nvmath-python 0.9.0__cp314-cp314-win_amd64.whl

nvmath/device/cusolverdx.py

@@ -0,0 +1,2690 @@
+# Copyright (c) 2026, NVIDIA CORPORATION & AFFILIATES. ALL RIGHTS RESERVED.
+#
+# SPDX-License-Identifier: Apache-2.0
+
+__all__ = [
+    "Solver",
+    "CholeskySolver",
+    "LUSolver",
+    "LUPivotSolver",
+    "TriangularSolver",
+    "QRFactorize",
+    "LQFactorize",
+    "QRMultiply",
+    "LQMultiply",
+    "LeastSquaresSolver",
+    "compile_solver_execute",
+]
+
+from collections.abc import Sequence
+from functools import cached_property
+from typing import Any, Literal
+
+import numpy as np
+
+from nvmath._utils import get_nvrtc_version
+from nvmath.bindings import mathdx
+from nvmath.internal.utils import docstring_decorator
+
+from .common import SHARED_DEVICE_DOCSTRINGS, check_code_type, parse_code_type, parse_sm
+from .common_backend import get_isa_version, get_lto
+from .common_cuda import (
+    Code,
+    ComputeCapability,
+)
+from .cusolverdx_backend import (
+    _ENABLE_CUSOLVERDX_0_3,
+    ALLOWED_ARRANGEMENT,
+    ALLOWED_CUSOLVERDX_FUNCTIONS,
+    ALLOWED_DATA_TYPE,
+    ALLOWED_DIAG,
+    ALLOWED_EXECUTION,
+    ALLOWED_FILL_MODE,
+    ALLOWED_REAL_NP_TYPES,
+    ALLOWED_SIDE,
+    ALLOWED_TRANSPOSE_MODE,
+    CUSOLVERDX_0_3_ALLOWED_FUNCTIONS,
+    DIAG_SUPPORTED_FUNCTIONS,
+    FILL_MODE_SUPPORTED_FUNCTIONS,
+    JOB_SUPPORT_MAP,
+    JOB_SUPPORTED_FUNCTIONS,
+    SIDE_SUPPORTED_FUNCTIONS,
+    generate_code,
+    generate_SOLVER,
+    get_int_traits,
+    get_str_trait,
+    validate,
+)
+from .types import Complex, complex64, complex128
+
+# ==========================
+# docs
+# ==========================
+
+SOLVER_DOCSTRING_SIZE_BASE = """Problem size specified as a sequence of 1 to 3 elements:
+``(M,)`` (treated as ``(M, M, 1)``), ``(M, N)`` (treated as ``(M, N, 1)``), or ``(M, N, K)``.""".replace("\n", " ")
+
+SOLVER_DOCSTRING = {
+    "function": f"""Solver function to be executed on execute() method. List of available options:
+{", ".join(f"``'{k}'`` ({v})" for k, v in ALLOWED_CUSOLVERDX_FUNCTIONS.items())}.
+Functions {", ".join(f"``'{k}'``" for k in CUSOLVERDX_0_3_ALLOWED_FUNCTIONS)} require libmathdx 0.3.2 or later.""".replace(
+        "\n", " "
+    ),
+    #
+    "size": f"""{SOLVER_DOCSTRING_SIZE_BASE} Please refer to cuSOLVERDx functionalities for detailed meaning.
+""".replace("\n", " "),
+    #
+    "arrangement": f"""Storage layout for matrices A and B, specified as a sequence of 2 elements ``(arr_A, arr_B)``.
+Each element can be one of: {", ".join(f"``'{v}'``" for v in ALLOWED_ARRANGEMENT)}.
+Defaults to ``("col_major", "col_major")``.""".replace("\n", " "),
+    #
+    "transpose_mode": f"""Transpose mode of matrix A. Refer to cuSOLVERDx documentation,
+as some functions do not support this option.
+Can be one of: {", ".join(f"``'{v}'``" for v in ALLOWED_TRANSPOSE_MODE)}.
+Defaults to ``'non_transposed'``.""".replace("\n", " "),
+    #
+    "side": f"""Side of matrix A in a multiplication operation.
+Required and supported only by functions: {", ".join(f"``'{v}'``" for v in SIDE_SUPPORTED_FUNCTIONS)}.
+Can be one of: {", ".join(f"``'{v}'``" for v in ALLOWED_SIDE)}.""".replace("\n", " "),
+    #
+    "diag": f"""Indicates whether the diagonal elements of matrix A are ones or not.
+Required and supported only by {", ".join(f"``'{v}'``" for v in DIAG_SUPPORTED_FUNCTIONS)} functions.
+Can be one of: {", ".join(f"``'{v}'``" for v in ALLOWED_DIAG)}.""".replace("\n", " "),
+    #
+    "fill_mode": f"""Indicates which part of matrix A is filled and should be used by function.
+Required and supported only by functions: {", ".join(f"``'{v}'``" for v in FILL_MODE_SUPPORTED_FUNCTIONS)}.
+Can be one of: {", ".join(f"``'{v}'``" for v in ALLOWED_FILL_MODE)}.""".replace("\n", " "),
+    #
+    "job": f"""Job type for eigenvalue computation.
+Required and supported only by functions: {", ".join(f"``'{v}'``" for v in JOB_SUPPORTED_FUNCTIONS)}.
+For ``'htev'``: {", ".join(f"``'{v}'``" for v in JOB_SUPPORT_MAP["htev"])}.
+For ``'heev'``: {", ".join(f"``'{v}'``" for v in JOB_SUPPORT_MAP["heev"])}.
+Requires libmathdx 0.3.2 or later.""".replace("\n", " "),
+    #
+    "batches_per_block": """Number of batches to compute in parallel in a single CUDA block.
+Can be a non-zero integer or the string ``'suggested'`` for automatic selection of an optimal value.
+We recommend using 1 for matrix A size larger than or equal to 16 x 16,
+and using ``'suggested'`` for smaller sizes to achieve optimal performance.
+Defaults to 1.""".replace("\n", " "),
+    #
+    "data_type": f"""The data type of the input matrices,
+can be one of: {", ".join(f"``'{v}'``" for v in ALLOWED_DATA_TYPE)}.
+Defaults to ``'real'``.""".replace("\n", " "),
+    #
+    "leading_dimensions": """The leading dimensions for input matrices A and B,
+specified as a sequence of 2 elements (``lda``, ``ldb``) or ``None``.
+If not provided, it will be automatically deduced from ``size`` and ``arrangement``.
+""".replace("\n", " "),
+    #
+    "block_dim": """The block dimension for launching the CUDA kernel,
+specified as a 1 to 3 integer sequence (x, y, z) where missing dimensions are assumed to be 1.
+Can be a sequence of 1 to 3 positive integers, the string ``'suggested'``
+for optimal value selection, or ``None`` for the default value.""".replace("\n", " "),
+    #
+    "execution": f"""A string specifying the execution method.
+Supported values: {", ".join(f"``'{v}'``" for v in ALLOWED_EXECUTION)}.""".replace("\n", " "),
+    #
+    "precision": f"""The computation precision specified as a numpy float dtype.
+Currently supports: {", ".join(f"``numpy.{v.__name__}``" for v in ALLOWED_REAL_NP_TYPES)}.""".replace("\n", " "),
+    #
+    "sm": SHARED_DEVICE_DOCSTRINGS["sm"],
+}
+
+# ==========================
+# Solver class
+# ==========================
+
+
+@docstring_decorator(SOLVER_DOCSTRING, skip_missing=False)
+class Solver:
+    """
+    A class that encapsulates a partial dense matrix factorization and solve
+    device function.
+
+    **Memory Layout Requirements:**
+
+    Matrices must be stored in shared memory according
+    to their arrangement and leading dimension (ld):
+
+    * **Column-major arrangement**: Matrix shape ``(batches_per_block, rows, cols)``
+        with strides ``(ld * cols, 1, ld)``
+    * **Row-major arrangement**: Matrix shape ``(batches_per_block, rows, cols)``
+        with strides ``(ld * rows, ld, 1)``
+
+    Args:
+        function (str): {function}
+
+        size (Sequence[int]): {size}
+
+        precision (type[np.floating]): {precision}
+
+        execution (str): {execution}
+
+        sm (ComputeCapability): {sm}
+
+        arrangement (Sequence[str], optional): {arrangement}
+
+        transpose_mode (str, optional): {transpose_mode}
+
+        side (str, optional): {side}
+
+        diag (str, optional): {diag}
+
+        fill_mode (str, optional): {fill_mode}
+
+        batches_per_block (int | Literal["suggested"], optional): {batches_per_block}
+
+        data_type (str, optional): {data_type}
+
+        leading_dimensions (Sequence[int], optional): {leading_dimensions}
+
+        block_dim (Sequence[int] | Literal["suggested"], optional): {block_dim}
+
+        job (str, optional): {job}
+
+    See Also:
+        The attributes of this class provide a 1:1
+        mapping with the CUDA C++ cuSOLVERDx APIs.
+        For further details, please refer to :cusolverdx_doc:`cuSOLVERDx documentation
+        <index.html>`.
+    """
+
+    def __init__(
+        self,
+        function: str,
+        size: Sequence[int],
+        precision: type[np.floating],
+        execution: str,
+        *,
+        sm=None,
+        arrangement: Sequence[str] | None = None,
+        transpose_mode: str | None = None,
+        side: str | None = None,
+        diag: str | None = None,
+        fill_mode: str | None = None,
+        batches_per_block: int | Literal["suggested"] | None = None,
+        data_type: str | None = None,
+        leading_dimensions: Sequence[int] | None = None,
+        block_dim: Sequence[int] | Literal["suggested"] | None = None,
+        job: str | None = None,
+    ):
+        if get_nvrtc_version() < (12, 6, 85):
+            raise RuntimeError("cuSOLVERDx requires CUDA Toolkit 12.6 Update 3 or later.")
+
+        sm = parse_sm(sm)
+
+        validate(
+            function=function,
+            size=size,
+            precision=precision,
+            execution=execution,
+            sm=sm,
+            arrangement=arrangement,
+            transpose_mode=transpose_mode,
+            side=side,
+            diag=diag,
+            fill_mode=fill_mode,
+            batches_per_block=batches_per_block,
+            data_type=data_type,
+            leading_dimensions=leading_dimensions,
+            block_dim=block_dim,
+            job=job,
+        )
+
+        if len(size) == 1:
+            size = (size[0], size[0], 1)
+        elif len(size) == 2:
+            size = (size[0], size[1], 1)
+        else:
+            size = (size[0], size[1], size[2])
+
+        if block_dim is not None and block_dim != "suggested":
+            if len(block_dim) == 1:
+                block_dim = (block_dim[0], 1, 1)
+            elif len(block_dim) == 2:
+                block_dim = (block_dim[0], block_dim[1], 1)
+            else:
+                block_dim = (block_dim[0], block_dim[1], block_dim[2])
+
+        self._function = function
+        self._size = size
+        self._precision = precision
+        self._execution = execution
+        self._sm = sm
+        self._arrangement = tuple(arrangement) if arrangement is not None else None
+        self._transpose_mode = transpose_mode
+        self._side = side
+        self._diag = diag
+        self._fill_mode = fill_mode
+        self._data_type = data_type
+        self._job = job
+
+        self._batches_per_block = batches_per_block if batches_per_block != "suggested" else None
+        self._leading_dimensions = tuple(leading_dimensions) if leading_dimensions is not None else None
+        self._block_dim = block_dim if block_dim != "suggested" and block_dim is not None else None
+
+        if batches_per_block == "suggested" or block_dim == "suggested":
+            # Update suggested fields
+            traits = _SolverTraits(self)
+
+            if batches_per_block == "suggested":
+                self._batches_per_block = traits.suggested_batches_per_block
+
+            if block_dim == "suggested":
+                self._block_dim = traits.suggested_block_dim
+
+    # ==========================
+    # Operator properties
+    # ==========================
+
+    @property
+    def function(self) -> str:
+        return self._function
+
+    @property
+    def size(self) -> tuple[int, int, int]:
+        return self._size
+
+    @property
+    def m(self) -> int:
+        m, _, _ = self.size
+        return m
+
+    @property
+    def n(self) -> int:
+        _, n, _ = self.size
+        return n
+
+    @property
+    def k(self) -> int:
+        _, _, k = self.size
+        return k
+
+    @property
+    def precision(self) -> type[np.floating]:
+        return self._precision
+
+    @property
+    def execution(self) -> str:
+        return self._execution
+
+    @property
+    def sm(self) -> ComputeCapability:
+        return self._sm
+
+    @property
+    def arrangement(self) -> Sequence[str]:
+        if self._arrangement is None:
+            # TODO: replace with libmathdx trait when supported
+            return ("col_major", "col_major")
+        if len(self._arrangement) == 1:
+            return (self._arrangement[0], "col_major")
+        return self._arrangement
+
+    @property
+    def a_arrangement(self) -> str:
+        arr_a, _ = self.arrangement
+        return arr_a
+
+    @property
+    def b_arrangement(self) -> str:
+        _, arr_b = self.arrangement
+        return arr_b
+
+    @property
+    def transpose_mode(self) -> str | None:
+        return self._transpose_mode
+
+    @property
+    def side(self) -> str | None:
+        return self._side
+
+    @property
+    def diag(self) -> str | None:
+        return self._diag
+
+    @property
+    def fill_mode(self) -> str | None:
+        return self._fill_mode
+
+    @property
+    def batches_per_block(self) -> int:
+        if self._batches_per_block is None:
+            # TODO: replace with libmathdx trait when supported
+            return 1
+        return self._batches_per_block
+
+    @property
+    def data_type(self) -> str:
+        if self._data_type is None:
+            # TODO: replace with libmathdx trait when supported
+            return "real"
+        return self._data_type
+
+    @property
+    def leading_dimensions(self) -> tuple:
+        if self._leading_dimensions is None:
+            # TODO: replace with libmathdx trait when supported
+            return self._calculate_default_leading_dimensions()
+
+        if len(self._leading_dimensions) == 1:
+            return (self._leading_dimensions[0], self._calculate_default_leading_dimensions()[1])
+        return self._leading_dimensions
+
+    @property
+    def lda(self) -> int:
+        lda, _ = self.leading_dimensions
+        return lda
+
+    @property
+    def ldb(self) -> int:
+        _, ldb = self.leading_dimensions
+        return ldb
+
+    @property
+    def block_dim(self) -> tuple:
+        if self._block_dim is None:
+            return self._traits.block_dim
+        return self._block_dim
+
+    @property
+    def block_size(self) -> int:
+        return self.block_dim[0] * self.block_dim[1] * self.block_dim[2]
+
+    @property
+    def job(self) -> str | None:
+        return self._job
+
+    # ==========================
+    # Trait properties
+    # ==========================
+
+    @cached_property
+    def _traits(self):
+        return _SolverTraits(self)
+
+    @property
+    def workspace_size(self) -> int:
+        if not _ENABLE_CUSOLVERDX_0_3:
+            raise RuntimeError("workspace size requires libmathdx 0.3.2")
+
+        return self._traits.workspace_size
+
+    @property
+    def value_type(self) -> type[np.floating] | Complex:
+        if self.data_type == "complex":
+            return complex64 if self.precision == np.float32 else complex128
+        return self.precision
+
+    @property
+    def info_type(self) -> type[np.signedinteger]:
+        return np.int32
+
+    @property
+    def ipiv_type(self) -> type[np.signedinteger]:
+        return np.int32
+
+    @property
+    def tau_type(self) -> type[np.floating] | Complex:
+        return self.value_type
+
+    # ==========================
+    # execute()
+    # ==========================
+
+    def execute(*args):
+        raise RuntimeError("execute is a device function and cannot be called on the host.")
