* The last axis of `indices` is how deep to index into `tensor` so  this index
  depth must be less than the rank of `tensor`: `indices.shape[-1] <= tensor.ndim`

if `indices.shape[-1] = tensor.rank` this Op indexes and updates scalar elements.
if `indices.shape[-1] < tensor.rank` it indexes and updates slices of the input
`tensor`.

        "properties": {
          "start": {
            "default": 0,

representing a sub-field or item. The string will follow one of these four formats: 'f:<name>', where <name> is the name of a field in a struct, or key in a map 'v:<value>', where <value> is the exact json formatted value of a list item 'i:<index>', where <index> is position of a item in a list 'k:<keys>', where <keys> is a map of  a list item's key fields to their unique values If a key maps to an empty Fields value, the field that key represents is part of the set.\n\nThe exact format is defined...

*   Enable JIT-compiled i64-indexed kernels on GPU for large tensors with more than 2**32 elements.
    *   Unary GPU kernels: Abs, Atanh, Acos, Acosh, Asin, Asinh, Atan, Cos, Cosh, Sin, Sinh, Tan, Tanh.