nvmath-python 1.0.0__cp314-cp314t-win_amd64.whl

nvmath/tensor/contract.py

@@ -0,0 +1,1900 @@
+# Copyright (c) 2025-2026, NVIDIA CORPORATION & AFFILIATES. ALL RIGHTS RESERVED.
+#
+# SPDX-License-Identifier: Apache-2.0
+
+import logging
+from collections.abc import Sequence
+from typing import Any, Literal, cast
+
+import numpy as np
+
+from .. import memory
+from .._internal.layout import is_contiguous_and_dense
+from .._internal.workspace import Workspace
+from .._utils import CudaDataType
+from ..bindings import cutensor
+from ..internal import tensor_wrapper, utils
+from ..internal.typemaps import DATA_TYPE_TO_NAME, NAME_TO_DATA_TYPE
+from ._configuration import ContractionOptions, ExecutionCUDA
+from ._internal import cutensor_utils, einsum_parser
+from ._internal.cutensor_config_ifc import ContractionPlanPreference
+from ._internal.typemaps import get_compute_type_name, get_default_compute_type_from_dtype_name, get_supported_compute_types
+
+__all__ = [
+    "BinaryContraction",
+    "TernaryContraction",
+    "ContractionPlanPreference",
+    "ComputeDesc",
+    "binary_contraction",
+    "ternary_contraction",
+    "Operator",
+    "tensor_qualifiers_dtype",
+]
+
+
+Operator = cutensor.Operator
+
+ComputeDesc = cutensor.ComputeDesc
+
+tensor_qualifiers_dtype = np.int32
+
+# As of cuTensor 2.5.0, only the following operators are supported in the contraction APIs
+OPERATORS_SUPPORTED = {Operator.OP_IDENTITY, Operator.OP_CONJ}
+
+SHARED_CONTRACTION_DOCUMENTATION = utils.COMMON_SHARED_DOC_MAP.copy()
+SHARED_CONTRACTION_DOCUMENTATION.update(
+    {
+        "expr": """\
+The einsum expression to perform the contraction.
+""".replace("\n", " "),
+        #
+        "a": """\
+A tensor representing the first operand to the tensor contraction. The currently supported types
+are :class:`numpy.ndarray`, :class:`cupy.ndarray`, and :class:`torch.Tensor`.""".replace("\n", " "),
+        #
+        "b": """\
+A tensor representing the second operand to the tensor contraction. The currently supported types
+are :class:`numpy.ndarray`, :class:`cupy.ndarray`, and :class:`torch.Tensor`.""".replace("\n", " "),
+        #
+        "c": """\
+A tensor representing the third operand to the tensor contraction. The currently supported types
+are :class:`numpy.ndarray`, :class:`cupy.ndarray`, and :class:`torch.Tensor`.""".replace("\n", " "),
+        #
+        "addend": """\
+(Optional) A tensor representing the operand to add to the tensor contraction result (fused operation in cuTensor).
+The currently supported types are :class:`numpy.ndarray`, :class:`cupy.ndarray`,
+and :class:`torch.Tensor`.""".replace("\n", " "),
+        #
+        "alpha": """\
+The scale factor for the tensor contraction term as a real or complex number. The default is
+:math:`1.0`.""".replace("\n", " "),
+        #
+        "beta": """\
+The scale factor for the tensor addition term as a real or complex number. A value for `beta` must be provided if
+the operand to be added is specified.""".replace("\n", " "),
+        #
+        "qualifiers": """\
+If desired, specify the operators as a :class:`numpy.ndarray` of dtype :class:`~nvmath.tensor.tensor_qualifiers_dtype`
+with the same length as the number of operands in the contraction expression plus one (for the operand to be added).
+All elements must be valid :class:`~nvmath.tensor.Operator` objects. See
+:ref:`matrix-tensor-qualifiers` for the motivation behind qualifiers.""".replace("\n", " "),
+        #
+        "options": """\
+Specify options for the tensor contraction as a :class:`~nvmath.tensor.ContractionOptions` object. Alternatively,
+a `dict` containing the parameters for the ``ContractionOptions`` constructor can also be provided. If not specified, the
+value will be set to the default-constructed ``ContractionOptions`` object.""".replace("\n", " "),
+        #
+        "execution": """\
+Specify execution space options for the tensor contraction as a :class:`ExecutionCUDA` object or a string 'cuda'.
+Alternatively, a `dict` containing 'name' key set to 'cuda' and the additional parameters for the ``ExecutionCUDA``
+constructor can also be provided. If not provided, the execution space will be selected to match operand's storage if
+the operands are on the GPU. If the operands are on the CPU and execution space is not provided, the execution space
+will be a default-constructed :class:`ExecutionCUDA` object with device_id = 0.""".replace("\n", " "),
+        #
+        "out": """\
+(Optional) The output tensor to store the result of the contraction. Must be a :class:`numpy.ndarray`, \
+:class:`cupy.ndarray`, or :class:`torch.Tensor` object and must be on the same device as the input operands. \
+If not specified, the result will be returned on the same device as the input operands.
+
+            .. experimental:: parameter
+
+""".strip(),
+        #
+        "result": """\
+The result of the specified contraction, which remains on the same device and belong to the
+same package as the input operands. """.replace("\n", " "),
+        #
+        "reset_operands_semantics": """\
+Semantics:
+            - Only the operands explicitly passed are updated. At least one operand
+              is required (all of them after :meth:`release_operands`), otherwise
+              a :class:`ValueError` is raised.
+
+            - This method validates each new operand against its corresponding one
+              set during the object's initialization.
+              An operand is compatible if all of the following requirements are met:
+
+              - The shapes, strides, and datatypes match those of the old one.
+              - The package (e.g., cupy, torch, numpy) matches that of the old one.
+              - The memory space (CPU or GPU) matches that of the old one.
+              - The device matches that of the old one, if the operand is on GPU.
+              - The pointer alignment must be the same or a multiple of the old one.
+
+            - If the execution space matches the memory space of the operand:
+              operand's reference is updated with no data copying.
+
+            - If the execution space does not match the memory space of the operand:
+              data is copied between different memory spaces.
+""",
+        #
+        "reset_operands_unchecked": utils._reset_operand_unchecked_docstring(
+            True, version_added="0.9.0", validation_examples="package match, data type match, pointer alignment match"
+        ),
+    }
+)
+
+
+class InvalidContractionState(Exception):
+    pass
+
+
+def _validate_contraction_preconditions(expr, a, b, c, d, qualifiers, options):
+    """
+    Validate preconditions for _ElementaryContraction initialization.
+
+    This function checks all preconditions that can be validated before
+    wrapping operands or allocating resources. It raises exceptions for invalid inputs.
+
+    Args:
+        expr: Einstein expression string
+        a: First input operand
+        b: Second input operand
+        c: Optional third operand (for binary) or offset (for ternary)
+        d: Optional offset operand (for ternary only)
+        qualifiers: Optional qualifiers array
+        options: Optional contraction options
+
+    Returns:
+        tuple: (num_inputs, inputs, output) from parsed expression
+    """
+    # Parse expression to determine number of inputs (validates expression format)
+    inputs, output = einsum_parser.parse_einsum_str(expr)
+    num_inputs = len(inputs)
+
+    # Validate number of inputs, c/d operand consistency with expression type
+    if num_inputs == 2:
+        if d is not None:
+            raise ValueError(f"Binary contraction '{expr}' (2 inputs) cannot have a 'd' operand. ")
+    elif num_inputs == 3:
+        if c is None:
+            raise ValueError(f"Ternary contraction '{expr}' (3 inputs) requires 'c' operand (third multiplicand). ")
+    else:
+        raise NotImplementedError(
+            f"Expression '{expr}' has {num_inputs} inputs. Only binary and ternary contractions are supported."
+        )
+
+    # Validate qualifiers structure (if provided)
+    if qualifiers is not None:
+        try:
+            qualifiers_array = np.asarray(qualifiers, dtype=np.int32)
+        except Exception as e:
+            raise TypeError(f"Qualifiers must be array-like and convertible to int32: {e}") from e
+
+        # Check array length
+        expected_len = num_inputs + 1  # one per input + one for offset
+        if qualifiers_array.size != expected_len:
+            contraction_type = "binary" if num_inputs == 2 else "ternary"
+            operand_names = "a, b, offset" if num_inputs == 2 else "a, b, c, offset"
+            raise ValueError(
+                f"The qualifiers must be a numpy array of length {expected_len} "
+                f"(one per operand: {operand_names}), got {qualifiers_array.size}. "
+                f"Expression '{expr}' is a {contraction_type} contraction."
+            )
+
+        # Validate offset qualifier must be identity
+        if qualifiers_array[num_inputs] != cutensor.Operator.OP_IDENTITY:
+            raise ValueError(f"The operand for the offset must be the identity operator, found {qualifiers_array[num_inputs]}")
+
+        # Validate all input qualifiers are supported operators
+        operand_names = ["a", "b", "c"][:num_inputs]
+        for op_name, qualifier in zip(operand_names, qualifiers_array[:-1], strict=False):
+            if qualifier not in OPERATORS_SUPPORTED:
+                raise ValueError(
+                    f"Each operator must be a valid cuTensor operator, "
+                    f"currently only support {OPERATORS_SUPPORTED}, "
+                    f"got {qualifier} for operand '{op_name}'."
