py-sadl 1.0.2__py3-none-any.whl

py_sadl-1.0.2.dist-info/METADATA

@@ -0,0 +1,338 @@
+Metadata-Version: 2.4
+Name: py-sadl
+Version: 1.0.2
+Summary: Simple Autograd Deep Learning: A minimal, readable deep learning framework built on NumPy
+Project-URL: Homepage, https://github.com/timcares/sadl
+Project-URL: Documentation, https://github.com/timcares/sadl#readme
+Project-URL: Repository, https://github.com/timcares/sadl
+Project-URL: Issues, https://github.com/timcares/sadl/issues
+Author: Tim Cares
+License: MIT
+License-File: LICENSE
+Keywords: autograd,automatic-differentiation,cupy,deep-learning,educational,machine-learning,neural-networks,numpy
+Classifier: Development Status :: 4 - Beta
+Classifier: Environment :: GPU :: NVIDIA CUDA
+Classifier: Intended Audience :: Developers
+Classifier: Intended Audience :: Education
+Classifier: Intended Audience :: Science/Research
+Classifier: License :: OSI Approved :: MIT License
+Classifier: Natural Language :: English
+Classifier: Operating System :: OS Independent
+Classifier: Programming Language :: Python :: 3
+Classifier: Programming Language :: Python :: 3.13
+Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
+Classifier: Topic :: Scientific/Engineering :: Mathematics
+Classifier: Typing :: Typed
+Requires-Python: >=3.13
+Requires-Dist: numpy>=2.4.0
+Provides-Extra: dev
+Requires-Dist: commitizen>=4.0.0; extra == 'dev'
+Requires-Dist: ipykernel>=7.1.0; extra == 'dev'
+Requires-Dist: ipywidgets>=8.1.8; extra == 'dev'
+Requires-Dist: jupyter>=1.1.1; extra == 'dev'
+Requires-Dist: matplotlib>=3.10.8; extra == 'dev'
+Requires-Dist: mypy>=1.13.0; extra == 'dev'
+Requires-Dist: notebook>=7.5.3; extra == 'dev'
+Requires-Dist: pre-commit>=4.0.0; extra == 'dev'
+Requires-Dist: pytest-cov>=6.0.0; extra == 'dev'
+Requires-Dist: pytest>=8.3.0; extra == 'dev'
+Requires-Dist: python-semantic-release>=9.0.0; extra == 'dev'
+Requires-Dist: ruff>=0.8.0; extra == 'dev'
+Provides-Extra: gpu
+Requires-Dist: cupy-cuda12x>=13.0.0; extra == 'gpu'
+Provides-Extra: gpu-cuda11
+Requires-Dist: cupy-cuda11x>=13.0.0; extra == 'gpu-cuda11'
+Description-Content-Type: text/markdown
+
+<p align="center">
+  <img src="assets/sadl_icon_light.png" alt="SADL Logo" width="200">
+</p>
+
+<h1 align="center">SADL: Simple Autograd Deep Learning</h1>
+
+<p align="center">
+  A minimal, readable deep learning framework built on NumPy and CuPy.<br>
+  Automatic differentiation, neural network primitives, and optimization in ~2000 lines of Python.
+</p>
+
+## Installation
+
+Using [uv](https://docs.astral.sh/uv/) for installation is recommended.
+
+(I had to name the pypi project `py-sadl` instead of `sadl`, because `sadl` was too similar to an existing project.)
+
+```bash
+# Install with uv (recommended)
+uv add py-sadl
+
+# With GPU support (CUDA 12.x)
+uv add py-sadl --extra gpu
+
+# With GPU support (CUDA 11.x)
+uv add py-sadl --extra gpu-cuda11
+```
+
+Alternatively, using pip:
+
+```bash
+# Install with pip
+pip install py-sadl
+
+# With GPU support
+pip install "py-sadl[gpu]"
+```
+
+## Quick Start
+
+```python
+import sadl
+
+# Create tensors
+x = sadl.tensor([[1.0, 2.0], [3.0, 4.0]], requires_grad=True)
+
+# Build a model
+model = sadl.Mlp([
+    sadl.Linear(dim_in=2, dim_out=4),
+    sadl.ReLU(),
+    sadl.Linear(dim_in=4, dim_out=1),
+])
+
+# Forward pass
+output = model(x)
+loss = output.sum()
+
+# Backward pass and optimization
+optimizer = sadl.SGD(list(model.parameters), lr=0.01)
+optimizer.backward(loss)
+optimizer.step()
+optimizer.zero_grad()
+```
+
+## Motivation
+
+Modern deep learning frameworks like PyTorch and TensorFlow are powerful but complex. Their codebases span millions of lines, making it difficult to understand how automatic differentiation and neural network training actually work at a fundamental level.
+
+SADL addresses this by providing a complete, functional deep learning framework that remains small enough to read and understand in its entirety. Every component, from tensor operations to backpropagation, is implemented transparently using standard NumPy operations.
+
+The goal is not to replace production frameworks, but to serve as an educational resource and a foundation for experimentation. Researchers and engineers can trace exactly how gradients flow through computations without navigating layers of abstraction.
+
+## Related Projects
+
+SADL joins a family of educational and minimal deep learning frameworks that have made autodiff more accessible:
+
+**[micrograd](https://github.com/karpathy/micrograd)** by Andrej Karpathy is an elegant, minimal autograd engine operating on scalar values. In roughly 150 lines of code, it demonstrates the core concepts of backpropagation with remarkable clarity. micrograd is an excellent starting point for understanding how gradients flow through computations.
+
+**[tinygrad](https://github.com/tinygrad/tinygrad)** by George Hotz takes a different approach, building a fully-featured deep learning framework with a focus on simplicity and hardware portability. tinygrad supports multiple backends and has grown into a serious alternative for running models on diverse hardware.
+
+SADL takes inspiration from both projects while pursuing its own path: building directly on NumPy's ndarray infrastructure. By subclassing `numpy.ndarray` and intercepting operations via `__array_ufunc__` and `__array_function__`, SADL achieves autograd without introducing a new tensor abstraction. This means existing NumPy code works unchanged, and the mental model stays close to the numerical computing patterns that researchers already know.
+
+## Design Principles
+
+### Build on NumPy
+
+SADL extends `numpy.ndarray` directly rather than wrapping arrays in custom containers. This means:
+
+- All NumPy operations work out of the box
+- No need to learn a new tensor API
+- Seamless interoperability with the scientific Python ecosystem
+- GPU support through CuPy with zero code changes
+
+### Mathematical Functions as First-Class Citizens
+
+Neural network layers are modeled as mathematical functions, matching how they appear in research papers. The `Function` abstract base class enforces a simple contract: implement `__call__` to define the forward pass. This creates a natural bridge between mathematical notation and code.
+
+```python
+class Sigmoid(Function):
+    def __call__(self, x: Tensor) -> Tensor:
+        return 1 / (xp.exp(-x) + 1)
+```
+
+### Explicit Over Implicit
+
+SADL favors explicit behavior over magic:
+
+- Gradients must be explicitly enabled with `requires_grad=True`
+- Parameters are a distinct type that always tracks gradients
+- The computation graph is visible and inspectable
+- Device transfers are explicit operations
+
+### Minimal but Complete
+
+The framework includes only what is necessary for training neural networks:
+
+- Tensor with autograd support
+- Parameter for learnable weights
+- Function base class for layers
+- Optimizer base class with SGD implementation
+- Serialization for model persistence
+
+Additional layers and optimizers can be built on these primitives without modifying core code.
+
+## How Autodiff Works
+
+SADL implements reverse-mode automatic differentiation (backpropagation) using a dynamic computation graph, similar to PyTorch.
+
+### The Computation Graph
+
+In SADL, **Tensors are the computation graph**. There is no separate graph data structure. Each Tensor stores a `src` attribute pointing to the Tensors it was created from. This forms a back-referencing graph where each node knows its parents, but parents do not know their children:
+
+```
+Forward computation:
+
+x ─┐
+   ├─► z ─► loss
+y ─┘
+
+Graph structure (src references):
+
+   loss
+    │
+    ▼
+    z
+   ╱ ╲
+  ▼   ▼
+  x   y
+```
+
+This is intentional. Deep learning frameworks optimize for backward traversal because that is what backpropagation requires. Starting from the loss, we follow `src` references backward through the graph to compute gradients. Forward references (parent to child) are unnecessary and would only consume memory.
+
+### Forward Pass: Building the Graph
+
+When operations are performed on Tensors with `requires_grad=True`, the graph builds itself automatically:
+
+1. The `Tensor` class overrides `__array_ufunc__` and `__array_function__` to intercept NumPy operations
+2. Each operation creates a new Tensor that stores:
+  - `src`: References to input tensors (the parents in the graph)
+  - `backward_fn`: The gradient function for this operation
+  - `op_ctx`: Any context needed for gradient computation (axis, masks, etc.)
+3. The graph grows dynamically as operations execute
+
+```python
+x = sadl.tensor([1.0, 2.0], requires_grad=True)  # leaf, src = ()
+y = sadl.tensor([3.0, 4.0], requires_grad=True)  # leaf, src = ()
+z = x * y      # z.src = (x, y), z.backward_fn = mul_backward
+loss = z.sum() # loss.src = (z,), loss.backward_fn = sum_backward
