nvmath-python 1.0.0__cp314-cp314t-win_amd64.whl

nvmath/linalg/advanced/helpers/matmul.py

@@ -0,0 +1,1316 @@
+# Copyright (c) 2025-2026, NVIDIA CORPORATION & AFFILIATES. ALL RIGHTS RESERVED.
+#
+# SPDX-License-Identifier: Apache-2.0
+
+from __future__ import annotations
+
+from enum import Enum
+from typing import TYPE_CHECKING, Literal
+
+if TYPE_CHECKING:
+    import torch
+
+import math
+
+import numpy as np
+
+from nvmath.internal.tensor_ifc import TensorHolder
+from nvmath.internal.tensor_wrapper import wrap_operand
+from nvmath.internal.utils import create_empty_tensor, get_or_create_stream, infer_object_package
+
+__all__ = [
+    "BlockScalingFormat",
+    "create_mxfp8_scale",
+    "invert_mxfp8_scale",
+    "apply_mxfp8_scale",
+    "unpack_fp4",
+    "quantize_to_fp4",
+    "get_block_scale_offset",
+    "to_block_scale",
+    "expand_block_scale",
+]
+
+# ====================================================
+# helper functions for FP8 and MXFP8
+# ====================================================
+
+
+def _c_strides(operand_shape: tuple[int, ...]) -> tuple[int, ...]:
+    ndim = len(operand_shape)
+    strides = [0] * ndim
+    stride = 1
+    for i in range(ndim - 1, -1, -1):
+        strides[i] = stride
+        stride *= operand_shape[i]
+    return tuple(strides)
+
+
+def _validate_operand_ndim_for_block_scaling(operand_shape: tuple[int, ...]):
+    ndim = len(operand_shape)
+    if ndim < 2:
+        raise ValueError(f"Operand shape must be at least 2D, got {ndim}D with shape {operand_shape}")
+    return ndim
+
+
+def _infer_blocked_axis(operand: TensorHolder):
+    # In general, strides are not reliable if corresponding extent is 1,
+    # but we expect (and check) operand last two extents to be divisible by
+    # 128 or 64.
+    return -1 if operand.strides[-2] > operand.strides[-1] else -2
+
+
+class BlockScalingFormat(str, Enum):
+    """
+    Block scaling format for microscaling data types.
+
+    Passed as the ``block_scaling_format`` argument to
+    :func:`get_block_scale_offset`, :func:`to_block_scale`, and
+    :func:`expand_block_scale`.
+    """
+
+    NVFP4 = "NVFP4"
+    MXFP8 = "MXFP8"
+
+
+# Per-format properties for microscaling (block-scaled) data types.
+# Used by validation helpers and expand_block_scale to dispatch on
+# block size, allowed dtypes, and scale interpretation without
+# per-format if/elif branches. To add a new format, add an entry here.
+_MICROSCALING_FORMAT_PROPERTIES: dict[BlockScalingFormat, dict] = {
+    BlockScalingFormat.NVFP4: {
+        "num_scalars_in_block": 16,
+        "operand_dtypes": ("float4_e2m1fn_x2",),
+        # NVFP4 scales are float8_e4m3fn values. uint8 is also accepted
+        # because the two dtypes have the same bit-width, and internally
+        # the scale bytes are reinterpreted as uint8 for the strided
+        # expansion anyway (see _expand_block_scale).
+        "scale_dtypes": ("float8_e4m3fn", "uint8"),
+        # Raw uint8 scale bytes are reinterpreted as float8_e4m3fn values.
+        "scale_interpretation": "float8_e4m3fn",
+    },
+    BlockScalingFormat.MXFP8: {
+        "num_scalars_in_block": 32,
+        "operand_dtypes": ("float8_e4m3fn", "float8_e5m2"),
+        "scale_dtypes": ("uint8",),
+        # Raw uint8 scale bytes are unsigned 8-bit biased exponents: 2^(value - 127).
+        "scale_interpretation": "ue8m0",
+    },
+}
+
+_COMPATIBLE_FORMATS_FOR_OPERAND_DTYPE: dict[str, set[BlockScalingFormat]] = {}
+for _fmt, _props in _MICROSCALING_FORMAT_PROPERTIES.items():
+    for _dt in _props["operand_dtypes"]:
+        _COMPATIBLE_FORMATS_FOR_OPERAND_DTYPE.setdefault(_dt, set()).add(_fmt)
+
+
+def _validate_operand_dtype_block_scaling_format_compatibility(
+    operand_dtype: str,
+    block_scaling_format: BlockScalingFormat,
+) -> None:
+    """Check that *block_scaling_format* is consistent with *operand_dtype*."""
+    compatible = _COMPATIBLE_FORMATS_FOR_OPERAND_DTYPE.get(operand_dtype)
+    if compatible is not None and block_scaling_format not in compatible:
+        compatible_str = {str(f) for f in compatible}
+        raise ValueError(
+            f"Operand dtype {operand_dtype} requires block_scaling_format in {compatible_str}, got '{block_scaling_format}'."
+        )
+
+
+def _validate_scale_dtype_block_scaling_format_compatibility(
+    scale_dtype: str,
+    block_scaling_format: BlockScalingFormat,
+    param_name: str = "scales",
+) -> None:
+    """Raise :exc:`TypeError` if *scale_dtype* is not valid for *block_scaling_format*."""
+    props = _MICROSCALING_FORMAT_PROPERTIES[block_scaling_format]
+    if scale_dtype not in props["scale_dtypes"]:
+        allowed = " or ".join(f"torch.{d}" for d in props["scale_dtypes"])
+        raise TypeError(f"For {block_scaling_format}, {param_name} must have dtype {allowed}, got {scale_dtype}")
+
+
+def _infer_operand_shape_axis(
+    operand_or_shape: torch.Tensor | tuple[int, ...],
+    axis: Literal[-1, -2] | None,
+) -> tuple[int, tuple[int, ...], Literal[-1, -2], str | None]:
+    """Resolve operand shape, ndim, axis, and (optionally) operand dtype.
+
+    Returns ``(ndim, operand_shape, axis, operand_dtype)`` where
+    *operand_dtype* is ``None`` when a plain shape tuple is passed.
+    """
+    import torch
+
+    if isinstance(operand_or_shape, tuple):
+        ndim = _validate_operand_ndim_for_block_scaling(operand_or_shape)
+        if axis is None:
+            raise ValueError(f"Axis must be specified when operand shape is provided, got {axis}")
+        return ndim, operand_or_shape, axis, None
+    elif isinstance(operand_or_shape, torch.Tensor):
+        operand = wrap_operand(operand_or_shape)
+        operand_shape: tuple[int, ...] = operand.shape  # type: ignore
+        ndim = _validate_operand_ndim_for_block_scaling(operand_shape)
+
+        # Infer blocked axis
+        inferred_axis = _infer_blocked_axis(operand)
+        if axis is None:
+            axis = inferred_axis
+        else:
+            if axis >= 0:
+                axis -= ndim
+            if axis != inferred_axis:
+                layout = "row-major" if inferred_axis == -1 else "col-major"
+                raise ValueError(f"Incorrect axis: {axis} for {layout} operand, expected {inferred_axis}.")
+
+        if operand.dtype == "float4_e2m1fn_x2":
+            operand_shape_list = list(operand_shape)
+            operand_shape_list[axis] *= 2
+            operand_shape = tuple(operand_shape_list)
+        return ndim, operand_shape, axis, operand.dtype
+    else:
+        raise ValueError(
+            f"Expected operand_shape to be a tuple or a torch.Tensor to infer the shape from, got {type(operand_or_shape)}."