+
+    # ==========================
+    # Private methods
+    # ==========================
+
+    def _generate_SOLVER(self, execution_api):
+        return generate_SOLVER(
+            function=self._function,
+            size=self._size,
+            precision=self._precision,
+            execution=self._execution,
+            sm=self._sm,
+            arrangement=self.arrangement,  # TODO: remove property when libmathdx supports traits
+            transpose_mode=self._transpose_mode,
+            side=self._side,
+            diag=self._diag,
+            fill_mode=self._fill_mode,
+            batches_per_block=self.batches_per_block,  # TODO: remove property when libmathdx supports traits
+            data_type=self.data_type,  # TODO: remove property when libmathdx supports traits
+            leading_dimensions=self.leading_dimensions,  # TODO: remove property when libmathdx supports traits
+            block_dim=self._block_dim,
+            execution_api=execution_api,
+            job=self._job,
+        )
+
+    def _calculate_default_leading_dimensions(self) -> tuple:
+        """
+        Calculates default leading dimensions based on problem
+        size (m, n, k) for different solver functions.
+        """
+
+        def calculate_ld(rows, cols, arr):
+            return rows if arr == "col_major" else cols
+
+        arr_a, arr_b = self.arrangement
+        m, n, k = self.size
+
+        match self.function:
+            case (
+                "potrf"
+                | "potrs"
+                | "posv"
+                | "getrf_no_pivot"
+                | "getrs_no_pivot"
+                | "gesv_no_pivot"
+                | "getrf_partial_pivot"
+                | "getrs_partial_pivot"
+                | "gesv_partial_pivot"
+            ):
+                return (calculate_ld(m, n, arr_a), calculate_ld(n, k, arr_b))
+            case "trsm":
+                assert self.side is not None
+                return (m if self.side == "left" else n, calculate_ld(m, n, arr_b))
+            case "geqrf" | "gelqf" | "htev" | "unglq" | "ungqr" | "heev":
+                return (calculate_ld(m, n, arr_a), 0)
+            case "unmqr":
+                return (calculate_ld(m if self.side == "left" else n, k, arr_a), calculate_ld(m, n, arr_b))
+            case "unmlq":
+                return (calculate_ld(k, m if self.side == "left" else n, arr_a), calculate_ld(m, n, arr_b))
+            case "gels":
+                return (calculate_ld(m, n, arr_a), calculate_ld(max(m, n), k, arr_b))  # mirrors cusolverdx behaviour
+            case "gtsv_no_pivot":
+                return (0, calculate_ld(n, k, arr_b))
+            case _:
+                raise NotImplementedError(f"Function {self.function} is not supported for leading dimension calculation")
+
+
+# ==========================
+# Internal traits class
+# ==========================
+
+
+class _SolverTraits:
+    def __init__(self, solver: Solver):
+        h = solver._generate_SOLVER("compiled_leading_dim")
+
+        try:
+            self.suggested_batches_per_block = int(
+                mathdx.cusolverdx_get_trait_int64(h.descriptor, mathdx.CusolverdxTraitType.SUGGESTED_BATCHES_PER_BLOCK)
+            )
+
+            self.suggested_block_dim = tuple(get_int_traits(h.descriptor, mathdx.CusolverdxTraitType.SUGGESTED_BLOCK_DIM, 3))
+            self.block_dim = tuple(get_int_traits(h.descriptor, mathdx.CusolverdxTraitType.BLOCK_DIM, 3))
+
+            if _ENABLE_CUSOLVERDX_0_3:
+                self.workspace_size = int(
+                    mathdx.cusolverdx_get_trait_int64(h.descriptor, mathdx.CusolverdxTraitType.WORKSPACE_SIZE)
+                )
+
+        except mathdx.LibMathDxError as e:
+            raise RuntimeError(
+                "Failed to compile the solver. This may indicate incompatible "
+                f"parameter combination. Please refer to the cuSOLVERDx documentation. Details: {e}"
+            ) from e
+
+
+# ==========================
+# Compile function
+# ==========================
+
+
+def compile_solver_execute(
+    solver: Solver,
+    code_type: Any,
+    execution_api: str,
+) -> tuple[Code, str]:
+    code_type = parse_code_type(code_type)
+    check_code_type(code_type, "cuSOLVERdx")
+
+    h = solver._generate_SOLVER(execution_api).descriptor
+
+    try:
+        code = generate_code(h, code_type.cc)
+    except mathdx.LibMathDxError as e:
+        raise RuntimeError(
+            "Failed to compile the solver. This may indicate incompatible "
+            f"parameter combination. Please refer to the cuSOLVERDx documentation. Details: {e}"
+        ) from e
+
+    lto = get_lto(code.descriptor)
+    isa_version = get_isa_version(code.descriptor)
+    symbol = get_str_trait(h, mathdx.CusolverdxTraitType.SYMBOL_NAME)
+
+    return Code(code_type, isa_version, lto), symbol
+
+
+# ==========================
+# Pythonic Adapters
+# ==========================
+
+
+def _calculate_strides(shape: tuple[int, int], ld: int, arrangement: str) -> tuple[int, int, int]:
+    return (ld * shape[1], 1, ld) if arrangement == "col_major" else (ld * shape[0], ld, 1)
+
+
+class _SolverProperties:
+    def __init__(self, source: Solver):
+        self._properties_source = source
+
+    @property
+    def size(self) -> tuple[int, int, int]:
+        return self._properties_source.size
+
+    @property
+    def m(self) -> int:
+        return self._properties_source.m
+
+    @property
+    def n(self) -> int:
+        return self._properties_source.n
+
+    @property
+    def k(self) -> int:
+        return self._properties_source.k
+
+    @property
+    def precision(self) -> type[np.floating]:
+        return self._properties_source.precision
+
+    @property
+    def execution(self) -> str:
+        return self._properties_source.execution
+
+    @property
+    def sm(self) -> ComputeCapability:
+        return self._properties_source.sm
+
+    @property
+    def arrangement(self) -> Sequence[str]:
+        return self._properties_source.arrangement
+
+    @property
+    def a_arrangement(self) -> str:
+        return self._properties_source.a_arrangement
+
+    @property
+    def b_arrangement(self) -> str:
+        return self._properties_source.b_arrangement
+
+    @property
+    def batches_per_block(self) -> int:
+        return self._properties_source.batches_per_block
+
+    @property
+    def data_type(self) -> str:
+        return self._properties_source.data_type
+
+    @property
+    def leading_dimensions(self) -> tuple:
+        return self._properties_source.leading_dimensions
+
+    @property
+    def lda(self) -> int:
+        return self._properties_source.lda
+
+    @property
+    def ldb(self) -> int:
+        return self._properties_source.ldb
+
+    @property
+    def block_dim(self) -> tuple:
+        return self._properties_source.block_dim
+
+    @property
+    def block_size(self) -> int:
+        return self._properties_source.block_size
+
+    @property
+    def value_type(self) -> type[np.floating] | Complex:
+        return self._properties_source.value_type
+
+
+class _LinearSolverProperties(_SolverProperties):
+    def __init__(self, source: Solver):
+        super().__init__(source)
+
+    @property
+    def info_type(self) -> type[np.signedinteger]:
+        return self._properties_source.info_type
+
+    @property
+    def info_shape(self) -> tuple[int]:
+        return (self.batches_per_block,)
+
+    @property
+    def info_strides(self) -> tuple[int]:
+        return (1,)
+
+    @property
+    def a_shape(self) -> tuple[int, int, int]:
+        return (self.batches_per_block, self.m, self.n)
+
+    @property
+    def b_shape(self) -> tuple[int, int, int]:
+        return (self.batches_per_block, self.n, self.k)
+
+    def a_strides(self, *, lda: int | None = None) -> tuple[int, int, int]:
+        lda = self.lda if lda is None else lda
+        return _calculate_strides(self.a_shape[1:], lda, self.a_arrangement)
+
+    def b_strides(self, *, ldb: int | None = None) -> tuple[int, int, int]:
+        ldb = self.ldb if ldb is None else ldb
+        return _calculate_strides(self.b_shape[1:], ldb, self.b_arrangement)
+
+    def a_size(self, *, lda: int | None = None) -> int:
+        return self.a_strides(lda=lda)[0] * self.a_shape[0]
+
+    def b_size(self, *, ldb: int | None = None) -> int:
+        return self.b_strides(ldb=ldb)[0] * self.b_shape[0]
+
+
+# ==========================
+# Cholesky Solver
+# ==========================
+
+ADAPTERS_API_LD_DOCSTRING = """
+Note: When provided in the constructor, leading dimensions are set at compile-time.
+To use runtime leading dimensions (avoiding recompilation for different leading dimensions),
+provide the leading dimension parameters directly to the device methods instead.
+""".replace("\n", " ")
+
+CHOLESKY_SOLVER_DOCSTRING = SOLVER_DOCSTRING.copy()
+del CHOLESKY_SOLVER_DOCSTRING["function"]
+del CHOLESKY_SOLVER_DOCSTRING["transpose_mode"]
+del CHOLESKY_SOLVER_DOCSTRING["side"]
+del CHOLESKY_SOLVER_DOCSTRING["diag"]
+del CHOLESKY_SOLVER_DOCSTRING["job"]
+
+CHOLESKY_SOLVER_DOCSTRING["size"] = f"""{SOLVER_DOCSTRING_SIZE_BASE}
+``M`` represents the dimension of the square matrix A (``M`` x ``M``) used in factorization, ``N`` must be equal to ``M``.
+``K`` represents the number of columns in the right-hand side matrix B
+(dimensions ``M`` x ``K``) for the solve operation.""".replace("\n", " ")
+CHOLESKY_SOLVER_DOCSTRING["leading_dimensions"] = SOLVER_DOCSTRING["leading_dimensions"] + ADAPTERS_API_LD_DOCSTRING
+CHOLESKY_SOLVER_DOCSTRING["fill_mode"] = f"""Indicates which part of matrix A is filled and should be used by function.
