Device#

Collection of device Ivy functions.

ivy.as_ivy_dev(device, /)[source]#

Convert device to string representation.

Parameters:

device (Union[Device, str]) – The device handle to convert to string.

Return type:

Device

Returns:

ret – Device string e.g. ‘cuda:0’.

Examples

>>> y = ivy.as_ivy_dev('cpu')
>>> print(y)
cpu
ivy.as_native_dev(device, /)[source]#

Convert device string representation to native device type.

Parameters:

device (Union[Device, NativeDevice]) – The device string to convert to native device handle. A native device handle can be passed in instead - in this case the unmodified parameter is returned.

Return type:

NativeDevice

Returns:

ret – Native device handle.

Examples

With ivy.Device input:

>>> ivy.set_backend("numpy")
>>> ivy.as_native_dev("cpu")
'cpu'
>>> ivy.set_backend("tensorflow")
>>> ivy.as_native_dev("tpu:3")
'/TPU:3'

With ivy.NativeDevice input:

>>> import torch
>>> device = torch.device("cuda")
>>> device
device(type='cuda')
>>> ivy.as_native_dev(device)
device(type='cuda')
ivy.clear_cached_mem_on_dev(device, /)[source]#

Clear memory cache on target device.

Parameters:

device (Union[Device, NativeDevice]) – The device string to convert to native device handle or native device handle.

Return type:

None

Examples

>>> import torch
>>> ivy.set_backend("torch")
>>> device = torch.device("cuda")
>>> ivy.clear_cached_mem_on_dev(device)
ivy.default_device(device=None, /, *, item=None, as_native=None)[source]#

Return the input device or the default device. If the as_native flag is set, the device will be converted to a native device. If the item is provided, the item’s device is returned. If the device is not provided, the last default device is returned. If a default device has not been set, the first gpu is returned if available, otherwise the cpu is returned.

Parameters:
  • device (Optional[Union[Device, NativeDevice]], default: None) – The device to be returned or converted.

  • item (Optional[Union[list, tuple, dict, Array, NativeArray]], default: None) – The item to get the device from.

  • as_native (Optional[bool], default: None) – Whether to convert the device to a native device.

Return type:

Union[Device, NativeDevice]

Returns:

ret – Device handle or string.

Examples

>>> ivy.default_device()
device(type='cpu')
>>> ivy.default_device("gpu:0")
'gpu:0'
>>> ivy.default_device(item=[], as_native=False)
'cpu'
>>> ivy.default_device(item=(), as_native=True)
device(type='cpu')
>>> ivy.default_device(item={"a": 1}, as_native=True)
device(type='cpu')
>>> x = ivy.array([1., 2., 3.])
>>> x = ivy.to_device(x, 'gpu:0')
>>> ivy.default_device(item=x, as_native=True)
device(type='gpu', id=0)
ivy.dev(x, /, *, as_native=False)[source]#

Get the native device handle for input array x.

Parameters:
  • x (Union[Array, NativeArray]) – array for which to get the device handle.

  • as_native (bool, default: False) – Whether or not to return the dev in native format. Default is False.

Return type:

Union[Device, NativeDevice]

Returns:

ret

Device handle for the array.

Examples

With ivy.Array input:

>>> x = ivy.array([3, 1, 4, 5])
>>> y = ivy.dev(x)
>>> print(y)
cpu

With ivy.NativeArray input:

>>> x = ivy.native_array([[2, 5, 4], [3, 1, 5]])
>>> y = ivy.dev(x, as_native=True)
>>> print(y)
cpu
ivy.dev_util(device, /)[source]#

Get the current utilization (%) for a given device.

Parameters:

device (Union[Device, NativeDevice]) – The device string of the device to query utilization for.

Return type:

float

Returns:

ret – The device utilization (%)

Example

>>> ivy.dev_util('cpu')
13.4
>>> ivy.dev_util('gpu:0')
7.8
>>> ivy.dev_util('cpu')
93.4
>>> ivy.dev_util('gpu:2')
57.4
>>> ivy.dev_util('cpu')
84.2
ivy.function_supported_devices(fn, recurse=True)[source]#

Return the supported devices of the current backend’s function. The function returns a dict containing the supported devices for the compositional and primary implementations in case of partial mixed functions.

Parameters:
  • fn (Callable) – The function to check for the supported device attribute

  • recurse (bool, default: True) – Whether to recurse into used ivy functions. Default is True.

