@spec cuda_device_count() :: integer()

Get the number of available CUDA devices.

@spec cuda_is_available() :: boolean()

Check if CUDA is available.

@spec cuda_max_memory_allocated(integer()) :: integer()

Get peak allocated CUDA memory in bytes for a device. Returns -1 if CUDA is not available.

@spec cuda_memory_allocated(integer()) :: integer()

Get currently allocated CUDA memory in bytes for a device. Returns -1 if CUDA is not available.

@spec cuda_memory_reserved(integer()) :: integer()

Get currently reserved CUDA memory in bytes for a device. Returns -1 if CUDA is not available.

@spec data_ptr(t()) :: integer()

Get the raw data pointer of a tensor as an integer.

This is intended for zero-copy tensor exchange with other frameworks (e.g., Nx/Torchx) that share the same process address space.

The returned pointer is only valid as long as the source tensor is alive.

@spec device(t()) :: ExTorch.Device.device()

Get the device of a tensor.

Arguments

• tensor (ExTorch.Tensor): Input tensor

@spec dim(t()) :: integer()

Get the total dimensions of a tensor.

Arguments

• tensor: Input tensor

@spec dtype(t()) :: ExTorch.DType.dtype()

Get the dtype of a tensor.

Arguments

• tensor (ExTorch.Tensor): Input tensor

@spec element_size(t()) :: integer()

Get the element size in bytes for the tensor's dtype.

@spec from_blob(
  integer(),
  tuple(),
  tuple()
) :: t()

See ExTorch.Tensor.from_blob/5

Available signature calls:

• from_blob(ptr, shape, blob_strides)

@spec from_blob(
  integer(),
  tuple(),
  tuple(),
  ExTorch.DType.dtype()
) :: t()

@spec from_blob(
  integer(),
  tuple(),
  tuple(),
  dtype: ExTorch.DType.dtype(),
  device: ExTorch.Device.device()
) :: t()

See ExTorch.Tensor.from_blob/5

Available signature calls:

• from_blob(ptr, shape, blob_strides, kwargs)
• from_blob(ptr, shape, blob_strides, dtype)

@spec from_blob(
  integer(),
  tuple(),
  tuple(),
  ExTorch.DType.dtype(),
  ExTorch.Device.device()
) :: t()

@spec from_blob(
  integer(),
  tuple(),
  tuple(),
  ExTorch.DType.dtype(),
  [{:device, ExTorch.Device.device()}]
) :: t()

Create a tensor from a raw data pointer (zero-copy).

The created tensor shares the memory pointed to by ptr. The caller must keep the source data alive for the lifetime of the returned tensor.

Arguments

• ptr - Raw memory address as an integer.
• shape - Tensor dimensions as a tuple.
• strides - Stride in each dimension as a tuple.
• dtype - Data type of the elements.
• device - Device where the memory lives (default: :cpu).

@spec is_complex(t()) :: boolean()

Returns true if the datatype of input is a complex data type. i.e., one of :complex64 or :complex128

Arguments

• tensor (ExTorch.Tensor): Input tensor.

Examples

# Complex tensors yield true
iex> a = ExTorch.rand({2, 2}, dtype: :complex64)
iex> ExTorch.Tensor.is_complex(a)
true

# Non-complex tensors yield false
iex> b = ExTorch.randint(3, {2, 2}, dtype: :int32)
iex> ExTorch.Tensor.is_complex(b)
false

@spec is_conj(t()) :: boolean()

Returns true if input is a conjugated tensor. i.e., its conjugate bit is set to true.

Arguments

• tensor (ExTorch.Tensor): Input tensor.

Examples

# Complex tensors have the conj bit set to false by default
iex> a = ExTorch.rand({3, 4}, dtype: :complex64)
iex> ExTorch.Tensor.is_conj(a)
false

# Conjugated tensor views have the conj bit set to true
iex> b = ExTorch.conj(a)
iex> ExTorch.Tensor.is_conj(b)
true

# Materialized conjugate tensors have the conj bit set to false
iex> c = ExTorch.resolve_conj(b)
iex> ExTorch.Tensor.is_conj(c)
false

@spec is_contiguous(t()) :: boolean()

Check if a tensor is contiguous in memory.

@spec is_floating_point(t()) :: boolean()

Returns true if the datatype of input is a floating data type. i.e., one of :float16, :float32, :float64 or :bfloat16.

Arguments

• tensor (ExTorch.Tensor): Input tensor.

