WebGPU backend roadmap for Q1_0 kernels?

[pmenic]

Hi, thanks for the great work on Bonsai-8B and the Q1_0 quantization.

I'd like to ask about plans for running Q1_0 models in the browser via WebGPU.

Current state (as far as I can see):

• The Q1_0 custom kernels exist for CUDA (.cu), Metal (.metal), and Vulkan (GLSL/SPIR-V), plus CPU.
• Upstream ggml-org/llama.cpp has an active WebGPU backend in development (multiple PRs merged in recent weeks, including Intel subgroup matrix support).
• However, the WebGPU backend upstream only targets standard quant types; there is no WGSL implementation of the Q1_0 kernel yet.

Questions:

1. Is a WGSL port of the Q1_0 kernel on your roadmap, or under consideration?
2. If not planned internally, would you be open to a community PR once the upstream WebGPU backend stabilizes?
3. Any preference between waiting for upstream WebGPU to mature vs. an earlier experimental port?

Running Bonsai-8B purely in-browser (no server, no install) would be a great showcase of the 1-bit quantization story — curious where this sits in your plans.

Thanks!

[khosravipasha]

Thanks :)

Yeah WebGPU backend in llama.cpp for Q1_0 would be amazing. Currently we are not actively working on it.
Of course would love a PR (can do here, or main llama.cpp or both :) )
I have not played around with the webgpu backend so don't know how good it is

One a related note Huggingface team made a WebGPU demo you can try here:
I believe its using their transformer.js library.
https://huggingface.co/spaces/webml-community/bonsai-webgpu