+```
+
+A more complex example:
+
+```
+a = tensor(...)      # leaf
+b = tensor(...)      # leaf
+c = tensor(...)      # leaf
+
+d = a + b            # d.src = (a, b)
+e = d * c            # e.src = (d, c)
+f = e.sum()          # f.src = (e,)
+
+Graph (following src backwards from f):
+
+    f
+    │
+    ▼
+    e
+   ╱ ╲
+  ▼   ▼
+  d   c
+ ╱ ╲
+▼   ▼
+a   b
+```
+
+### Backward Pass: Computing Gradients
+
+When `optimizer.backward(loss)` is called:
+
+1. **Topological Sort**: The graph is traversed from the loss tensor to find all nodes, ordered so that each node appears after all nodes that depend on it. This uses an iterative stack-based algorithm to avoid recursion limits on deep graphs.
+2. **Gradient Propagation**: Starting from the loss (seeded with gradient 1.0), each node's `backward_fn` is called with:
+  - The input tensors (`src`)
+  - Which inputs need gradients (`compute_grad`)
+  - The upstream gradient (`grad_out`)
+  - Operation context (`op_ctx`)
+3. **Gradient Accumulation**: Gradients flow backward through the graph. When a tensor is used in multiple operations, gradients are summed.
+4. **Graph Cleanup**: After backpropagation, the graph structure is cleared to free memory. Parameter gradients are retained for the optimizer step.
+
+### Gradient Operations Registry
+
+Each supported operation has a corresponding backward function registered in `grad_ops.py` with metadata (op type inspired by [tinygrad](https://github.com/tinygrad/tinygrad)):
+
+```python
+@register_grad_op(
+    op_type=OpType.ELEMENTWISE,
+    op_inputs=OpInputs.BINARY,
+    forward_names=("mul", "multiply"),
+)
+@broadcastable
+def mul_backward(*inputs, compute_grad, grad_out):
+    x, y = inputs
+    grad_x = y * grad_out if compute_grad[0] else None
+    grad_y = x * grad_out if compute_grad[1] else None
+    return grad_x, grad_y
+```
+
+The `@broadcastable` decorator handles gradient reduction when inputs were broadcast during the forward pass.
+
+### Supported Operations
+
+Unary: `abs`, `negative`, `sqrt`, `square`, `exp`, `log`, `sin`, `cos`
+
+Binary: `add`, `subtract`, `multiply`, `divide`, `power`, `matmul`, `maximum`, `minimum`
+
+Reductions: `sum`, `mean`, `max`, `min`
+
+## Architecture
+
+```
+sadl/
+├── __init__.py     # Public API re-exports
+├── backend.py      # NumPy/CuPy abstraction
+├── disk.py         # Saving and loading data to/from disk
+├── tensor.py       # Tensor, Parameter, serialization
+├── grad_ops.py     # Gradient operation registry
+├── function.py     # Neural network layers
+├── optimizer.py    # Optimizer base class, SGD, backpropagation
+├── ops.py          # Array creation and device utilities
+└── utils.py        # Device transfer utilities
+```
+
+### Key Components
+
+**Tensor**: Subclass of `numpy.ndarray` with additional attributes for autograd. Intercepts NumPy operations to build the computation graph.
+
+**Parameter**: Tensor subclass for learnable weights. Always requires gradients and retains them after backward pass for gradient accumulation.
+
+**Function**: Abstract base class for neural network layers. Provides parameter traversal, device management, and train/inference mode switching.
+
+**Optimizer**: Abstract base class that owns the backward pass. Performs topological sort, gradient computation, and graph cleanup.
+
+**GradOp Registry**: Dictionary mapping operation names to backward functions. New operations can be registered with a decorator.
+
+## Serialization
+
+SADL uses a custom binary format (`.sadl` files) for efficient tensor storage:
+
+- 4-byte magic header for format validation
+- Version byte for forward compatibility
+- Compact encoding of dtype, shape, and raw data
+- Support for single tensors or ordered dictionaries of tensors
+
+## Contributing
+
+See [CONTRIBUTING.md](CONTRIBUTING.md) for development setup, commands, and guidelines.
+
+## Code of Conduct
+
+See [CODE_OF_CONDUCT.md](CODE_OF_CONDUCT.md) for behavior guidelines. The file was created using the [Contributor Covenant](https://www.contributor-covenant.org).
+
+## Future Plans
+
+- Static graph compilation for repeated computations
+- Additional layers and components (convolution, batch normalization, attention)
+- More optimizers (Adam, AdamW, RMSprop)
+- XLA compilation backend for TPU support
+- Automatic mixed precision training
+- Distributed training primitives
+
+## See Also
+
+- `docs/API_REFERENCE.md`: Complete API documentation