+Can be one of: {", ".join(f"``'{v}'``" for v in ALLOWED_FILL_MODE)}.""".replace("\n", " ")
+
+
+@docstring_decorator(CHOLESKY_SOLVER_DOCSTRING, skip_missing=False)
+class CholeskySolver(_LinearSolverProperties):
+    """
+    A class that encapsulates Cholesky factorization and solve device functions
+    for symmetric positive definite matrices.
+
+    **Available operations:**
+
+    * factorize: Computes the Cholesky factorization
+      A = L @ L^H (lower) or A = U^H @ U (upper),
+      where L is a lower triangular matrix and U is an upper triangular matrix.
+      The choice depends on the fill_mode parameter.
+    * solve: Solves the system Ax = B using a previously computed Cholesky factorization
+
+    **Memory Layout Requirements:**
+
+    Matrices must be stored in shared memory according
+    to their arrangement and leading dimension (ld):
+
+    **For matrix A (M x N):**
+
+    * **Column-major arrangement**: Matrix shape ``(batches_per_block, M, N)``
+      with strides ``(lda * N, 1, lda)``
+    * **Row-major arrangement**: Matrix shape ``(batches_per_block, M, N)``
+      with strides ``(lda * M, lda, 1)``
+
+    **For matrix B (N x K):**
+
+    * **Column-major arrangement**: Matrix shape ``(batches_per_block, N, K)``
+      with strides ``(ldb * K, 1, ldb)``
+    * **Row-major arrangement**: Matrix shape ``(batches_per_block, N, K)``
+      with strides ``(ldb * N, ldb, 1)``
+
+    Args:
+        size (Sequence[int]): {size}
+
+        precision (type[np.floating]): {precision}
+
+        execution (str): {execution}
+
+        sm (ComputeCapability): {sm}
+
+        fill_mode (str): {fill_mode}
+
+        arrangement (Sequence[str], optional): {arrangement}
+
+        batches_per_block (int | Literal["suggested"], optional): {batches_per_block}
+
+        data_type (str, optional): {data_type}
+
+        leading_dimensions (Sequence[int], optional): {leading_dimensions}
+
+        block_dim (Sequence[int] | Literal["suggested"], optional): {block_dim}
+
+    See Also:
+        For further details, please refer to the cuSOLVERDx documentation:
+
+        * :cusolverdx_doc:`factorize <get_started/functions/potrf.html>`
+        * :cusolverdx_doc:`solve <get_started/functions/potrs.html>`
+    """
+
+    # ==========================
+    # Constructor
+    # ==========================
+
+    def __init__(
+        self,
+        size: Sequence[int],
+        precision: type[np.floating],
+        execution: str,
+        fill_mode: str,
+        *,
+        sm=None,
+        arrangement: Sequence[str] | None = None,
+        batches_per_block: int | Literal["suggested"] | None = None,
+        data_type: str | None = None,
+        leading_dimensions: Sequence[int] | None = None,
+        block_dim: Sequence[int] | Literal["suggested"] | None = None,
+    ):
+        if fill_mode is None:
+            raise ValueError("fill_mode must be provided for CholeskySolver")
+
+        def construct(function):
+            return Solver(
+                function=function,
+                size=size,
+                precision=precision,
+                execution=execution,
+                sm=sm,
+                arrangement=arrangement,
+                fill_mode=fill_mode,
+                batches_per_block=batches_per_block,
+                data_type=data_type,
+                leading_dimensions=leading_dimensions,
+                block_dim=block_dim,
+            )
+
+        self._factorize = construct("potrf")
+        self._solve = construct("potrs")
+
+        super().__init__(self._solve)
+
+        if self.m != self.n:
+            raise ValueError("A must be a square matrix for CholeskySolver.")
+
+    # ==========================
+    # Property methods
+    # ==========================
+
+    @property
+    def fill_mode(self) -> str:
+        return self._solve.fill_mode
+
+    # ==========================
+    # Device function methods
+    # ==========================
+
+    def factorize(self, a, info, lda=None) -> None:
+        """
+        Computes the Cholesky factorization of a symmetric positive definite matrix A.
+
+        This device function computes A = L @ L^H (if fill_mode = ``'lower'``)
+        or A = U^H @ U (if fill_mode = ``'upper'``). Uses cuSOLVERDx ``'potrf'``.
+
+        If ``lda`` is provided, uses runtime version with the specified
+        leading dimension. If ``lda`` is not provided (``None``),
+        uses compile-time version with
+        default or constructor-provided leading dimensions.
+
+        For more details, see: :cusolverdx_doc:`get_started/functions/potrf.html`
+
+        Args:
+            a: Pointer to an array in shared memory, storing
+               the matrix according to the specified
+               arrangement and leading dimension (see :meth:`__init__`).
+               On entry, contains the symmetric positive definite matrix.
+               On exit, contains the triangular factor L (lower) or U (upper).
+            info: Pointer to a 1D array of ``int32``. On exit, ``info[batch_id] = 0``
+                  indicates success for that batch, ``info[batch_id] != 0``
+                  indicates the matrix is not positive definite.
+            lda: Optional runtime leading dimension of matrix A.
+                 If not specified, the compile-time ``lda`` is used.
+        """
+        raise RuntimeError("factorize is a device function and cannot be called on the host.")
+
+    def solve(self, a, b, lda=None, ldb=None) -> None:
+        """
+        Solves a system of linear equations Ax = B using the Cholesky factorization.
+
+        This device function uses the previously computed
+        factorization A = L @ L^H (lower) or A = U^H @ U (upper)
+        to solve the system. Uses cuSOLVERDx ``'potrs'``.
+
+        If ``lda`` and ``ldb`` are provided, uses
+        runtime version with the specified leading dimensions.
+        If not provided (``None``), uses
+        compile-time version with default or constructor-provided leading dimensions.
+
+        For more details, see: :cusolverdx_doc:`get_started/functions/potrs.html`
+
+        Args:
+            a: Pointer to an array in shared memory, storing
+               the triangular factor L (lower) or U (upper)
+               from the Cholesky factorization, according
+               to the specified arrangement and leading dimension (see :meth:`__init__`).
+            b: Pointer to an array in shared memory, storing the
+               matrix according to the specified
+               arrangement and leading dimension (see :meth:`__init__`).
+               The matrix is overwritten in place with the solution matrix x.
+            lda: Optional runtime leading dimension of matrix A.
+                 The ``lda`` and ``ldb`` must be specified together.
+                 If not specified, the compile-time ``lda`` is used.
+            ldb: Optional runtime leading dimension of matrix B.
+                 The ``lda`` and ``ldb`` must be specified together.
+                 If not specified, the compile-time ``ldb`` is used.
+        """
+        raise RuntimeError("solve is a device function and cannot be called on the host.")
+
+
+# ==========================
+# LU Solver
+# ==========================
+
+LU_SOLVER_DOCSTRING = SOLVER_DOCSTRING.copy()
+del LU_SOLVER_DOCSTRING["function"]
+del LU_SOLVER_DOCSTRING["fill_mode"]
+del LU_SOLVER_DOCSTRING["side"]
+del LU_SOLVER_DOCSTRING["diag"]
+del LU_SOLVER_DOCSTRING["job"]
+
+LU_SOLVER_DOCSTRING["size"] = f"""{SOLVER_DOCSTRING_SIZE_BASE}
+``M`` and ``N`` represent the dimensions of the matrix A used in factorization.
+``K`` represents the number of columns in the right-hand side matrix B (dimensions ``N`` x ``K``) for the ``solve`` operation.
+To use :meth:`solve`, ``N`` must be equal to ``M``,
+otherwise an exception will be thrown when ``solver.solve()`` is used.""".replace("\n", " ")
+LU_SOLVER_DOCSTRING["leading_dimensions"] = SOLVER_DOCSTRING["leading_dimensions"] + ADAPTERS_API_LD_DOCSTRING
+LU_SOLVER_DOCSTRING["transpose_mode"] = f"""Transpose mode of matrix A for the solve operation.
+Can be one of: {", ".join(f"``'{v}'``" for v in ALLOWED_TRANSPOSE_MODE)}.
+Defaults to ``'non_transposed'``.""".replace("\n", " ")
+
+
+@docstring_decorator(LU_SOLVER_DOCSTRING, skip_missing=False)
+class LUSolver(_LinearSolverProperties):
+    """
+    A class that encapsulates cuSOLVERDx LU factorization without pivoting
+    and linear solver for general matrices.
+
+    **Available operations:**
+
+    * factorize: Computes the LU factorization A = L @ U,
+      where L is a unit lower triangular matrix and U is an upper triangular matrix.
+    * solve: Solves the system Ax = B using a previously computed LU factorization.
+
+    **Memory Layout Requirements:**
+
+    Matrices must be stored in shared memory according
+    to their arrangement and leading dimension (ld):
+
+    **For matrix A (M x N):**
+
+    * **Column-major arrangement**: Matrix shape ``(batches_per_block, M, N)``
+      with strides ``(lda * N, 1, lda)``
+    * **Row-major arrangement**: Matrix shape ``(batches_per_block, M, N)``
+      with strides ``(lda * M, lda, 1)``
+
+    **For matrix B (N x K):**
+
+    * **Column-major arrangement**: Matrix shape ``(batches_per_block, N, K)``
+      with strides ``(ldb * K, 1, ldb)``
+    * **Row-major arrangement**: Matrix shape ``(batches_per_block, N, K)``
+      with strides ``(ldb * N, ldb, 1)``
+
+    .. note::
+        If a nonsingular matrix A is diagonal dominant, then it is safe to factorize
+        without pivoting. If a matrix is not diagonal dominant, then pivoting is usually
+        required to ensure numerical stability (see :class:`LUPivotSolver`).
+
+    Args:
+        size (Sequence[int]): {size}
+
+        precision (type[np.floating]): {precision}
+
+        execution (str): {execution}
+
+        sm (ComputeCapability): {sm}
+
+        transpose_mode (str, optional): {transpose_mode}
+
+        arrangement (Sequence[str], optional): {arrangement}
+
+        batches_per_block (int | Literal["suggested"], optional): {batches_per_block}
+
+        data_type (str, optional): {data_type}
+
+        leading_dimensions (Sequence[int], optional): {leading_dimensions}
+
+        block_dim (Sequence[int] | Literal["suggested"], optional): {block_dim}
+
+    See Also:
+        For further details, please refer to the cuSOLVERDx documentation:
+
+          * :cusolverdx_doc:`factorize <get_started/functions/getrf.html>`
+          * :cusolverdx_doc:`solve <get_started/functions/getrs.html>`
+
+    """
+
+    # ==========================
+    # Constructor
+    # ==========================
+
+    def __init__(
+        self,
+        size: Sequence[int],
+        precision: type[np.floating],
+        execution: str,
+        *,
+        sm=None,
+        transpose_mode: str = "non_transposed",
+        arrangement: Sequence[str] | None = None,
+        batches_per_block: int | Literal["suggested"] | None = None,
+        data_type: str | None = None,
+        leading_dimensions: Sequence[int] | None = None,
+        block_dim: Sequence[int] | Literal["suggested"] | None = None,
+    ):
+        def construct(function, transpose_mode):
+            return Solver(
+                function=function,
+                size=size,
+                precision=precision,
+                execution=execution,
+                sm=sm,
+                arrangement=arrangement,
+                batches_per_block=batches_per_block,
+                data_type=data_type,
+                leading_dimensions=leading_dimensions,
+                block_dim=block_dim,
+                transpose_mode=transpose_mode,
+            )
+
+        if transpose_mode is None:
+            raise ValueError("transpose_mode must be provided for LUSolver")
+
+        self._factorize = construct("getrf_no_pivot", "non_transposed")
+        self._transpose_mode = transpose_mode
+        self._solve = construct("getrs_no_pivot", transpose_mode) if self._factorize.m == self._factorize.n else None
+
+        super().__init__(self._factorize)
+
+    # ==========================
+    # Property methods
+    # ==========================
+
+    @property
+    def transpose_mode(self) -> str:
+        return self._transpose_mode
+
+    # ==========================
+    # Device function methods
+    # ==========================
+
+    def factorize(self, a, info, lda=None) -> None:
+        """
+        Computes the LU factorization of a general matrix A without pivoting.
+
+        This device function computes A = L @ U, where L is a unit lower triangular matrix
+        and U is an upper triangular matrix. This variant is suitable for diagonally
+        dominant matrices or when pivoting is not required.
+        Uses cuSOLVERDx ``'getrf_no_pivot'``.
+
+        If ``lda`` is provided, uses runtime version with the specified leading dimension.
+        If ``lda`` is not provided (``None``), uses compile-time version with default
+        or constructor-provided leading dimensions.
+
+        .. note::
+            The ``transpose_mode`` parameter does not affect factorization.
+            This operation always treats the input matrix as-is (non-transposed).
+
+        For more details, see: :cusolverdx_doc:`get_started/functions/getrf.html`
+
+        Args:
+            a: Pointer to an array in shared memory, storing the batched matrix according
+               to the specified arrangement and leading dimension (see :meth:`__init__`).
+               The matrix is overwritten in place.