+                )
+
+        # Validate qualifiers against operand dtypes (e.g., OP_CONJ requires complex)
+        operands = [a, b, c][:num_inputs]
+        for op_name, operand, qualifier in zip(operand_names, operands, qualifiers_array[:-1], strict=False):
+            if qualifier == cutensor.Operator.OP_CONJ:
+                # Check if operand has dtype attribute
+                if not hasattr(operand, "dtype"):
+                    raise TypeError(
+                        f"Operand '{op_name}' must be array-like with a dtype attribute "
+                        f"(numpy.ndarray, cupy.ndarray, or torch.Tensor)"
+                    )
+
+                # Check for complex dtype
+                dtype_str = str(operand.dtype)
+                if "complex" not in dtype_str:
+                    raise ValueError(
+                        f"Cannot apply OP_CONJ (conjugate) operator to operand '{op_name}' "
+                        f"with dtype '{operand.dtype}'. Conjugate operator requires complex dtype."
+                    )
+
+    # Validate options structure (if provided)
+    if options is not None:
+        # Extract compute_type from options (could be dict or ContractionOptions object)
+        if isinstance(options, dict):
+            compute_type = options.get("compute_type")
+        else:
+            compute_type = getattr(options, "compute_type", None)
+
+        # Validate compute_type is None or int
+        if compute_type is not None and not isinstance(compute_type, int):
+            raise ValueError(f"Invalid compute type: {compute_type}. compute_type must be None or an integer.")
+
+    return num_inputs, inputs, output
+
+
+@utils.docstring_decorator(SHARED_CONTRACTION_DOCUMENTATION, skip_missing=False)
+class _ElementaryContraction:
+    """
+    Pairwise contraction:
+        out = a @ b + c
+    Ternary contraction:
+        out = a @ b @ c + d
+    """
+
+    def __init__(
+        self,
+        expr,
+        a,
+        b,
+        *,
+        c=None,
+        d=None,
+        out=None,
+        qualifiers=None,
+        options=None,
+        execution: ExecutionCUDA | str | dict[str, Any] | None = None,
+        stream: utils.AnyStream | int | None = None,
+    ):
+        """Binary & Ternary Contraction"""
+
+        # Initialize valid_state to True here because this flag is currenttly only
+        # used to check if the object has been free'd or not.
+        # As far as freeing the object is concerned, upon initialization,
+        # the object is already valid.
+        self.valid_state = True
+
+        # Track whether the user has called release_operands().
+        self._operands_released = False
+
+        # ========================================================================
+        # Validate preconditions right away, if it fails, no state changes will be made
+        # ========================================================================
+        self.num_inputs, inputs, output = _validate_contraction_preconditions(expr, a, b, c, d, qualifiers, options)
+        self.expr = expr
+
+        # ========================================================================
+        # Process options (needed for logging and configuration)
+        # ========================================================================
+        self.options: ContractionOptions = utils.check_or_create_options(  # type: ignore[assignment]
+            ContractionOptions, options, "elementary contraction options"
+        )
+        self.logger = self.options.logger if self.options.logger is not None else logging.getLogger()
+        log_info = self.logger.isEnabledFor(logging.INFO)
+        log_debug = self.logger.isEnabledFor(logging.DEBUG)
+
+        # ========================================================================
+        # Wrap and validate operands (a, b, and c, d if needed)
+        # ========================================================================
+        self.a, self.b = tensor_wrapper.wrap_operands([a, b])
+        self.input_operand_class = self.a.__class__
+        self.input_package = utils.get_operands_package([self.a, self.b])
+
+        # Determine internal package (numpy -> cuda conversion)
+        if self.input_package == "numpy":
+            self.internal_package = "cuda"
+        else:
+            self.internal_package = self.input_package
+        tensor_wrapper.maybe_register_package(self.internal_package)
+
+        # Wrap optional operands c, d and validate package consistency
+        wrapped_operands = [self.a, self.b]
+        self.c_provided = c is not None
+        self.d_provided = d is not None
+
+        self.c: tensor_wrapper.TensorHolder[Any] | None
+        self.d: tensor_wrapper.TensorHolder[Any] | None
+        for op_name, op in zip(["c", "d"], [c, d], strict=False):
+            if op is not None:
+                op = tensor_wrapper.wrap_operand(op)
+                if op.name != self.input_package:
+                    raise ValueError(
+                        f"operand has package '{op.name}' but expected '{self.input_package}'. "
+                        f"All operands must be from the same tensor package."
+                    )
+                wrapped_operands.append(op)
+            setattr(self, op_name, op)
+
+        # ========================================================================
+        # Setup qualifiers
+        # ========================================================================
+        # If here, preconditions were validated, we can safely create the qualifiers
+        # or use the ones provided.
+        if qualifiers is None:
+            self.qualifiers = np.full(self.num_inputs + 1, cutensor.Operator.OP_IDENTITY, dtype=np.int32)  # size of enum value
+        else:
+            # validation of qualifiers done during preconditions' check
+            self.qualifiers = np.asarray(qualifiers, dtype=np.int32)
+
+        # ========================================================================
+        # Setup execution environment and device management
+        # ========================================================================
+        self.input_device_id = utils.get_operands_device_id(wrapped_operands)
+        self.blocking = self.options.blocking is True or self.input_device_id == "cpu"
+        if self.blocking:
+            self.call_prologue = "This call is blocking and will return only after the operation is complete."
+        else:
+            self.call_prologue = (
+                "This call is non-blocking and will return immediately after the operation is launched on the device."
+            )
+
+        self.execution = utils.check_or_create_one_of_options(
+            (ExecutionCUDA,),
+            execution,
+            "execution options",
+            default_name="cuda",
+        )
+
+        if log_info:
+            self.logger.info(
+                f"The input tensor's memory space is {self.input_device_id}, and the execution space "
+                f"is device {self.execution.device_id}."
+            )
+
+        # Determine execution device and create stream
+        if self.input_device_id == "cpu":
+            self.execution_device_id = self.execution.device_id
+            stream_holder = utils.get_or_create_stream(self.execution_device_id, stream, self.internal_package)
+            # Transfer CPU operands to execution device
+            self.a = self.a.to(self.execution_device_id, stream_holder)
+            self.b = self.b.to(self.execution_device_id, stream_holder)
+            if self.c is not None:
+                self.c = self.c.to(self.execution_device_id, stream_holder)
+            if self.d is not None:
+                self.d = self.d.to(self.execution_device_id, stream_holder)
+            if log_debug:
+                self.logger.debug(f"The input tensors have been copied to the execution device: {self.execution_device_id}.")
+        else:
+            self.execution_device_id = self.input_device_id
+            stream_holder = utils.get_or_create_stream(self.execution_device_id, stream, self.internal_package)
+
+        # ========================================================================
+        # Determine data types and compute configuration
+        # ========================================================================
+        # TODO: cutensor supports R_64F (A) C_64F (B) C_64F (C) combination (and inverse)
+        # https://docs.nvidia.com/cuda/cutensor/latest/api/cutensor.html#cutensorcreatecontraction
+        self.data_type = utils.get_operands_dtype(wrapped_operands)
+        self.cuda_data_type = NAME_TO_DATA_TYPE[self.data_type]
+
+        # Parse compute descriptor
+        if self.options.compute_type is None:
+            self.compute_type = get_default_compute_type_from_dtype_name(self.data_type)
+        else:
+            # Type validation done in preconditions, here check compatibility with data type
+            if self.options.compute_type not in get_supported_compute_types(self.data_type):
+                raise ValueError(f"Invalid compute type: {self.options.compute_type} for data type: {self.data_type}")
+            self.compute_type = self.options.compute_type
+            if self.compute_type == ComputeDesc.COMPUTE_8XINT8():
+                # TODO: remove the check once cutensor requirement is bumped to 2.6.0
+                version = cutensor.get_version()
+                if version == 20500:
+                    raise RuntimeError(
+                        "The 8XINT8 compute type is not supported in cuTensor 2.5.0 due to a known bug. "
+                        "Please upgrade cuTensor to a later version."
+                    )
+
+        if log_info:
+            self.logger.info(f"The compute type is: {get_compute_type_name(self.compute_type)}.")
+            qualifiers_message = (
+                f"The contraction qualifiers are: "
+                f"A = {cutensor.Operator(self.qualifiers[0]).name}, "
+                f"B = {cutensor.Operator(self.qualifiers[1]).name}, "
+                f"C = {cutensor.Operator(self.qualifiers[2]).name}"
+            )
+            if self.num_inputs == 3:
+                qualifiers_message += f", D = {cutensor.Operator(self.qualifiers[3]).name}"
+            self.logger.info(qualifiers_message)
+        # ========================================================================
+        # Parse einsum modes and setup output tensor
+        # ========================================================================
+        self.input_modes, self.output_modes, _, size_dict = einsum_parser.parse_elementary_einsum(
+            inputs, output, self.a, self.b, self.c
+        )[:4]
+
+        self.output_shape = [size_dict[mode] for mode in self.output_modes]
+
+        # self.out is the output tensor that will be used for the execution
+        # self.out_return is the output tensor that will be returned by the execute method
+        self.output_provided = out is not None
+        self.result_strides: Sequence[int] | None = None
+        self.result_layout: Literal["optimized", "natural", "C", "F"] | None = None
+
+        if self.output_provided:
+            out = tensor_wrapper.wrap_operand(out)
+            if out.name != self.input_package:
+                raise ValueError(f"The output operand out must be a {self.input_package} tensor")
+            if out.device_id != self.input_device_id:
+                raise ValueError("The output operand out must be on the same device as the input operands.")
