Friend, this is great, your Tenmo really could become a competitor for PyTorch. I especially liked what you wrote about pure Mojo, you really dug into the low level, even the tricky stuff. But is that really enough? You’ve reached a peak, but PyTorch already has a lot of momentum. It might be tough for your implementation to catch up, so maybe start by imitating PyTorch common behavior as much as possible, so people can switch easily, then gradually add your own API.
Also, MNIST numbers look good for a pure-Mojo CPU run, but bigger models and GPU workloads will really show where it stands. Docs and a short guide for switching would also help a lot. Overall really impressive work, keep it up!
Friend, this is great, your Tenmo really could become a competitor for PyTorch. I especially liked what you wrote about pure Mojo, you really dug into the low level, even the tricky stuff. But is that really enough? You’ve reached a peak, but PyTorch already has a lot of momentum. It might be tough for your implementation to catch up, so maybe start by imitating PyTorch common behavior as much as possible, so people can switch easily, then gradually add your own API. Also, MNIST numbers look good for a pure-Mojo CPU run, but bigger models and GPU workloads will really show where it stands. Docs and a short guide for switching would also help a lot. Overall really impressive work, keep it up!