flashinfer.cake_mhc_nondet_train.cake_mhc_nondet_train_backward_workspace

flashinfer.cake_mhc_nondet_train.cake_mhc_nondet_train_backward_workspace(T: int, device: device, *, hidden: int = 4096) → Tensor

Allocate a zero-initialized backward workspace of this package’s layout (see cake_mhc_nondet_train_backward_workspace_bytes()).

Parameters:
Returns:

A zero-filled uint8 tensor of that many bytes on device.

Return type:

torch.Tensor