pub fn slice_flat(tensor: &Tensor, start: usize, end: usize) -> Result<Tensor>Expand description
Extracts [start, end) of a tensor’s flattened data as a genuinely smaller 1-D
tensor.
The whole point of ZeRO is that a rank only retains its own slice, so this really
allocates end - start elements rather than rescaling the original. (The source
tensor is read in full here because partitioning happens once, at setup, from a
materialised parameter map.)
§Errors
Returns an error when the range is out of bounds or the tensor is not
f32-readable.