Skip to content

numojo.core.accelerator.device

Device (numojo.core.accelerator.device)

This module defines the Device struct, which represents an execution device for array and matrix operations. It supports CPU and GPU devices, with GPU backends for NVIDIA CUDA, AMD ROCm, and Apple Metal.

Aliases

cpu

comptime cpu

Value: Device.CPU

cuda

comptime cuda

Value: Device.CUDA

rocm

comptime rocm

Value: Device.ROCM

mps

comptime mps

Value: Device.MPS

Structs

Device

struct Device

Memory convention: memory_only
Implements: AnyType, Copyable, Equatable, ImplicitlyCopyable, ImplicitlyDestructible, Movable, Representable, Stringable, Writable

Represents an execution device for array and matrix operations.

A Device identifies where computation should run, analogous to torch.device in PyTorch. Each device has a type ("cpu" or "gpu"), an optional backend name ("cuda", "rocm", or "mps" for GPUs), and a zero-based device index.

Use the predefined comptime constants for common devices: - Device.CPU — CPU execution. - Device.CUDA — NVIDIA CUDA GPU. - Device.ROCM — AMD ROCm GPU. - Device.MPS — Apple Metal GPU.

Devices can also be constructed from torch-style strings: 

var dev = Device("cuda:0")
var cpu = Device("cpu")

Fields

  • type (String): Device type: "cpu" or "gpu".
  • name (String): Backend identifier: "" for CPU, "cuda" | "rocm" | "mps" for GPU.
  • id (Int): Zero-based device index on the backend.

Aliases

CPU
comptime CPU

Value: Device._unchecked_init("cpu", "", 0)

CPU device.

CUDA
comptime CUDA

Value: Device._unchecked_init("gpu", "cuda", 0)

NVIDIA CUDA GPU device.

ROCM
comptime ROCM

Value: Device._unchecked_init("gpu", "rocm", 0)

AMD ROCm GPU device.

MPS
comptime MPS

Value: Device._unchecked_init("gpu", "mps", 0)

Apple Metal GPU device.

__del__is_trivial
comptime __del__is_trivial

Value: False

__move_ctor_is_trivial
comptime __move_ctor_is_trivial

Value: True

__copy_ctor_is_trivial
comptime __copy_ctor_is_trivial

Value: False

Methods

__init__
Overload 1
__init__(out self)

static

Initialize a default CPU device.

Args:

  • self (Self) [out]

Returns:

  • Self
Overload 2
__init__(out self, text: String)

static

Initialize a device by parsing a torch-style device string.

Supported formats: "cpu", "cuda", "cuda:0", "rocm", "rocm:1", "mps", "mps:0", "gpu".

Args:

  • text (String): A device string to parse.
  • self (Self) [out]

Returns:

  • Self

Raises

Error on invalid device string format.

Overload 3
__init__(out self, type: String, name: String, id: Int)

static

Initialize a device with explicit type, name, and index.

Validates the arguments and falls back to CPU if the requested GPU backend is not available on the current system.

Args:

  • type (String): Device type, must be "cpu" or "gpu".
  • name (String): Backend name ("" for CPU; "cuda", "rocm", or "mps" for GPU).
  • id (Int): Zero-based device index (must be 0 for CPU, >= 0 for GPU).
  • self (Self) [out]

Returns:

  • Self
__eq__
__eq__(self, other: Self) -> Bool

Check equality with another device.

Args:

  • self (Self)
  • other (Self): The device to compare against.

Returns:

  • Bool
__ne__
__ne__(self, other: Self) -> Bool

Check inequality with another device.

Args:

  • self (Self)
  • other (Self): The device to compare against.

Returns:

  • Bool
__str__
__str__(self) -> String

Return a human-readable string representation.

Args:

  • self (Self)

Returns:

  • String
__repr__
__repr__(self) -> String

Return the canonical string representation.

Args:

  • self (Self)

Returns:

  • String
write_to
write_to[W: Writer](self, mut writer: W)

Write the string representation to a writer.

Parameters:

  • W (Writer): The writer type.

Args:

  • self (Self)
  • writer (W) [mut]: The writer to write to.
is_cpu
is_cpu(self) -> Bool

Check if this is a CPU device.

Args:

  • self (Self)

Returns:

  • Bool
is_gpu
is_gpu(self) -> Bool

Check if this is a GPU device.

Args:

  • self (Self)

Returns:

  • Bool
is_available
is_available(self) -> Bool

Check if this device is available on the current system.

Args:

  • self (Self)

Returns:

  • Bool
default_device
default_device() -> Self

static

Return the best available device: GPU if present, otherwise CPU.

Returns:

  • Self
available_gpu
available_gpu() -> String

static

Return the name of the best available GPU backend.

Checks in order: CUDA → ROCm → MPS.

Returns:

  • String

Raises

NumojoError if no GPU accelerator is detected.

available_devices
available_devices() -> String

static

List all available devices on the current system.

Returns:

  • String
parse_device_string
parse_device_string(text: String) -> Self

static

Parse a torch-style device string into a Device.

Supported formats: - "cpu" - "cuda", "cuda:0", "cuda:1", ... - "rocm", "rocm:0", "rocm:1", ... - "mps", "mps:0", "mps:1", ... - "gpu" (resolves to best available GPU backend)

Args:

  • text (String): The device string to parse.

Returns:

  • Self

Raises

Error when "gpu" is specified but no GPU backend is available.

Functions

is_accelerator_available

is_accelerator_available[device: Device]() -> Bool

Check at compile time whether the given device's GPU accelerator exists.

Parameters:

  • device (Device): The device to check.

Returns:

  • Bool