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 Deploy2023
Run any open-source LLM, embedding, speech, image, or multimodal model behind one OpenAI-compatible API — swap GPT for an open model in a single line. Routes across vLLM, llama.cpp, GGML, and TensorRT, scaling from a laptop to a multi-node GPU cluster.
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.
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.
AI Image2023
Runs Stable Diffusion XL behind a Midjourney-style interface, hiding samplers, model swaps, and LoRA weights. A built-in GPT2 expander rewrites prompts into richer styling, and it works fully offline on as little as 4GB of Nvidia VRAM.
Chatbot2023
Runs open-weight LLMs (Llama, Gemma, Qwen, GGUF) offline on your machine, with an optional bridge to OpenAI/Anthropic/Mistral. Exposes an OpenAI-compatible API at localhost:1337, so SDK code built for OpenAI switches by changing one base URL.
AI Audio2023
Generates expressive multilingual speech from text, with sub-word control over prosody and emotion via inline tags like [whisper] or [angry]. Handles multi-speaker, multi-turn dialogue; the weights ship under a research-only license.
AI Audio2023
Terminal CLI for on-device Whisper ASR using Hugging Face Transformers + Optimum, with optional Flash Attention 2, batching, and diarization support — focused on high-throughput transcription on NVIDIA GPUs and Apple Silicon (mps).
AI Agent2023
Routes LLM and agent decisions through semantic similarity instead of waiting for full generations, useful for intent routing, tool selection, guardrails, and multimodal handling.
AI Audio2023
Performs speaker diarization (who spoke when) with pyannote-audio: combines voice-activity detection, speaker-change and overlapped-speech detection to produce time-stamped speaker segments; compatible with Hugging Face Endpoints and ASR pipelines.
AI Train2023
Collection of runnable model implementations — LLaMA, Mistral, Stable Diffusion, Whisper, CLIP, plus LoRA fine-tuning — ported to the MLX array framework so they run natively on Apple silicon's unified memory rather than CUDA.
AI Infra2023
A selective State Space Model architecture and PyTorch implementation for linear-time sequence modeling. Hardware-aware, designed for information-dense tasks (e.g. language modeling), with pretrained weights on Hugging Face; requires CUDA-enabled PyTorch.