py_sadl-1.0.2.dist-info/RECORD

@@ -0,0 +1,13 @@
+sadl/__init__.py,sha256=t-lx9xcwdU9egY5zUXOMjLI-Wz1t9lTzZYJnS5VWlf4,1214
+sadl/backend.py,sha256=G7nZ1V7qm7myfgBSQV7hWIgk4ohVW_SfA_6BvvwvVdw,1060
+sadl/disk.py,sha256=Sy9Fg6EbovTQvawUn-DMO8e_YecqLZXwtTf99K7w4WA,4807
+sadl/function.py,sha256=eJgSAO91KvwNfGiL1-AUZHWiAr6P77P_uvnpV5F8rLw,14196
+sadl/grad_ops.py,sha256=rRJ2HhXzmqdIBfdd_yuKUaZ_f_ZkXQeGOuZsCk9EPho,37549
+sadl/ops.py,sha256=-NptddSsHAredwjcZUiGm50GLSatXYHGcXiS3qfA6Qw,1868
+sadl/optimizer.py,sha256=qPscSJyqbWTNcFUIJDmaVaYFVpuqKVESDp6Dn8wSoPQ,13361
+sadl/tensor.py,sha256=My-GJ1wQxclTbFjobG-hCn2VXQYBLW17NUeDOL6DKi4,18160
+sadl/utils.py,sha256=M_0r0s14w3JdvA5Q_IxeRitXA8W5Fm3-1LUBGJVGMv4,1126
+py_sadl-1.0.2.dist-info/METADATA,sha256=S9cULasg-M7LKX6U17OwpMcNXMiRDigqYf1SICMoONk,13065
+py_sadl-1.0.2.dist-info/WHEEL,sha256=WLgqFyCfm_KASv4WHyYy0P3pM_m7J5L9k2skdKLirC8,87
+py_sadl-1.0.2.dist-info/licenses/LICENSE,sha256=zOSEe9MBxdTcRzOau3ZZlnaBKkhmbLJtskkyONfSeIk,1066
+py_sadl-1.0.2.dist-info/RECORD,,

py_sadl-1.0.2.dist-info/WHEEL

@@ -0,0 +1,4 @@
+Wheel-Version: 1.0
+Generator: hatchling 1.28.0
+Root-Is-Purelib: true
+Tag: py3-none-any

py_sadl-1.0.2.dist-info/licenses/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 Tim Cares
+
+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.