+        )
+
+
+def _validate_shape_axes_block_scaling_format(
+    operand_or_shape: torch.Tensor | tuple[int, ...],
+    axis: Literal[-1, -2] | None,
+    block_scaling_format: BlockScalingFormat,
+) -> tuple[tuple[int, ...], Literal[-1, -2], Literal[-1, -2], BlockScalingFormat, int]:
+    """
+    Infers operand_shape and blocked/unblocked axes from ``operand_or_shape`` when a tensor
+    is passed; validates ``block_scaling_format`` and related ``num_scalars_in_block``.
+
+    The accepted operand_shape, axis, and block_scaling_format are validated, so that axis
+    is one of the last two dimensions of the operand, and operand's shape extents are
+    multiples of cuBLAS tile size and num_scalars_in_block.
+    """
+
+    ndim, operand_shape, axis, operand_dtype = _infer_operand_shape_axis(operand_or_shape, axis)
+
+    if operand_dtype is not None:
+        _validate_operand_dtype_block_scaling_format_compatibility(operand_dtype, block_scaling_format)
+
+    neg_axis = axis if axis < 0 else axis - ndim
+    if neg_axis == -1:
+        unblocked_axis, blocked_axis = -2, -1  # type: tuple[Literal[-1, -2], Literal[-1, -2]]
+    elif neg_axis == -2:
+        unblocked_axis, blocked_axis = -1, -2
+    else:
+        raise ValueError(f"Axis must point to one of the last two dimensions of the operand, got {axis}")
+
+    num_scalars_in_block = _MICROSCALING_FORMAT_PROPERTIES[block_scaling_format]["num_scalars_in_block"]
+
+    if any(extent <= 0 for extent in operand_shape):
+        raise ValueError(f"Operand shape must have positive extents, got {operand_shape}")
+
+    tile_height = 128
+    tile_width = 4
+    scalars_per_tile_row = tile_width * num_scalars_in_block
+    unblocked_extent = operand_shape[unblocked_axis]
+    blocked_extent = operand_shape[blocked_axis]
+
+    # Validate dimension requirements for NVFP4/MXFP8 tiling
+    if unblocked_extent % tile_height != 0:
+        raise ValueError(
+            f"The extent at axis {unblocked_axis} must be a positive multiple of {tile_height} "
+            f"(cuBLASLt uses {tile_height}x{tile_width} scale tiles), "
+            f"got operand_shape[{unblocked_axis}] = {unblocked_extent}"
+        )
+
+    if blocked_extent % scalars_per_tile_row != 0:
+        raise ValueError(
+            f"The extent at axis {blocked_axis} must be a positive multiple of {scalars_per_tile_row} "
+            f"(cuBLASLt uses {tile_height}x{tile_width} "
+            f"scale tiles where each tile row covers {num_scalars_in_block} elements along blocked "
+            f"dimension: {tile_width} groups * {num_scalars_in_block} elements in a group), "
+            f"got operand_shape[{blocked_axis}] = {blocked_extent}"
+        )
+    return operand_shape, unblocked_axis, blocked_axis, block_scaling_format, num_scalars_in_block
+
+
+def _scales_2d_matrix_tiled_layout(
+    operand_shape: tuple[int, ...],
+    unblocked_axis: int,
+    blocked_axis: int,
+    num_scalars_in_block: int,
+    broadcast_block: bool = True,
+) -> tuple[tuple[int, ...], tuple[int, ...], tuple[int, ...]]:
+    """
+    Returns logical 2D shape for a matrix of scales and corresponding 5D (or 6D when
+    broadcasting) tensor shape and strides that account for the tiling and interleaved
+    layout of the tile.
+    """
+
+    tile_width = 4
+    tile_col_shape = (4, 32)
+    tile_height = 128  # prod(tile_y)
+
+    unblocked_extent = operand_shape[unblocked_axis]
+    blocked_extent = operand_shape[blocked_axis]
+
+    num_blocks_in_a_row = blocked_extent // num_scalars_in_block
+    num_tiles_in_a_row = num_blocks_in_a_row // tile_width
+    num_tiles_in_col = unblocked_extent // tile_height
+
+    # For operand of shape (H, W), the scale tensor is logically
+    # a 2D matrix of shape (H, W // num_blocks_in_a_row) (if axis = -1)
+    # or (H // num_blocks_in_a_row, W) (if axis = -2).
+    # The scale matrix is tiled with tile of shape 128x4.
+    # The tile has interleaved layout that splits column into
+    # 4 groups of 32 elements, so actual tile shape is (4, 32, 4).
+    # To account for the interleaved layout,
+    # we need to split the 2D matrix into 5D tensor.
+    shape_unblocked = num_tiles_in_col, tile_col_shape[0], tile_col_shape[1]
+    squeezed_unblocked = unblocked_extent
+    assert squeezed_unblocked == math.prod(shape_unblocked)
+    shape_blocked = num_tiles_in_a_row, tile_width
+    squeezed_blocked = num_blocks_in_a_row
+    assert squeezed_blocked == math.prod(shape_blocked)
+    # In total, we get 5D tensor with following dense strides:
+    strides_unblocked = (
+        512 * num_tiles_in_a_row,  # 1 * tile_size * num_tiles_in_a_row
+        4,  # 1 * tile_width
+        16,  # 1 * tile_width * tile_col_shape[0]
+    )
+    strides_blocked = (
+        512,  # 1 * tile_width * tile_col_shape[0] * tile_col_shape[1] = tile_size
+        1,  # fastest changing dim
+    )
+    # Broadcast every scale to the whole block if needed.
+    if broadcast_block:
+        shape_blocked = shape_blocked + (num_scalars_in_block,)  # type: ignore
+        strides_blocked = strides_blocked + (0,)  # type: ignore
+        squeezed_blocked = blocked_extent
+        assert math.prod(shape_blocked) == blocked_extent
+    # Finally, we combine the two parts into a 5D (or 6D when broadcasting) tensor.
+    if blocked_axis == -1:
+        shape = shape_unblocked + shape_blocked
+        strides = strides_unblocked + strides_blocked  # type: ignore
+        logical_shape = (squeezed_unblocked, squeezed_blocked)
+    else:
+        assert blocked_axis == -2
+        shape = shape_blocked + shape_unblocked
+        strides = strides_blocked + strides_unblocked
+        logical_shape = (squeezed_blocked, squeezed_unblocked)
+    return shape, strides, logical_shape
+
+
+def _scales_nd_matrix_tiled_layout(
+    operand_shape: tuple[int, ...],
+    unblocked_axis: Literal[-1, -2],
+    blocked_axis: Literal[-1, -2],
+    num_scalars_in_block: int,
+    broadcast_block: bool,
+) -> tuple[tuple[int, ...], tuple[int, ...], tuple[int, ...]]:
+    """
+    Returns logical 2D shape for a matrix of scales (or 3D to account for batching) and
+    corresponding (1 if batching) + 5D (+ 1 if broadcasting) tensor shape and strides that
+    account for the tiling and interleaved layout of the tile.
+    """
+
+    matrix_shape, matrix_strides, matrix_logical_shape = _scales_2d_matrix_tiled_layout(
+        operand_shape, unblocked_axis, blocked_axis, num_scalars_in_block, broadcast_block
+    )
+
+    operand_ndim = len(operand_shape)
+    if operand_ndim >= 2:
+        batch_shape = operand_shape[:-2]
+        batch_size = math.prod(batch_shape)
+        batch_stride = math.prod(operand_shape[-2:]) // num_scalars_in_block
+        matrix_shape = (batch_size, *matrix_shape)
+        matrix_strides = (batch_stride, *matrix_strides)
+        matrix_logical_shape = batch_shape + matrix_logical_shape
+
+    return matrix_shape, matrix_strides, matrix_logical_shape
+
+
+def _validate_tensor(x, where, tensor_name="tensor", dtype=None):
+    """
+    Validate the tensor package and dtype.
