Successfully trained very small model. Optimized kernel a lot (especially attention backward). Now I have to optimize kernels even more, documented code, and train full model.