Compute exponential base e minus 1. max ulp err = 0.99746 */

Create a :obj:`QuantizedEmbedding` layer from an :obj:`Embedding` layer.

Parse the json hostfile that contains both the hostnames to ssh into and the ips to communicate over when using the ring backend. It can also

Create a :obj:`QQLinear` layer from a :obj:`Linear` layer.

Compute matrix product with matrix-level gather */

Calculate the gradient wrt to the weights of the following calculation y = gather_mm(x, w.T, lhs_indices, rhs_indices, sorted) Note the transpose ab

Get CPU architecture string at runtime

Compute padded dimensions for tiled layout Tiles are 128 rows × 4 columns, must allocate full tiles

Compute an FFT twiddle factor

Return true if we are currently performing a function transformation in * order to keep the graph when evaluating tracer arrays. */

The size of the array's datatype in bytes. */

* The counterpoint of `accept_connections`. Basically connect to each of the * provided addresses. */

Create a tracer copy of a primal for use in vjp/jvp. If the primal is a stale Copy from a previous transform call (not an active tracer), peel it off

Helper that merges two arrays in the graph by setting the parents of the source to point to the destination. The arrays are assumed to be coming from

MLX naive: expand masks and use regular matmul.

The number of dimensions of the array. */

In MLX a singleton dim (shape[dim] == 1) can have any stride, but in cuDNN whether a tensor is contiguous is determined with: shape[dim] == shape[dim