Stream: wasi-nn

Topic: GPU support through WASI


view this post on Zulip Ransford Hyman (Apr 02 2023 at 15:29):

I'm curious to know does WASI need to implement capabilities to execute GPU code? All of the efforts around WASM/WASI has been around CPU compute, but a great deal of ML/AI is done on GPU. So wandering where the community stands in regards to this effort :smile:

view this post on Zulip Andrew Brown (Apr 05 2023 at 08:29):

wasi-nn has an execution-target that includes GPUs. A while back I experimented with inferences on a GPU (with OpenVINO, FWIW) and things seemed to work fine.

Neural Network proposal for WASI. Contribute to WebAssembly/wasi-nn development by creating an account on GitHub.

view this post on Zulip Ayako Akasaka (May 05 2023 at 07:22):

Sorry for jumping in. I also have a question about execution-target
Now I want to use NPU(DSP) for AI/ML task. It seems external_delegation is used for TFLite. So I want to say it doesn't seem GPU, TPU, but extra AI/ML accelerator.
In this case, which execution-target should I use? gpu or tpu ?
I am asking it because I am trying to push this implementation to WAMR.

Neural Network proposal for WASI. Contribute to WebAssembly/wasi-nn development by creating an account on GitHub.
Add VX delegation as an external delegation of TFLite. So that several NPU/GPU (from VeriSilicon, NXP, Amlogic) can be controlled via WASI-NN. Test Code can work with the X86 simulator.

view this post on Zulip Andrew Brown (May 08 2023 at 17:18):

@Ayako Akasaka, if you feel like it is a sufficiently different execution target than the ones that currently exist and TFLite also has explicit support for it, you could submit a PR to add it to the wasi-nn execution-target list.

view this post on Zulip Ayako Akasaka (May 17 2023 at 11:35):

@Andrew Brown Thank you for the reply. I will consider creating PR for it


Last updated: Dec 23 2024 at 13:07 UTC