+
+    Args:
+        x: wrapped tensor object
+
+        where: name of the function that is performing the validation
+
+        tensor_name: name of the tensor to use in the error messages
+
+        dtype: if not None, check that the object dtype matches the specified dtype
+
+    """
+    package = infer_object_package(x)
+    if package != "torch":
+        raise ValueError(
+            f"Only torch.Tensor is currently supported by function '{where}'; the "
+            f"specified {tensor_name} belongs to '{package}' package."
+        )
+    x = wrap_operand(x)
+    if dtype is not None and x.dtype != dtype:
+        raise ValueError(
+            f"The function '{where}' requires the specified {tensor_name} to have dtype "
+            f"'{dtype}', whereas it has dtype '{x.dtype}'."
+        )
+    return x
+
+
+def _validate_mxfp8_scale(mx_scales, where, x=None):
+    if x is not None:
+        x = _validate_tensor(x, where)
+    mx_scales = _validate_tensor(mx_scales, where, tensor_name="mx_scales tensor", dtype="uint8")
+
+    if x is None:
+        return mx_scales
+
+    if mx_scales.shape != (x.size // 32,):
+        raise ValueError(
+            f"The shape of mx_scales {mx_scales.shape} is not compatible with a tensor of shape {x.shape}. "
+            f"The expected mx_scales shape is {(x.size // 32,)}."
+        )
+    return mx_scales, x
+
+
+def create_mxfp8_scale(x, exponent, stream=None):
+    """
+    .. experimental:: function
+
+    Create MXFP8 block scale with the same value for the whole tensor ``x``.
+
+    Args:
+        x: The tensor to create the block scale for
+
+        exponent: An integer from [-127, 128] range. Effective scale will be ``2^exponent``.
+
+        stream: Optional stream to create the block scale on.
+            Defaults to the stream of ``x``.
+
+    Returns:
+        An MXFP8 block scale factors tensor (1D, cuBLAS-compatible interleaved layout) to be
+        used with MXFP8 computations.
+    """
+    x = _validate_tensor(x, "create_mxfp8_scale")
+
+    if not -127 <= exponent <= 128:
+        raise ValueError("The exponent should be an integer from [-127, 128] range.")
+
+    stream_holder = None if x.device_id == "cpu" else get_or_create_stream(x.device_id, stream, x.name)
+    mx_scales = create_empty_tensor(
+        x.__class__, (x.size // 32,), "uint8", device_id=x.device_id, stream_holder=stream_holder, verify_strides=False
+    )
+    mx_scales.tensor[:] = exponent + 127
+    return mx_scales.tensor
+
+
+def invert_mxfp8_scale(mx_scales):
+    """
+    .. experimental:: function
+
+    Compute a reciprocal of MXFP8 block scale.
+
+    Args:
+        mx_scales: MXFP8 block scale tensor (UE8M0 format).
+
+    Returns:
+        Tensor of the same shape as ``mx_scales`` with exponents replaced by the
+        reciprocals.
+    """
+    _validate_mxfp8_scale(mx_scales, "invert_mxfp8_scale")
+
+    ret = mx_scales.clone()
+    ret[ret == 255] = 254  # Prevent the overflow
+    # remove the bias (127), negate, add the bias back: -(scale - 127) + 127
+    ret[:] = 254 - ret
+    return ret
+
+
+def _idx_batch_offset(
+    operand_shape: tuple[int, ...], index: tuple[int, ...] | tuple[torch.Tensor, ...], num_scalars_in_block: int
+) -> int | torch.Tensor:
+    ndim = len(operand_shape)
+    if ndim == 2:
+        return 0
+
+    assert ndim > 2
+    batch_strides = _c_strides(operand_shape)[:-2]
+    batch_index = index[:-2]
+    batch_offset = sum(i * stride for i, stride in zip(batch_index, batch_strides, strict=True))  # type: ignore
+    batch_offset //= num_scalars_in_block
+    return batch_offset
+
+
+def get_block_scale_offset(
+    index: tuple[int, ...] | tuple[torch.Tensor, ...],
+    operand_or_shape: torch.Tensor | tuple[int, ...],
+    block_scaling_format: BlockScalingFormat,
+    *,
+    axis: Literal[-1, -2] | None = None,
+) -> int | torch.Tensor:
+    """
+    .. experimental:: function
+
+    Computes offset of a block scale factor in the 1D interleaved scales tensor.
+
+    Matmul (cuBLAS) expects scale factors in specific `interleaved layout
+    <https://docs.nvidia.com/cuda/cublas/index.html#d-block-scaling-factors-layout>`_.
+
+    This function aims to abstract away the interleaved layout details, offering indexing
+    that more directly corresponds to the operand's shape.
+
+    Example:
+        Suppose that you are doing an NVFP4 matmul ``a @ b`` with ``a`` of shape (M=128,
+        K=128). For matrix ``a``, a single scale is applied to consecutive 16 elements
+        blocks in a row (axis=-1). Therefore, to find the scale applied to ``a[y, x]``, we
+        first need to adjust the x index to the index of the 16-element block it belongs to,
+        which is ``block_idx = x // 16``. Then, calling: ``get_block_scale_offset((y,
+        block_idx), a, BlockScalingFormat.NVFP4)`` will return the offset of the scale
+        applied to ``a[y, x]`` (and all other elements in the same 16-element block).
+
+        The schematic below shows matrix ``a`` with the 16-element blocks annotated.
+        Asterisks mark two target blocks:
+
+        - elements in ``a`` at indices from (5, 32) to (5, 47), correspond to the same block
+          (K-group 2) and map to the same offset ``get_block_scale_offset((5, 2), a,
+          BlockScalingFormat.NVFP4) == 82``
+        - elements in ``a`` at indices from (5, 80) to (5, 95), correspond to the same block
+          (K-group 5) and map to the same offset ``get_block_scale_offset((5, 5), a,
+          BlockScalingFormat.NVFP4) == 593``
+
+        .. code-block:: text
+
+                  | K-grp 0  | K-grp 1  | K-grp 2  | K-grp 3  | K-grp 4  | K-grp 5  | ...
+                  | [0..15]  | [16..31] | [32..47] | [48..63] | [64..79] | [80..95] | ...
+                  +----------+----------+----------+----------+----------+----------+---
+            row 0 |          |          |          |          |          |          |
+             ...  |          |          |          |          |          |          |
+            row 5 |          |          |    *     |          |          |    *     |
+             ...  |          |          |          |          |          |          |
+            row127|          |          |          |          |          |          |
+                  +----------+----------+----------+----------+----------+----------+---
+                                          (5,2)                            (5,5)
+
+    Note:
+        As far as computing the block scale offset, the only difference between MXFP8 and
+        NVFP4 is the number of elements in a block (32 for MXFP8, 16 for NVFP4).
+
+    Args:
+        index: A tuple of indices with length equal to ``len(operand_shape)``. Can be:
+
+            - A tuple of integers for single-element query, e.g., ``(10, 20)``
+            - A tuple of tensors for batch query, e.g., ``(xs, ys)`` where ``xs`` and ``ys``
+              are tensors of the same shape
+
+        operand_or_shape: Operand tensor (that the scales apply to) or
+            the operand's logical (non-packed, non-blocked) shape.
+
+        block_scaling_format: The block scaling format of the operand:
+            :attr:`BlockScalingFormat.NVFP4` or :attr:`BlockScalingFormat.MXFP8`.
