Are there any plans to have a mojo library for both training and inference, like pytorch or tensorflow? Now that mojo reached 1.0 and that we can write gpu kernels in pure mojo, this looks like a natural next step.
I know there are packages made by the community, but I would expect a library to be provided by the Modular itself, given the AI first philosophy of the company.
MAX will eventually grow training capabilities, but that’s going to take a while. Right now Modular has a rare chance to fix the mistakes of the past. Pytorch and JAX have one glaring issue for training, which is that they aren’t fault tolerant, which, in my opinion, becomes actively irresponsible at our current state of “1 Megawatt per rack” where you don’t want losing an NVMe drive to knock out 72 GPUs.
Unfortunately, this means that Modular has to start basically from scratch and redo everything, which means it will be a while before that investment can be justified.