+            self.out_return = out
+            if out.device_id == self.execution_device_id:
+                self.out = out
+            else:
+                self.out = out.to(self.execution_device_id, stream_holder)
+                if log_debug:
+                    self.logger.debug(f"The output tensor is copied to the execution device: {self.execution_device_id}.")
+        else:
+            self.out = self.out_return = None
+
+            # Determine the result layout to use
+            if self.options.result_layout == "auto":
+                if self.num_inputs == 2 and c is None:
+                    self.result_layout = "optimized"
+                elif self.num_inputs == 3 and d is None:
+                    # Optimized layout is binary-only, so ternary defaults to row-major.
+                    self.result_layout = "C"
+                else:
+                    # Follow the addend layout to satisfy cuTENSOR stride requirements.
+                    self.result_layout = "natural"
+            else:
+                self.result_layout = cast(Literal["C", "F", "optimized"], self.options.result_layout)
+                if self.result_layout == "optimized":
+                    if self.num_inputs == 3:
+                        raise ValueError("Optimized result layout is not supported for ternary contraction.")
+                    elif c is not None:
+                        raise ValueError("Optimized result layout is not supported when an addend is specified.")
+
+            # Compute the result strides based on the result layout
+            if self.result_layout == "optimized":
+                assert self.num_inputs == 2, "Internal Error: optimized layout unexpectedly used for non-binary contraction."
+                self.result_strides = cutensor_utils.get_optimized_strides(
+                    self.input_modes, self.output_modes, self.output_shape
+                )
+                self.logger.debug(f"The optimized result strides are: {self.result_strides}")
+            elif self.result_layout in {"C", "F"}:
+                self.result_strides = cutensor_utils.compute_strides(self.output_shape, order=self.result_layout)
+            elif self.result_layout == "natural":
+                if self.num_inputs == 2:
+                    assert self.c is not None
+                    self.result_strides = self.c.strides
+                else:
+                    assert self.d is not None
+                    self.result_strides = self.d.strides
+                if not is_contiguous_and_dense(self.output_shape, self.result_strides):
+                    raise ValueError(
+                        "When output is not specified and an addend is provided, the addend must be contiguous in memory."
+                    )
+            else:
+                raise AssertionError("Internal Error: Invalid result layout.")
+
+        # ========================================================================
+        # Setup memory management
+        # ========================================================================
+        self.allocator = (
+            self.options.allocator
+            if self.options.allocator is not None
+            else memory._MEMORY_MANAGER[self.internal_package](self.execution_device_id, self.logger)
+        )
+        self.memory_limit = utils.get_memory_limit_from_device_id(self.options.memory_limit, self.execution_device_id)
+        self.tensor_descs = {}
+
+        # ========================================================================
+        # Create cuTensor handle and descriptors
+        # ========================================================================
+        with utils.device_ctx(self.execution_device_id):
+            if self.options.handle is not None:
+                self.own_handle = False
+                self.handle = self.options.handle
+                self.logger.info(f"The library handle has been set to the specified value: {self.handle}.")
+            else:
+                self.own_handle = True
+                self.handle = cutensor.create()
+                self.logger.info(f"The library handle has been created: {self.handle}.")
+
+        operands_names: tuple[str, ...]
+        if self.num_inputs == 2:
+            addend_name = "c"
+            operands_names = ("a", "b", "out") if c is None else ("a", "b", "c", "out")
+        else:
+            addend_name = "d"
+            operands_names = ("a", "b", "c", "out") if d is None else ("a", "b", "c", "d", "out")
+
+        # Create tensor descriptors for all relevant operands
+        for op_name in operands_names:
+            op = getattr(self, op_name)
+            if op is None:
+                assert op_name == "out", "Internal Error: out should be None if not provided."
+                # NOTE: Even when result_strides are not C order, the output pointer
+                # alignment (corresponding to the first element of the output tensor)
+                # from cupy/torch/ndbuffer's default allocator are still always
+                # aligned to at least 256 bytes.
+                self.tensor_descs[op_name] = cutensor_utils.TensorDescriptor.from_metadata(
+                    self.handle, self.output_shape, self.data_type, strides=self.result_strides
+                )
+            else:
+                self.tensor_descs[op_name] = cutensor_utils.TensorDescriptor.from_tensor_holder(self.handle, op)
+            if log_debug:
+                self.logger.debug(
+                    f"The tensor descriptor for operand {op_name} with shape {self.tensor_descs[op_name].shape}, "
+                    f"strides {self.tensor_descs[op_name].strides}, dtype {self.tensor_descs[op_name].dtype}, "
+                    f"and pointer alignment {self.tensor_descs[op_name].alignment} has been created."
+                )
+
+        if addend_name not in operands_names:
+            # If addend is not specified, we can reuse the output descriptor for the addend
+            self.tensor_descs[addend_name] = self.tensor_descs["out"]
+
+        # ========================================================================
+        # Create contraction descriptor
+        # ========================================================================
+        if self.num_inputs == 2:
+            self.contraction_desc = cutensor.create_contraction(
+                self.handle,
+                self.tensor_descs["a"].ptr,
+                self.input_modes[0],
+                self.qualifiers[0],
+                self.tensor_descs["b"].ptr,
+                self.input_modes[1],
+                self.qualifiers[1],
+                self.tensor_descs["c"].ptr,
+                self.output_modes,  # NOTE: currently assuming c has the same output modes as the out
+                self.qualifiers[2],  # only identity operator is supported for c
+                self.tensor_descs["out"].ptr,
+                self.output_modes,
+                self.compute_type,
+            )
+        else:
+            self.contraction_desc = cutensor.create_contraction_trinary(
+                self.handle,
+                self.tensor_descs["a"].ptr,
+                self.input_modes[0],
+                self.qualifiers[0],
+                self.tensor_descs["b"].ptr,
+                self.input_modes[1],
+                self.qualifiers[1],
+                self.tensor_descs["c"].ptr,
+                self.input_modes[2],
+                self.qualifiers[2],
+                self.tensor_descs["d"].ptr,
+                self.output_modes,
+                self.qualifiers[3],  # only identity operator is supported for d
+                self.tensor_descs["out"].ptr,
+                self.output_modes,
+                self.compute_type,
+            )
+
+        # ========================================================================
+        # Query scalar type and initialize planning/workspace state
+        # ========================================================================
+        scalar_dtype = cutensor.get_operation_descriptor_attribute_dtype(cutensor.OperationDescriptorAttribute.SCALAR_TYPE)
+        scalar_dtype_buffer = np.empty(1, dtype=scalar_dtype)
+        cutensor.operation_descriptor_get_attribute(
+            self.handle,
+            self.contraction_desc,
+            cutensor.OperationDescriptorAttribute.SCALAR_TYPE,
+            scalar_dtype_buffer.ctypes.data,
+            scalar_dtype_buffer.itemsize,
+        )
+        self.scalar_type = CudaDataType(scalar_dtype_buffer.item())
+
+        self.alpha = np.empty(1, dtype=DATA_TYPE_TO_NAME[self.scalar_type])
+        self.beta = np.empty(1, dtype=DATA_TYPE_TO_NAME[self.scalar_type])
+
+        # Initialize planning-related members
+        self.contraction_planned = False
+        self.plan_preference_ptr = cutensor.create_plan_preference(self.handle, cutensor.Algo.DEFAULT, cutensor.JitMode.NONE)
+        self._plan_preference = ContractionPlanPreference(self)
+
+        # Initialize remaining members
+        # Workspace lifecycle is managed by `Workspace` (allocate-on-demand,
+        # reuse, exception cleanup, stream-ordered free).
+        self.workspace = Workspace(
+            self.allocator,
+            self.logger,
+            label="contraction workspace",
+            device_id=self.execution_device_id,
+        )
+        self.last_compute_event = None
+        self.plan_ptr = None
+
+        if log_info:
+            self.logger.info(f"The {self.__class__.__name__} object has been created.")
+
+    def _check_valid_contraction(self, *args, **kwargs):
+        """
+        Check if the ElementaryContraction object is alive and well.
+        """
+        if not self.valid_state:
+            raise InvalidContractionState("The ElementaryContraction object cannot be used after resources are free'd")
+
+    def _check_valid_operands(self, *args, **kwargs):
+        """
+        Check if the operands are available for the operation.
+        """
+        what = kwargs["what"]
+        if self._operands_released:
+            raise RuntimeError(
+                f"{what} cannot be performed after the operands have been released. "
+                f"Use reset_operands() to provide new operands before performing the {what.lower()}."
+            )
+
+    def _free_plan_resources(self, exception: Exception | None = None) -> bool:
+        """
+        Free resources allocated in planning.
+        """
+
+        if self.plan_ptr is not None:
+            cutensor.destroy_plan(self.plan_ptr)
+            self.plan_ptr = None
+
+        if self.plan_preference_ptr is not None:
+            cutensor.destroy_plan_preference(self.plan_preference_ptr)
+            self.plan_preference_ptr = None
+
+        self.contraction_planned = False
+        return True
+
+    def _check_planned(self, *args, **kwargs):
+        what = kwargs["what"]
+        if not self.contraction_planned:
+            raise RuntimeError(f"{what} cannot be performed before plan() has been called.")
