Module ops 

Array operations as free functions with MLX’s names: the mirrors of the Array methods.

Every operation whose first argument is an array is a method, and the method is the canonical, documented form: x.sum(-1), x.matmul(&w.t()), m.select(&a, 0.0). This module holds the same operations as free functions with MLX’s names and the same parameters (the receiver becomes the first argument), so MLX code ports line by line: sum(&x, -1), matmul(&x, &w.t()), where_(&m, &a, 0.0). A few operations exist only here because their first argument is not one array: concatenate, stack, where_ and the constructors (zeros, ones, full, arange).

Operations return Array, never Result: invalid arguments give a failed array carrying crate::Error::Op with the operation’s name and the caller’s source location, and every array computed from it fails with the same error, surfacing at Array::check, Array::eval, Array::item or Array::to_vec.

use nuvec_core::{Array, array, ops};

let x = array![[1.0, 2.0], [3.0, 4.0]];
let a = x.exp().sum(-1);
let b = ops::sum(&ops::exp(&x), -1);
assert_eq!(a.to_vec::<f32>().unwrap(), b.to_vec::<f32>().unwrap());

Functions

abs

Element-wise abs; see Array::abs.

add

Element-wise add; see Array::add. Either operand may be a weak Rust number.

addmm

alpha · (a @ b) + beta · c; see Array::addmm.

all

Whether every element is non-zero over axes; see Array::all.

all_keepdims

Whether every element is non-zero over axes, keeping them as size-1 dimensions; see Array::all_keepdims.

any

Whether any element is non-zero over axes; see Array::any.

any_keepdims

Whether any element is non-zero over axes, keeping them as size-1 dimensions; see Array::any_keepdims.

arange

Evenly spaced values start, start + step, … below stop (above it for a negative step), as MLX: int32 when every argument is an integer (int64 when they do not fit), float32 when any is a float. Use Array::astype for another dtype.

argmax

Indices (uint32) of the maxima along one axis (..: the flattened array); see Array::argmax.

argmax_keepdims

argmax keeping the reduced axis; see Array::argmax_keepdims.

argmin

Indices (uint32) of the minima along one axis (..: the flattened array); see Array::argmin.

argmin_keepdims

argmin keeping the reduced axis; see Array::argmin_keepdims.

argpartition

The indices (uint32) that partition the array; see Array::argpartition.

argsort

The indices (uint32) that sort the array along one axis; see Array::argsort.

astype

The elements converted to dtype; see Array::astype.

broadcast_to

The array broadcast to shape; see Array::broadcast_to.

concatenate

arrays joined along axis (MLX’s concatenate); dtypes are promoted.

conjugate

Element-wise conjugate; see Array::conjugate.

contiguous

Element-wise contiguous; see Array::contiguous.

conv1d

A 1-D convolution of [N, L, C_in] with [C_out, K, C_in / groups]; see Array::conv1d.

copy

Element-wise copy; see Array::copy.

cos

Element-wise cos; see Array::cos.

cummax

The running maximum along one axis (..: the flattened array); see Array::cummax.

cummax_with

The running maximum with ScanOptions; see Array::cummax_with.

cummin

The running minimum along one axis (..: the flattened array); see Array::cummin.

cummin_with

The running minimum with ScanOptions; see Array::cummin_with.

cumprod

The cumulative product along one axis (..: the flattened array); see Array::cumprod.

cumprod_with

The cumulative product with ScanOptions; see Array::cumprod_with.

cumsum

The cumulative sum along one axis (..: the flattened array); see Array::cumsum.

cumsum_with

The cumulative sum with ScanOptions; see Array::cumsum_with.

depends

x, evaluated only after deps (MLX’s depends): evaluating the result evaluates the dependencies first, without using their values. Its gradient goes to x.

dequantize

The weights of quantize restored, in dtype (default: the scales’ dtype for affine, bfloat16 for the floating-point modes); MLX’s dequantize. See also Quantized::dequantize.

divide

Element-wise divide; see Array::divide. Either operand may be a weak Rust number.

equal

Element-wise equal; see Array::equal. Either operand may be a weak Rust number.

erfinv

Element-wise erfinv; see Array::erfinv.

exp

Element-wise exp; see Array::exp.

expand_dims

Size-1 dimensions inserted at axes (positions in the result); see Array::expand_dims.

floor_divide

Element-wise floor_divide; see Array::floor_divide. Either operand may be a weak Rust number.

from_fp8

E4M3 bytes as dtype floats (MLX’s from_fp8); see Array::from_fp8.

full

An array of shape with every element value: a Rust number is weak, as in MLX (0.5 gives float32, 1 int32), and an array keeps its dtype (Array::from(7u8) gives uint8).

full_like

value (a weak scalar takes a’s dtype) in a’s shape; see Array::full_like.

gather_mm

Products of gathered matrices; see Array::gather_mm.

gather_qmm

Quantized products of gathered matrices (mixture-of-experts layers); see Array::gather_qmm.

greater

Element-wise greater; see Array::greater. Either operand may be a weak Rust number.