+            Internally, it is validated to be consistent with the operand dtype, and a
+            :exc:`ValueError` is raised if not.
+
+        axis: The blocked dimension of the operand tensor.
+            For example, for NVFP4/MXFP8 matmul, A is blocked in rows (``axis = -1``), and B
+            is blocked in columns (``axis = -2``). Depending on ``operand_or_shape``:
+
+            - if a *shape* is passed to ``operand_or_shape``, then ``axis`` is required
+            - if an *operand* is passed to ``operand_or_shape``, then ``axis`` can be
+              omitted and the blocked dimension is inferred from the operand's layout.
+
+    Returns:
+        An integer (if ``index`` contains integers) or a tensor of integers (if ``index``
+        contains tensors), indicating the offset(s) to the MXFP8/NVFP4 block scale
+        factor(s). The returned offset points to a block scale factor that is applied to:
+
+        - for axis == -2: ``operand[*index[-2:],
+          block_size*index[-2]:block_size*(index[-2]+1), index[-1]]``.
+        - for axis == -1: ``operand[*index[-2:], index[-1],
+          block_size*index[-1]:block_size*(index[-1]+1)]``.
+
+        where the block size is 32 for MXFP8 and 16 for NVFP4.
+
+    Note:
+        In typical use-cases, there should be no need to manually modify MXFP8 scales. The
+        scales returned as ``"d_out_scale"`` by one matmul, can be directly reused as input
+        scales for another matmul.
+
+    Hint:
+        - To apply the interleaved scales (e.g. as returned by matmul's ``d_out_scale``) to
+          the operand, use :func:`apply_mxfp8_scale` instead.
+        - To specify scales as ND tensor and copy them to cuBLAS-compatible interleaved
+          layout, use :func:`to_block_scale` instead.
+
+    """
+    # infer operand_shape, unblocked_axis, blocked_axis, num_scalars_in_block
+    # validate operand_shape's dim and extents for divisibility by cuBLAS tile
+    # and num_scalars_in_block
+    operand_shape, unblocked_axis, blocked_axis, _, num_scalars_in_block = _validate_shape_axes_block_scaling_format(
+        operand_or_shape, axis, block_scaling_format
+    )
+
+    # We need to validate index against the operand_shape
+    ndim = len(operand_shape)
+    if len(index) != ndim:
+        raise ValueError("Index length must match the number of dimensions of the operand.")
+
+    blocked_dim = operand_shape[blocked_axis]
+    unblocked_idx = index[unblocked_axis]
+    blocked_group_idx = index[blocked_axis]
+
+    # Tile dimensions
+    TILE_OUTER = 128
+    TILE_INNER_GROUPS = 4
+
+    # Determine tile coordinates
+    sf_outer = unblocked_idx // TILE_OUTER
+    sf_inner = (blocked_group_idx // TILE_INNER_GROUPS) * TILE_INNER_GROUPS
+    sf_inner_dim = blocked_dim // num_scalars_in_block
+
+    # Compute tile offset (global layout)
+    # see https://docs.nvidia.com/cuda/cublas/index.html#d-block-scaling-factors-layout
+    tile_offset = (sf_inner + sf_outer * sf_inner_dim) * 128
+
+    # Compute intra-tile offset
+    intra_outer = unblocked_idx % TILE_OUTER
+    intra_inner = blocked_group_idx % TILE_INNER_GROUPS
+    intra_offset = (intra_outer % 32) * 16 + (intra_outer // 32) * 4 + intra_inner
+
+    batch_offset = _idx_batch_offset(operand_shape, index, num_scalars_in_block)
+
+    return batch_offset + tile_offset + intra_offset
+
+
+def _expand_block_scale(
+    scales_1d: torch.Tensor,
+    operand_or_shape: torch.Tensor | tuple[int, ...],
+    block_scaling_format: BlockScalingFormat,
+    axis: Literal[-1, -2] | None,
+    device: Literal["cuda", "cpu"] | None = None,
+) -> torch.Tensor:
+    """
+    For a documentation, see the public function :func:`expand_block_scale`.
+    This one does most of the work but does not handle conversion of the expanded
+    scales to output dtype, instead it returns scales as uint8.
+    """
+    import torch
+
+    if not isinstance(scales_1d, torch.Tensor):
+        raise TypeError(f"scales_1d must be a torch.Tensor, got {type(scales_1d)}")
+
+    scales_wrapped = wrap_operand(scales_1d)
+    scales_ndim = len(scales_wrapped.shape)
+
+    if scales_ndim != 1:
+        raise ValueError(f"scales_1d must be a 1D tensor, got {scales_ndim}D tensor with shape {scales_wrapped.shape}")
+
+    if scales_wrapped.strides[0] != 1:
+        raise ValueError(f"scales_1d must be 1D contiguous tensor, got non-unit stride: {scales_wrapped.strides}")
+
+    operand_shape, unblocked_axis, blocked_axis, block_scaling_format, num_scalars_in_block = (
+        _validate_shape_axes_block_scaling_format(operand_or_shape, axis, block_scaling_format)
+    )
+
+    _validate_scale_dtype_block_scaling_format_compatibility(scales_wrapped.dtype, block_scaling_format, "scales_1d")
+
+    expected_num_scales = math.prod(operand_shape) // num_scalars_in_block
+    num_scales = scales_wrapped.shape[0]
+    if num_scales != expected_num_scales:
+        raise ValueError(
+            f"The scale tensor must have {expected_num_scales} elements: "
+            f"{expected_num_scales} = math.prod({operand_shape}) // {num_scalars_in_block},"
+            f" got: {num_scales}."
+        )
+
+    matrix_shape, matrix_strides, matrix_logical_shape = _scales_nd_matrix_tiled_layout(
+        operand_shape, unblocked_axis, blocked_axis, num_scalars_in_block, True
+    )
+
+    if device is not None and device not in ("cuda", "cpu"):
+        raise ValueError(f"device must be 'cuda', 'cpu', or None, got '{device}'")
+    target_device = torch.device(device) if device is not None else scales_wrapped.tensor.device
+
+    scales_on_device = scales_wrapped.tensor.to(target_device)
+    scales_on_device = scales_on_device.view(torch.uint8)
+
+    expanded = torch.as_strided(
+        scales_on_device,
+        size=matrix_shape,
+        stride=matrix_strides,
+    ).reshape(matrix_logical_shape)
+    return expanded
+
+
+def _convert_to_output_dtype(
+    tensor: torch.Tensor,
+    output_dtype: Literal["smallest"] | torch.dtype,
+) -> torch.Tensor:
+    import torch
+
+    def _float_rank(dtype):
+        if dtype == torch.float16:
+            return 0
+        elif dtype == torch.float32:
+            return 1
+        else:
+            return 2
+
+    if output_dtype == tensor.dtype:
+        return tensor
+
+    smallest_dtype = _smallest_dtype_that_fits(tensor)
+
+    if output_dtype == "smallest":
+        return tensor.type(smallest_dtype)
+    else:
+        if _float_rank(output_dtype) < _float_rank(smallest_dtype):
+            raise ValueError(
+                f"Result requires at least {smallest_dtype}; requested {output_dtype} would overflow or underflow."
+            )
+        return tensor.type(output_dtype)
+
+
+def _convert_uint8_ue8m0_scale_to_float64(mx_scales: torch.Tensor) -> torch.Tensor:
+    import torch
+
+    assert mx_scales.dtype == torch.uint8
+    # Use float64 for scale computation to be safe always,
+    # this way we avoid overflow for corner case
+    # when exponent is 128 (2^128 overflows float32)
+    return 2 ** (mx_scales.type(torch.float64) - 127)
+
+
+def _smallest_dtype_that_fits(out_64):
+    """
+    (Private) Return the smallest torch dtype that can
+    hold the values in out_64 without overflow or underflow.