+
+    @property
+    def plan_preference(self):
+        """
+        An accessor to configure or query the contraction planning phase
+        attributes.
+
+        Returns:
+            A :class:`ContractionPlanPreference` object, whose attributes can be set (or
+            queried) to configure the planning phase.
+
+        .. seealso::
+            :class:`ContractionPlanPreference`, :meth:`plan`.
+        """
+        return self._plan_preference
+
+    @utils.precondition(_check_valid_contraction)
+    def reset_operands(self, *, a=None, b=None, c=None, d=None, out=None, stream: utils.AnyStream | int | None = None):
+        if self.num_inputs == 2 and d is not None:
+            raise RuntimeError("Internal Error: For pairwise contractions, d can not be set.")
+
+        # if the operands have been released, all operands must be provided
+        if self._operands_released:
+            all_provided = (
+                a is not None
+                and b is not None
+                and (c is not None or not self.c_provided)
+                and (d is not None or not self.d_provided)
+                and (out is not None or not self.output_provided)
+            )
+            if not all_provided:
+                raise ValueError("After release_operands(), all operands must be provided.")
+        elif all(arg is None for arg in (a, b, c, d, out)):
+            # All arguments are None: there is nothing to update, so reject the call.
+            raise ValueError("reset_operands() requires at least one operand to be provided.")
+
+        stream_holder = None  # lazy initialization
+        for op_name, op in zip(["a", "b", "c", "d", "out"], [a, b, c, d, out], strict=False):
+            # if op is None, we keep the original value of the operand,
+            # so skip the rest of the logic for this operand
+            if op is None:
+                continue
+
+            if (op_name == "c" and not self.c_provided) or (op_name == "d" and not self.d_provided):
+                raise ValueError(
+                    f"operand {op_name} was not specified during the initialization "
+                    f"of the ElementaryContraction object and therefore can not be reset "
+                    f"to a concrete tensor."
+                )
+            op = tensor_wrapper.wrap_operand(op)
+            if op.name != self.input_package:
+                raise ValueError(f"The operand {op_name} must be a {self.input_package} tensor")
+            if op.device_id != self.input_device_id:
+                raise ValueError(
+                    f"The operand {op_name} must be on the same device "
+                    f"as the operands provided during the initialization of the "
+                    f"ElementaryContraction object."
+                )
+
+            tensor_desc = self.tensor_descs[op_name]
+
+            error_pattern = (
+                f"The operand {op_name} must have the same {{attr}} "
+                f"as the one specified during the initialization of the "
+                f"ElementaryContraction object."
+            )
+
+            if tensor_desc.shape != op.shape:
+                raise ValueError(error_pattern.format(attr="shape"))
+
+            if tensor_desc.dtype != op.dtype:
+                raise ValueError(error_pattern.format(attr="dtype"))
+
+            if tensor_desc.strides != op.strides:
+                raise ValueError(error_pattern.format(attr="strides"))
+
+            if op_name == "out":
+                self.out_return = op
+
+            if op.device_id != self.execution_device_id:
+                if stream_holder is None:
+                    stream_holder = utils.get_or_create_stream(self.execution_device_id, stream, self.internal_package)
+                if op_name == "out" and self.out is not None:
+                    # if out_name is "out" and we own a valid self.out,
+                    # we can directly reuse it here
+                    op = self.out
+                else:
+                    op = op.to(self.execution_device_id, stream_holder)
+            elif cutensor_utils.compute_pointer_alignment(op.data_ptr) % self.tensor_descs[op_name].alignment:
+                # If the operand is in the same memory space as the
+                # execution space (copy not needed), we need to check if the pointer
+                # alignment is the same or a multiple of the original pointer alignments.
+                # If the operand is in host memory, this check can be skipped as we
+                # internally copy the operand to the execution space and
+                # device ptr alignment should be guaranteed to be the same.
+                raise ValueError(
+                    f"The pointer alignment of the operand {op_name} must be the same or a multiple of the corresponding "
+                    f"pointer alignment specified during the initialization of the {self.__class__.__name__} object."
+                )
+
+            setattr(self, op_name, op)
+            self.logger.info(f"operand {op_name} has been reset to the new operand provided.")
+        self._operands_released = False
+        return
+
+    @utils.precondition(_check_valid_contraction)
+    def _release_operands(self):
+        if self._operands_released:
+            self.logger.info("Operands have already been released; nothing to do.")
+            return
+
+        # We release the internal tensor references held by the wrappers
+        # for the user-provided operands and/or their GPU mirrors.
+        # The TensorHolder wrappers themselves are kept alive so that
+        # reset_operands_unchecked can reuse them without re-wrapping.
+        self.a.tensor = None
+        self.b.tensor = None
+        if self.c_provided:
+            self.c.tensor = None
+        if self.d_provided:
+            self.d.tensor = None
+        if self.output_provided:
+            # self.out_return always holds the user's tensor.
+            # self.out holds the user's tensor when operand memory space
+            # matches execution (same device; same object as self.out_return),
+            # or an internal device mirror when they differ (inputs on CPU).
+            # In both cases, release the inner tensor references.
+            self.out_return.tensor = None
+            self.out.tensor = None
+        self._operands_released = True
+        self.logger.info("User-provided operands have been released.")
+
+    def _reset_operands_unchecked(
+        self, *, a=None, b=None, c=None, d=None, out=None, stream: utils.AnyStream | int | None = None
+    ):
+        # Since we have the caller's compatibility guarantee for the inputs, we can leverage
+        # the metadata stored during initialization to know if the inputs were on the GPU.
+        # If the memory space of the inputs matches the execution space, namely the inputs
+        # during initialization resided on the GPU, then we know/expect the newly provided
+        # inputs must also reside on the GPU. The TensorHolder wrappers are always alive
+        # (release_operands only clears .tensor), so we can unconditionally swap the inner
+        # tensor — no wrap_operand() needed.
+        if self.input_device_id != "cpu":
+            if a is not None:
+                self.a.tensor = a
+            if b is not None:
+                self.b.tensor = b
+            if c is not None:
+                assert self.c is not None
+                self.c.tensor = c
+            if d is not None:
+                assert self.d is not None
+                self.d.tensor = d
+            if out is not None:
+                self.out.tensor = out
+                self.out_return.tensor = out
+
+            self._operands_released = False
+            return
+
+        # if we are here, it means the inputs were on the CPU
+        # so we need to distinguish between the case where the operands
+        # were released and the case where they were not released.
+        if self._operands_released:
+            # CPU inputs after release: device mirrors were freed, need re-allocation.
+            stream_holder = utils.get_or_create_stream(self.execution_device_id, stream, self.internal_package)
+            if a is not None:
+                self.a = tensor_wrapper.wrap_operand(a).to(self.execution_device_id, stream_holder)
+            if b is not None:
+                self.b = tensor_wrapper.wrap_operand(b).to(self.execution_device_id, stream_holder)
+            if c is not None:
+                self.c = tensor_wrapper.wrap_operand(c).to(self.execution_device_id, stream_holder)
+            if d is not None:
+                self.d = tensor_wrapper.wrap_operand(d).to(self.execution_device_id, stream_holder)
+            if out is not None:
+                out_wrapped = tensor_wrapper.wrap_operand(out)
+                self.out = out_wrapped.to(self.execution_device_id, stream_holder)
+                self.out_return = out_wrapped
+        else:
+            # CPU inputs, not released: device mirrors are valid, copy into them.
+            stream_holder = utils.get_or_create_stream(self.execution_device_id, stream, self.internal_package)
+            if a is not None:
+                self.a.copy_(tensor_wrapper.wrap_operand(a), stream_holder=stream_holder)
+            if b is not None:
+                self.b.copy_(tensor_wrapper.wrap_operand(b), stream_holder=stream_holder)
+            if c is not None:
+                assert self.c is not None
+                self.c.copy_(tensor_wrapper.wrap_operand(c), stream_holder=stream_holder)
+            if d is not None:
+                assert self.d is not None
+                self.d.copy_(tensor_wrapper.wrap_operand(d), stream_holder=stream_holder)
+            if out is not None:
+                out_wrapped = tensor_wrapper.wrap_operand(out)
+                self.out.copy_(out_wrapped, stream_holder=stream_holder)
+                self.out_return = out_wrapped
+
+        self._operands_released = False
+
+    @utils.precondition(_check_valid_contraction)
+    @utils.atomic(_free_plan_resources, method=True)
+    def plan(self, *, stream: utils.AnyStream | int | None = None):
+        """
+        Plan the tensor contraction. The planning phase can be optionally configured through
+        the property :attr:`plan_preference` (an object of type
+        :class:`ContractionPlanPreference`).
+
+        Args:
+            stream: {stream}
+
+        .. seealso::
+            :attr:`plan_preference`, :class:`ContractionPlanPreference`.
+
+        Note:
+            If the :attr:`plan_preference` has been updated, a :meth:`plan` call is
+            required to apply the changes.
+        """
+        log_info = self.logger.isEnabledFor(logging.INFO)
+        log_debug = self.logger.isEnabledFor(logging.DEBUG)
+
+        # A new plan needs to be created at each plan() call
+        if self.plan_ptr is not None:
+            cutensor.destroy_plan(self.plan_ptr)
+            self.plan_ptr = None
+            if log_debug:
+                self.logger.debug("The previous plan has been destroyed.")
