This looks exciting, and I'm looking forward to its progress. I have questions about GPU infrastructure in particular. It seems like the GPU backend is primarily through MLIR. Does that target CUDA? Vulkan compute shaders? WebGPU? Metal shaders? I note that IREE has the first three of those listed as targets, but am not clear on exactly where that fits in to the rest of the ecosystem.

I have an idea of a dream GPU infrastructure based on compute shaders, where you compile your problem into GPU IR suitable for your operating system (SPIR-V, Metal, DXIL) and have a very lightweight runtime that just runs those. For the ahead of time compilation case, you wouldn't need to ship GPU compilers in a deployed application, though you would want that for more rapid iteration when doing research and exploratory programming.

From what I can see so far, this seems basically orthogonal to what Modular and Mojo are trying to do, but it's entirely possible I'm missing something. I'm wondering if anyone is actually building this (IREE and MediaPipe are the closest things in the space I'm aware of), and if not, why not.