Examples

# Floating-type tensors yield true
iex> a = ExTorch.rand({2, 2}, dtype: :float16)
iex> ExTorch.Tensor.is_floating_point(a)
true

# Other type of tensors yield false
iex> b = ExTorch.rand({2, 2}, dtype: :complex128)
iex> ExTorch.Tensor.is_floating_point(b)
false

@spec is_nonzero(t()) :: boolean()

Returns true if the input is a single element tensor which is not equal to zero after type conversions.

Arguments

• tensor (ExTorch.Tensor): Input tensor.

Examples

iex> ExTorch.Tensor.is_nonzero(ExTorch.tensor([0.0]))
false
iex> ExTorch.Tensor.is_nonzero(ExTorch.tensor([1.5]))
true
iex> ExTorch.Tensor.is_nonzero(ExTorch.tensor([false]))
false
iex> ExTorch.Tensor.is_nonzero(ExTorch.tensor([3]))
true
iex> ExTorch.Tensor.is_nonzero(ExTorch.tensor([1, 2, 3]))
** (ErlangError) Erlang error: "Boolean value of Tensor with more than one value is ambiguous"
    (extorch 0.1.0-pre0) ExTorch.Native.is_nonzero(#Tensor<
[1, 2, 3]
[size: {3}, dtype: :byte, device: :cpu, requires_grad: false]>)

@spec item(t()) :: ExTorch.Scalar.t()

Returns the value of this tensor as a standard Elixir value.

This only works for tensors with one element. For other cases, see ExTorch.Tensor.to_list/1.

Arguments

• tensor (ExTorch.Tensor): Input tensor.

Examples

iex> x = ExTorch.tensor([false])
iex> ExTorch.Tensor.item(x)
false

iex> x = ExTorch.tensor([-3.5])
iex> ExTorch.Tensor.item(x)
-3.5

iex> x = ExTorch.tensor([ExTorch.Complex.complex(-2, 1)])
iex> ExTorch.Tensor.item(x)
-2.0 + 1.0j

iex> x = ExTorch.tensor([:nan])
iex> ExTorch.Tensor.item(x)
:nan

@spec layout(t()) :: ExTorch.Layout.layout()

Get the layout of a tensor.

Arguments

• tensor (ExTorch.Tensor): Input tensor

@spec memory_format(t()) :: ExTorch.MemoryFormat.memory_format()

Get the memory_format of a tensor.

Arguments

• tensor (ExTorch.Tensor): Input tensor

@spec ndim(t()) :: integer()

Alias to dim/1

@spec ndimension(t()) :: integer()

Alias to dim/1

@spec numel(t()) :: integer()

Returns the total number of elements in the input tensor.

Arguments

• tensor (ExTorch.Tensor): Input tensor.

Examples

iex> x = ExTorch.empty({3, 4, 5})
iex> ExTorch.Tensor.numel(x)
60

@spec repr(t()) :: binary()

See ExTorch.Tensor.repr/2

Available signature calls:

• repr(tensor)

@spec repr(t(),
  edgeitems: integer(),
  linewidth: integer(),
  precision: integer(),
  sci_mode: boolean() | nil,
  threshold: float()
) :: binary()

@spec repr(t(), ExTorch.Utils.PrintOptions.t()) :: binary()

Get a human readable representation of a tensor.

Arguments

• tensor (ExTorch.Tensor): Input tensor

Keyword args

• precision: Number of digits of precision for floating point output. Default: 4

• threshold: Total number of array elements which trigger summarization rather than full repr. Default: 1000.

• edgeitems: Number of array items in summary at beginning and end of each dimension. Default: 3.

• linewidth: The number of characters per line for the purpose of inserting line breaks (default = 80). Thresholded matrices will ignore this parameter.

• sci_mode: Enable (true) or disable (false) scientific notation. If nil (default) is specified, the value is defined by the formatter. This value is automatically chosen by the framework.

@spec requires_grad(t()) :: boolean()

Get the requires_grad status of a tensor.

Arguments

• tensor (ExTorch.Tensor): Input tensor

@spec size(t()) :: tuple()

Get the size of a tensor.

Arguments

• tensor: Input tensor

@spec strides(t()) :: [integer()]

Get the strides of a tensor.

Returns a list of integers representing the step size in each dimension.

