Source
Introducing @huggingface/kernels: 200+ WebGPU Kernels for Local AI
Hugging Face has released @huggingface/kernels and a collection of 207 WebGPU kernels to optimize local AI inference in web browsers. This release includes Fleet, a benchmarking tool to crowdsource GPU performance and correctness data.
Why it matters
By providing optimized operations for diverse hardware, this reduces the performance variability of browser-based AI, making local AI applications faster and more reliable for users.
The details
Each kernel repository contains a manifest, correctness tests, benchmark cases, and WGSL shader templates. The @huggingface/kernels library allows developers to load these kernels via Hub repository IDs and version numbers. Hugging Face is also collaborating with the ONNX Runtime team to upstream these improvements.
What's next
Hugging Face plans to connect these kernels to higher-level model tooling and continue expanding the coverage of available operations.
Show entities and relationshipsHide entities and relationships
In this article
People
Companies
Key connections
Hugging Face owns Fleet
Nico Martin works at Hugging Face
Joshua Xenova works at Hugging Face
Show 6 more connectionsShow fewer connections
Hugging Face Hub is related to CUDA
Hugging Face Hub is related to ROCm
Hugging Face Hub is related to Metal
Hugging Face Hub is related to WebGPU
@huggingface/kernels uses WGSL
Get the weekly recap
The stories like this one, picked and explained — once a week, straight to your inbox.