+    """
+    import torch
+
+    nonzero = out_64[out_64 != 0]
+    if nonzero.numel() == 0:
+        min_abs = 0
+        max_abs = 0
+    else:
+        abs_vals = torch.abs(nonzero)
+        min_abs = torch.min(abs_vals).item()
+        max_abs = torch.max(abs_vals).item()
+
+    finfo16 = torch.finfo(torch.float16)
+    finfo32 = torch.finfo(torch.float32)
+
+    if max_abs <= finfo16.max and (min_abs >= finfo16.tiny or min_abs == 0):
+        return torch.float16
+    if max_abs <= finfo32.max and (min_abs >= finfo32.tiny or min_abs == 0):
+        return torch.float32
+    return torch.float64
+
+
+def apply_mxfp8_scale(
+    x: torch.Tensor,
+    scales_1d: torch.Tensor,
+    output_dtype: Literal["smallest"] | torch.dtype = "smallest",
+) -> torch.Tensor:
+    """
+    .. experimental:: function
+
+    Apply MXFP8 block scale factors to a tensor ``x``.
+
+    Args:
+        x: The tensor to which the scaling should be applied.
+            Currently it must be a ``torch.Tensor``.
+
+        scales_1d: The block scale factors (stored in cuBLAS-compatible interleaved layout)
+            to apply. Its shape must be compatible with ``x``, and currently it must also be
+            a ``torch.Tensor``.
+
+        output_dtype: Output dtype. If provided, must be a floating-point
+            ``torch.dtype`` (float16, float32, or float64) and must be at least as wide as
+            the smallest dtype that can represent the result, or :exc:`ValueError` is
+            raised. If 'smallest' (default), the smallest dtype that can represent the
+            result is automatically chosen.
+
+    Returns:
+        A tensor with values of ``x`` with scales applied, in the chosen or provided dtype.
+
+    Raises:
+        ValueError: When the result will over/underflow the requested dtype.
+
+    Behavior:
+        The operation is computed in float64. Then, the function determines the smallest
+        dtype (float16, float32, or float64) that can represent the result without overflow
+        or underflow. If ``output_dtype`` was passed, it must be at least as wide as that
+        minimum otherwise :exc:`ValueError` is raised; if ``output_dtype='smallest'``, that
+        minimum is used. The result is finally cast to the chosen dtype and returned.
+
+    Note:
+        This function is not intended for production usage due to its relatively low
+        performance and high memory consumption. Prefer
+        :attr:`~nvmath.linalg.advanced.MatmulOptions.result_type` to request non-FP8 output.
+    """
+    import torch
+
+    if output_dtype != "smallest" and output_dtype not in (torch.float16, torch.float32, torch.float64):
+        raise TypeError("output_dtype must be 'smallest' or one of torch.float16, torch.float32, torch.float64.")
+
+    # Use float64 for scale computation to be safe always,
+    # this way we avoid overflow for corner case
+    # when exponent is 128 (2^128 overflows float32)
+    expanded_scales = _expand_block_scale(scales_1d, x, BlockScalingFormat.MXFP8, axis=None)
+    expanded_scales = _convert_uint8_ue8m0_scale_to_float64(expanded_scales)
+    # Explicitly cast x to float64, promotion is not guaranteed, see:
+    # https://docs.pytorch.org/docs/stable/tensor_attributes.html#type-promotion-doc,
+    # "Promotion for shell dtypes is not defined".
+    out_64 = x.type(torch.float64) * expanded_scales
+    return _convert_to_output_dtype(out_64, output_dtype)
+
+
+# ====================================================
+# helper functions for FP4 E2M1 encoding/decoding
+# ====================================================
+
+
+_FP4_DECODE_VALUES = (0.0, 0.5, 1.0, 1.5, 2.0, 3.0, 4.0, 6.0, -0.0, -0.5, -1.0, -1.5, -2.0, -3.0, -4.0, -6.0)
+_FP4_BOUNDARIES = [0.25, 0.75, 1.25, 1.75, 2.5, 3.5, 5.0]
+_FP4_MAG_CODES = [0x0, 0x1, 0x2, 0x3, 0x4, 0x5, 0x6, 0x7]
+
+
+def _get_fp4_lookup_table(device):
+    """
+    (Private) Create FP4 lookup table as
+    a torch tensor on the specified device.
+
+    Args:
+        device: The device to create the lookup table on.
+
+    Returns:
+        A torch tensor with dtype torch.float32 and shape (16,)
+        on the specified device.
+    """
+    import torch
+
+    return torch.tensor(
+        _FP4_DECODE_VALUES,
+        dtype=torch.float32,
+        device=device,
+    )
+
+
+def _quantize_to_fp4_codes_bucketize(x: torch.Tensor, device) -> torch.Tensor:
+    """
+    Quantize float32 tensor to nearest FP4 value using torch.bucketize on the 8
+    distinct FP4 magnitudes.
+    """
+    import torch
+
+    if x.dtype != torch.float32:
+        raise ValueError(f"x must be float32, got {x.dtype}")
+
+    boundaries = torch.tensor(_FP4_BOUNDARIES, dtype=torch.float32, device=device)
+    mag_codes = torch.tensor(_FP4_MAG_CODES, dtype=torch.uint8, device=device)
+
+    # Example: x = [1.2, -2.3, 0.1]
+
+    # sign: True where negative (including -0.0).  [False, True, False]
+    sign = torch.signbit(x)
+
+    # mag: absolute value.  [1.2, 2.3, 0.1]
+    mag = x.abs()
+
+    # Binary-search each magnitude into the boundary bins:
+    #   https://docs.pytorch.org/docs/stable/generated/torch.bucketize.html
+    #   boundaries = [0.25, 0.75, 1.25, 1.75, 2.5, 3.5, 5.0]
+    #   Default (right=False): boundary[i-1] < value <= boundary[i]
+    #   1.2: 0.75 < 1.2 <= 1.25  -> bucket 2
+    #   2.3: 1.75 < 2.3 <= 2.5   -> bucket 4
+    #   0.1: 0.1  <= 0.25        -> bucket 0
+    #   Result: [2, 4, 0]
+    bucket = torch.bucketize(mag, boundaries)
+
+    # Map bucket index to the 3-bit magnitude code (lower 3 bits of FP4).
+    #   mag_codes = [0x0, 0x1, 0x2, 0x3, 0x4, 0x5, 0x6, 0x7]
+    #   bucket [2, 4, 0] -> code [0x2, 0x4, 0x0]
+    #   These represent magnitudes [1.0, 2.0, 0.0].
+    code = mag_codes[bucket]
+
+    # Set bit 3 (the FP4 sign bit) for negative values.
+    #   sign.to(torch.uint8): [F, T, F] -> [0, 1, 0]
+    #   << 3: shift into bit 3: [0x0, 0x8, 0x0]
+    #   OR with magnitude code:
+    #   code [0x2, 0x4, 0x0] | [0x0, 0x8, 0x0] = [0x2, 0xC, 0x0]
+    #   Final FP4 values: [1.0, -2.0, 0.0]
+    code = code | (sign.to(torch.uint8) << 3)
+    return code
+
+
+def quantize_to_fp4(
+    x: torch.Tensor,
+    axis: Literal[-1, -2],
+) -> torch.Tensor:
+    """
+    .. experimental:: function
+
+    Quantize a torch tensor to ``torch.float4_e2m1fn_x2`` dtype.