+
+        if log_info:
+            self.logger.info("= PLANNING PHASE =")
+        stream_holder = utils.get_or_create_stream(self.execution_device_id, stream, self.internal_package)
+        required_workspace_size_buffer = np.empty(
+            1, dtype=cutensor.get_plan_attribute_dtype(cutensor.PlanAttribute.REQUIRED_WORKSPACE)
+        )
+
+        with utils.cuda_call_ctx(stream_holder, self.blocking, timing=log_info) as (
+            self.last_compute_event,
+            elapsed,
+        ):
+            self.plan_ptr = cutensor.create_plan(
+                self.handle, self.contraction_desc, self.plan_preference_ptr, self.memory_limit
+            )
+            cutensor.plan_get_attribute(
+                self.handle,
+                self.plan_ptr,
+                cutensor.PlanAttribute.REQUIRED_WORKSPACE,
+                required_workspace_size_buffer.ctypes.data,
+                required_workspace_size_buffer.itemsize,
+            )
+
+        if log_info and elapsed.data is not None:
+            self.logger.info(f"The planning phase took {elapsed.data:.3f} ms to complete.")
+
+        self.workspace.set_size(required_workspace_size_buffer.item())
+        if log_info:
+            self.logger.info(f"The required workspace size for the contraction operation is {self.workspace.size}.")
+        self.contraction_planned = True
+
+    @utils.precondition(_check_valid_contraction)
+    @utils.precondition(_check_planned, "Execution")
+    @utils.precondition(_check_valid_operands, "Execution")
+    def execute(self, *, alpha=1.0, beta=None, release_workspace=False, stream: utils.AnyStream | int | None = None):
+        """
+        Execute a prepared tensor contraction.
+
+        Args:
+            alpha: {alpha}
+
+            beta: {beta}
+
+            release_workspace: {release_workspace}
+
+            stream: {stream}
+
+        Returns:
+           {result}
+        """
+        if beta is None:
+            if self.num_inputs == 2 and self.c is not None:
+                raise ValueError("beta must be set when c is specified in a binary contraction")
+            elif self.num_inputs == 3 and self.d is not None:
+                raise ValueError("beta must be set when d is specified in a ternary contraction")
+            beta = 0.0
+        else:
+            if self.num_inputs == 2 and self.c is None:
+                raise ValueError("For binary contraction, beta can only be set if c is specified")
+            elif self.num_inputs == 3 and self.d is None:
+                raise ValueError("For ternary contraction, beta can only be set if d is specified")
+
+        log_info = self.logger.isEnabledFor(logging.INFO)
+        log_debug = self.logger.isEnabledFor(logging.DEBUG)
+
+        self.alpha[0] = alpha
+        self.beta[0] = beta
+
+        stream_holder = utils.get_or_create_stream(self.execution_device_id, stream, self.internal_package)
+
+        # The buffer to be used for cutensor execution
+        # If out was provided during initialization, this would have been created
+        # regardless of where it resided (CPU or GPU).
+        # If out was not provided during initialization, we need to create a new buffer
+        if self.out is None:
+            assert not self.output_provided, (
+                "Internal Error: out cannot be None if the output was provided during initialization."
+            )
+            self.out = self.a.__class__.empty(
+                self.output_shape,
+                device_id=self.execution_device_id,
+                dtype=self.data_type,
+                stream_holder=stream_holder,
+                strides=self.result_strides,
+            )
+            alignment = cutensor_utils.compute_pointer_alignment(self.out.data_ptr)
+            if alignment % self.tensor_descs["out"].alignment:
+                raise RuntimeError(
+                    f"Output tensor pointer alignment is incompatible with the contraction plan: expected a multiple "
+                    f"of {self.tensor_descs['out'].alignment} bytes, but detected {alignment} bytes at execution time. "
+                    f"This can happen if the default allocator for {self.internal_package} was changed to one that "
+                    f"does not provide the alignment assumed by the contraction plan."
+                )
+
+            if log_debug:
+                self.logger.debug(
+                    f"The output tensor of type {type(self.out.tensor)} with shape {self.out.shape} has been created."
+                )
+
+        if log_info:
+            self.logger.info("= EXECUTION PHASE =")
+            self.logger.info("Starting tensor contraction calculation...")
+            self.logger.info(f"{self.call_prologue}")
+            self.logger.info(f"The specified stream for execute() is {stream_holder.obj}.")
+
+        with self.workspace.allocate_perhaps(
+            stream_holder,
+            get_last_event=lambda: self.last_compute_event,
+        ) as workspace:
+            with utils.cuda_call_ctx(stream_holder, self.blocking, timing=log_info) as (
+                self.last_compute_event,
+                elapsed,
+            ):
+                if self.num_inputs == 2:
+                    cutensor.contract(
+                        self.handle,
+                        self.plan_ptr,
+                        self.alpha.ctypes.data,
+                        self.a.data_ptr,
+                        self.b.data_ptr,
+                        self.beta.ctypes.data,
+                        self.c.data_ptr if self.c is not None else self.out.data_ptr,
+                        self.out.data_ptr,
+                        workspace.raw_ptr,
+                        workspace.size,
+                        stream_holder.ptr,
+                    )
+                else:
+                    assert self.c is not None, "Internal Error: ternary contraction requires operand c."
+                    cutensor.contract_trinary(
+                        self.handle,
+                        self.plan_ptr,
+                        self.alpha.ctypes.data,
+                        self.a.data_ptr,
+                        self.b.data_ptr,
+                        self.c.data_ptr,
+                        self.beta.ctypes.data,
+                        self.d.data_ptr if self.d is not None else self.out.data_ptr,
+                        self.out.data_ptr,
+                        workspace.raw_ptr,
+                        workspace.size,
+                        stream_holder.ptr,
+                    )
+
+            if release_workspace:
+                workspace.release(self.last_compute_event)
+                self.last_compute_event = None
+
+        if log_info and elapsed.data is not None:
+            self.logger.info(f"The tensor contraction calculation took {elapsed.data:.3f} ms to complete.")
+
+        if self.output_provided:
+            if self.execution_device_id != self.input_device_id:
+                self.out_return.copy_(self.out, stream_holder)
+            return self.out_return.tensor
+        else:
+            if self.execution_device_id != self.input_device_id:
+                return self.out.to(self.input_device_id, stream_holder).tensor
+            else:
+                out = self.out
+                # release the output tensor as the ownership is transferred to the caller
+                self.out = None
+                return out.tensor
+
+    @utils.precondition(_check_valid_contraction)
+    def free(self):
+        """Free tensor contraction resources.
+
+        It is recommended that the contraction object be used
+        within a context, but if it is not possible then this method must be
+        called explicitly to ensure that the tensor contraction resources
+        (especially internal library objects) are properly cleaned up.
+        """
+
+        if not self.valid_state:
+            return
+
+        class_name = self.__class__.__name__
+        try:
+            self.workspace.release(self.last_compute_event)
+            self.last_compute_event = None
+
+            self._free_plan_resources()
+
+            # Free handle if we own it.
+            if self.handle is not None and self.own_handle:
+                cutensor.destroy(self.handle)
+                self.handle, self.own_handle = None, False
+
+            if self.contraction_desc is not None:
+                cutensor.destroy_operation_descriptor(self.contraction_desc)
+                self.contraction_desc = None
+
+            self.tensor_descs = None
+
+            # Set all attributes to None except for logger and valid_state
+            _keep = {"logger", "valid_state"}
+            for attr in list(vars(self)):
+                if attr not in _keep:
+                    setattr(self, attr, None)
+
+        except Exception as e:
+            self.logger.critical(f"Internal error: only part of the {class_name} object's resources have been released.")
+            self.logger.critical(str(e))
+            raise e
+        finally:
+            self.valid_state = False
+
+        self.logger.info(f"The {class_name} object's resources have been released.")
+
+    def __enter__(self):
+        return self
+
+    def __exit__(self, exc_type, exc_value, traceback):
+        self.free()
+
+
+@utils.docstring_decorator(SHARED_CONTRACTION_DOCUMENTATION, skip_missing=False)
+class BinaryContraction(_ElementaryContraction):
+    """
+    Create a stateful object encapsulating the specified binary tensor contraction
+    :math:`\\alpha a @ b + \\beta c` and the required resources to perform the operation.
+    A stateful object can be used to amortize the cost of preparation (planning in the
+    case of binary tensor contraction) across multiple executions (also see the
+    :ref:`Stateful APIs<host api types>` section).
+
+    The function-form API :func:`binary_contraction` is a convenient alternative to using
+    stateful objects for *single* use (the user needs to perform just one tensor
+    contraction, for example), in which case there is no possibility of amortizing
+    preparatory costs. The function-form APIs are just convenience wrappers around
+    the stateful object APIs.
+
+    Using the stateful object typically involves the following steps:
+
+    1. **Problem Specification**: Initialize the object with a defined operation and
+       options.
+    2. **Preparation**: Use :meth:`plan` to determine the best algorithmic implementation
+       for this specific binary tensor contraction operation.
+    3. **Execution**: Perform the tensor contraction computation with :meth:`execute`.