sadl/__init__.py

@@ -0,0 +1,74 @@
+"""SADL: Simple Autograd Deep Learning.
+
+A minimal, readable deep learning framework built on NumPy/CuPy.
+"""
+
+from importlib.metadata import PackageNotFoundError, version
+
+from . import grad_ops
+
+try:
+    __version__ = version("py-sadl")
+except PackageNotFoundError:
+    __version__ = "0.0.0.dev0"  # Fallback for uninstalled package
+from .backend import (
+    BACKEND,
+    TensorDevice,
+    xp,
+)
+from .disk import (
+    load,
+    save,
+)
+from .function import (
+    Function,
+    Linear,
+    Mlp,
+    ReLU,
+    Sigmoid,
+)
+from .ops import (
+    copy_to_device,
+    ones_like,
+    zeros_like,
+)
+from .optimizer import (
+    SGD,
+    Optimizer,
+)
+from .tensor import (
+    Parameter,
+    Tensor,
+    get_current_global_grad_mode,
+    no_grad,
+    no_grad_fn,
+    set_global_grad_mode,
+    tensor,
+)
+
+__all__ = [
+    "BACKEND",
+    "SGD",
+    "Function",
+    "Linear",
+    "Mlp",
+    "Optimizer",
+    "Parameter",
+    "ReLU",
+    "Sigmoid",
+    "Tensor",
+    "TensorDevice",
+    "__version__",
+    "copy_to_device",
+    "get_current_global_grad_mode",
+    "grad_ops",
+    "load",
+    "no_grad",
+    "no_grad_fn",
+    "ones_like",
+    "save",
+    "set_global_grad_mode",
+    "tensor",
+    "xp",
+    "zeros_like",
+]

sadl/backend.py

@@ -0,0 +1,45 @@
+import logging
+from typing import TYPE_CHECKING, Any, Literal
+
+logger = logging.getLogger(__name__)
+
+
+TensorDevice = Literal["cpu"] | int
+
+
+if TYPE_CHECKING:
+    import numpy
+
+    ModuleType = numpy.ndarray[Any, Any] | Any  # numpy or cupy module
+
+
+def _validate_cupy_available() -> None:
+    """Validate that CuPy is available with working CUDA devices.
+
+    Raises:
+        RuntimeError: If CUDA is unavailable or no devices are found.
+    """
+    try:
+        _device_count = xp.cuda.runtime.getDeviceCount()
+    except Exception as exc:
+        raise RuntimeError("Cupy is installed but CUDA is unavailable") from exc
+
+    if _device_count < 1:
+        raise RuntimeError("Cupy is installed but no CUDA devices are available")
+
+
+try:
+    import cupy as xp
+
+    _validate_cupy_available()
+
+    BACKEND = "cupy"
+    logger.debug("Using cupy as backend")
+except (ImportError, RuntimeError):
+    import numpy as xp
+
+    BACKEND = "numpy"
+    logger.warning("Cupy backend unavailable; falling back to numpy (cpu)")
+
+
+__all__ = ["BACKEND", "TensorDevice", "xp"]

