# Copyright 2024, UChicago Argonne, LLC
# All Rights Reserved
# Software Name: NEML2 -- the New Engineering material Model Library, version 2
# By: Argonne National Laboratory
# OPEN SOURCE LICENSE (MIT)
#
# Permission is hereby granted, free of charge, to any person obtaining a copy
# of this software and associated documentation files (the "Software"), to deal
# in the Software without restriction, including without limitation the rights
# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
# copies of the Software, and to permit persons to whom the Software is
# furnished to do so, subject to the following conditions:
#
# The above copyright notice and this permission notice shall be included in
# all copies or substantial portions of the Software.
#
# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
# THE SOFTWARE.
"""Scalar — a physically-meaningful 0-base-shape tensor.
`Scalar` is a wrapper (not a `torch.Tensor` alias) so cross-type operator
dispatch (`Scalar * SR2 → SR2`) is deterministic via Python's reflected-operator
protocol. With a bare `torch.Tensor` on the left, Python would never invoke
`SR2.__rmul__`.
What Scalar overrides on top of :class:`PrimitiveTensor`:
- **Literal-friendly ``__init__``**: ``Scalar(2.5)`` and ``Scalar([1, 2, 3])``
work directly; literals are coerced to ``torch.float64`` (the default
precision for typed-wrapper algebra).
- **``float64`` factory defaults**: ``Scalar.zeros(n)``, ``Scalar.ones(n)``,
``Scalar.full(n, fill_value=...)`` default to ``torch.float64`` rather than
torch's global ``float32`` default.
- **``arange``**: Scalar-only classmethod mirroring the torch creation API. To
build a ramp between two endpoints — the v3 form of v2's ``dynamic_linspace`` /
``intmd_linspace`` / ``base_linspace`` — use the
:func:`~neml2.types.functions.linspace` / :func:`~neml2.types.functions.logspace`
free functions with region-view endpoints (``x.dynamic_batch`` / ``x.sub_batch``).
- **``+`` / ``-`` with Python literals**: ``s + 1.5`` and ``s - 1`` are valid;
the other primitives reject this (a uniform additive offset is rare and
ambiguous on a (3,3) or (6,) tensor). Multiply / divide by literal are
inherited from :meth:`PrimitiveTensor._scale`.
- **``__rsub__`` / ``__rtruediv__``** for the reverse-with-literal cases.
- **``__pow__``, ``__abs__``**: torch-backed forwarders, Scalar-only.
Everything else (``__neg__``, region views, ``zeros``/``ones`` shape semantics)
comes from the base layers.
"""
from __future__ import annotations
from dataclasses import dataclass
from typing import ClassVar, TypeVar, overload
import torch
from neml2.types._base import TensorWrapper, align_sub_batch
from neml2.types._primitive import PrimitiveTensor
from neml2.types._pytree import register
_WrapperT = TypeVar("_WrapperT", bound=TensorWrapper)
[docs]
@dataclass(frozen=True, eq=False)
class Scalar(PrimitiveTensor):
"""Wraps a `torch.Tensor` of base shape ``()`` (i.e., one number per batch entry)."""
data: torch.Tensor
sub_batch_ndim: int = 0
sub_batch_state: tuple = ()
sub_batch_meta: tuple = ()
k_ndim: int = 0
k_state: tuple = ()
k_pairing: tuple = ()
BASE_NDIM: ClassVar[int] = 0
BASE_SHAPE: ClassVar[tuple[int, ...]] = ()
def __init__(
self,
data,
sub_batch_ndim: int = 0,
sub_batch_state: tuple = (),
sub_batch_meta: tuple = (),
k_ndim: int = 0,
k_state: tuple = (),
k_pairing: tuple = (),
*,
dtype: torch.dtype | None = None,
device: torch.device | str | None = None,
) -> None:
# Accept TensorWrapper input via the same auto-unwrap policy the
# base ``__post_init__`` enforces for dataclass-generated __init__s
# (Vec, SR2, etc.). This branch is the Scalar-specific equivalent;
# without it, ``Scalar(other_scalar, ...)`` falls through to
# ``torch.as_tensor(other_scalar)`` which doesn't know about
# TensorWrapper and raises. The outer call's metadata wins, matching
# the base policy.
if isinstance(data, TensorWrapper):
if not isinstance(data, Scalar):
raise TypeError(
f"Cannot wrap a {type(data).__name__} as a Scalar; "
"wrapper types must match. Pass `inner.data` instead of the wrapper."
)
data = data.data
if dtype is not None or device is not None:
data = data.to(dtype=dtype, device=device)
elif isinstance(data, torch.Tensor):
if dtype is not None or device is not None:
data = data.to(dtype=dtype, device=device)
else:
data = torch.as_tensor(data, dtype=dtype or torch.float64, device=device)
# The class is `@dataclass(frozen=True)` so direct attribute writes
# are forbidden; route through `object.__setattr__` to seat the
# dataclass-declared fields.
object.__setattr__(self, "data", data)
object.__setattr__(self, "sub_batch_ndim", sub_batch_ndim)
object.__setattr__(self, "sub_batch_state", sub_batch_state)
object.__setattr__(self, "sub_batch_meta", sub_batch_meta)
object.__setattr__(self, "k_ndim", k_ndim)
object.__setattr__(self, "k_state", k_state)
object.__setattr__(self, "k_pairing", k_pairing)
self.__post_init__()
[docs]
@classmethod
def from_value(cls, x: float | int, *, like: TensorWrapper) -> Scalar:
"""Construct a Scalar inheriting dtype/device from an existing wrapper."""
