Hugging Face ecosystem for Local AI/ ML
summary
Live demo shows a 1.7B model in browser using WebGPU/TVM and llama.cpp to parse unstructured text into JSON and to call a Hugging Face Space for image generation; a tiny orchestration library tiny agents (<100 lines) wires local or remote models to tools (MCP, Playwright, Spaces). Hub stats given: 230,000 LLMs on Hub, ~5–6 billion monthly pulls and ~7 PB/day. Covers the conversion checklist (transformers → llama.cpp/ONNX → verify outputs), common failures (tokenizers, RoPE, chat templates) and quantization advice (use largest model that fits, avoid <Q4).
AI-written summary · about
chapters (11)
- 0:00Introduction to talk
- 1:08Running LLMs in browser
- 5:06Tools for on-device ML
- 6:56Demo with Tiny Agents
- 16:35Use cases for local models
- 18:00Hugging Face ecosystem overview
- 20:07Common model conversion issues
- 22:13Hugging Face contributions
- 23:37Model adoption trends
- 24:40Future of local AI
- 25:22Q&A session
related talks
19mBuilding Machine Learning Applications using Hugging Face
34mExploring Hugging Face: A Hub for Machine Learning Models
28mUnlock the Power of LLM: Build a HuggingFace Agent
49mMachine Learning 2.0 With Hugging Face
33mMachine Learning 2.0 with Hugging Face
20mSelf-hosting AI LLMs - a beginners guide
14mKeynote: Building AI with Open Source & Hugging Face- 14mKeynote: Building AI with Open Source & Hugging Face