| 129 | |
| 130 | @staticmethod |
| 131 | def backward(ctx, *output_grads): |
| 132 | ctx.input_tensors = [x.detach().requires_grad_(True) for x in ctx.input_tensors] |
| 133 | with torch.enable_grad(): |
| 134 | # Fixes a bug where the first op in run_function modifies the |
| 135 | # Tensor storage in place, which is not allowed for detach()'d |
| 136 | # Tensors. |
| 137 | shallow_copies = [x.view_as(x) for x in ctx.input_tensors] |
| 138 | output_tensors = ctx.run_function(*shallow_copies) |
| 139 | input_grads = torch.autograd.grad( |
| 140 | output_tensors, |
| 141 | ctx.input_tensors + ctx.input_params, |
| 142 | output_grads, |
| 143 | allow_unused=True, |
| 144 | ) |
| 145 | del ctx.input_tensors |
| 146 | del ctx.input_params |
| 147 | del output_tensors |
| 148 | return (None, None) + input_grads |
| 149 | |
| 150 | |
| 151 | def timestep_embedding(timesteps, dim, max_period=10000, repeat_only=False): |