+    4. **Resource Management**: Ensure all resources are released either by explicitly
+       calling :meth:`free` or by managing the stateful object within a context manager.
+
+    Detailed information on what's happening in the various phases described above can be
+    obtained by passing in a :class:`logging.Logger` object to :class:`ContractionOptions`
+    or by setting the appropriate options in the root logger object,
+    which is used by default:
+
+        >>> import logging
+        >>> logging.basicConfig(
+        ...     level=logging.INFO,
+        ...     format="%(asctime)s %(levelname)-8s %(message)s",
+        ...     datefmt="%m-%d %H:%M:%S",
+        ... )
+
+    A user can select the desired logging level and, in general, take advantage of all of
+    the functionality offered by the Python `logging` module.
+
+    Args:
+        a: {a}
+
+        b: {b}
+
+        c: {addend}
+
+        out: {out}
+
+        qualifiers: {qualifiers}
+
+        stream: {stream}
+
+        options: {options}
+
+        execution: {execution}
+
+    .. seealso::
+        :attr:`plan_preference`, :meth:`plan`, :meth:`reset_operands`,
+        :meth:`release_operands`, :meth:`execute`
+
+    Examples:
+
+        >>> import numpy as np
+        >>> import nvmath
+
+        Create two 3-D float64 ndarrays on the CPU:
+
+        >>> M, N, K = 32, 32, 32
+        >>> a = np.random.rand(M, N, K)
+        >>> b = np.random.rand(N, K, M)
+
+        We will define a binary tensor contraction operation.
+
+        Create a BinaryContraction object encapsulating the problem specification above:
+
+        >>> contraction = nvmath.tensor.BinaryContraction("ijk,jkl->il", a, b)
+
+        Options can be provided above to control the behavior of the operation using the
+        `options` argument (see :class:`ContractionOptions`).
+
+        Next, plan the operation. Optionally, preferences can
+        be specified for planning:
+
+        >>> contraction.plan()
+
+        Now execute the binary tensor contraction, and obtain the result `r1` as a NumPy
+        ndarray.
+
+        >>> r1 = contraction.execute()
+
+        Finally, free the object's resources. To avoid having to explicitly making this
+        call, it's recommended to use the BinaryContraction object as a context manager
+        as shown below, if possible.
+
+        >>> contraction.free()
+
+        Note that all :class:`BinaryContraction` methods execute on the current
+        stream by default. Alternatively, the `stream` argument can be used to run a
+        method on a specified stream.
+
+        Let's now look at the same problem with CuPy ndarrays on the GPU.
+
+        Create a 3-D float64 CuPy ndarray on the GPU:
+
+        >>> import cupy as cp
+        >>> a = cp.random.rand(M, N, K)
+        >>> b = cp.random.rand(N, K, M)
+
+        Create an BinaryContraction object encapsulating the problem specification
+        described earlier and use it as a context manager.
+
+        >>> with nvmath.tensor.BinaryContraction("ijk,jkl->il", a, b) as contraction:
+        ...     contraction.plan()
+        ...
+        ...     # Execute the operation to get the first result.
+        ...     r1 = contraction.execute()
+        ...
+        ...     # Update operands A and B in-place (see reset_operands() for an
+        ...     # alternative).
+        ...     a[:] = cp.random.rand(M, K)
+        ...     b[:] = cp.random.rand(K, N)
+        ...
+        ...     # Execute the operation to get the new result.
+        ...     r2 = contraction.execute()
+
+
+        All the resources used by the object are released at the end of the block.
+
+        Further examples can be found in the `nvmath/examples/tensor/contraction
+        <https://github.com/NVIDIA/nvmath-python/tree/main/examples/tensor/contraction>`_
+        directory.
+    """
+
+    def __init__(
+        self,
+        expr,
+        a,
+        b,
+        *,
+        c=None,
+        out=None,
+        qualifiers=None,
+        stream: utils.AnyStream | int | None = None,
+        options: ContractionOptions | dict[str, Any] | None = None,
+        execution: ExecutionCUDA | str | dict[str, Any] | None = None,
+    ):
+        # Check here for a valid expr because it is a precondition for the constructor
+        # of the base class where it is used to extract the number of operands.
+        if not isinstance(expr, str) or expr.count(",") != 1:
+            raise ValueError("Binary contraction requires a string with exactly 2 comma-separated operands")
+
+        super().__init__(expr, a, b, c=c, out=out, qualifiers=qualifiers, stream=stream, options=options, execution=execution)
+
+    def reset_operands(self, *, a=None, b=None, c=None, out=None, stream: utils.AnyStream | int | None = None):
+        """
+        Reset one or more operands held by this :class:`BinaryContraction` instance.
+
+        .. versionchanged:: 0.9
+            All parameters are now keyword-only.
+
+        Args:
+            a: {a}
+
+            b: {b}
+
+            c: {addend}
+
+            out: {out}
+
+            stream: {stream}
+
+        {reset_operands_semantics}
+
+        Examples:
+
+            >>> import cupy as cp
+            >>> import nvmath
+
+            Create two 3-D float64 ndarrays on the GPU:
+
+            >>> M, N, K = 128, 128, 256
+            >>> a = cp.random.rand(M, K)
+            >>> b = cp.random.rand(K, N)
+
+            Create an binary contraction object as a context manager
+
+            >>> with nvmath.tensor.BinaryContraction("ij,jk->ik", a, b) as contraction:
+            ...     # Plan the operation.
+            ...     contraction.plan()
+            ...
+            ...     # Execute the contraction to get the first result.
+            ...     r1 = contraction.execute()
+            ...
+            ...     # Reset the operands to new CuPy ndarrays.
+            ...     a1 = cp.random.rand(M, K)
+            ...     b1 = cp.random.rand(K, N)
+            ...     contraction.reset_operands(a=a1, b=b1)
+            ...
+            ...     # Execute to get the new result corresponding to the updated operands.
+            ...     r2 = contraction.execute()
+
+            With :meth:`reset_operands`, minimal overhead is achieved as problem
+            specification and planning are only performed once.
+
+            For the particular example above, explicitly calling :meth:`reset_operands` is
+            equivalent to updating the operands in-place, i.e, replacing
+            ``contraction.reset_operands(a=a1, b=b1)`` with ``a[:]=a1`` and ``b[:]=b1``.
+            Note that updating the operand in-place should be adopted with caution as it can
+            only yield the expected result under the additional constraint below:
+
+                - The operand is on the GPU (more precisely, the operand memory space should
+                  be accessible from the execution space).
+
+            For more details, please refer to `inplace update example
+            <https://github.com/NVIDIA/nvmath-python/tree/main/examples/tensor/contraction/example06_stateful_inplace.py>`_.
+
+        .. seealso::
+            :meth:`reset_operands_unchecked`, :meth:`release_operands`
+        """
+        super().reset_operands(a=a, b=b, c=c, out=out, stream=stream)
+
+    def reset_operands_unchecked(self, *, a=None, b=None, c=None, out=None, stream: utils.AnyStream | int | None = None):
+        """
+        {reset_operands_unchecked}
+        """
+        super()._reset_operands_unchecked(a=a, b=b, c=c, out=out, stream=stream)
+
+    def release_operands(self):
+        """
+        {release_operands}
+        """
+        self._release_operands()
+
+
+@utils.docstring_decorator(SHARED_CONTRACTION_DOCUMENTATION, skip_missing=False)
+class TernaryContraction(_ElementaryContraction):
+    """
+    Create a stateful object encapsulating the specified ternary tensor contraction
+    :math:`\\alpha a @ b @ c + \\beta d` and the required resources to perform the
+    operation.
+    A stateful object can be used to amortize the cost of preparation (planning in the
+    case of ternary tensor contraction) across multiple executions (also see the
+    :ref:`Stateful APIs<host api types>` section).
+
+    The function-form API :func:`ternary_contraction` is a convenient alternative to using
+    stateful objects for *single* use (the user needs to perform just one tensor
+    contraction, for example), in which case there is no possibility of amortizing
+    preparatory costs. The function-form APIs are just convenience wrappers around
+    the stateful object APIs.
+
+    Using the stateful object typically involves the following steps:
+
+    1. **Problem Specification**: Initialize the object with a defined operation and
+       options.
+    2. **Preparation**: Use :meth:`plan` to determine the best algorithmic implementation
+       for this specific ternary tensor contraction operation.
+    3. **Execution**: Perform the tensor contraction computation with :meth:`execute`.
+    4. **Resource Management**: Ensure all resources are released either by explicitly
+       calling :meth:`free` or by managing the stateful object within a context manager.
+
+    Detailed information on what's happening in the various phases described above can be
+    obtained by passing in a :class:`logging.Logger` object to :class:`ContractionOptions`
+    or by setting the appropriate options in the root logger object,
+    which is used by default:
+
+        >>> import logging
+        >>> logging.basicConfig(
+        ...     level=logging.INFO,
+        ...     format="%(asctime)s %(levelname)-8s %(message)s",
+        ...     datefmt="%m-%d %H:%M:%S",
+        ... )
+
+    A user can select the desired logging level and, in general, take advantage of all of
+    the functionality offered by the Python `logging` module.
