Inquiry regarding documentation and status of wasi-gpu for non-browser applications

[guimvmatos]

Hi everyone,

I have been working with the WebAssembly Component Model for some time now, specifically using WASI, WIT, and WAC to compose applications written in C and Rust. My current workflow involves running these components via Wasmtime.

I am very interested in extending my current applications to leverage the host's GPU for computational tasks (specifically for kernel execution and performance modeling), but I want to do this strictly in a non-browser/standalone environment.

I noticed this repository is under active development. Could you point me to any available documentation, examples, or even draft WIT files that demonstrate how to grant a component access to the GPU using the current Component Model standards?

Specifically, I am looking for:

• The recommended way to define GPU requirements in a .wit world.
• How to link these requirements to the host's native GPU (e.g., using wgpu or wasmtime-gpu) during the composition/instantiation phase.

Thank you for your great work on this specification!

[MendyBerger]

Hi @guimvmatos, thanks for filing this issue.

The recommended way to define GPU requirements in a .wit world.

In WebGPU you define requirements when calling request-device through required-features and required-limits

How to link these requirements to the host's native GPU (e.g., using wgpu or wasmtime-gpu) during the composition/instantiation phase.

I think you're looking for wasi-gfx-runtime. wasi-gfx-runtime adds wasi:webgpu support to wasmtime.

More information, docs, and examples, can be found in the wasi-gfx org.

Let me know if this answers your question.