A constant node with the specified shape that holds increasing values starting from 0 along the iota dimension.

A constant node for a unidimensional array of increasing values starting from 0.

Create a new builder with the associated name, the name is only used for debugging purposes.

Create a new array shape.

Creates a new tokenizer, this takes as input the path for the bpe rank file.

Create a new array shape.

Hacky parser for the npy header, a typical example would be: {'descr': '<f8', 'fortran_order': False, 'shape': (128,), }

Reads a npy file and return the stored multi-dimensional array as a literal.

Reads a npz file and returns the stored multi-dimensional arrays together with their names.

Reads a npz file and returns the stored multi-dimensional arrays for some specified names.

A node that computes the minimum value across the specified dimensions.

Create a new literal containing the same data but using a different shape. This returns an error if the number of elements in the literal is different

A node that when executed generates values using a random normal distribution.

A node that when executed generates values using a random uniform distribution.

Create a literal from a scalar value, the resulting literal has zero dimensions and stores a single element.

Sigmoid activation function. This computes the element-wise sigmoid.

Transfer and return a value for the given shape from the outfeed queue.

Build a tuple from multiple operands.

The kind of elements that are computed by this operand, shortcut for `primitive_type`.

An error node, using the 'unknown error' error type.