return cls(x, dtype=like.dtype, device=like.device)
# ---- torch-analogue factories with float64 defaults ----
#
# Override the inherited PrimitiveTensor factories so Scalars default to
# ``torch.float64`` (matching ``Scalar.__init__``) rather than torch's
# global float32 default. To build a ramp between two endpoints use the
# :func:`~neml2.types.functions.linspace` / ``logspace`` free functions with
# region-view endpoints (e.g. ``linspace(Scalar(a).sub_batch, Scalar(b).sub_batch, n)``).
[docs]
@classmethod
def zeros(
cls,
*shape: int,
dtype: torch.dtype | None = None,
device: torch.device | str | None = None,
) -> Scalar:
return cls(torch.zeros(*shape, dtype=dtype or torch.float64, device=device))
[docs]
@classmethod
def ones(
cls,
*shape: int,
dtype: torch.dtype | None = None,
device: torch.device | str | None = None,
) -> Scalar:
return cls(torch.ones(*shape, dtype=dtype or torch.float64, device=device))
[docs]
@classmethod
def full(
cls,
*shape: int,
fill_value: float,
dtype: torch.dtype | None = None,
device: torch.device | str | None = None,
) -> Scalar:
return cls(torch.full(shape, fill_value, dtype=dtype or torch.float64, device=device))
[docs]
@classmethod
def arange(
cls,
start: float,
end: float | None = None,
step: float = 1,
*,
dtype: torch.dtype | None = None,
device: torch.device | str | None = None,
) -> Scalar:
"""Like ``torch.arange``: ``arange(N)`` -> ``[0, …, N-1]``, ``arange(a, b, s)``
-> ``[a, a+s, …]`` up to (excluding) ``b``."""
if end is None:
return cls(torch.arange(start, dtype=dtype or torch.float64, device=device))
return cls(torch.arange(start, end, step, dtype=dtype or torch.float64, device=device))
# ---- arithmetic with Python literals (add/sub) ----
#
# ``__mul__`` / ``__truediv__`` are inherited via :meth:`PrimitiveTensor._scale`
# which already handles ``float`` / ``int`` literals. ``__add__`` / ``__sub__``
# need explicit overrides because the base ``_binary`` deliberately rejects
# literals (a uniform additive offset is rare and ambiguous on a (3,3) or
# (6,) tensor — on Scalar it's the natural thing).
def __add__(self, other) -> Scalar:
if isinstance(other, (float, int)):
return self._rewrap(self.data + other, sub_batch_ndim=self.sub_batch_ndim)
return self._binary(other, lambda a, b: a + b)
def __radd__(self, other) -> Scalar:
return self.__add__(other)
def __sub__(self, other) -> Scalar:
if isinstance(other, (float, int)):
return self._rewrap(self.data - other, sub_batch_ndim=self.sub_batch_ndim)
return self._binary(other, lambda a, b: a - b)
def __rsub__(self, other) -> Scalar:
if isinstance(other, (float, int)):
return self._rewrap(other - self.data, sub_batch_ndim=self.sub_batch_ndim)
return NotImplemented
@overload
def __mul__(self, other: Scalar | float | int) -> Scalar: ...
@overload
def __mul__(self, other: _WrapperT) -> _WrapperT: ...
def __mul__(self, other):
# The Scalar × Scalar and Scalar × literal cases are handled by the
# inherited ``_scale``. The wrapper-promotion case (``Scalar * SR2 ->
# SR2``) is delegated explicitly here so it type-checks at the
# ``-> _WrapperT`` overload — without this branch we'd return
# ``NotImplemented`` and rely on the wrapper's ``__rmul__``, which
# works at runtime but isn't visible to type-checkers staring at the
# ``Scalar * wrapper`` expression.
if isinstance(other, TensorWrapper) and not isinstance(other, Scalar):
return other * self
return self._scale(other, lambda a, b: a * b)
def __rmul__(self, other) -> Scalar:
return self.__mul__(other)
def __rtruediv__(self, other) -> Scalar:
if isinstance(other, (float, int)):
# ``other / self`` lowers to ``other * reciprocal(self)``, whose
# reverse-mode backward saves its OUTPUT -- which AOTInductor cannot
# lower under strict + dynamic-batch export (pytorch/pytorch#187907).
# On the AD path route through the input-recompute reciprocal; keep
# the plain divide off it so the value path is byte-identical.
data = self.data
if data.requires_grad:
from neml2.types.functions import reciprocal_ad # noqa: PLC0415
return self._rewrap(other * reciprocal_ad(data), sub_batch_ndim=self.sub_batch_ndim)
return self._rewrap(other / data, sub_batch_ndim=self.sub_batch_ndim)
return NotImplemented
# ---- Scalar-only transcendentals / unary ops ----
#
# ``__neg__`` is inherited from PrimitiveTensor — its body is the same.
def __abs__(self) -> Scalar:
return self._rewrap(torch.abs(self.data), sub_batch_ndim=self.sub_batch_ndim)
def __pow__(self, n: float | int) -> Scalar:
return self._rewrap(torch.pow(self.data, n), sub_batch_ndim=self.sub_batch_ndim)
register(Scalar)
# Tell PrimitiveTensor's generic ``_binary`` dispatch which class to recognise as
# Scalar for the (Self × Scalar) interop branch. This avoids an import cycle in
# ``_primitive.py`` (PrimitiveTensor doesn't know about Scalar at class-definition
# time, but every code path that triggers the Scalar branch must have already
# imported Scalar to construct one).
PrimitiveTensor._SCALAR_CLS = Scalar
# Unused import suppressed by being referenced here.
_ = align_sub_batch
NEML2