Return type:

Union[Tuple, dict]

Returns:

ret – Tuple or dict containing the supported devices of the function

Examples

>>> import ivy
>>> ivy.set_backend('numpy')
>>> print(ivy.function_supported_devices(ivy.ones))
('cpu',)
>>> ivy.set_backend('torch')
>>> x = ivy.function_supported_devices(ivy.ones)
>>> x = sorted(x)
('cpu', 'gpu')
ivy.function_unsupported_devices(fn, recurse=True)[source]#

Return the unsupported devices of the current backend’s function. The function returns a dict containing the unsupported devices for the compositional and primary implementations in case of partial mixed functions.

Parameters:
  • fn (Callable) – The function to check for the unsupported device attribute

  • recurse (bool, default: True) – Whether to recurse into used ivy functions. Default is True.

Return type:

Union[Tuple, dict]

Returns:

ret – Tuple or dict containing the unsupported devices of the function

Examples

>>> print(ivy.function_unsupported_devices(ivy.ones))
('tpu',)
ivy.get_all_ivy_arrays_on_dev(device, /)[source]#

Get all ivy arrays which are currently alive on the specified device.

Parameters:

device (Union[Device, NativeDevice]) – The device handle from which to get the arrays

Return type:

Container

Returns:

ret – Container with the arrays found for the specified device [identity, array]

Examples

>>> x = ivy.array([1,0,2])
>>> y = ivy.dev(x)
>>> z = ivy.get_all_ivy_arrays_on_dev(y)
>>> print(z)
{139740789224448:ivy.array([1,0,2])},
ivy.gpu_is_available()[source]#

Determine whether a GPU is available to use, with the backend framework.

Return type:

bool

Returns:

ret – Boolean, as to whether a gpu is available.

Examples

>>> print(ivy.gpu_is_available())
False
ivy.handle_soft_device_variable(*args, fn, **kwargs)[source]#
ivy.num_cpu_cores(*, logical=True)[source]#

Determine the number of cores available in the cpu.

Parameters:

logical (bool, default: True) – Whether request is for number of physical or logical cores available in CPU

Return type:

int

Returns:

ret – Number of cores available in CPU

Examples

>>> print(ivy.num_cpu_cores(logical=False))
2
ivy.num_gpus()[source]#

Determine the number of available GPUs, with the backend framework.

Return type:

int

Returns:

ret – Number of available GPUs.

Examples

>>> print(ivy.num_gpus())
1
ivy.num_ivy_arrays_on_dev(device, /)[source]#

Return the number of arrays which are currently alive on the specified device.

Parameters:

device (Union[Device, NativeDevice]) – The device handle from which to count the arrays

Return type:

int

Returns:

ret – Number of arrays on the specified device

Examples

>>> x1 = ivy.array([-1, 0, 5.2])
>>> x2 = ivy.array([-1, 0, 5.2, 4, 5])
>>> y = ivy.num_ivy_arrays_on_dev(ivy.default_device())
>>> print(y)
2
>>> x1 = ivy.native_array([-1, 0, 5.2])
>>> y = ivy.num_ivy_arrays_on_dev(ivy.default_device())
>>> print(y)
0
>>> x = ivy.Container(x1=ivy.array([-1]),
...                   x2=ivy.native_array([-1]))
>>> y = ivy.num_ivy_arrays_on_dev(ivy.default_device())
>>> print(y)
1
ivy.percent_used_mem_on_dev(device, /, *, process_specific=False)[source]#

Get the percentage used memory for a given device string. In case of CPU, the used RAM is returned.

Parameters:
  • device (Union[Device, NativeDevice]) – The device string to convert to native device handle.

  • process_specific (bool, default: False) – Whether the check the memory used by this python process alone. Default is False.

Return type:

float

Returns:

ret – The percentage used memory on the device.

Examples

>>> x = ivy.percent_used_mem_on_dev("cpu", process_specific = False)
>>> print(x)
94.036902561555
>>> x = ivy.percent_used_mem_on_dev("cpu", process_specific = True)
>>> print(x)
0.7024003467681645
>>> x = ivy.as_native_dev("gpu:0")
>>> y = ivy.percent_used_mem_on_dev(x, process_specific = False)
>>> print(y)
0.7095597456708771
ivy.print_all_ivy_arrays_on_dev(*, device=None, attr_only=True)[source]#

Print the shape and dtype for all ivy arrays which are currently alive on the specified device.

