Why aren’t float64 tensors supported for MPS?
I think the main reason is the lack of native float64 support, meaning that the GPU acceleration for the tensors would have to emulate the data type. This would likely lead to poorer performance than use cases would require.
There have been some projects taking a stab at this path like GitHub - philipturner/metal-float64: Emulating double-precision arithmetic on Apple GPUs · GitHub but due to the amount of work required and the potentially small payoff it doesn’t seem to have gathered too much steam in the past.