+               On exit, contains the factors L and U from the factorization A = L @ U.
+               The unit diagonal elements of L are not stored.
+            info: Pointer to a 1D array of ``int32``.
+                  On exit, ``info[batch_id] = 0`` indicates success for that batch,
+                  ``info[batch_id] = i > 0`` indicates ``U(i,i)`` is exactly zero,
+                  meaning the factorization has been completed but the factor U
+                  is singular and division by zero will occur if it is used to solve
+                  a system of equations.
+            lda: Optional runtime leading dimension of matrix A.
+                 If not specified, the compile-time ``lda`` is used.
+        """
+        raise RuntimeError("factorize is a device function and cannot be called on the host.")
+
+    def solve(self, a, b, lda=None, ldb=None) -> None:
+        """
+        Solves a system of linear equations Ax = B
+        using the LU factorization without pivoting.
+        The ``a`` operand must be a square matrix (``M == N``),
+        otherwise this function will throw an exception.
+
+        This device function uses the previously computed factorization A = L @ U
+        to solve the system. Uses cuSOLVERDx ``'getrs_no_pivot'``.
+
+        If ``lda`` and ``ldb`` are provided, uses runtime version with
+        the specified leading dimensions. If not provided (``None``),
+        uses compile-time version with default or constructor-provided
+        leading dimensions.
+
+        .. note::
+            The ``transpose_mode`` parameter (set in constructor) determines which
+            system is solved: A*x=B (``'non_transposed'``), A^T*x=B (``'transposed'``),
+            or A^H*x=B (``'conj_transposed'`` for complex matrices).
+
+        For more details, see: :cusolverdx_doc:`get_started/functions/getrs.html`
+
+        Args:
+            a: Pointer to an array in shared memory, storing the batched factors L and U
+               from the LU factorization, according to the specified arrangement
+               and leading dimension (see :meth:`__init__`).
+               The unit diagonal elements of L are not stored.
+               See the :meth:`factorize` documentation for details.
+            b: Pointer to an array in shared memory, storing the batched matrix according
+               to the specified arrangement and leading dimension (see :meth:`__init__`).
+               The matrix is overwritten in place with the solution matrix x.
+            lda: Optional runtime leading dimension of matrix A.
+                 The ``lda`` and ``ldb`` must be specified together.
+                 If not specified, the compile-time ``lda`` is used.
+            ldb: Optional runtime leading dimension of matrix B.
+                 The ``lda`` and ``ldb`` must be specified together.
+                 If not specified, the compile-time ``ldb`` is used.
+        """
+        raise RuntimeError("solve is a device function and cannot be called on the host.")
+
+
+# ==========================
+# Triangular Solver
+# ==========================
+
+TRIANGULAR_SOLVER_DOCSTRING = SOLVER_DOCSTRING.copy()
+del TRIANGULAR_SOLVER_DOCSTRING["function"]
+del TRIANGULAR_SOLVER_DOCSTRING["job"]
+
+TRIANGULAR_SOLVER_DOCSTRING["size"] = f"""{SOLVER_DOCSTRING_SIZE_BASE}
+``M`` and ``N`` represent the dimensions of matrices A and B.
+When ``side='left'``, A is ``M`` x ``M``, otherwise when ``side='right'``, A is ``N`` x ``N``.
+B is always ``M`` x ``N``.""".replace("\n", " ")
+
+TRIANGULAR_SOLVER_DOCSTRING["side"] = f"""Side of matrix A in the triangular solve operation (required for TRSM).
+Can be one of: {", ".join(f"``'{v}'``" for v in ALLOWED_SIDE)}.
+If ``side='left'``, solves op(A) * X = B where A is ``M`` x ``M``.
+If ``side='right'``, solves X * op(A) = B where A is ``N`` x ``N``.""".replace("\n", " ")
+
+TRIANGULAR_SOLVER_DOCSTRING["fill_mode"] = f"""Indicates which part of triangular matrix A is filled and should be used.
+Can be one of: {", ".join(f"``'{v}'``" for v in ALLOWED_FILL_MODE)}.
+For lower fill mode, only the diagonal and lower triangular part of A is processed, the upper part is untouched.
+For upper fill mode, only the diagonal and upper triangular part of A is processed, the lower part is untouched.""".replace(
+    "\n", " "
+)
+
+TRIANGULAR_SOLVER_DOCSTRING["diag"] = f"""Indicates whether the diagonal elements of matrix A are unity or not.
+Can be one of: {", ".join(f"``'{v}'``" for v in ALLOWED_DIAG)}.
+For unit diagonal mode, the diagonal elements of A are unity and are not accessed.
+For non-unit diagonal mode, the diagonal elements of A are used in the computation.""".replace("\n", " ")
+
+TRIANGULAR_SOLVER_DOCSTRING["transpose_mode"] = f"""Transpose mode for operation op(A) applied to matrix A.
+Can be one of: {", ".join(f"``'{v}'``" for v in ALLOWED_TRANSPOSE_MODE)}.
+Defaults to ``'non_transposed'``.""".replace("\n", " ")
+
+TRIANGULAR_SOLVER_DOCSTRING["leading_dimensions"] = SOLVER_DOCSTRING["leading_dimensions"] + ADAPTERS_API_LD_DOCSTRING
+
+
+@docstring_decorator(TRIANGULAR_SOLVER_DOCSTRING, skip_missing=False)
+class TriangularSolver(_SolverProperties):
+    """
+    A class that encapsulates triangular matrix-matrix solve device function (``'trsm'``).
+
+    TRSM (TRiangular Solve for Matrix) solves a triangular linear system
+    with multiple right-hand sides:
+
+    * op(A) * X = B (if ``side='left'``)
+    * X * op(A) = B (if ``side='right'``)
+
+    where:
+
+    * A is the input batched triangular matrix stored in lower or upper mode
+    * B is the batched right-hand side matrix, overwritten by the result X on exit
+    * Operation op(A) indicates if matrix A is ``'non_transposed'``,
+      ``'transposed'`` (for real data type),
+      or ``'conj_transposed'`` (for complex data type)
+
+    **Memory Layout Requirements:**
+
+    Matrices must be stored in shared memory according
+    to their arrangement and leading dimension (ld):
+
+    **For matrix A (M x M) with ``side='left'``:**
+
+    * **Column-major arrangement**: Matrix shape ``(batches_per_block, M, M)``
+      with strides ``(lda * M, 1, lda)``
+    * **Row-major arrangement**: Matrix shape ``(batches_per_block, M, M)``
+      with strides ``(lda * M, lda, 1)``
+
+    **For matrix A (N x N) with ``side='right'``:**
+
+    * **Column-major arrangement**: Matrix shape ``(batches_per_block, N, N)``
+      with strides ``(lda * N, 1, lda)``
+    * **Row-major arrangement**: Matrix shape ``(batches_per_block, N, N)``
+      with strides ``(lda * N, lda, 1)``
+
+    **For matrix B (M x N):**
+
+    * **Column-major arrangement**: Matrix shape ``(batches_per_block, M, N)``
+      with strides ``(ldb * N, 1, ldb)``
+    * **Row-major arrangement**: Matrix shape ``(batches_per_block, M, N)``
+      with strides ``(ldb * M, ldb, 1)``
+
+    .. note::
+        The TRSM function is temporarily exposed in cuSolverDx library
+        and will be moved to cuBLASDx library in a future release.
+
+    Args:
+        size (Sequence[int]): {size}
+
+        precision (type[np.floating]): {precision}
+
+        execution (str): {execution}
+
+        sm (ComputeCapability): {sm}
+
+        side (str): {side}
+
+        fill_mode (str): {fill_mode}
+
+        diag (str): {diag}
+
+        transpose_mode (str, optional): {transpose_mode}
+
+        arrangement (Sequence[str], optional): {arrangement}
+
+        batches_per_block (int | Literal["suggested"], optional): {batches_per_block}
+
+        data_type (str, optional): {data_type}
+
+        leading_dimensions (Sequence[int], optional): {leading_dimensions}
+
+        block_dim (Sequence[int] | Literal["suggested"], optional): {block_dim}
+
+    See Also:
+        For further details, please refer to the cuSOLVERDx documentation:
+
+        * :cusolverdx_doc:`trsm <get_started/functions/trsm.html>`
+    """
+
+    # ==========================
+    # Constructor
+    # ==========================
+
+    def __init__(
+        self,
+        size: Sequence[int],
+        precision: type[np.floating],
+        execution: str,
+        side: str,
+        fill_mode: str,
+        diag: str,
+        transpose_mode: str = "non_transposed",
+        *,
+        sm=None,
+        arrangement: Sequence[str] | None = None,
+        batches_per_block: int | Literal["suggested"] | None = None,
+        data_type: str | None = None,
+        leading_dimensions: Sequence[int] | None = None,
+        block_dim: Sequence[int] | Literal["suggested"] | None = None,
+    ):
+        self._solve: Solver = Solver(
+            function="trsm",
+            size=size,
+            precision=precision,
+            execution=execution,
+            sm=sm,
+            side=side,
+            fill_mode=fill_mode,
+            diag=diag,
+            transpose_mode=transpose_mode,
+            arrangement=arrangement,
+            batches_per_block=batches_per_block,
+            data_type=data_type,
+            leading_dimensions=leading_dimensions,
+            block_dim=block_dim,
+        )
+
+        super().__init__(self._solve)
+
+        if not isinstance(self._solve.side, str):
+            raise ValueError("side must be provided for TriangularSolver")
+
+        if not isinstance(self._solve.fill_mode, str):
+            raise ValueError("fill_mode must be provided for TriangularSolver")
+
+        if not isinstance(self._solve.diag, str):
+            raise ValueError("diag must be provided for TriangularSolver")
+
+    # ==========================
+    # Property methods
+    # ==========================
+
+    @property
+    def side(self) -> str:
+        assert self._solve.side is not None
+        return self._solve.side
+
+    @property
+    def fill_mode(self) -> str:
+        assert self._solve.fill_mode is not None
+        return self._solve.fill_mode
+
+    @property
+    def diag(self) -> str:
+        assert self._solve.diag is not None
+        return self._solve.diag
+
+    @property
+    def transpose_mode(self) -> str:
+        assert self._solve.transpose_mode is not None
+        return self._solve.transpose_mode
+
+    @property
+    def a_shape(self) -> tuple[int, int, int]:
+        dim = self.m if self.side == "left" else self.n
+        return (self.batches_per_block, dim, dim)
+
+    @property
+    def b_shape(self) -> tuple[int, int, int]:
+        return (self.batches_per_block, self.m, self.n)
+
+    def a_strides(self, *, lda: int | None = None) -> tuple[int, int, int]:
+        lda = self.lda if lda is None else lda
+        return _calculate_strides(self.a_shape[1:], lda, self.a_arrangement)
+
+    def b_strides(self, *, ldb: int | None = None) -> tuple[int, int, int]:
+        ldb = self.ldb if ldb is None else ldb
+        return _calculate_strides(self.b_shape[1:], ldb, self.b_arrangement)
+
+    def a_size(self, *, lda: int | None = None) -> int:
+        return self.a_strides(lda=lda)[0] * self.a_shape[0]
+
+    def b_size(self, *, ldb: int | None = None) -> int:
+        return self.b_strides(ldb=ldb)[0] * self.b_shape[0]
+
+    # ==========================
+    # Device function methods
+    # ==========================
+
+    def solve(self, a, b, lda=None, ldb=None) -> None:
+        """
+        Solves a triangular linear system with multiple right-hand sides:
+            ``op(A) * X = B`` (if ``side='left'``)
+            ``X * op(A) = B`` (if ``side='right'``)
+
+        This device function solves a triangular system where A is a triangular matrix.
+        Uses cuSOLVERDx ``'trsm'``. The operation is in-place: result X overwrites B.
+
+        If ``lda`` and ``ldb`` are provided, uses runtime version with the
+        specified leading dimensions. If not provided (``None``), uses compile-time
+        version with default or constructor-provided leading dimensions.
+
+        For more details, see: :cusolverdx_doc:`get_started/functions/trsm.html`
+
+        Args:
+            a: Pointer to an array in shared memory, storing the batched triangular matrix
+               according to the specified arrangement
+               and leading dimension (see :meth:`__init__`).
+               The ``fill_mode`` parameter denotes which
+               part of the matrix is used (the other part is ignored).
+               For unit diagonal mode (``diag='unit'``),
+               diagonal elements are unity and not accessed.
+            b: Pointer to an array in shared memory,
+               storing the batched ``M`` x ``N`` right-hand side
+               matrix according to the specified arrangement
+               and leading dimension (see :meth:`__init__`).
+               The operation is in-place: result X overwrites B.
+            lda: Optional runtime leading dimension for matrix A.
+                 The ``lda`` and ``ldb`` must be specified together.
+                 If not specified, the compile-time ``lda`` is used.
+            ldb: Optional runtime leading dimension for matrix B.
+                 The ``lda`` and ``ldb`` must be specified together.
