webnn is neat. I’m not super excited for it too much since last time I looked into it, it for sure had some issues. They seem to be addressed though. ONNX is not a terrible interface to work with either and has good platform support so that wont be an issue. I do prefer wgpu solutions however. While they don’t work with NPUs (for obvious reasons :D) they are pretty much a “program once run anywhere” solution since it supports metal, dx12 and vulkan. (Only recently got fp16 support though so most things are still rough)
but for higher perf needs I can see webnn being a lot more useful
(Not to pull rank, but my mail profile can be tracked to Netscape Navigator, across multiple OSs 😁)
Yeah, I don’t think I like llamafile, reusing some weights between models, and smaller updates, sounds like a better idea.
What I’d like to see is a unified WebNN support, for CPU, GPU, and NPU: WebNN Overview
(Not to pull rank, but my mail profile can be tracked to Netscape Navigator, across multiple OSs 😁)
webnn is neat. I’m not super excited for it too much since last time I looked into it, it for sure had some issues. They seem to be addressed though. ONNX is not a terrible interface to work with either and has good platform support so that wont be an issue. I do prefer wgpu solutions however. While they don’t work with NPUs (for obvious reasons :D) they are pretty much a “program once run anywhere” solution since it supports metal, dx12 and vulkan. (Only recently got fp16 support though so most things are still rough)
but for higher perf needs I can see webnn being a lot more useful
just means older then I am ;D