+
+    Args:
+        a: {a}
+
+        b: {b}
+
+        c: {c}
+
+        d: {addend}
+
+        out: {out}
+
+        qualifiers: {qualifiers}
+
+        stream: {stream}
+
+        options: {options}
+
+        execution: {execution}
+
+    .. seealso::
+        :attr:`plan_preference`, :meth:`plan`, :meth:`reset_operands`,
+        :meth:`release_operands`, :meth:`execute`
+
+    Examples:
+
+        >>> import numpy as np
+        >>> import nvmath
+
+        Create three 3-D float64 ndarrays on the CPU:
+
+        >>> M, N, K = 32, 32, 32
+        >>> a = np.random.rand(M, N, K)
+        >>> b = np.random.rand(N, K, M)
+        >>> c = np.random.rand(M, N)
+
+        We will define a ternary tensor contraction operation.
+
+        Create a TernaryContraction object encapsulating the problem specification above:
+
+        >>> expr = "ijk,jkl,ln->in"
+        >>> contraction = nvmath.tensor.TernaryContraction(expr, a, b, c)
+
+        Options can be provided above to control the behavior of the operation using the
+        `options` argument (see :class:`ContractionOptions`).
+
+        Next, plan the operation. Optionally, preferences can
+        be specified for planning:
+
+        >>> contraction.plan()
+
+        Now execute the ternary tensor contraction, and obtain the result `r1` as
+        a NumPy ndarray.
+
+        >>> r1 = contraction.execute()
+
+        Finally, free the object's resources. To avoid having to explicitly making this
+        call, it's recommended to use the TernaryContraction object as a context manager
+        as shown below, if possible.
+
+        >>> contraction.free()
+
+        Note that all :class:`TernaryContraction` methods execute on the current
+        stream by default. Alternatively, the `stream` argument can be used to run a
+        method on a specified stream.
+
+        Let's now look at the same problem with CuPy ndarrays on the GPU.
+
+        Create a 3-D float64 CuPy ndarray on the GPU:
+
+        >>> import cupy as cp
+        >>> a = cp.random.rand(M, N, K)
+        >>> b = cp.random.rand(N, K, M)
+        >>> c = cp.random.rand(M, N)
+
+        Create an TernaryContraction object encapsulating the problem specification
+        described earlier and use it as a context manager.
+
+        >>> expr = "ijk,jkl,ln->in"
+        >>> with nvmath.tensor.TernaryContraction(expr, a, b, c) as contraction:
+        ...     contraction.plan()
+        ...
+        ...     # Execute the operation to get the first result.
+        ...     r1 = contraction.execute()
+        ...
+        ...     # Update operands A, B and C in-place (see reset_operands() for an
+        ...     # alternative).
+        ...     a[:] = cp.random.rand(M, N, K)
+        ...     b[:] = cp.random.rand(N, K, M)
+        ...     c[:] = cp.random.rand(M, N)
+        ...
+        ...     # Execute the operation to get the new result.
+        ...     r2 = contraction.execute()
+
+
+        All the resources used by the object are released at the end of the block.
+
+        Further examples can be found in the `nvmath/examples/tensor/contraction
+        <https://github.com/NVIDIA/nvmath-python/tree/main/examples/tensor/contraction>`_
+        directory.
+    """
+
+    def __init__(
+        self,
+        expr,
+        a,
+        b,
+        c,
+        *,
+        d=None,
+        out=None,
+        qualifiers=None,
+        stream: utils.AnyStream | int | None = None,
+        options: ContractionOptions | dict[str, Any] | None = None,
+        execution: ExecutionCUDA | str | dict[str, Any] | None = None,
+    ):
+        # Check here for a valid expr because it is a precondition for the constructor
+        # of the base class where it is used to extract the number of operands.
+        if not isinstance(expr, str) or expr.count(",") != 2:
+            raise ValueError("Ternary contraction requires a string with exactly 3 comma-separated operands")
+
+        super().__init__(
+            expr, a, b, c=c, d=d, out=out, qualifiers=qualifiers, stream=stream, options=options, execution=execution
+        )
+
+    def reset_operands(self, *, a=None, b=None, c=None, d=None, out=None, stream: utils.AnyStream | int | None = None):
+        """
+        Reset one or more operands held by this :class:`TernaryContraction` instance.
+
+        .. versionchanged:: 0.9
+            All parameters are now keyword-only.
+
+        Args:
+            a: {a}
+
+            b: {b}
+
+            c: {c}
+
+            d: {addend}
+
+            out: {out}
+
+            stream: {stream}
+
+        {reset_operands_semantics}
+
+        Examples:
+
+            >>> import cupy as cp
+            >>> import nvmath
+
+            Create two 3-D float64 ndarrays on the GPU:
+
+            >>> M, N, K = 12, 16, 32
+            >>> a = cp.random.rand(M, M, N)
+            >>> b = cp.random.rand(N, K)
+            >>> c = cp.random.rand(K, K)
+
+            Create an ternary contraction object as a context manager
+
+            >>> expr = "ijk,kl,lm->ijm"
+            >>> with nvmath.tensor.TernaryContraction(expr, a, b, c) as contraction:
+            ...     # Plan the operation.
+            ...     contraction.plan()
+            ...
+            ...     # Execute the contraction to get the first result.
+            ...     r1 = contraction.execute()
+            ...
+            ...     # Reset the operands to new CuPy ndarrays.
+            ...     a1 = cp.random.rand(M, M, N)
+            ...     b1 = cp.random.rand(N, K)
+            ...     c1 = cp.random.rand(K, K)
+            ...     contraction.reset_operands(a=a1, b=b1, c=c1)
+            ...
+            ...     # Execute to get the new result corresponding to the updated operands.
+            ...     r2 = contraction.execute()
+
+            With :meth:`reset_operands`, minimal overhead is achieved as problem
+            specification and planning are only performed once.
+
+            For the particular example above, explicitly calling :meth:`reset_operands`
+            is equivalent to updating the operands in-place, i.e, replacing
+            ``contraction.reset_operands(a=a1, b=b1, c=c1)`` with ``a[:]=a1``
+            and ``b[:]=b1`` and ``c[:]=c1``. Note that updating the operand in-place
+            should be adopted with caution as it can only yield the expected result
+            under the additional constraint below:
+
+                - The operand is on the GPU (more precisely, the operand memory space should
+                  be accessible from the execution space).
+
+            For more details, please refer to `inplace update example
+            <https://github.com/NVIDIA/nvmath-python/tree/main/examples/tensor/contraction/example06_stateful_inplace.py>`_.
+
+        .. seealso::
+            :meth:`reset_operands_unchecked`, :meth:`release_operands`
+        """
+        super().reset_operands(a=a, b=b, c=c, d=d, out=out, stream=stream)
+
+    def reset_operands_unchecked(
+        self, *, a=None, b=None, c=None, d=None, out=None, stream: utils.AnyStream | int | None = None
+    ):
+        """
+        {reset_operands_unchecked}
+        """
+        super()._reset_operands_unchecked(a=a, b=b, c=c, d=d, out=out, stream=stream)
+
+    def release_operands(self):
+        """
+        {release_operands}
+        """
+        self._release_operands()
+
+
+@utils.docstring_decorator(SHARED_CONTRACTION_DOCUMENTATION, skip_missing=False)
+def binary_contraction(
+    expr,
+    a,
+    b,
+    *,
+    c=None,
+    alpha=1.0,
+    beta=None,
+    out=None,
+    qualifiers=None,
+    stream: utils.AnyStream | int | None = None,
+    options: ContractionOptions | dict[str, Any] | None = None,
+    execution: ExecutionCUDA | str | dict[str, Any] | None = None,
+):
+    """
+    Evaluate the Einstein summation convention for binary contraction on the operands.
+
+    Explicit as well as implicit form is supported for the Einstein summation expression.
+
+    Additionally, the binary contraction can be performed with
+    an additional operand, which is added to the result with a scale factor.
+
+    This function-form is a wrapper around the stateful
+    :class:`BinaryContraction` object APIs and is meant for *single* use (the user needs
+    to perform just one binary contraction, for example), in which case there is
+    no possibility of amortizing preparatory costs.
+
+    Detailed information on what's happening within this function can be obtained by passing
+    in a :class:`logging.Logger` object to :class:`ContractionOptions` or by setting the
+    appropriate options in the root logger object, which is used by default:
+
+        >>> import logging
+        >>> logging.basicConfig(
+        ...     level=logging.INFO,
+        ...     format="%(asctime)s %(levelname)-8s %(message)s",
+        ...     datefmt="%m-%d %H:%M:%S",
+        ... )
+
+    A user can select the desired logging level and, in general, take advantage of all of
+    the functionality offered by the Python `logging` module.
+
+    Args:
+        expr: {expr}
+
+        a: {a}
+
+        b: {b}
+
+        c: {addend}
+
+        alpha: {alpha}
+
+        beta: {beta}
+
+        out: {out}
+
+        qualifiers: {qualifiers}
+
+        stream: {stream}
+
+        options: {options}
+
+        execution: {execution}
+
+    Returns:
+        {result}
+
+    .. seealso::
+        :class:`BinaryContraction`, :func:`ternary_contraction`,
+        :class:`TernaryContraction`, :class:`ContractionOptions`,
+        :class:`ContractionPlanPreference`
+
+        For tensor network contraction with arbitrary number of operands including
+        contraction path finding, see cuQuantum:
+
+        - :external+cuquantum:py:func:`cuquantum.tensornet.contract`
+        - :external+cuquantum:py:class:`cuquantum.tensornet.Network`
+
+    Examples:
+
+        >>> import cupy as cp
+        >>> import nvmath
+
+        Create three float32 ndarrays on the GPU:
+
+        >>> M, N = 32, 64
+        >>> a = cp.random.rand(M, M, N, N, dtype=cp.float32)
+        >>> b = cp.random.rand(N, N, N, N, dtype=cp.float32)
+        >>> c = cp.random.rand(M, M, N, N, dtype=cp.float32)
+
+        Perform the operation :math:`\\alpha \\sum a[i,j,k,l] * b[k,l,m,n] +
+        \\beta c[i,j,m,n]` using :func:`binary_contraction`.
