Build a Linear(len(grad_values), 1) and set its weight gradient to grad_values as a single row, leaving the bias gradient at zero. Clip the model's global gradient norm to max_norm using PyTorch's own clipping utility, then return the weight gradient afterwards as a 1-D tensor.

Clipping rescales all gradients by one shared factor when the global norm exceeds the threshold, and leaves them untouched when it does not -- it is not a per-element clamp.

Input

grad_values = [3.0, 4.0]
max_norm = 1.0

Output

tensor([0.6000, 0.8000])