Machine Learning with Wasm (wasi-nn)
summary
WASI-NN provides a WASI-NN loader API that treats models as opaque virtual I/O objects and proxies inference to native backends; reference implementation uses Wasmtime + OpenVINO, with AssemblyScript and Rust bindings and a live pizza image classification demo (AlexNet, MobileNet, Inception). Design favors loader first (builder deferred), supports CPU AVX512/GPU/TPU/FPGA backends, and reports measured 7–32x speedup versus a pure Wasm-only execution.
AI-written summary · about
chapters (13)
- 0:00Intro
- 0:16Agenda
- 1:03What is WASI?
- 1:39What is the Bytecode Alliance?
- 2:22Why wasi-nn?
- 4:08WebAssembly Environments & Use Cases
- 6:15Loader vs. Builder API
- 9:27ML Model as a virtualized I/O type
- 11:00API Walkthrough
- 11:57Reference implementation on Wasmtime
- 13:08Demo: wasi-nn
- 18:25Performance Comparison (cont.)
- 20:26Call for action
related talks
08mLightning Talk: High Performance Neural Network Inferencing Using WASI NN
27mWasm Is Becoming the Runtime for LLMs
10mDeploying machine learning in wasm to CPU, GPU and NPU with the new wa...
33mBringing AI to the Heterogeneous Edge with WebAssembly & ONNX
37mGetting Started with AI and WebAssembly
33mEfficient and Cross-Platform LLM Inference in the Heterogenous Cloud
25mLeveraging Wasm for Portable AI Inference Across GPUs, CPUs, OS & Cloud-Nativ...
41mWebAssembly Based AI as a Service