> with the C++ library onnxruntime emerging as the quickest route to hardware optimization. However, this library presents substantial challenges due to its 20 distinct execution providers
> Mojo's current appeal lies [...] in its ability to improve the optimization process for inference at a lower level
Will Mojo be supporting the same targets as the ONNX runtime? The appeal of ONNX was never the quickness aspect of it to me, but rather the hardware support. I think a lot of people are in the same boat, and won't be ready to switch over until esoteric Rockchip and PowerPC support is eked out.
Offering a form of IR between these formats isn't a bad idea, but it sorta seems like it backsteps into the same complexity problem if it wants to compete with offerings like ONNX.
Extending Python is a great idea but existing libraries will not benefit automatically. They need to be rewritten since all the benefits come from the new syntax. Because of this, bolting it onto Python seems more like a marketing ploy. Hopefully I misunderstood.
4 comments
[ 3.6 ms ] story [ 14.0 ms ] thread> Mojo's current appeal lies [...] in its ability to improve the optimization process for inference at a lower level
Will Mojo be supporting the same targets as the ONNX runtime? The appeal of ONNX was never the quickness aspect of it to me, but rather the hardware support. I think a lot of people are in the same boat, and won't be ready to switch over until esoteric Rockchip and PowerPC support is eked out.
Offering a form of IR between these formats isn't a bad idea, but it sorta seems like it backsteps into the same complexity problem if it wants to compete with offerings like ONNX.