+
+    The function supports 1D, 2D, and higher-dimensional input torch tensors with dtype
+    float32. It quantizes each float32 value to the nearest representable FP4 value and
+    packs two 4-bit codes per byte, halving the packed dimension. The packing direction is
+    controlled by ``axis``:
+
+    - ``axis=-1``: Packs consecutive elements along the last dimension. Input shape ``(...,
+      Q)`` produces output ``(..., Q//2)`` with row-major layout (last stride = 1).
+
+    - ``axis=-2``: Packs consecutive elements along the second-to-last dimension. Input
+      shape ``(..., P, Q)`` produces output ``(..., P//2, Q)`` with column-major layout
+      (second-to-last stride = 1).
+
+    Args:
+        x: Torch tensor with dtype float32 (1D, 2D, or higher-dimensional).
+
+        axis: The axis along which to pack. Must be ``-1`` (last dimension)
+            or ``-2`` (second-to-last dimension).
+
+    Returns:
+        Torch tensor with dtype ``torch.float4_e2m1fn_x2`` on the same device as the input.
+
+    .. important::
+        The packed dimension must have even size.
+
+    Note:
+        This helper quantizes a single tensor and is suitable for understanding how packing
+        for ``torch.float4_e2m1fn_x2`` works in practice and/or for experimenting with FP4
+        GEMMs outside of typical deep-learning workflows. It is not fully optimized for
+        performance but should be adequate for most common use cases. For production
+        whole-model quantization, consider tools such as `torchao
+        <https://github.com/pytorch/ao>`_ or `bitsandbytes
+        <https://github.com/bitsandbytes-foundation/bitsandbytes>`_.
+
+    .. seealso::
+        :func:`unpack_fp4` — decode packed FP4 values back to float32.
+    """
+    import torch
+
+    # preconditions
+    # -------------
+    if not isinstance(x, torch.Tensor):
+        raise TypeError(f"x must be a torch.Tensor, got {type(x)}")
+    if x.dtype != torch.float32:
+        raise ValueError(f"x must be float32, got {x.dtype}")
+    if axis not in (-1, -2):
+        raise ValueError(f"axis must be -1 or -2, got {axis}")
+    if x.ndim < 1:
+        raise ValueError(f"x must be at least 1D, got {x.ndim}D")
+    if x.ndim == 1 and axis != -1:
+        raise ValueError(f"axis must be -1 for 1D tensors, got {axis}")
+
+    if (packed_dim := x.shape[axis]) % 2 != 0:
+        raise ValueError(f"Packed dimension must be even, got {packed_dim}")
+
+    device = x.device
+    codes = _quantize_to_fp4_codes_bucketize(x, device)
+
+    if axis == -1:
+        low = codes[..., 0::2]
+        high = codes[..., 1::2]
+        packed = low | (high << 4)
+        return packed.view(torch.float4_e2m1fn_x2)
+
+    # axis == -2: column-wise packing
+    low = codes[..., 0::2, :]
+    high = codes[..., 1::2, :]
+    packed_codes = low | (high << 4)
+    # Transpose so columns become contiguous rows, force a physical copy,
+    # reinterpret as fp4x2 (requires stride[-1]==1), then transpose back.
+    # The result has the original (…, rows, cols) shape with column-major
+    # memory layout, as requested by the caller.
+    result = packed_codes.mT.contiguous().view(torch.float4_e2m1fn_x2).mT
+    return result
+
+
+def _decode_fp4_1d_tensor_to_float32(fp4_tensor: torch.Tensor) -> torch.Tensor:
+    """
+    (Private) Decode a 1D FP4 tensor to float32.
+
+    Args:
+        fp4_tensor: 1D FP4 tensor with shape (K//2,)
+
+    Returns:
+        Float32 tensor with unpacked shape (K,) and contiguous layout.
+    """
+    import torch
+
+    assert fp4_tensor.ndim == 1, f"fp4_tensor must be 1D, got {fp4_tensor.ndim}D"
+    assert fp4_tensor.dtype == torch.float4_e2m1fn_x2, f"fp4_tensor must be float4_e2m1fn_x2, got {fp4_tensor.dtype}"
+
+    # Validate stride (must be contiguous)
+    stride = fp4_tensor.stride()
+    assert stride[0] == 1, f"1D FP4 tensor must be contiguous (stride=1), but got stride={stride}"
+
+    # Create FP4 lookup table
+    fp4_lookup = _get_fp4_lookup_table(fp4_tensor.device)
+
+    # View as uint8 and extract 4-bit FP4 codes [0..15] from each byte.
+    # low needs & 0xF to mask off bits [4..7]; high doesn't because
+    # >> 4 on uint8 already zeros bits [4..7].
+    fp4_as_uint8 = fp4_tensor.view(torch.uint8)
+    low_code = (fp4_as_uint8 & 0xF).int()
+    high_code = (fp4_as_uint8 >> 4).int()
+
+    # Map codes to float values via lookup table
+    vals_low = fp4_lookup[low_code]
+    vals_high = fp4_lookup[high_code]
+
+    # Interleave: (K//2,) -> (K,)
+    # stack along last dim pairs [low_i, high_i], reshape flattens to
+    # [low0, high0, low1, high1, ...].
+    return torch.stack([vals_low, vals_high], dim=-1).reshape(-1)
+
+
+def _decode_fp4_2d_plus_tensor_to_float32(fp4_tensor: torch.Tensor, row_wise_packing: bool) -> torch.Tensor:
+    """
+    (Private) Decode a 2-D or higher FP4 tensor to float32.
+
+    Args:
+        fp4_tensor: FP4 tensor with shape (..., M, K//2) for row-wise
+            packing or (..., K//2, N) for column-wise packing, where ... represents zero or
+            more batch dimensions.
+
+        row_wise_packing: If True, uses row-wise packing (trailing
+            dimension is packed). If False, uses column-wise packing (second-to-last
+            dimension is packed). Determined by the caller from tensor strides.
+
+    Returns:
+        Float32 tensor with unpacked shape (..., M, K) for row-wise or (..., K, N) for
+        column-wise packing.
+    """
+    import torch
+
+    assert fp4_tensor.ndim >= 2, f"fp4_tensor must be at least 2D, got {fp4_tensor.ndim}D"
+    assert fp4_tensor.dtype == torch.float4_e2m1fn_x2, f"fp4_tensor must be float4_e2m1fn_x2, got {fp4_tensor.dtype}"
+
+    batch_shape = fp4_tensor.shape[:-2]
+    device = fp4_tensor.device
+
+    # Create FP4 lookup table
+    fp4_lookup = _get_fp4_lookup_table(device)
+
+    # View as uint8 and extract 4-bit values (vectorized across all dims).
+    # low needs & 0xF to mask off bits [4..7]; high doesn't because
+    # >> 4 on uint8 already zeros bits [4..7].