+                 If not specified, the compile-time ``ldb`` is used.
+        """
+        raise RuntimeError("solve is a device function and cannot be called on the host.")
+
+
+# ==========================
+# LU Pivot Solver
+# ==========================
+
+
+@docstring_decorator(LU_SOLVER_DOCSTRING, skip_missing=False)
+class LUPivotSolver(_LinearSolverProperties):
+    """
+    A class that encapsulates cuSOLVERDx LU factorization with partial pivoting
+    and linear solver for general matrices.
+
+    **Available operations:**
+
+    * factorize: Computes the LU factorization P @ A = L @ U with partial pivoting,
+      where P is a permutation matrix, L is a
+      lower triangular matrix and U is an upper triangular matrix.
+    * solve: Solves the system Ax = B using a previously
+      computed LU factorization with partial pivoting
+
+    **Memory Layout Requirements:**
+
+    Matrices must be stored in shared memory according
+    to their arrangement and leading dimension (ld):
+
+    **For matrix A (M x N):**
+
+    * **Column-major arrangement**: Matrix shape ``(batches_per_block, M, N)``
+      with strides ``(lda * N, 1, lda)``
+    * **Row-major arrangement**: Matrix shape ``(batches_per_block, M, N)``
+      with strides ``(lda * M, lda, 1)``
+
+    **For matrix B (N x K):**
+
+    * **Column-major arrangement**: Matrix shape ``(batches_per_block, N, K)``
+      with strides ``(ldb * K, 1, ldb)``
+    * **Row-major arrangement**: Matrix shape ``(batches_per_block, N, K)``
+      with strides ``(ldb * N, ldb, 1)``
+
+    .. note::
+        This solver uses partial pivoting for improved numerical stability and is
+        suitable for general matrices. If your matrix is diagonally dominant,
+        you may consider using :class:`LUSolver` which does not use pivoting
+        and may be faster.
+
+    Args:
+        size (Sequence[int]): {size}
+
+        precision (type[np.floating]): {precision}
+
+        execution (str): {execution}
+
+        sm (ComputeCapability): {sm}
+
+        transpose_mode (str, optional): {transpose_mode}
+
+        arrangement (Sequence[str], optional): {arrangement}
+
+        batches_per_block (int | Literal["suggested"], optional): {batches_per_block}
+
+        data_type (str, optional): {data_type}
+
+        leading_dimensions (Sequence[int], optional): {leading_dimensions}
+
+        block_dim (Sequence[int] | Literal["suggested"], optional): {block_dim}
+
+    See Also:
+        For further details, please refer to the cuSOLVERDx documentation:
+
+          * :cusolverdx_doc:`factorize <get_started/functions/getrf.html>`
+          * :cusolverdx_doc:`solve <get_started/functions/getrs.html>`
+
+    """
+
+    # ==========================
+    # Constructor
+    # ==========================
+
+    def __init__(
+        self,
+        size: Sequence[int],
+        precision: type[np.floating],
+        execution: str,
+        *,
+        sm=None,
+        transpose_mode: str = "non_transposed",
+        arrangement: Sequence[str] | None = None,
+        batches_per_block: int | Literal["suggested"] | None = None,
+        data_type: str | None = None,
+        leading_dimensions: Sequence[int] | None = None,
+        block_dim: Sequence[int] | Literal["suggested"] | None = None,
+    ):
+        def construct(function, transpose_mode):
+            return Solver(
+                function=function,
+                size=size,
+                precision=precision,
+                execution=execution,
+                sm=sm,
+                arrangement=arrangement,
+                batches_per_block=batches_per_block,
+                data_type=data_type,
+                leading_dimensions=leading_dimensions,
+                block_dim=block_dim,
+                transpose_mode=transpose_mode,
+            )
+
+        if transpose_mode is None:
+            raise ValueError("transpose_mode must be provided for LUPivotSolver")
+
+        self._factorize = construct("getrf_partial_pivot", "non_transposed")
+        self._solve = construct("getrs_partial_pivot", transpose_mode) if self._factorize.m == self._factorize.n else None
+        self._transpose_mode = transpose_mode
+
+        super().__init__(self._factorize)
+
+    # ==========================
+    # Property methods
+    # ==========================
+
+    @property
+    def transpose_mode(self) -> str:
+        return self._transpose_mode
+
+    @property
+    def ipiv_type(self) -> type[np.signedinteger]:
+        return self._factorize.ipiv_type
+
+    @property
+    def ipiv_shape(self) -> tuple[int, int]:
+        return (self.batches_per_block, min(self.m, self.n))
+
+    @property
+    def ipiv_strides(self) -> tuple[int, int]:
+        return (self.ipiv_shape[1], 1)
+
+    @property
+    def ipiv_size(self) -> int:
+        return self.ipiv_shape[0] * self.ipiv_shape[1]
+
+    # ==========================
+    # Device function methods
+    # ==========================
+
+    def factorize(self, a, ipiv, info, lda=None) -> None:
+        """
+        Computes the LU factorization of a general matrix A with partial pivoting.
+
+        This device function computes P @ A = L @ U, where P is a permutation matrix,
+        L is a unit lower triangular matrix and U is an upper triangular matrix.
+        This variant uses partial pivoting for improved numerical stability
+        and is suitable for general matrices. Uses cuSOLVERDx ``'getrf_partial_pivot'``.
+
+        If ``lda`` is provided, uses runtime version with the specified leading dimension.
+        If ``lda`` is not provided (``None``), uses compile-time version with default
+        or constructor-provided leading dimensions.
+
+        .. note::
+            The ``transpose_mode`` parameter does not affect factorization.
+            This operation always treats the input matrix as-is (non-transposed).
+
+        For more details, see: :cusolverdx_doc:`get_started/functions/getrf.html`
+
+        Args:
+            a: Pointer to an array in shared memory, storing the batched matrix according
+               to the specified arrangement and leading dimension (see :meth:`__init__`).
+               The matrix is overwritten in place.
+               On exit, contains the factors L and U from the factorization P @ A = L @ U.
+               The unit diagonal elements of L are not stored.
+            ipiv: Pointer to a 1D array of ``int32``, storing pivot indices.
+                  The array has size min(M, N) for each batch.
+                  On exit, ``ipiv[batch_id * min(M, N) + i]`` indicates that row i
+                  was interchanged with row ``ipiv[batch_id * min(M, N) + i] - 1``
+                  in the batch_id-th batch of A.
+            info: Pointer to a 1D array of ``int32``. On exit, ``info[batch_id] = 0``
+                  indicates success for that batch, ``info[batch_id] = i > 0``
+                  indicates ``U(i,i)`` is exactly zero,
+                  meaning the factorization has been completed but the factor U
+                  is singular and division by zero will occur if it is used to solve
+                  a system of equations.
+            lda: Optional runtime leading dimension of matrix A.
+                 If not specified, the compile-time ``lda`` is used.
+        """
+        raise RuntimeError("factorize is a device function and cannot be called on the host.")
+
+    def solve(self, a, ipiv, b, lda=None, ldb=None) -> None:
+        """
+        Solves a system of linear equations Ax = B
+        using the LU factorization with partial pivoting.
+        The ``a`` operand must be a square matrix (``M == N``),
+        otherwise this function will throw an exception.
+
+        This device function uses the previously computed factorization P @ A = L @ U
+        to solve the system. Uses cuSOLVERDx ``'getrs_partial_pivot'``.
+
+        If ``lda`` and ``ldb`` are provided, uses runtime version with
+        the specified leading dimensions. If not provided (``None``),
+        uses compile-time version with default or constructor-provided
+        leading dimensions.
+
+        .. note::
+            The ``transpose_mode`` parameter (set in constructor) determines which
+            system is solved: A*x=B (``'non_transposed'``), A^T*x=B (``'transposed'``),
+            or A^H*x=B (``'conj_transposed'`` for complex matrices).
+
+        For more details, see: :cusolverdx_doc:`get_started/functions/getrs.html`
+
+        Args:
+            a: Pointer to an array in shared memory, storing the batched factors L and U
+               from the LU factorization with partial pivoting, according
+               to the specified arrangement and leading dimension (see :meth:`__init__`).
+               The unit diagonal elements of L are not stored.
+               See the :meth:`factorize` documentation for details.
+            ipiv: Pointer to a 1D array of ``int32`` in shared or global memory
+                  storing pivot indices.
+                  The array has size min(M, N) for each batch. The ipiv array should contain
+                  the pivot information from the :meth:`factorize` call.
+                  ``ipiv[batch_id * min(M, N) + i]`` indicates that row i was interchanged
+                  with row ``ipiv[batch_id * min(M, N) + i]`` in the batch_id-th batch of A.
+            b: Pointer to an array in shared memory, storing the batched matrix according
+               to the specified arrangement and leading dimension (see :meth:`__init__`).
+               The matrix is overwritten in place with the solution matrix x.
+            lda: Optional runtime leading dimension of matrix A.
+                 The ``lda`` and ``ldb`` must be specified together.
+                 If not specified, the compile-time ``lda`` is used.
+            ldb: Optional runtime leading dimension of matrix B.
+                 The ``lda`` and ``ldb`` must be specified together.
+                 If not specified, the compile-time ``ldb`` is used.
+        """
+        raise RuntimeError("solve is a device function and cannot be called on the host.")
+
+
+# ==========================
+# QR/LQ Factorizers Base
+# ==========================
+
+ORTOGHONAL_FACTORIZER_DOCSTRING = SOLVER_DOCSTRING.copy()
+del ORTOGHONAL_FACTORIZER_DOCSTRING["function"]
+del ORTOGHONAL_FACTORIZER_DOCSTRING["fill_mode"]
+del ORTOGHONAL_FACTORIZER_DOCSTRING["side"]
+del ORTOGHONAL_FACTORIZER_DOCSTRING["diag"]
+del ORTOGHONAL_FACTORIZER_DOCSTRING["transpose_mode"]
+del ORTOGHONAL_FACTORIZER_DOCSTRING["job"]
+
+ORTOGHONAL_FACTORIZER_DOCSTRING["size"] = f"""{SOLVER_DOCSTRING_SIZE_BASE}
+``M`` and ``N`` represent the dimensions of the matrix A used in factorization.
+``K`` is ignored if specified.""".replace("\n", " ")
+
+ORTOGHONAL_FACTORIZER_DOCSTRING["arrangement"] = (
+    """Storage layout for matrix A.
+Can be one of: ``'col_major'``, ``'row_major'``. Defaults to ``'col_major'``.""".replace("\n", " ")
+    + " "
+    + ADAPTERS_API_LD_DOCSTRING
+)
+
+ORTOGHONAL_FACTORIZER_DOCSTRING["leading_dimension"] = (
+    """The leading dimension for input matrix A, or ``None``.
+If not provided, it will be automatically deduced from ``size`` and ``arrangement``.""".replace("\n", " ")
+    + " "
+    + ADAPTERS_API_LD_DOCSTRING
+)
+
+
+class _OrthogonalFactorizerProperties:
+    def __init__(self, source: Solver):
+        self._properties_source = source
+
+    @property
+    def size(self) -> tuple[int, int, int]:
+        return self._properties_source.size
+
+    @property
+    def m(self) -> int:
+        return self._properties_source.m
+
+    @property
+    def n(self) -> int:
+        return self._properties_source.n
+
+    @property
+    def precision(self) -> type[np.floating]:
+        return self._properties_source.precision
+
+    @property
+    def execution(self) -> str:
+        return self._properties_source.execution
+
+    @property
+    def sm(self) -> ComputeCapability:
+        return self._properties_source.sm
+
+    @property
+    def a_arrangement(self) -> str:
+        return self._properties_source.a_arrangement
+
+    @property
+    def batches_per_block(self) -> int:
+        return self._properties_source.batches_per_block
+
+    @property
+    def data_type(self) -> str:
+        return self._properties_source.data_type
+
+    @property
+    def lda(self) -> int:
+        return self._properties_source.lda
+
+    @property
+    def block_dim(self) -> tuple:
+        return self._properties_source.block_dim
+
+    @property
+    def block_size(self) -> int:
+        return self._properties_source.block_size
+
+    @property
+    def value_type(self) -> type[np.floating] | Complex:
+        return self._properties_source.value_type
+
+    @property
+    def tau_type(self) -> type[np.floating] | Complex:
+        return self._properties_source.tau_type
+
+    @property
+    def tau_shape(self) -> tuple[int, int]:
+        return (self.batches_per_block, min(self.m, self.n))
+
+    @property
+    def tau_strides(self) -> tuple[int, int]:
+        return (self.tau_shape[1], 1)
+
+    @property
+    def a_shape(self) -> tuple[int, int, int]:
+        return (self.batches_per_block, self.m, self.n)
+
+    def a_strides(self, *, lda: int | None = None) -> tuple[int, int, int]:
+        lda = self.lda if lda is None else lda
+        return _calculate_strides(self.a_shape[1:], lda, self.a_arrangement)
+
+    def a_size(self, *, lda: int | None = None) -> int:
+        return self.a_strides(lda=lda)[0] * self.a_shape[0]
+
+    @property
+    def tau_size(self) -> int:
+        return self.tau_shape[0] * self.tau_shape[1]
+
+
+# ==========================
+# QR Factorize
+# ==========================
+
+
+@docstring_decorator(ORTOGHONAL_FACTORIZER_DOCSTRING, skip_missing=False)
+class QRFactorize(_OrthogonalFactorizerProperties):
+    """
+    A class that encapsulates QR orthogonal factorization device function
+    for general matrices using Householder reflections.
