LLMs, VLMs, diffusion and retrieval models. Integrates with Hugging Face for day-0 model support, provides YAML-driven recipes, DTensor/FSDP2 parallelism and NVIDIA-optimized kernels (Transformer Engine, DeepEP, FlexAttn).
Enables bidirectional checkpoint conversion between Hugging Face and Megatron formats and provides a PyTorch-native training library with tensor/pipeline parallelism, FP8/BF16 mixed precision, SFT and PEFT (LoRA) support for large and multimodal models.
Physics-aware simulated sensor dataset for training and evaluating autonomous-vehicle perception and control models. Includes multimodal sensor streams with physical-scene annotations intended for tasks that require grounding in real-world dynamics.
Turns commodity WiFi Channel State Information into spatial sensing: 17-keypoint pose estimation, presence detection, and contactless breathing/heart-rate monitoring through walls, with no camera. Runs on a mesh of ESP32-S3 nodes (~$9 each).
Reimplements the vLLM inference engine from scratch in ~1,200 lines of readable Python, matching its offline throughput on small models. Prefix caching, tensor parallelism, torch.compile, and CUDA graphs are all kept legible.
Provides a unified Python interface to collect data, train visual/dynamics world models, and evaluate them with model-predictive control across many standardized environments. Includes reference baselines, planning solvers, dataset converters, and LanceDB-backed formats for reproducible experiments. Best suited for researchers benchmarking world-model algorithms.
Forecasts financial candlesticks (OHLCV K-lines) with a decoder-only transformer pre-trained on 12B+ records from 45 exchanges. A tokenizer turns market data into discrete tokens, enabling price/volatility forecasting and synthetic K-line generation.
Model-compression toolkit for large LLMs/VLMs that integrates quantization (FP8/INT4/etc.), speculative decoding, token pruning and deployment hooks—designed for end-to-end performance on single/multi-GPU inference workflows and research-to-prod model optimization.
Runs named-entity recognition, text classification, structured-JSON parsing and relation extraction from one 205M-parameter encoder in a single CPU forward pass, using schemas with per-field regex validators. A larger 1B model is available via API.
An agentic framework that analyzes, plans, and executes multi-step video understanding and editing workflows using multimodal LLM-driven agents—features intent decomposition, graph-based workflow orchestration, and automated shot planning for long-form video tasks.