+    fp4_as_uint8 = fp4_tensor.view(torch.uint8)
+    low_code = (fp4_as_uint8 & 0xF).int()
+    high_code = (fp4_as_uint8 >> 4).int()
+
+    # Lookup decoded values (vectorized)
+    vals_low = fp4_lookup[low_code]
+    vals_high = fp4_lookup[high_code]
+
+    if row_wise_packing:
+        # Row-wise: (..., M, K//2) -> (..., M, K) with row-major layout
+        # Expand last dimension by 2
+        expanded_shape = batch_shape + (fp4_tensor.shape[-2], fp4_tensor.shape[-1] * 2)
+        result = torch.zeros(expanded_shape, dtype=torch.float32, device=device)
+        result[..., 0::2] = vals_low
+        result[..., 1::2] = vals_high
+    else:
+        # Column-wise: (..., K//2, N) -> (..., K, N) with column-major layout
+        # Expand second-to-last dimension by 2
+        rows, cols = fp4_tensor.shape[-2:]
+        expanded_shape = batch_shape + (rows * 2, cols)
+
+        # Create column-major layout for batched tensors
+        matrix_size = (rows * 2) * cols
+        batch_size = int(np.prod(batch_shape)) if batch_shape else 1
+        storage = torch.zeros(batch_size * matrix_size, dtype=torch.float32, device=device)
+        # Stride: batch elements contiguous, then column-major per matrix
+        output_stride = (matrix_size,) + (1, rows * 2)
+        result = torch.as_strided(storage, size=(batch_size,) + (rows * 2, cols), stride=output_stride)
+        # Flatten batch dims in vals_low/vals_high to match result shape
+        vals_low_flat = vals_low.reshape(batch_size, rows, cols)
+        vals_high_flat = vals_high.reshape(batch_size, rows, cols)
+        result[:, 0::2, :] = vals_low_flat
+        result[:, 1::2, :] = vals_high_flat
+        result = result.reshape(expanded_shape)
+
+    return result
+
+
+def unpack_fp4(fp4_tensor: torch.Tensor, axis: Literal[-1, -2]) -> torch.Tensor:
+    """
+    .. experimental:: function
+
+    Unpack an N-D torch tensor with dtype ``torch.float4_e2m1fn_x2`` to float32.
+
+    Since each byte stores two FP4 values, the output tensor has one dimension doubled along
+    ``axis``.
+
+    - ``axis=-1``: The last dimension is the packed axis. Input shape ``(..., Q)`` with
+      row-major layout produces output ``(..., 2*Q)``.
+    - ``axis=-2``: The second-to-last dimension is the packed axis. Input shape ``(..., P,
+      Q)`` with column-major layout produces output ``(..., 2*P, Q)``.
+
+    Args:
+        fp4_tensor: FP4 tensor with dtype ``torch.float4_e2m1fn_x2``.
+
+        axis: The axis along which the tensor was packed. Must be ``-1``
+            (last dimension) or ``-2`` (second-to-last dimension).
+
+    Returns:
+        A torch tensor with dtype float32 with the unpacked shape on the same device as the
+        input.
+
+    .. seealso::
+        :func:`quantize_to_fp4` — quantize and pack float32 values to FP4.
+    """
+    import torch
+
+    if not isinstance(fp4_tensor, torch.Tensor):
+        raise TypeError(f"fp4_tensor must be a torch.Tensor, got {type(fp4_tensor)}")
+    if fp4_tensor.ndim < 1:
+        raise ValueError(f"fp4_tensor must be at least 1D, got {fp4_tensor.ndim}D")
+    if fp4_tensor.dtype != torch.float4_e2m1fn_x2:
+        raise ValueError(f"fp4_tensor must be float4_e2m1fn_x2, got {fp4_tensor.dtype}")
+    if axis not in (-1, -2):
+        raise ValueError(f"axis must be -1 or -2, got {axis}")
+    if fp4_tensor.ndim == 1 and axis != -1:
+        raise ValueError(f"axis must be -1 for 1D tensors, got {axis}")
+
+    row_wise_packing = axis == -1
+
+    if fp4_tensor.ndim == 1:
+        return _decode_fp4_1d_tensor_to_float32(fp4_tensor)
+    return _decode_fp4_2d_plus_tensor_to_float32(fp4_tensor, row_wise_packing)
+
+
+def to_block_scale(
+    scale_tensor: torch.Tensor,
+    operand_or_shape: torch.Tensor | tuple[int, ...],
+    block_scaling_format: BlockScalingFormat,
+    *,
+    axis: Literal[-1, -2] | None = None,
+    out: torch.Tensor | None = None,
+) -> torch.Tensor:
+    """
+    .. experimental:: function
+
+    Copy ND scale tensor to flat tensor accounting for the tiled layout required by
+    cuBLASLt.
+
+    Matmul (cuBLAS) expects scale factors in specific `interleaved layout
+    <https://docs.nvidia.com/cuda/cublas/index.html#d-block-scaling-factors-layout>`_.
+
+    This function aims to abstract away the interleaved layout details, offering a way to
+    specify scales as ND tensor with shape corresponding to the operand's shape and copy
+    them to cuBLAS-compatible interleaved layout.
+
+    Example:
+        Suppose that you are doing an NVFP4 matmul ``a @ b`` with ``a`` of shape (M=128,
+        K=128). For matrix ``a``, a single scale is applied to consecutive 16 elements
+        blocks in a row (axis=-1). You can specify the block scales as a ND tensor with
+        shape ``(M, K // 16)`` such that scale from ``scale_tensor[i, j]`` will be applied
+        to the block of elements ``a[i, j*16:j*16+16]`` and then call
+        ``to_block_scale(scale_tensor, a, BlockScalingFormat.NVFP4)``, which will return a
+        1D interleaved scale tensor that can be passed as quantization scales for the
+        matmul.
+
+    Note:
+        As far as computing the block scale offset, the only difference between MXFP8 and
+        NVFP4 is the number of elements in a block (32 for MXFP8, 16 for NVFP4).
+
+    Args:
+        scale_tensor: ND scale tensor with dtype:
+
+            - for NVFP4: ``torch.float8_e4m3fn`` or ``torch.uint8`` (interpreted as
+              ``torch.float8_e4m3fn``)
+            - for MXFP8: ``torch.uint8`` (interpreted as ``UE8M0``)
+
+        operand_or_shape: Operand tensor (that the scales apply to) or
+            the operand's logical (non-packed, non-blocked) shape.
+
+        block_scaling_format: The block scaling format of the operand:
+            :attr:`BlockScalingFormat.NVFP4` or :attr:`BlockScalingFormat.MXFP8`.
+            Internally, it is validated to be consistent with the operand dtype, and a
+            :exc:`ValueError` is raised if not.
+
+        axis: The blocked dimension of the operand tensor.
+            For example, for NVFP4/MXFP8 matmul, A is blocked in rows (``axis = -1``), and B
+            is blocked in columns (``axis = -2``). Depending on ``operand_or_shape``:
+
+            - if a *shape* is passed to ``operand_or_shape``, then ``axis`` is required
+            - if an *operand* is passed to ``operand_or_shape``, then ``axis`` can be
+              omitted and the blocked dimension is inferred from the operand's layout.
+
+        out: Output tensor to copy the scales to. If ``None``, a new tensor is created.
+
+    Returns:
+        Flat ``out`` tensor containing the scales copied to match cuBLAS-compatible
+        interleaved layout. The `out` dtype is the same as the `scale_tensor` dtype.
+    """
+    import torch
+
+    if not isinstance(scale_tensor, torch.Tensor):
+        raise TypeError(f"scale_tensor must be a torch.Tensor, got {type(scale_tensor)}")
+
+    scale_wrapped = wrap_operand(scale_tensor)
+    num_scales = scale_wrapped.size
+
+    operand_shape, unblocked_axis, blocked_axis, block_scaling_format, num_scalars_in_block = (
+        _validate_shape_axes_block_scaling_format(operand_or_shape, axis, block_scaling_format)
+    )
+
+    _validate_scale_dtype_block_scaling_format_compatibility(scale_wrapped.dtype, block_scaling_format, "scale_tensor")
+
+    num_scales = scale_wrapped.size
+    expected_num_scales = math.prod(operand_shape) // num_scalars_in_block
+    if num_scales != expected_num_scales:
+        raise ValueError(
+            f"For operand of shape {operand_shape}, and block_scaling_format {block_scaling_format}, "
+            f"the scale_tensor must have shape {expected_num_scales}, got {num_scales} (shape:{scale_wrapped.shape})."