+
+    **Available operation:**
+
+    * factorize: Computes the QR factorization A = Q @ R,
+      where Q is a unitary M x M matrix
+      and R is an upper triangular matrix (if M >= N)
+      or upper trapezoidal matrix (if M < N).
+
+    The factorization uses Householder reflection transformations and does not explicitly
+    form the unitary matrix Q. Instead, Q is represented as a product of Householder vectors
+    stored in the input matrix A along with the tau array.
+
+    **Memory Layout Requirements:**
+
+    Matrices must be stored in shared memory according
+    to their arrangement and leading dimension (ld):
+
+    **For matrix A (M x N):**
+
+    * **Column-major arrangement**: Matrix shape ``(batches_per_block, M, N)``
+      with strides ``(lda * N, 1, lda)``
+    * **Row-major arrangement**: Matrix shape ``(batches_per_block, M, N)``
+      with strides ``(lda * M, lda, 1)``
+
+    Args:
+        size (Sequence[int]): {size}
+
+        precision (type[np.floating]): {precision}
+
+        execution (str): {execution}
+
+        sm (ComputeCapability): {sm}
+
+        arrangement (str, optional): {arrangement}
+
+        batches_per_block (int | Literal["suggested"], optional): {batches_per_block}
+
+        data_type (str, optional): {data_type}
+
+        leading_dimension (int, optional): {leading_dimension}
+
+        block_dim (Sequence[int] | Literal["suggested"], optional): {block_dim}
+
+    See Also:
+        For further details, please refer to the cuSOLVERDx documentation:
+
+        * :cusolverdx_doc:`factorize (geqrf) <get_started/functions/geqrf.html>`
+    """
+
+    # ==========================
+    # Constructor
+    # ==========================
+
+    def __init__(
+        self,
+        size: Sequence[int],
+        precision: type[np.floating],
+        execution: str,
+        *,
+        sm=None,
+        arrangement: str | None = None,
+        batches_per_block: int | Literal["suggested"] | None = None,
+        data_type: str | None = None,
+        leading_dimension: int | None = None,
+        block_dim: Sequence[int] | Literal["suggested"] | None = None,
+    ):
+        self._factorize = Solver(
+            function="geqrf",
+            size=size,
+            precision=precision,
+            execution=execution,
+            sm=sm,
+            arrangement=(arrangement,) if arrangement is not None else None,
+            batches_per_block=batches_per_block,
+            data_type=data_type,
+            leading_dimensions=(leading_dimension,) if leading_dimension is not None else None,
+            block_dim=block_dim,
+        )
+        super().__init__(self._factorize)
+
+    # ==========================
+    # Device function methods
+    # ==========================
+
+    def factorize(self, a, tau, lda=None) -> None:
+        """
+        Computes the QR factorization of a general matrix A using Householder reflections.
+
+        This device function computes A = Q @ R, where Q is a unitary M x M matrix
+        and R is an upper triangular matrix (if M >= N)
+        or upper trapezoidal matrix (if M < N).
+        Uses cuSOLVERDx ``'geqrf'``.
+
+        If ``lda`` is provided, uses runtime version with the
+        specified leading dimension. If ``lda`` is not provided (``None``),
+        uses compile-time version with
+        default or constructor-provided leading dimensions.
+
+        Matrix Q is not explicitly formed. Instead, Q is represented as a product of
+        min(M, N) Householder vectors: Q = H(0) * H(1) * ... * H(min(M, N) - 1).
+
+        Each Householder vector has the form H(i) = I - tau[i] * v * v^H, where:
+
+        * v is a vector of size M for each batch
+        * v[0:i-1] = 0, v[i] = 1
+        * v[i+1:M] is stored on exit in A[i+1:M, i]
+
+        For more details, see: :cusolverdx_doc:`get_started/functions/geqrf.html`
+
+        Args:
+            a: Pointer to an array in shared memory, storing
+               the batched matrix according
+               to the specified arrangement and leading dimension (see :meth:`__init__`).
+               The matrix is overwritten in place.
+               On exit, the upper triangular or upper trapezoidal part (including diagonal)
+               contains the matrix R. The elements below the diagonal, with the array tau,
+               represent the unitary matrix Q as a product of Householder vectors.
+            tau: Pointer to a 1D array of size min(M, N) for each batch.
+                 Contains the scalar factors of the Householder reflections.
+                 The tau array, together with the Householder vectors stored in A,
+                 defines the unitary matrix Q.
+            lda: Optional runtime leading dimension of matrix A.
+                 If not specified, the compile-time ``lda`` is used.
+        """
+        raise RuntimeError("factorize is a device function and cannot be called on the host.")
+
+
+# ==========================
+# LQ Factorize
+# ==========================
+
+
+@docstring_decorator(ORTOGHONAL_FACTORIZER_DOCSTRING, skip_missing=False)
+class LQFactorize(_OrthogonalFactorizerProperties):
+    """
+    A class that encapsulates LQ orthogonal factorization device function
+    for general matrices using Householder reflections.
+
+    **Available operation:**
+
+    * factorize: Computes the LQ factorization A = L @ Q,
+      where L is a lower triangular matrix (if M <= N)
+      or lower trapezoidal matrix (if M > N)
+      and Q is a unitary N x N matrix.
+
+    The factorization uses Householder reflection transformations and does not explicitly
+    form the unitary matrix Q. Instead, Q is represented as a product of Householder vectors
+    stored in the input matrix A along with the tau array.
+
+    **Memory Layout Requirements:**
+
+    Matrices must be stored in shared memory according
+    to their arrangement and leading dimension (ld):
+
+    **For matrix A (M x N):**
+
+    * **Column-major arrangement**: Matrix shape ``(batches_per_block, M, N)``
+      with strides ``(lda * N, 1, lda)``
+    * **Row-major arrangement**: Matrix shape ``(batches_per_block, M, N)``
+      with strides ``(lda * M, lda, 1)``
+
+    .. note::
+        The LQ factorization is essentially equivalent to the
+        QR factorization of A^T (or A^H for complex types).
+
+    Args:
+        size (Sequence[int]): {size}
+
+        precision (type[np.floating]): {precision}
+
+        execution (str): {execution}
+
+        sm (ComputeCapability): {sm}
+
+        arrangement (str, optional): {arrangement}
+
+        batches_per_block (int | Literal["suggested"], optional): {batches_per_block}
+
+        data_type (str, optional): {data_type}
+
+        leading_dimension (int, optional): {leading_dimension}
+
+        block_dim (Sequence[int] | Literal["suggested"], optional): {block_dim}
+
+    See Also:
+        For further details, please refer to the cuSOLVERDx documentation:
+
+        * :cusolverdx_doc:`factorize (gelqf) <get_started/functions/gelqf.html>`
+    """
+
+    # ==========================
+    # Constructor
+    # ==========================
+
+    def __init__(
+        self,
+        size: Sequence[int],
+        precision: type[np.floating],
+        execution: str,
+        *,
+        sm=None,
+        arrangement: str | None = None,
+        batches_per_block: int | Literal["suggested"] | None = None,
+        data_type: str | None = None,
+        leading_dimension: int | None = None,
+        block_dim: Sequence[int] | Literal["suggested"] | None = None,
+    ):
+        self._factorize = Solver(
+            function="gelqf",
+            size=size,
+            precision=precision,
+            execution=execution,
+            sm=sm,
+            arrangement=(arrangement,) if arrangement is not None else None,
+            batches_per_block=batches_per_block,
+            data_type=data_type,
+            leading_dimensions=(leading_dimension,) if leading_dimension is not None else None,
+            block_dim=block_dim,
+        )
+        super().__init__(self._factorize)
+
+    # ==========================
+    # Device function methods
+    # ==========================
+
+    def factorize(self, a, tau, lda=None) -> None:
+        """
+        Computes the LQ factorization of a general matrix A using Householder reflections.
+
+        This device function computes A = L @ Q, where L is a lower triangular matrix
+        (if M <= N) or lower trapezoidal matrix (if M > N),
+        and Q is a unitary N x N matrix.
+        Uses cuSOLVERDx ``'gelqf'``.
+
+        If ``lda`` is provided, uses runtime version with the
+        specified leading dimension. If ``lda`` is not provided (``None``),
+        uses compile-time version with
+        default or constructor-provided leading dimensions.
+
+        The LQ factorization is essentially the same as the QR factorization of A^T
+        (or A^H for complex data types).
+
+        Matrix Q is not explicitly formed. Instead, Q is represented as a product of
+        min(M, N) Householder vectors: Q = H(min(M, N) - 1)^H * ... * H(1)^H * H(0)^H.
+
+        Each Householder vector has the form H(i) = I - tau[i] * v * v^H, where:
+
+        * v is a vector of size N for each batch
+        * v[0:i-1] = 0, v[i] = 1
+        * conjugate(v[i+1:N]) is stored on exit in A[i, i+1:N]
+
+        For more details, see: :cusolverdx_doc:`get_started/functions/gelqf.html`
+
+        Args:
+            a: Pointer to an array in shared memory, storing
+               the batched M x N matrix according
+               to the specified arrangement and leading dimension (see :meth:`__init__`).
+               The matrix is overwritten in place.
+               On exit, the lower triangular or lower trapezoidal part (including diagonal)
+               contains the matrix L. The elements above the diagonal, with the array tau,
+               represent the unitary matrix Q as a product of Householder vectors.
+            tau: Pointer to a 1D array of size min(M, N) for each batch.
+                 Contains the scalar factors of the Householder reflections.
+                 The tau array, together with the Householder vectors stored in A,
+                 defines the unitary matrix Q.
+            lda: Optional runtime leading dimension of matrix A.
+                 If not specified, the compile-time ``lda`` is used.
+        """
+        raise RuntimeError("factorize is a device function and cannot be called on the host.")
+
+
+# ==========================
+# QR Multiplier (UNMQR)
+# ==========================
+
+QR_MULTIPLIER_DOCSTRING = SOLVER_DOCSTRING.copy()
+del QR_MULTIPLIER_DOCSTRING["function"]
+del QR_MULTIPLIER_DOCSTRING["fill_mode"]
+del QR_MULTIPLIER_DOCSTRING["diag"]
+del QR_MULTIPLIER_DOCSTRING["job"]
+
+QR_MULTIPLIER_DOCSTRING["size"] = f"""{SOLVER_DOCSTRING_SIZE_BASE}
+``M`` and ``N`` represent the dimensions of matrix C.
+``K`` represents the number of Householder reflections from the QR factorization.
+If ``side='left'``, then ``K <= M`` and A is ``M`` x ``K``.
+If ``side='right'``, then ``K <= N`` and A is ``N`` x ``K``.""".replace("\n", " ")
+
+QR_MULTIPLIER_DOCSTRING["side"] = f"""Side of matrix Q in the multiplication operation.
+Can be one of: {", ".join(f"``'{v}'``" for v in ALLOWED_SIDE)}.
+If ``side='left'``, computes op(Q) * C where Q is ``M`` x ``M``.
+If ``side='right'``, computes C * op(Q) where Q is ``N`` x ``N``.""".replace("\n", " ")
+
+QR_MULTIPLIER_DOCSTRING["transpose_mode"] = f"""Transpose mode for operation op(Q) applied to matrix Q.
+Can be one of: {", ".join(f"``'{v}'``" for v in ALLOWED_TRANSPOSE_MODE)}.
+Defaults to ``'non_transposed'``.""".replace("\n", " ")
+
+QR_MULTIPLIER_DOCSTRING["leading_dimensions"] = SOLVER_DOCSTRING["leading_dimensions"] + ADAPTERS_API_LD_DOCSTRING
+
+
+@docstring_decorator(QR_MULTIPLIER_DOCSTRING, skip_missing=False)
+class QRMultiply(_SolverProperties):
+    """
+    A class that encapsulates the multiplication of a matrix C by the unitary matrix Q
+    from a QR factorization (UNMQR operation).
