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.
AI Deploy2023
Open-source LLM inference and serving engine built around PagedAttention, which manages the KV cache like OS virtual memory to cut waste and raise throughput. Supports continuous batching, KV cache sharing, quantization, and an OpenAI-compatible API.
AI Audio2023
Web UI to train and run retrieval-based voice conversion models from small datasets (≈10 minutes), featuring top-1 feature retrieval to avoid timbre leakage, model fusion, real-time conversion, vocal separation, and multi-hardware support.
AI Train2023
Streamlines post-training and fine-tuning for large language and multimodal models with a single YAML-driven pipeline. Supports LoRA/QLoRA, full fine-tuning, preference tuning, RL methods, multi-GPU/FSDP/DeepSpeed, and many model backends (Hugging Face, local checkpoints).
AI Train2023
Connects a frozen vision encoder to a language model via visual instruction tuning, yielding an open multimodal assistant that follows image-grounded instructions. Released checkpoints span 7B-34B and approach GPT-4V on vision-language benchmarks.
AI Train2023
Fine-tunes 100+ LLMs and VLMs from one config file or a no-code web UI, unifying LoRA, QLoRA, full tuning, DPO, PPO, KTO and ORPO behind a single interface. Bundles GaLore, Unsloth, FlashAttention-2 and 2-8bit quantization to fit a single 24GB GPU.
AI Infra2023
Runs 70B-class LLM inference on a single 4GB GPU without quantization and supports Llama3.1 405B on 8GB VRAM. Uses layer-splitting and block-wise model compression (4/8-bit) to reduce disk load and can speed up inference loading by up to ~3x; integrates with Hugging Face models.
AI Image2023
Reference implementation for Stability AI's diffusion models: SDXL base/refiner/Turbo for text-to-image, plus Stable Video Diffusion, SV3D, and SV4D for image-to-video and 4D synthesis. A modular engine separates samplers, guiders, and conditioners.
Enables real-time (≥30 fps) 1080p novel-view synthesis by representing scenes as optimized anisotropic 3D Gaussians plus a visibility-aware splatting renderer; provides the paper's reference implementation, pretrained models and viewers — high-quality training requires CUDA GPU and significant VRAM.
Builds a GPT-style LLM in PyTorch step by step — tokenizer, attention, pretraining, and finetuning — with no external LLM frameworks. Companion code to a Manning book, with bonus chapters on LoRA and modern Llama/Qwen-style architectures.
AI Train2023
Trains LLMs with RLHF at scale by splitting actor, critic, reward, and reference models across separate GPU groups via Ray, with vLLM-accelerated generation and DeepSpeed ZeRO-3. Supports PPO, GRPO, REINFORCE++, DPO, plus async and agentic multi-turn RL.