+        )
+
+    if out is None:
+        out = torch.empty(
+            num_scales,
+            device=scale_wrapped.tensor.device,
+            dtype=scale_wrapped.tensor.dtype,
+        )
+    else:
+        if out.ndim != 1:
+            raise ValueError(f"out must be a 1D tensor, got {out.ndim}D tensor with shape {out.shape}")
+        if out.shape[0] != num_scales:
+            raise ValueError(
+                f"Flat scale tensor (out) and ND scale_tensor must have the same number "
+                f"of elements, got {out.shape[0]} and {num_scales}"
+            )
+        if out.dtype != scale_wrapped.tensor.dtype:
+            raise ValueError(
+                f"Flat scale tensor (out) and ND scale_tensor must "
+                f"have the same dtype, got {out.dtype} and {scale_wrapped.dtype}"
+            )
+
+    matrix_shape, matrix_strides, matrix_logical_shape = _scales_nd_matrix_tiled_layout(
+        operand_shape, unblocked_axis, blocked_axis, num_scalars_in_block, False
+    )
+
+    if scale_wrapped.shape != matrix_logical_shape:
+        raise ValueError(
+            f"For operand of shape {operand_shape}, block_scaling_format {block_scaling_format}, "
+            f"blocked along axis {axis}, "
+            f"the scale_tensor must have shape {matrix_logical_shape}, got {scale_wrapped.shape}."
+        )
+
+    torch.as_strided(
+        out,
+        size=matrix_shape,
+        stride=matrix_strides,
+    ).copy_(scale_wrapped.tensor.view(matrix_shape))
+    return out
+
+
+def expand_block_scale(
+    scales_1d: torch.Tensor,
+    operand_or_shape: torch.Tensor | tuple[int, ...],
+    block_scaling_format: BlockScalingFormat,
+    *,
+    axis: Literal[-1, -2] | None = None,
+    output_dtype: Literal["smallest"] | torch.dtype = "smallest",
+    device: Literal["cuda", "cpu"] | None = None,
+) -> torch.Tensor:
+    """
+    .. experimental:: function
+
+    Expand NVFP4/MXFP8 block scales from 1D cuBLAS-compatible interleaved array to the full
+    operand shape.
+
+    Matmul (cuBLAS) expects and returns the block scale factors in specific `interleaved
+    layout
+    <https://docs.nvidia.com/cuda/cublas/index.html#d-block-scaling-factors-layout>`_.
+
+    This function takes that 1D interleaved scale array (either provided as input or
+    returned by cuBLASLt for NVFP4/MXFP8 output) and expands it to a full ND tensor with
+    shape ``operand_or_shape`` where each element gets its corresponding scale value. This
+    can be useful, for example, to manually dequantize the result of a matmul, by
+    elementwise multiplication of the expanded scales with the result.
+
+    Args:
+        scales_1d: 1D tensor of scale values with dtype:
+
+            - for NVFP4: ``torch.float8_e4m3fn``, or ``torch.uint8`` (interpreted as
+              ``torch.float8_e4m3fn``)
+            - for MXFP8: ``torch.uint8``, interpreted as exponent (``UE8M0``)
+
+            The scales are expected to be stored in cuBLAS-compatible interleaved layout
+            (e.g. as returned by matmul's ``d_out_scale``). The number of elements in the
+            tensor must be equal to the number of elements in the operand tensor, divided by
+            the number of elements in a block (for NVFP4: 16, for MXFP8: 32).
+
+        operand_or_shape: Operand tensor or its logical (non-packed, non-blocked) shape.
+            The scales are expanded to match this shape.
+
+        block_scaling_format: The block scaling format of the operand:
+            :attr:`BlockScalingFormat.NVFP4` or :attr:`BlockScalingFormat.MXFP8`.
+            Internally, it is validated to be consistent with the operand dtype, and a
+            :exc:`ValueError` is raised if not.
+
+        axis: The blocked dimension of the operand tensor.
+            For example, for NVFP4/MXFP8 matmul, A is blocked in rows (``axis = -1``), and B
+            is blocked in columns (``axis = -2``). Depending on ``operand_or_shape``:
+
+            - if a *shape* is passed to ``operand_or_shape``, then ``axis`` is required
+            - if an *operand* is passed to ``operand_or_shape``, then ``axis`` can be
+              omitted and the blocked dimension is inferred from the operand's layout.
+
+        output_dtype: Output dtype.
+            If provided, must be a torch's dtype:
+
+            - for NVFP4: ``float8_e4m3fn``, ``float16``, ``float32``, or ``float64``
+            - for MXFP8: ``uint8`` (exponent ``UE8M0``), ``float16``, ``float32``, or
+              ``float64``
+
+            It must be wide enough to represent the result, or :exc:`ValueError` is raised.
+            If 'smallest' (default), the smallest of accepted dtypes that can represent the
+            result is automatically chosen (for MXFP8: ``uint8`` interpreted as exponent
+            (``UE8M0``), for NVFP4: ``float8_e4m3fn``).
+
+        device: Device for the output tensor. When ``None`` (default), the
+            device is inferred from ``scales_1d``. When specified, must be ``"cuda"`` or
+            ``"cpu"``.
+
+    Returns:
+        Tensor with shape ``operand_or_shape`` (and dtype as specified by ``output_dtype``)
+        containing expanded scales, on the target device. Each element contains the scale
+        value that applies to the corresponding position in the FP4/FP8 matrix.
+
+    Note:
+        For computing a single scale index rather than expanding all scales, use
+        :func:`get_block_scale_offset` instead.
+    """
+    import torch
+
+    _COMMON_EXPAND_OUTPUT_DTYPES = (torch.float16, torch.float32, torch.float64)
+
+    expanded = _expand_block_scale(scales_1d, operand_or_shape, block_scaling_format, axis, device)
+    assert expanded.dtype == torch.uint8
+
+    scale_interpretation = _MICROSCALING_FORMAT_PROPERTIES[block_scaling_format]["scale_interpretation"]
+
+    if scale_interpretation == "float8_e4m3fn":
+        expanded = expanded.view(torch.float8_e4m3fn)
+        if output_dtype == "smallest" or output_dtype == torch.float8_e4m3fn:
+            return expanded
+        elif output_dtype in _COMMON_EXPAND_OUTPUT_DTYPES:
+            return expanded.type(output_dtype)
+        else:
+            supported_dtypes_str = ", ".join([str(dt) for dt in (torch.float8_e4m3fn,) + _COMMON_EXPAND_OUTPUT_DTYPES])
+            raise TypeError(f"output_dtype must be 'smallest' or one of {supported_dtypes_str}. Got {output_dtype}")
+    elif scale_interpretation == "ue8m0":
+        if output_dtype == "smallest" or output_dtype == torch.uint8:
+            return expanded
+        elif output_dtype in _COMMON_EXPAND_OUTPUT_DTYPES:
+            # TODO: This could be optimized - we can assess the exponents
+            # keeping uint8 type and directly convert to output dtype.
+            expanded = _convert_uint8_ue8m0_scale_to_float64(expanded)
+            return _convert_to_output_dtype(expanded, output_dtype)
+        else:
+            supported_dtypes_str = ", ".join([str(dt) for dt in (torch.uint8,) + _COMMON_EXPAND_OUTPUT_DTYPES])
+            raise TypeError(f"output_dtype must be 'smallest' or one of {supported_dtypes_str}. Got {output_dtype}")
+    else:
+        raise AssertionError(
+            f"Unknown scale_interpretation '{scale_interpretation}' for block_scaling_format '{block_scaling_format}'"
+        )