+
+    **Memory Layout Requirements:**
+
+    Matrices must be stored in shared memory according
+    to their arrangement and leading dimension (ld):
+
+    **For matrix A (containing Householder vectors):**
+
+    * If ``side='left'``: A is ``M`` x ``K``
+    * If ``side='right'``: A is ``N`` x ``K``
+    * **Column-major arrangement**: Matrix shape ``(batches_per_block, rows, K)``
+      with strides ``(lda * K, 1, lda)``
+    * **Row-major arrangement**: Matrix shape ``(batches_per_block, rows, K)``
+      with strides ``(lda * rows, lda, 1)``
+
+    **For matrix C (M x N):**
+
+    * **Column-major arrangement**: Matrix shape ``(batches_per_block, M, N)``
+      with strides ``(ldb * N, 1, ldb)``
+    * **Row-major arrangement**: Matrix shape ``(batches_per_block, M, N)``
+      with strides ``(ldb * M, ldb, 1)``
+
+    Args:
+        size (Sequence[int]): {size}
+
+        precision (type[np.floating]): {precision}
+
+        execution (str): {execution}
+
+        sm (ComputeCapability): {sm}
+
+        side (str): {side}
+
+        transpose_mode (str, optional): {transpose_mode}
+
+        arrangement (Sequence[str], optional): {arrangement}
+
+        batches_per_block (int | Literal["suggested"], optional): {batches_per_block}
+
+        data_type (str, optional): {data_type}
+
+        leading_dimensions (Sequence[int], optional): {leading_dimensions}
+
+        block_dim (Sequence[int] | Literal["suggested"], optional): {block_dim}
+
+    See Also:
+        For further details, please refer to the cuSOLVERDx documentation:
+
+        * :cusolverdx_doc:`unmqr <get_started/functions/unmqr.html>`
+    """
+
+    # ==========================
+    # Constructor
+    # ==========================
+
+    def __init__(
+        self,
+        size: Sequence[int],
+        precision: type[np.floating],
+        execution: str,
+        side: str,
+        *,
+        sm=None,
+        transpose_mode: str = "non_transposed",
+        arrangement: Sequence[str] | None = None,
+        batches_per_block: int | Literal["suggested"] | None = None,
+        data_type: str | None = None,
+        leading_dimensions: Sequence[int] | None = None,
+        block_dim: Sequence[int] | Literal["suggested"] | None = None,
+    ):
+        self._multiply: Solver = Solver(
+            function="unmqr",
+            size=size,
+            precision=precision,
+            execution=execution,
+            sm=sm,
+            side=side,
+            transpose_mode=transpose_mode,
+            arrangement=arrangement,
+            batches_per_block=batches_per_block,
+            data_type=data_type,
+            leading_dimensions=leading_dimensions,
+            block_dim=block_dim,
+        )
+
+        super().__init__(self._multiply)
+
+        if not isinstance(self._multiply.transpose_mode, str):
+            raise ValueError("transpose_mode must be provided for QRMultiply")
+
+        if not isinstance(self._multiply.side, str):
+            raise ValueError("side must be provided for QRMultiply")
+
+    # ==========================
+    # Property methods
+    # ==========================
+
+    @property
+    def side(self) -> str:
+        assert self._multiply.side is not None
+        return self._multiply.side
+
+    @property
+    def transpose_mode(self) -> str:
+        assert self._multiply.transpose_mode is not None
+        return self._multiply.transpose_mode
+
+    @property
+    def tau_type(self) -> type[np.floating] | Complex:
+        return self.value_type
+
+    @property
+    def tau_shape(self) -> tuple[int, int]:
+        return (self.batches_per_block, self.k)
+
+    @property
+    def tau_strides(self) -> tuple[int, int]:
+        return (self.tau_shape[1], 1)
+
+    @property
+    def a_shape(self) -> tuple[int, int, int]:
+        rows = self.m if self.side == "left" else self.n
+        return (self.batches_per_block, rows, self.k)
+
+    @property
+    def c_shape(self) -> tuple[int, int, int]:
+        return (self.batches_per_block, self.m, self.n)
+
+    def a_strides(self, *, lda: int | None = None) -> tuple[int, int, int]:
+        lda = self.lda if lda is None else lda
+        return _calculate_strides(self.a_shape[1:], lda, self.a_arrangement)
+
+    def c_strides(self, *, ldc: int | None = None) -> tuple[int, int, int]:
+        ldc = self.ldb if ldc is None else ldc
+        return _calculate_strides(self.c_shape[1:], ldc, self.b_arrangement)
+
+    def a_size(self, *, lda: int | None = None) -> int:
+        return self.a_strides(lda=lda)[0] * self.a_shape[0]
+
+    def c_size(self, *, ldc: int | None = None) -> int:
+        return self.c_strides(ldc=ldc)[0] * self.c_shape[0]
+
+    @property
+    def tau_size(self) -> int:
+        return self.tau_shape[0] * self.tau_shape[1]
+
+    # ==========================
+    # Device function methods
+    # ==========================
+
+    def multiply(self, a, tau, c, lda=None, ldc=None) -> None:
+        """
+        Multiplies matrix C by the unitary matrix Q from a QR factorization.
+
+        This device function computes:
+            ``op(Q) * C`` (if ``side='left'``)
+            ``C * op(Q)`` (if ``side='right'``)
+
+        where Q is the unitary matrix from the QR factorization, represented
+        by Householder vectors stored in A and the tau array.
+        Uses cuSOLVERDx ``'unmqr'``. The result overwrites matrix C.
+
+        If ``lda`` and ``ldc`` are provided, uses runtime version with the
+        specified leading dimensions. If not provided (``None``), uses compile-time
+        version with default or constructor-provided leading dimensions.
+
+        For more details, see: :cusolverdx_doc:`get_started/functions/unmqr.html`
+
+        Args:
+            a: Pointer to an array in shared memory, storing the batched matrix containing
+               Householder vectors from the QR factorization, according to the specified
+               arrangement and leading dimension (see :meth:`__init__`).
+               The elements below the diagonal of A, with the array tau, represent the
+               unitary matrix Q as a product of Householder reflections.
+               If ``side='left'``, A is ``M`` x ``K``.
+               If ``side='right'``, A is ``N`` x ``K``.
+            tau: Pointer to a 1D array of size K for each batch, containing the scalar
+                 factors of the Householder reflections from the QR factorization.
+                 The tau array, together with the Householder vectors in A, defines Q.
+            c: Pointer to an array in shared memory,
+               storing the batched ``M`` x ``N`` matrix
+               according to the specified arrangement
+               and leading dimension (see :meth:`__init__`).
+               The operation is in-place: result overwrites C.
+            lda: Optional runtime leading dimension for matrix A.
+                 The ``lda`` and ``ldc`` must be specified together.
+                 If not specified, the compile-time ``lda`` is used.
+            ldc: Optional runtime leading dimension for matrix C.
+                 The ``lda`` and ``ldc`` must be specified together.
+                 If not specified, the compile-time ``ldc`` is used.
+        """
+        raise RuntimeError("multiply is a device function and cannot be called on the host.")
+
+
+# ==========================
+# LQ Multiplier (UNMLQ)
+# ==========================
+
+LQ_MULTIPLIER_DOCSTRING = SOLVER_DOCSTRING.copy()
+del LQ_MULTIPLIER_DOCSTRING["function"]
+del LQ_MULTIPLIER_DOCSTRING["fill_mode"]
+del LQ_MULTIPLIER_DOCSTRING["diag"]
+del LQ_MULTIPLIER_DOCSTRING["job"]
+
+LQ_MULTIPLIER_DOCSTRING["size"] = f"""{SOLVER_DOCSTRING_SIZE_BASE}
+``M`` and ``N`` represent the dimensions of matrix C.
+``K`` represents the number of Householder reflections from the LQ factorization.
+If ``side='left'``, then ``K <= M`` and A is ``K`` x ``M``.
+If ``side='right'``, then ``K <= N`` and A is ``K`` x ``N``.""".replace("\n", " ")
+
+LQ_MULTIPLIER_DOCSTRING["side"] = f"""Side of matrix Q in the multiplication operation.
+Can be one of: {", ".join(f"``'{v}'``" for v in ALLOWED_SIDE)}.
+If ``side='left'``, computes op(Q) * C where Q is ``M`` x ``M``.
+If ``side='right'``, computes C * op(Q) where Q is ``N`` x ``N``.""".replace("\n", " ")
+
+LQ_MULTIPLIER_DOCSTRING["transpose_mode"] = f"""Transpose mode for operation op(Q) applied to matrix Q.
+Can be one of: {", ".join(f"``'{v}'``" for v in ALLOWED_TRANSPOSE_MODE)}.
+Defaults to ``'non_transposed'``.""".replace("\n", " ")
+
+LQ_MULTIPLIER_DOCSTRING["leading_dimensions"] = SOLVER_DOCSTRING["leading_dimensions"] + ADAPTERS_API_LD_DOCSTRING
+
+
+@docstring_decorator(LQ_MULTIPLIER_DOCSTRING, skip_missing=False)
+class LQMultiply(_SolverProperties):
+    """
+    A class that encapsulates the multiplication of a matrix C by the unitary matrix Q
+    from an LQ factorization (UNMLQ operation).
+
+    **Memory Layout Requirements:**
+
+    Matrices must be stored in shared memory according
+    to their arrangement and leading dimension (ld):
+
+    **For matrix A (containing Householder vectors):**
+
+    * If ``side='left'``: A is ``K`` x ``M``
+    * If ``side='right'``: A is ``K`` x ``N``
+    * **Column-major arrangement**: Matrix shape ``(batches_per_block, K, cols)``
+      with strides ``(lda * cols, 1, lda)``
+    * **Row-major arrangement**: Matrix shape ``(batches_per_block, K, cols)``
+      with strides ``(lda * K, lda, 1)``
+
+    **For matrix C (M x N):**
+
+    * **Column-major arrangement**: Matrix shape ``(batches_per_block, M, N)``
+      with strides ``(ldc * N, 1, ldc)``
+    * **Row-major arrangement**: Matrix shape ``(batches_per_block, M, N)``
+      with strides ``(ldc * M, ldc, 1)``
+
+    Args:
+        size (Sequence[int]): {size}
+
+        precision (type[np.floating]): {precision}
+
+        execution (str): {execution}
+
+        sm (ComputeCapability): {sm}
+
+        side (str): {side}
+
+        transpose_mode (str, optional): {transpose_mode}
+
+        arrangement (Sequence[str], optional): {arrangement}
+
+        batches_per_block (int | Literal["suggested"], optional): {batches_per_block}
+
+        data_type (str, optional): {data_type}
+
+        leading_dimensions (Sequence[int], optional): {leading_dimensions}
+
+        block_dim (Sequence[int] | Literal["suggested"], optional): {block_dim}
+
+    See Also:
+        For further details, please refer to the cuSOLVERDx documentation:
+
+        * :cusolverdx_doc:`unmlq <get_started/functions/unmlq.html>`
+    """
+
+    # ==========================
+    # Constructor
+    # ==========================
+
+    def __init__(
+        self,
+        size: Sequence[int],
+        precision: type[np.floating],
+        execution: str,
+        side: str,
+        *,
+        sm=None,
+        transpose_mode: str = "non_transposed",
+        arrangement: Sequence[str] | None = None,
+        batches_per_block: int | Literal["suggested"] | None = None,
+        data_type: str | None = None,
+        leading_dimensions: Sequence[int] | None = None,
+        block_dim: Sequence[int] | Literal["suggested"] | None = None,
+    ):
+        self._multiply: Solver = Solver(
+            function="unmlq",
+            size=size,
+            precision=precision,
+            execution=execution,
+            sm=sm,
+            side=side,
+            transpose_mode=transpose_mode,
+            arrangement=arrangement,
+            batches_per_block=batches_per_block,
+            data_type=data_type,
+            leading_dimensions=leading_dimensions,
+            block_dim=block_dim,
+        )
+
+        super().__init__(self._multiply)
+
+        if not isinstance(self._multiply.transpose_mode, str):
+            raise ValueError("transpose_mode must be provided for LQMultiply")
+
+        if not isinstance(self._multiply.side, str):
+            raise ValueError("side must be provided for LQMultiply")
+
+    # ==========================
+    # Property methods
+    # ==========================
+
+    @property
+    def side(self) -> str:
+        assert self._multiply.side is not None
+        return self._multiply.side
+
+    @property
+    def transpose_mode(self) -> str:
+        assert self._multiply.transpose_mode is not None
+        return self._multiply.transpose_mode
+
+    @property
+    def tau_type(self) -> type[np.floating] | Complex:
+        return self.value_type
+
+    @property
+    def tau_shape(self) -> tuple[int, int]:
+        return (self.batches_per_block, self.k)
+
+    @property
+    def tau_strides(self) -> tuple[int, int]:
+        return (self.tau_shape[1], 1)
+
+    @property
+    def a_shape(self) -> tuple[int, int, int]:
+        cols = self.m if self.side == "left" else self.n
+        return (self.batches_per_block, self.k, cols)
+
+    @property
+    def c_shape(self) -> tuple[int, int, int]:
+        return (self.batches_per_block, self.m, self.n)
+
+    def a_strides(self, *, lda: int | None = None) -> tuple[int, int, int]:
+        lda = self.lda if lda is None else lda
+        return _calculate_strides(self.a_shape[1:], lda, self.a_arrangement)
+
+    def c_strides(self, *, ldc: int | None = None) -> tuple[int, int, int]:
+        ldc = self.ldb if ldc is None else ldc
+        return _calculate_strides(self.c_shape[1:], ldc, self.b_arrangement)
+
+    def a_size(self, *, lda: int | None = None) -> int:
+        return self.a_strides(lda=lda)[0] * self.a_shape[0]
+
+    def c_size(self, *, ldc: int | None = None) -> int:
+        return self.c_strides(ldc=ldc)[0] * self.c_shape[0]
+
+    @property
+    def tau_size(self) -> int:
+        return self.tau_shape[0] * self.tau_shape[1]
+
+    # ==========================
+    # Device function methods
+    # ==========================
+
+    def multiply(self, a, tau, c, lda=None, ldc=None) -> None:
+        """
+        Multiplies matrix C by the unitary matrix Q from an LQ factorization.