greater_equal

Element-wise greater_equal; see Array::greater_equal. Either operand may be a weak Rust number.

less

Element-wise less; see Array::less. Either operand may be a weak Rust number.

less_equal

Element-wise less_equal; see Array::less_equal. Either operand may be a weak Rust number.

log

Element-wise log; see Array::log.

logaddexp

Element-wise logaddexp; see Array::logaddexp. Either operand may be a weak Rust number.

logcumsumexp

The running log(Σ exp) along one axis (..: the flattened array); see Array::logcumsumexp.

logcumsumexp_with

The running log(Σ exp) with ScanOptions; see Array::logcumsumexp_with.

logsumexp

log(sum(exp(a))), computed stably, over axes; see Array::logsumexp.

logsumexp_keepdims

log(sum(exp(a))), computed stably, over axes, keeping them as size-1 dimensions; see Array::logsumexp_keepdims.

matmul

Matrix product; see Array::matmul.

max

The maximum over axes; see Array::max.

max_keepdims

The maximum over axes, keeping them as size-1 dimensions; see Array::max_keepdims.

maximum

Element-wise maximum; see Array::maximum. Either operand may be a weak Rust number.

mean

The mean over axes; see Array::mean.

mean_keepdims

The mean over axes, keeping them as size-1 dimensions; see Array::mean_keepdims.

min

The minimum over axes; see Array::min.

min_keepdims

The minimum over axes, keeping them as size-1 dimensions; see Array::min_keepdims.

minimum

Element-wise minimum; see Array::minimum. Either operand may be a weak Rust number.

multiply

Element-wise multiply; see Array::multiply. Either operand may be a weak Rust number.

negative

Element-wise negative; see Array::negative.

not_equal

Element-wise not_equal; see Array::not_equal. Either operand may be a weak Rust number.

ones

An array of ones of shape.

ones_like

Ones of a’s shape and dtype; see Array::ones_like.

partition

The array partitioned around position kth along one axis; see Array::partition.

power

Element-wise power; see Array::power. Either operand may be a weak Rust number.

prod

The product over axes; see Array::prod.

prod_keepdims

The product over axes, keeping them as size-1 dimensions; see Array::prod_keepdims.

put_along_axis

A copy with values placed at indices along axis; see Array::put_along_axis.

quantize

w quantized along its last axis; see Array::quantize.

quantized_matmul

x · dequantize(w)ᵀ (or without the transpose); see Array::quantized_matmul.

repeat

Each element repeated repeats times along one axis; see Array::repeat.

reshape

The same elements in another shape (one -1 inferred); see Array::reshape.

scatter_add_along_axis

A copy with values added at indices along axis (repeated indices accumulate); see Array::scatter_add_along_axis.

sigmoid

Element-wise sigmoid; see Array::sigmoid.

sign

Element-wise sign; see Array::sign.

sin

Element-wise sin; see Array::sin.

slice_update

A copy of a with the region at index replaced by update; see Array::slice_update.

softmax

exp(a) / sum(exp(a)) over axes; see Array::softmax.

sort

The array sorted along one axis; see Array::sort.

split

The array split into parts equal parts along axis; see Array::split.

split_at

The array split before each of indices along axis; see Array::split_at.

split_n

The array split into N equal parts along axis; see Array::split_n.

sqrt

Element-wise sqrt; see Array::sqrt.

squeeze

Size-1 dimensions removed (..: all of them); see Array::squeeze.

stack

arrays (one shape) stacked along a new axis axis (MLX’s stack).

std

The standard deviation over axes; see Array::std.

std_keepdims

The standard deviation over axes, keeping them as size-1 dimensions; see Array::std_keepdims.

std_with

The standard deviation over axes with VarOptions; see Array::std_with.

stop_gradient

Element-wise stop_gradient; see Array::stop_gradient.

subtract

Element-wise subtract; see Array::subtract. Either operand may be a weak Rust number.

sum

The sum over axes; see Array::sum.

sum_keepdims

The sum over axes, keeping them as size-1 dimensions; see Array::sum_keepdims.

take

The elements at indices along one axis (..: the flattened array); see Array::take.

take_along_axis

The elements at indices along axis; see Array::take_along_axis.

tanh

Element-wise tanh; see Array::tanh.

to_fp8

Floats as E4M3 bytes (MLX’s to_fp8); see Array::to_fp8.

topk

The k largest elements along one axis; see Array::topk.

transpose

The axes reordered (..: reversed); see Array::transpose.

var

The variance over axes; see Array::var.

var_keepdims

The variance over axes, keeping them as size-1 dimensions; see Array::var_keepdims.

var_with

The variance over axes with VarOptions; see Array::var_with.

view

x’s bits as dtype (MLX’s view); see Array::view.

where_

Element-wise selection: a where condition is true, else b (MLX’s where); see Array::select. Any operand may be a weak Rust number.

zeros

An array of zeros of shape.

zeros_like

Zeros of a’s shape and dtype; see Array::zeros_like.