@spec to(t()) :: t()

See ExTorch.Tensor.to/6

Available signature calls:

• to(input)

@spec to(
  t(),
  ExTorch.DType.dtype() | nil
) :: t()

@spec to(t(),
  dtype: ExTorch.DType.dtype() | nil,
  device: ExTorch.Device.device() | nil,
  non_blocking: boolean(),
  copy: boolean(),
  memory_format: ExTorch.MemoryFormat.memory_format()
) :: t()

See ExTorch.Tensor.to/6

Available signature calls:

• to(input, kwargs)
• to(input, dtype)

@spec to(
  t(),
  ExTorch.DType.dtype() | nil,
  ExTorch.Device.device() | nil
) :: t()

@spec to(
  t(),
  ExTorch.DType.dtype() | nil,
  device: ExTorch.Device.device() | nil,
  non_blocking: boolean(),
  copy: boolean(),
  memory_format: ExTorch.MemoryFormat.memory_format()
) :: t()

See ExTorch.Tensor.to/6

Available signature calls:

• to(input, dtype, kwargs)
• to(input, dtype, device)

@spec to(
  t(),
  ExTorch.DType.dtype() | nil,
  ExTorch.Device.device() | nil,
  non_blocking: boolean(),
  copy: boolean(),
  memory_format: ExTorch.MemoryFormat.memory_format()
) :: t()

@spec to(
  t(),
  ExTorch.DType.dtype() | nil,
  ExTorch.Device.device() | nil,
  boolean()
) :: t()

See ExTorch.Tensor.to/6

Available signature calls:

• to(input, dtype, device, non_blocking)
• to(input, dtype, device, kwargs)

@spec to(
  t(),
  ExTorch.DType.dtype() | nil,
  ExTorch.Device.device() | nil,
  boolean(),
  boolean()
) :: t()

@spec to(
  t(),
  ExTorch.DType.dtype() | nil,
  ExTorch.Device.device() | nil,
  boolean(),
  copy: boolean(),
  memory_format: ExTorch.MemoryFormat.memory_format()
) :: t()

See ExTorch.Tensor.to/6

Available signature calls:

• to(input, dtype, device, non_blocking, kwargs)
• to(input, dtype, device, non_blocking, copy)

@spec to(
  t(),
  ExTorch.DType.dtype() | nil,
  ExTorch.Device.device() | nil,
  boolean(),
  boolean(),
  [{:memory_format, ExTorch.MemoryFormat.memory_format()}]
) :: t()

@spec to(
  t(),
  ExTorch.DType.dtype() | nil,
  ExTorch.Device.device() | nil,
  boolean(),
  boolean(),
  ExTorch.MemoryFormat.memory_format()
) :: t()

Performs ExTorch.Tensor dtype and/or device conversion.

Arguments

• input (ExTorch.Tensor) - the input tensor to convert.

Optional arguments

• dtype (ExTorch.DType or nil) - the dtype to convert the input tensor into. If nil, then it will be preserved from input. Default: nil.
• device (ExTorch.Device or nil) - the device to move the input tensor into. If nil, then it will be preserved from input. Default: nil.
• non_blocking (boolean) - when true, it tries to convert asynchronously with respect to the host if possible, e.g., converting a CPU Tensor with pinned memory to a CUDA Tensor. Default: false.
• copy (boolean) - If true, a new ExTorch.Tensor is created even when input already matches the desired conversion. Default: false.
• memory_format (ExTorch.MemoryFormat) - the desired memory format of the returned tensor. Default: :preserve_format.

Notes

• If the input already has the correct ExTorch.dtype and ExTorch.device, then input is returned. Otherwise, the returned tensor is a copy of input with the desired dtype and device.
• Unlike PyTorch, to does not accept another tensor as parameter, please use an explicit call to to(input, dtype: other.dtype, device: other.device) instead.

Examples

iex> a = ExTorch.randn({3, 3})
#Tensor<
[[ 0.5770, -0.8079, -0.4308],
 [-0.2186,  0.4031, -1.4976],
 [ 1.2380, -0.4259,  2.0745]]
[
  size: {3, 3},
  dtype: :float,
  device: :cpu,
  requires_grad: false
]>

# Change tensor dtype, preserving device
iex> ExTorch.Tensor.to(a, dtype: :complex64)
#Tensor<
[[ 0.5770+0.j, -0.8079+0.j, -0.4308+0.j],
 [-0.2186+0.j,  0.4031+0.j, -1.4976+0.j],
 [ 1.2380+0.j, -0.4259+0.j,  2.0745+0.j]]
[
  size: {3, 3},
  dtype: :complex_float,
  device: :cpu,
  requires_grad: false
]>

# Change tensor device
iex> ExTorch.Tensor.to(a, device: :cuda)
#Tensor<
[[ 0.5770, -0.8079, -0.4308],
 [-0.2186,  0.4031, -1.4976],
 [ 1.2380, -0.4259,  2.0745]]
[
  size: {3, 3},
  dtype: :float,
  device: {:cuda, 0},
  requires_grad: false
]>

@spec to_list(t()) :: list()

Convert a tensor into a list.

Arguments

• tensor (ExTorch.Tensor): Input tensor