Apple already has their own environment in CoreML and the neural engine. They have ways to train/run models on it in highly optimized formats and unified memory

Apple already has their own environment in CoreML and the neural engine. They have ways to train/run models on it in highly optimized formats and unified memory. PyTorch came from PCs running on GPUs with their own pool of dedicated memory and a slower link to main memory. I know people have been speeding it up a lot on Apple processors, but I’m not sure it’s a good indicator of how ready the iPhone is for running models for end user programs.