TensorDesc

class tensorrt.plugin.TensorDesc(shape_expr: ShapeExprs | None = None, dtype: DataType | None = None, format: TensorFormat | None = None, scale: float | None = None)

Descriptor for a tensor A TensorDesc never contains nor refers to any tensor data.

Parameters:

  • shape_expr (ShapeExprs) – The data with which to initialize the tensor.

  • dtype (trt.DataType) – The data type of the tensor.

  • format (trt.TensorFormat) – Format (layout) of the tensor.

  • scale (float) – Scale for INT8 data type.

Creates a TensorDesc with constant shape expressions
1tensor = trt.TensorDesc((10, 2, 32, 32), dtype=trt.float32)
Creates a TensorDesc from shape expression of another TensorDesc
1tensor = trt.from_shape_expr(other.shape_expr, dtype=trt.float32)
aliased() → TensorDesc
Returns:

Returns a TensorDesc which has identical properties and is aliased to this tensor (would result in a Tensor during enqueue sharing the same data buffer). Returned TensorDesc is immutable.

Communicate that output tensor has identical properties to the input tensor
1@tensorrt.plugin.register("my::plugin")
2def _(inp: tensorrt.plugin.TensorDesc) -> tensorrt.plugin.TensorDesc:
3    return inp.aliased()
property dtype: DataType

Data type of the tensor.

property format: TensorFormat

The format of the tensor.

Warning

Only accessible when TensorDesc.has_shape is true.

get_aliased() → TensorDesc
Returns:

Returns a TensorDesc for the tensor which this tensor is aliased to. Returns None is this tensor is not aliased to any other tensor.

property has_shape: bool

True if this tensor has concrete shape information, False otherwise.

property has_shape_expr: bool

True if this tensor has symbolic shape expressions, False otherwise.

property is_aliased: bool

True if this tensor is aliased to another tensor, False otherwise.

property is_dynamic: bool

True if this tensor has at least one dynamic dimension, False otherwise.

like() → TensorDesc
Returns:

Returns a TensorDesc which has identical properties to this tensor, and is mutable.

Communicate that output tensor has identical properties to the input tensor
1@tensorrt.plugin.register("my::plugin")
2def _(inp: tensorrt.plugin.TensorDesc) -> tensorrt.plugin.TensorDesc:
3    return inp.like()
property ndim: int

Number of dimensions

numel() → int
Returns:

Returns an int with the number of elements of the tensor.

Warning

Should only be called when TensorDesc.has_shape is true. If a symbolic expression for the number of elements is required, query TensorDesc.shape_expr.numel().

property scale: float

Scale for INT8 data type.

Warning

Only accessible when TensorDesc.has_shape is true.

property shape: Shape

The (concrete) shape of the tensor.

Warning

Only accessible when TensorDesc.has_shape is true.

property shape_expr: ShapeExprs

Symbolic expressions for the tensor shape.

class tensorrt.plugin.SizeTensorDesc(opt: ShapeExpr, upper_bound: ShapeExpr)

Extends TensorDesc

Descriptor for a size tensor: a scalar of either INT32 or INT64 data type used to express the extent of a data-dependent dimension.

Parameters:

  • opt (ShapeExpr) – Symbolic expression for the extent of this size tensor to use in the autotune process of the engine build

  • upper_bound (ShapeExpr) – Symbolic expression for the upper-bound of this size tensor

Note

It is recommended to construct a size tensor using size_tensor() instead of using this constructor directly

expr() → SizeTensorShapeExpr

Symbolic expression for this size tensor

property index: int

Output index at which this size tensor resides

property opt: ShapeExpr

Symbolic expression for the extent of this size tensor to use in the autotune process of the engine build

property upper_bound: ShapeExpr

Symbolic expression for the upper-bound of this size tensor