Parameters:
  • device (Optional[Union[Device, NativeDevice]], default: None) – The device on which to print the arrays

  • attr_only (bool, default: True) – Whether or not to only print the shape and dtype attributes of the array

Return type:

None

Examples

>>> x = ivy.array([[1,0,2], [3,2,1]])
>>> y = ivy.dev(x)
>>> ivy.print_all_ivy_arrays_on_dev(y)
((3,), 'int32')
((3,), 'int32')
>>> x = ivy.array([[1,0,2], [3,2,1]])
>>> y = ivy.dev(x)
>>> ivy.print_all_ivy_arrays_on_dev(y, attr_only = False)
[1,0,2]
[3,2,1]
ivy.set_default_device(device, /)[source]#

Sets the default device to the argument provided in the function.

Parameters:

device (Union[Device, NativeDevice]) – The device to be set as the default device.

Return type:

None

Returns:

ret – The new default device.

Examples

>>> ivy.default_device()
'cpu'
>>> ivy.set_backend('jax')
>>> ivy.set_default_device('gpu:0')
>>> ivy.default_device()
'gpu:0'
>>> ivy.set_backend('torch')
>>> ivy.set_default_device('gpu:1')
>>> ivy.default_device()
'gpu:1
>>> ivy.set_backend('tensorflow')
>>> ivy.set_default_device('tpu:0)
>>> ivy.default_device()
'tpu:0
>>> ivy.set_backend('paddle')
>>> ivy.set_default_device('cpu)
>>> ivy.default_device()
'cpu'
>>> ivy.set_backend('mxnet')
>>> ivy.set_default_device('cpu')
>>> ivy.default_device()
'cpu'
ivy.set_soft_device_mode(mode)[source]#

Set the mode of whether to move input arrays to ivy.default_device() before performing an operation.

Return type:

None

Parameter#

mode

boolean whether to move input arrays

Examples

>>> ivy.set_soft_device_mode(False)
>>> ivy.soft_device_mode
False
>>> ivy.set_soft_device_mode(True)
>>> ivy.soft_device_mode
True
ivy.set_split_factor(factor, /, *, device=None)[source]#

Set the global split factor for a given device, which can be used to scale batch splitting chunk sizes for the device across the codebase.

Parameters:
  • factor (float) – The factor to set the device-specific split factor to.

  • device (Optional[Union[Device, NativeDevice]], default: None) – The device to set the split factor for. Sets the default device by default.

Return type:

None

Examples

>>> print(ivy.default_device())
cpu
>>> ivy.set_split_factor(0.5)
>>> print(ivy.split_factors)
{'cpu': 0.5}
>>> import torch
>>> ivy.set_backend("torch")
>>> device = torch.device("cuda")
>>> ivy.set_split_factor(0.3, device=device)
>>> print(ivy.split_factors)
{device(type='cuda'): 0.3}
>>> ivy.set_split_factor(0.4, device="tpu")
>>> print(ivy.split_factors)
{'tpu': 0.4}
>>> import torch
>>> ivy.set_backend("torch")
>>> device = torch.device("cuda")
>>> ivy.set_split_factor(0.2)
>>> ivy.set_split_factor(0.3, device='gpu')
>>> print(ivy.split_factors)
{'cpu': 0.2, 'gpu': 0.3}
ivy.split_factor(device=None, /)[source]#

Get a device’s global split factor, which can be used to scale the device’s batch splitting chunk sizes across the codebase.

If the global split factor is set for a given device,

returns the split factor value for the device from the split factors dictionary

If the global split factor for a device is not configured,

returns the default value which is 0.0

Parameters:

device (Optional[Union[Device, NativeDevice]], default: None) – The device to query the split factor for. Sets the default device by default.

Return type:

float

Returns:

ret – The split factor for the specified device.

Examples

>>> x = ivy.split_factor()
>>> print(x)
0.0
>>> y = ivy.split_factor("gpu:0")
>>> print(y)
0.0
ivy.split_func_call(func, inputs, mode, /, *, max_chunk_size=None, chunk_size=None, input_axes=0, output_axes=None, stop_gradients=False, device=None)[source]#

Call a function by splitting its inputs along a given axis, and calling the function in chunks, rather than feeding the entire input array at once. This can be useful to reduce memory usage of the device the arrays are on.

Parameters:
  • func (Callable) – The function to be called.