+
+        This device function computes:
+            ``op(Q) * C`` (if ``side='left'``)
+            ``C * op(Q)`` (if ``side='right'``)
+
+        where Q is the unitary matrix from the LQ factorization, represented
+        by Householder vectors stored in A and the tau array.
+        Uses cuSOLVERDx ``'unmlq'``. The result overwrites matrix C.
+
+        If ``lda`` and ``ldc`` are provided, uses runtime version with the
+        specified leading dimensions. If not provided (``None``), uses compile-time
+        version with default or constructor-provided leading dimensions.
+
+        For more details, see: :cusolverdx_doc:`get_started/functions/unmlq.html`
+
+        Args:
+            a: Pointer to an array in shared memory, storing the batched matrix containing
+               Householder vectors from the LQ factorization, according to the specified
+               arrangement and leading dimension (see :meth:`__init__`).
+               The elements above the diagonal of A, with the array tau, represent the
+               unitary matrix Q as a product of Householder reflections.
+               If ``side='left'``, A is ``K`` x ``M``.
+               If ``side='right'``, A is ``K`` x ``N``.
+            tau: Pointer to a 1D array of size K for each batch, containing the scalar
+                 factors of the Householder reflections from the LQ factorization.
+                 The tau array, together with the Householder vectors in A, defines Q.
+            c: Pointer to an array in shared memory,
+               storing the batched ``M`` x ``N`` matrix
+               according to the specified arrangement
+               and leading dimension (see :meth:`__init__`).
+               The operation is in-place: result overwrites C.
+            lda: Optional runtime leading dimension for matrix A.
+                 The ``lda`` and ``ldc`` must be specified together.
+                 If not specified, the compile-time ``lda`` is used.
+            ldc: Optional runtime leading dimension for matrix C.
+                 The ``lda`` and ``ldc`` must be specified together.
+                 If not specified, the compile-time ``ldc`` is used.
+        """
+        raise RuntimeError("multiply is a device function and cannot be called on the host.")
+
+
+# ==========================
+# Least Squares Solver
+# ==========================
+
+LEAST_SQUARES_SOLVER_DOCSTRING = SOLVER_DOCSTRING.copy()
+del LEAST_SQUARES_SOLVER_DOCSTRING["function"]
+del LEAST_SQUARES_SOLVER_DOCSTRING["fill_mode"]
+del LEAST_SQUARES_SOLVER_DOCSTRING["side"]
+del LEAST_SQUARES_SOLVER_DOCSTRING["diag"]
+
+LEAST_SQUARES_SOLVER_DOCSTRING["size"] = f"""{SOLVER_DOCSTRING_SIZE_BASE}
+``M`` and ``N`` represent the dimensions of matrix A (``M`` x ``N``).
+``K`` represents the number of columns in the right-hand side matrix B.""".replace("\n", " ")
+
+LEAST_SQUARES_SOLVER_DOCSTRING["transpose_mode"] = f"""Transpose mode for operation op(A) applied to matrix A.
+Can be one of: {", ".join(f"``'{v}'``" for v in ALLOWED_TRANSPOSE_MODE)}.
+Defaults to ``'non_transposed'``.""".replace("\n", " ")
+
+LEAST_SQUARES_SOLVER_DOCSTRING["leading_dimensions"] = SOLVER_DOCSTRING["leading_dimensions"] + ADAPTERS_API_LD_DOCSTRING
+
+
+@docstring_decorator(LEAST_SQUARES_SOLVER_DOCSTRING, skip_missing=False)
+class LeastSquaresSolver(_SolverProperties):
+    """
+    A class that encapsulates least squares solver device function (``'gels'``).
+    GELS (GEneral Least Square) solves overdetermined
+    or underdetermined least squares problems:
+
+    .. math::
+       \\min \\| op(A) * X - B \\|_2
+
+    using the QR or LQ factorization of A, and overwriting B with the solution X.
+
+    The configurations supported by GELS are:
+
+    **Memory layout requirements:**
+
+    Matrices must be stored in shared memory according
+    to their arrangement and leading dimension (ld):
+
+    **For matrix A (M x N):**
+
+    * **Column-major arrangement**: Matrix shape ``(batches_per_block, M, N)``
+      with strides ``(lda * N, 1, lda)``
+    * **Row-major arrangement**: Matrix shape ``(batches_per_block, M, N)``
+      with strides ``(lda * M, lda, 1)``
+
+    **Matrix B and X** are stored in the same buffer (B is overwritten by X in place).
+    The buffer has shape ``(batches_per_block, max(M, N), K)``. The second leading
+    dimension ``ldb`` refers to this shared B/X buffer
+    and must satisfy ``ldb >= max(M, N)``.
+    Logical shapes differ by :attr:`transpose_mode`.
+    Use :attr:`b_shape` and :attr:`x_shape` for B and X.
+
+    **Matrix B** (right-hand side):
+
+    * **Logical shape**: :attr:`b_shape` - ``(batches_per_block, M, K)`` if non-transposed,
+      ``(batches_per_block, N, K)`` if transposed.
+    * **Column-major arrangement**: strides ``(ldb * K, 1, ldb)``.
+    * **Row-major arrangement**: strides ``(ldb * max(M, N), ldb, 1)``.
+
+    **Matrix X** (solution):
+
+    * **Logical shape**: :attr:`x_shape` - ``(batches_per_block, N, K)`` if non-transposed,
+      ``(batches_per_block, M, K)`` if transposed.
+    * **Column-major arrangement**: strides ``(ldb * K, 1, ldb)``.
+    * **Row-major arrangement**: strides ``(ldb * max(M, N), ldb, 1)``.
+
+    .. note::
+        GELS is an in-place function. Matrix A is overwritten by the QR or LQ factorization,
+        and matrix B is overwritten by the solution X. Both B and X use the single buffer
+        of shape ``(max(M, N), K)`` per batch.
+
+    Args:
+        size (Sequence[int]): {size}
+
+        precision (type[np.floating]): {precision}
+
+        execution (str): {execution}
+
+        sm (ComputeCapability): {sm}
+
+        transpose_mode (str, optional): {transpose_mode}
+
+        arrangement (Sequence[str], optional): {arrangement}
+
+        batches_per_block (int | Literal["suggested"], optional): {batches_per_block}
+
+        data_type (str, optional): {data_type}
+
+        leading_dimensions (Sequence[int], optional): {leading_dimensions}
+
+        block_dim (Sequence[int] | Literal["suggested"], optional): {block_dim}
+
+    See Also:
+        For further details, please refer to the cuSOLVERDx documentation:
+
+        * :cusolverdx_doc:`gels <get_started/functions/gels.html>`
+    """
+
+    # ==========================
+    # Constructor
+    # ==========================
+
+    def __init__(
+        self,
+        size: Sequence[int],
+        precision: type[np.floating],
+        execution: str,
+        *,
+        sm=None,
+        transpose_mode: str = "non_transposed",
+        arrangement: Sequence[str] | None = None,
+        batches_per_block: int | Literal["suggested"] | None = None,
+        data_type: str | None = None,
+        leading_dimensions: Sequence[int] | None = None,
+        block_dim: Sequence[int] | Literal["suggested"] | None = None,
+    ):
+        self._solve: Solver = Solver(
+            function="gels",
+            size=size,
+            precision=precision,
+            execution=execution,
+            sm=sm,
+            transpose_mode=transpose_mode,
+            arrangement=arrangement,
+            batches_per_block=batches_per_block,
+            data_type=data_type,
+            leading_dimensions=leading_dimensions,
+            block_dim=block_dim,
+        )
+
+        super().__init__(self._solve)
+
+    # ==========================
+    # Property methods
+    # ==========================
+
+    @property
+    def transpose_mode(self) -> str:
+        assert self._solve.transpose_mode is not None
+        return self._solve.transpose_mode
+
+    @property
+    def tau_type(self) -> type[np.floating] | Complex:
+        return self._solve.tau_type
+
+    @property
+    def tau_shape(self) -> tuple[int, int]:
+        return (self.batches_per_block, min(self.m, self.n))
+
+    @property
+    def tau_strides(self) -> tuple[int, int]:
+        return (self.tau_shape[1], 1)
+
+    @property
+    def a_shape(self) -> tuple[int, int, int]:
+        return (self.batches_per_block, self.m, self.n)
+
+    @property
+    def b_shape(self) -> tuple[int, int, int]:
+        return (self.batches_per_block, self.m if self.transpose_mode == "non_transposed" else self.n, self.k)
+
+    @property
+    def x_shape(self) -> tuple[int, int, int]:
+        return (self.batches_per_block, self.n if self.transpose_mode == "non_transposed" else self.m, self.k)
+
+    def a_strides(self, *, lda: int | None = None) -> tuple[int, int, int]:
+        lda = self.lda if lda is None else lda
+        return _calculate_strides(self.a_shape[1:], lda, self.a_arrangement)
+
+    def bx_strides(self, *, ldb: int | None = None) -> tuple[int, int, int]:
+        ldb = self.ldb if ldb is None else ldb
+        return _calculate_strides((max(self.m, self.n), self.k), ldb, self.b_arrangement)
+
+    def a_size(self, *, lda: int | None = None) -> int:
+        return self.a_strides(lda=lda)[0] * self.a_shape[0]
+
+    def bx_size(self, *, ldb: int | None = None) -> int:
+        return self.bx_strides(ldb=ldb)[0] * self.batches_per_block
+
+    @property
+    def tau_size(self) -> int:
+        return self.tau_shape[0] * self.tau_shape[1]
+
+    # ==========================
+    # Device function methods
+    # ==========================
+
+    def solve(self, a, tau, b, lda=None, ldb=None) -> None:
+        """
+        Solves a least squares problem using QR or LQ factorization.
+
+        This device function solves:
+
+        .. math::
+           \\min \\| op(A) * X - B \\|_2
+
+        using the QR or LQ factorization of A, and overwrites B with the solution X.
+        Uses cuSOLVERDx ``'gels'``. The operation is in-place: matrix A is overwritten
+        by the factorization, and matrix B is overwritten by the solution X.
+
+        If ``lda`` and ``ldb`` are provided, uses runtime version with the
+        specified leading dimensions. If not provided (``None``), uses compile-time
+        version with default or constructor-provided leading dimensions.
+
+        .. note::
+            The choice between QR and LQ factorization depends on the problem dimensions
+            and transpose mode:
+
+            * If ``op(A)`` is ``'non_transposed'`` and ``M >= N``: uses QR factorization
+            * If ``op(A)`` is ``'non_transposed'`` and ``M < N``: uses LQ factorization
+            * If ``op(A)`` is ``'transposed'`` or ``'conj_transposed'``
+              and ``M >= N``: uses LQ factorization
+            * If ``op(A)`` is ``'transposed'`` or ``'conj_transposed'``
+              and ``M < N``: uses QR factorization
+
+        For more details, see: :cusolverdx_doc:`get_started/functions/gels.html`
+
+        Args:
+            a: Pointer to an array in shared memory, storing the batched matrix
+               according to the specified arrangement
+               and leading dimension (see :meth:`__init__`).
+               The matrix is overwritten in place by the QR or LQ factorization.
+            tau: Pointer to a 1D array of size min(M, N) for each batch.
+                 Contains the scalar factors of the Householder reflections.
+                 The tau array, together with the Householder vectors stored in A,
+                 defines the unitary matrix Q.
+            b: Pointer to an array in shared memory,
+               storing the batched right-hand side matrix
+               according to the specified arrangement
+               and leading dimension (see :meth:`__init__`).
+               The storage size is ``max(M, N) x K`` per batch.
+               The operation is in-place: result X overwrites B.
+            lda: Optional runtime leading dimension for matrix A.
+                 The ``lda`` and ``ldb`` must be specified together.
+                 If not specified, the compile-time ``lda`` is used.
+            ldb: Optional runtime leading dimension for matrix B.
+                 The ``lda`` and ``ldb`` must be specified together.
+                 If not specified, the compile-time ``ldb`` is used.
+        """
+        raise RuntimeError("solve is a device function and cannot be called on the host.")