+        The result `r` is also a CuPy float32 ndarray:
+
+        >>> r = nvmath.tensor.binary_contraction(
+        ...     "ijkl,klmn->ijmn", a, b, c=c, alpha=1.23, beta=0.74
+        ... )
+
+        The result is equivalent to:
+
+        >>> r = 1.23 * cp.einsum("ijkl,klmn->ijmn", a, b) + 0.74 * c
+
+        Options can be provided to customize the operation:
+
+        >>> compute_type = nvmath.bindings.cutensor.ComputeDesc.COMPUTE_3XTF32()
+        >>> o = nvmath.tensor.ContractionOptions(compute_type=compute_type)
+        >>> r = nvmath.tensor.binary_contraction("ijkl,klmn->ijmn", a, b, options=o)
+
+        See `ContractionOptions` for the complete list of available options.
+
+        The package current stream is used by default, but a stream can be explicitly
+        provided to the binary contraction operation. This can be done if the operands
+        are computed on a different stream, for example:
+
+        >>> s = cp.cuda.Stream()
+        >>> with s:
+        ...     a = cp.random.rand(M, M, N, N)
+        ...     b = cp.random.rand(N, N, N, N)
+        >>> r = nvmath.tensor.binary_contraction("ijkl,klmn->ijmn", a, b, stream=s)
+
+        The operation above runs on stream `s` and is ordered with respect to the input
+        computation.
+
+        Create NumPy ndarrays on the CPU.
+
+        >>> import numpy as np
+        >>> a = np.random.rand(M, M, N, N)
+        >>> b = np.random.rand(N, N, N, N)
+
+        Provide the NumPy ndarrays to :func:`binary_contraction`, with the result
+        also being a NumPy ndarray:
+
+        >>> r = nvmath.tensor.binary_contraction("ijkl,klmn->ijmn", a, b)
+
+    Notes:
+        - This function is a convenience wrapper around :class:`BinaryContraction` and is
+          specifically meant for *single* use.
+
+    Further examples can be found in the `nvmath/examples/tensor/contraction
+    <https://github.com/NVIDIA/nvmath-python/tree/main/examples/tensor/contraction>`_
+    directory.
+    """
+    if c is None and beta is not None:
+        raise ValueError("beta can only be set if c is specified in a binary contraction")
+    elif c is not None and beta is None:
+        raise ValueError("beta must be set when c is specified in a binary contraction")
+    with BinaryContraction(
+        expr, a, b, c=c, out=out, qualifiers=qualifiers, stream=stream, options=options, execution=execution
+    ) as contraction:
+        contraction.plan(stream=stream)
+        out = contraction.execute(alpha=alpha, beta=beta, stream=stream)
+    return out
+
+
+@utils.docstring_decorator(SHARED_CONTRACTION_DOCUMENTATION, skip_missing=False)
+def ternary_contraction(
+    expr,
+    a,
+    b,
+    c,
+    *,
+    d=None,
+    alpha=1.0,
+    beta=None,
+    out=None,
+    qualifiers=None,
+    stream: utils.AnyStream | int | None = None,
+    options: ContractionOptions | dict[str, Any] | None = None,
+    execution: ExecutionCUDA | str | dict[str, Any] | None = None,
+):
+    """
+    Evaluate the Einstein summation convention for ternary contraction on the operands.
+
+    Explicit as well as implicit form is supported for the Einstein summation expression.
+
+    Additionally, the ternary contraction can be performed with
+    an additional operand, which is added to the result with a scale factor.
+
+    This function-form is a wrapper around the stateful
+    :class:`TernaryContraction` object APIs and is meant for *single* use (the user needs
+    to perform just one ternary contraction, for example), in which case there is
+    no possibility of amortizing preparatory costs.
+
+    Detailed information on what's happening within this function can be obtained by passing
+    in a :class:`logging.Logger` object to :class:`ContractionOptions` or by setting the
+    appropriate options in the root logger object, which is used by default:
+
+        >>> import logging
+        >>> logging.basicConfig(
+        ...     level=logging.INFO,
+        ...     format="%(asctime)s %(levelname)-8s %(message)s",
+        ...     datefmt="%m-%d %H:%M:%S",
+        ... )
+
+    A user can select the desired logging level and, in general, take advantage of all of
+    the functionality offered by the Python `logging` module.
+
+    Args:
+        expr: {expr}
+
+        a: {a}
+
+        b: {b}
+
+        c: {c}
+
+        d: {addend}
+
+        alpha: {alpha}
+
+        beta: {beta}
+
+        out: {out}
+
+        qualifiers: {qualifiers}
+
+        stream: {stream}
+
+        options: {options}
+
+        execution: {execution}
+
+    Returns:
+        {result}
+
+    .. seealso::
+        :class:`TernaryContraction`, :func:`binary_contraction`,
+        :class:`BinaryContraction`, :class:`ContractionOptions`,
+        :class:`ContractionPlanPreference`
+
+        For tensor network contraction with arbitrary number of operands including
+        contraction path finding, see cuQuantum:
+
+        - :external+cuquantum:py:func:`cuquantum.tensornet.contract`
+        - :external+cuquantum:py:class:`cuquantum.tensornet.Network`
+
+    Examples:
+
+        >>> import cupy as cp
+        >>> import nvmath
+
+        Create three float32 ndarrays on the GPU:
+
+        >>> M, N, K = 16, 24, 32
+        >>> a = cp.random.rand(M, M, dtype=cp.float32)
+        >>> b = cp.random.rand(M, N, K, dtype=cp.float32)
+        >>> c = cp.random.rand(N, K, M, dtype=cp.float32)
+        >>> d = cp.random.rand(M, M, dtype=cp.float32)
+
+        Perform the operation :math:`\\alpha \\sum a[i,j] * b[j,k,l] * c[k,l,m] +
+        \\beta d[i,m]` using :func:`ternary_contraction`.
+        The result `r` is also a CuPy float32 ndarray:
+
+        >>> r = nvmath.tensor.ternary_contraction(
+        ...     "ij,jkl,klm->im", a, b, c, d=d, alpha=0.63, beta=0.22
+        ... )
+
+        The result is equivalent to:
+
+        >>> r = 0.63 * cp.einsum("ij,jkl,klm->im", a, b, c) + 0.22 * d
+
+        Options can be provided to customize the operation:
+
+        >>> compute_type = nvmath.bindings.cutensor.ComputeDesc.COMPUTE_3XTF32()
+        >>> o = nvmath.tensor.ContractionOptions(compute_type=compute_type)
+        >>> r = nvmath.tensor.ternary_contraction("ij,jkl,klm->im", a, b, c, options=o)
+
+        See `ContractionOptions` for the complete list of available options.
+
+        The package current stream is used by default, but a stream can be explicitly
+        provided to the ternary contraction operation. This can be done if the operands
+        are computed on a different stream, for example:
+
+        >>> s = cp.cuda.Stream()
+        >>> with s:
+        ...     a = cp.random.rand(M, M, dtype=cp.float32)
+        ...     b = cp.random.rand(M, N, K, dtype=cp.float32)
+        ...     c = cp.random.rand(N, K, M, dtype=cp.float32)
+        >>> r = nvmath.tensor.ternary_contraction("ij,jkl,klm->im", a, b, c, stream=s)
+
+        The operation above runs on stream `s` and is ordered with respect to the input
+        computation.
+
+        Create NumPy ndarrays on the CPU.
+
+        >>> import numpy as np
+        >>> a = np.random.rand(M, M)
+        >>> b = np.random.rand(M, N, K)
+        >>> c = np.random.rand(N, K, M)
+
+        Provide the NumPy ndarrays to :func:`ternary_contraction`, with the result
+        also being a NumPy ndarray:
+
+        >>> r = nvmath.tensor.ternary_contraction("ij,jkl,klm->im", a, b, c)
+
+    Notes:
+        - This function is a convenience wrapper around :class:`TernaryContraction` and is
+          specifically meant for *single* use.
+
+    Further examples can be found in the `nvmath/examples/tensor/contraction
+    <https://github.com/NVIDIA/nvmath-python/tree/main/examples/tensor/contraction>`_
+    directory.
+    """
+    if d is None and beta is not None:
+        raise ValueError("beta can only be set if d is specified in a ternary contraction")
+    elif d is not None and beta is None:
+        raise ValueError("beta must be set when d is specified in a ternary contraction")
+    with TernaryContraction(
+        expr, a, b, c, d=d, out=out, qualifiers=qualifiers, stream=stream, options=options, execution=execution
+    ) as contraction:
+        contraction.plan(stream=stream)
+        out = contraction.execute(alpha=alpha, beta=beta, stream=stream)
+    return out