sadl/disk.py

@@ -0,0 +1,147 @@
+"""Code for serializing and deserializing data."""
+
+from __future__ import annotations
+
+import struct
+from collections import OrderedDict
+from typing import Any
+
+import numpy as np
+
+from .tensor import Tensor, tensor
+
+_SADL_MAGIC = b"SADL"
+_SADL_VERSION = 1
+
+
+def _dtype_to_str(dtype: Any) -> str:
+    """Convert numpy/cupy dtype to string representation."""
+    return str(np.dtype(dtype).name)
+
+
+def _str_to_dtype(dtype_str: str) -> Any:
+    """Convert string back to numpy dtype."""
+    return np.dtype(dtype_str)
+
+
+def save(data: Tensor | OrderedDict[str, Tensor], file_path: str) -> None:
+    """Save Tensor data to disk using custom binary format.
+
+    Args:
+        data (Tensor | OrderedDict[str, Tensor]): The data to save,
+            can either be a single Tensor or an OrderedDict with strings
+            as keys and Tensors as values.
+        file_path (str): The file path to which to store the data. Must
+            end with ".sadl".
+
+    Raises:
+        ValueError: If an OrderedDict with non-Tensor values is passed.
+        ValueError: If file_path doesn't end with ".sadl".
+    """
+    if not file_path.endswith(".sadl"):
+        raise ValueError('file_path must end with ".sadl"')
+
+    # Normalize to OrderedDict
+    if isinstance(data, Tensor):
+        tensors = OrderedDict([("__single__", data)])
+    else:
+        if not all(isinstance(v, Tensor) for v in data.values()):
+            raise ValueError("If an OrderedDict is passed, all values must be Tensors.")
+        tensors = data
+
+    with open(file_path, "wb") as f:
+        # Write header
+        f.write(_SADL_MAGIC)
+        f.write(struct.pack("<B", _SADL_VERSION))  # uint8 version
+        f.write(struct.pack("<I", len(tensors)))  # uint32 num tensors
+
+        # Write each tensor
+        for key, tensor in tensors.items():
+            # Convert to numpy (CPU) for serialization
+            arr = np.asarray(tensor)
+            # Ensure C-contiguous
+            if not arr.flags["C_CONTIGUOUS"]:
+                arr = np.ascontiguousarray(arr)
+
+            # Key
+            key_bytes = key.encode("utf-8")
+            f.write(struct.pack("<I", len(key_bytes)))  # uint32 key length
+            f.write(key_bytes)
+
+            # Dtype
+            dtype_str = _dtype_to_str(arr.dtype)
+            dtype_bytes = dtype_str.encode("utf-8")
+            f.write(struct.pack("<B", len(dtype_bytes)))  # uint8 dtype length
+            f.write(dtype_bytes)
+
+            # Shape
+            f.write(struct.pack("<B", arr.ndim))  # uint8 ndim
+            f.writelines(struct.pack("<Q", dim) for dim in arr.shape)  # uint64 per dimension
+
+            # Raw data
+            f.write(arr.tobytes())
+
+
+def load(file_path: str) -> Tensor | OrderedDict[str, Tensor]:
+    """Load Tensor data from disk.
+
+    Args:
+        file_path (str): The file path from which to read the data. Must
+            end with ".sadl".
+
+    Raises:
+        ValueError: If file_path doesn't end with ".sadl".
+        ValueError: If file has invalid magic bytes or unsupported version.
+
+    Returns:
+        Tensor | OrderedDict[str, Tensor]: The loaded data. Returns a single
+            Tensor if one was saved, otherwise an OrderedDict.
+    """
+    if not file_path.endswith(".sadl"):
+        raise ValueError('file_path must end with ".sadl"')
+
+    with open(file_path, "rb") as f:
+        # Read and validate header
+        magic = f.read(4)
+        if magic != _SADL_MAGIC:
+            raise ValueError(f"Invalid file format. Expected SADL magic bytes, got {magic!r}")
+
+        version = struct.unpack("<B", f.read(1))[0]
+        if version != _SADL_VERSION:
+            raise ValueError(f"Unsupported version {version}. Expected {_SADL_VERSION}")
+
+        num_tensors = struct.unpack("<I", f.read(4))[0]
+
+        # Read tensors
+        tensors: OrderedDict[str, Tensor] = OrderedDict()
+        for _ in range(num_tensors):
+            # Key
+            key_length = struct.unpack("<I", f.read(4))[0]
+            key = f.read(key_length).decode("utf-8")
+
+            # Dtype
+            dtype_length = struct.unpack("<B", f.read(1))[0]
+            dtype_str = f.read(dtype_length).decode("utf-8")
+            dtype = _str_to_dtype(dtype_str)
+
+            # Shape
+            ndim = struct.unpack("<B", f.read(1))[0]
+            shape = tuple(struct.unpack("<Q", f.read(8))[0] for _ in range(ndim))
+
+            # Data
+            num_bytes = int(np.prod(shape)) * dtype.itemsize
+            data_bytes = f.read(num_bytes)
+            arr = np.frombuffer(data_bytes, dtype=dtype).reshape(shape)
+
+            tensors[key] = tensor(arr)
+
+        # Return single tensor if that's what was saved
+        if len(tensors) == 1 and "__single__" in tensors:
+            return tensors["__single__"]
+        return tensors
+
+
+__all__ = [
+    "load",
+    "save",
+]