  • inputs (Union[Array, NativeArray]) – A list of inputs to pass into the function.

  • mode (str) – The mode by which to unify the return values, must be one of [ concat | mean | sum ]

  • max_chunk_size (Optional[int], default: None) – The maximum size of each of the chunks to be fed into the function.

  • chunk_size (Optional[int], default: None) – The size of each of the chunks to be fed into the function. Specifying this arg overwrites the global split factor. Default is None.

  • input_axes (Union[int, Iterable[int]], default: 0) – The axes along which to split each of the inputs, before passing to the function. Default is 0.

  • output_axes (Optional[Union[int, Iterable[int]]], default: None) – The axes along which to concat each of the returned outputs. Default is same as fist input axis.

  • stop_gradients (bool, default: False) – Whether to stop the gradients for each computed return. Default is False.

  • device (Optional[Union[Device, NativeDevice]], default: None) – The device to set the split factor for. Sets the default device by default.

Return type:

Union[Array, NativeArray]

Returns:

ret – The return from the function, following input splitting and re-concattenation.

ivy.to_device(x, device, /, *, stream=None, out=None)[source]#

Move the input array x to the desired device, specified by device string.

Parameters:
  • x (Union[Array, NativeArray]) – input array to be moved to the desired device

  • device (Union[Device, NativeDevice]) – device to move the input array x to

  • stream (Optional[Union[int, Any]], default: None) – stream object to use during copy. In addition to the types supported in array.__dlpack__(), implementations may choose to support any library-specific stream object with the caveat that any code using such an object would not be portable.

  • out (Optional[Array], default: None) – optional output array, for writing the result to. It must have a shape that the inputs broadcast to.

Return type:

Array

Returns:

ret – input array x placed on the desired device

Examples

>>> x = ivy.array([1., 2., 3.])
>>> x = ivy.to_device(x, 'cpu')
>>> print(x.device)
cpu
ivy.total_mem_on_dev(device, /)[source]#

Get the total amount of memory (in GB) for a given device string. In case of CPU, the total RAM is returned.

Parameters:

device (Union[Device, NativeDevice]) – The device string to convert to native device handle.

Return type:

float

Returns:

ret – The total memory on the device in GB.

Examples

>>> x = ivy.total_mem_on_dev("cpu")
>>> print(x)
53.66700032
>>> x = ivy.total_mem_on_dev("gpu:0")
>>> print(x)
8.589934592
ivy.tpu_is_available()[source]#

Determine whether a TPU is available to use, with the backend framework.

Return type:

bool

Returns:

ret – Boolean, as to whether a tpu is available.

Examples

>>> ivy.set_backend("torch")
>>> print(ivy.tpu_is_available())
False
ivy.unset_default_device()[source]#

Reset the default device to “cpu”.

Return type:

None

Examples

>>> ivy.set_default_device("gpu:0")
>>> ivy.default_device()
"gpu:0"
>>> ivy.unset_default_device()
>>> ivy.default_device()
"cpu"
ivy.unset_soft_device_mode()[source]#

Reset the mode of moving input arrays to ivy.default_device() before performing an operation.

Return type:

None

Examples

>>> ivy.set_soft_device_mode(False)
>>> ivy.soft_device_mode
False
>>> ivy.unset_soft_device_mode()
>>> ivy.soft_device_mode
True
ivy.used_mem_on_dev(device, /, *, process_specific=False)[source]#

Get the used memory (in GB) for a given device string. In case of CPU, the used RAM is returned.

Parameters:
  • device (Union[Device, NativeDevice]) – The device string to convert to native device handle.

  • process_specific (bool, default: False) – Whether to check the memory used by this python process alone. Default is False.

Return type:

float

Returns:

ret – The used memory on the device in GB.

Examples

>>> x = ivy.used_mem_on_dev("cpu", process_specific = False)
>>> print(x)
6.219563008
>>> x = ivy.used_mem_on_dev("cpu", process_specific = True)
>>> print(x)
0.902400346
>>> y = ivy.used_mem_on_dev("gpu:0", process_specific = False)
>>> print(y)
0.525205504
class ivy.DefaultDevice(device, /)[source]#

Ivy Device Class.

class ivy.Profiler(save_dir)[source]#

The profiler class is used to profile the execution of some code.

Parameters:

save_dir (str) – The directory to save the profile data to.

This should have hopefully given you an overview of the device submodule, if you have any questions, please feel free to reach out on our discord!