Your question is Custom PyTorch Autograd for Hardware Constraints. Take a moment with it on the right.
Talk me through your thinking if you like. When you're confident, submit your answer and I'll grade it like a real screen (7/10 or better passes).
How would you write a custom PyTorch autograd function to simulate hardware-specific constraints during the forward pass for TETRAMEM?