torch-einops-kit 0.1.0__tar.gz → 0.1.2__tar.gz

{torch_einops_kit-0.1.0 → torch_einops_kit-0.1.2}/PKG-INFO

@@ -1,18 +1,32 @@
 Metadata-Version: 2.4
 Name: torch-einops-kit
-Version: 0.1.0
+Version: 0.1.2
 Summary: Typed tensor-shaping, masking, padding, device-routing, and checkpoint utilities for PyTorch and `einops`. A superset of `lucidrains/torch-einops-utils` with similar utilities from other lucidrains packages. Adds strict typing, extensive tests, and comprehensive docstrings.
+Keywords: artificial intelligence,einops,machine learning,pytorch,torch
 Author: Phil Wang, Hunter Hogan
 Author-email: Phil Wang <lucidrains@gmail.com>, Hunter Hogan <HunterHogan@pm.me>
 License-Expression: CC-BY-NC-4.0
+Classifier: Development Status :: 4 - Beta
+Classifier: Environment :: Console
+Classifier: Environment :: GPU :: NVIDIA CUDA
+Classifier: Intended Audience :: Developers
+Classifier: Intended Audience :: Science/Research
+Classifier: Natural Language :: English
+Classifier: Operating System :: OS Independent
+Classifier: Programming Language :: Python :: 3.10
+Classifier: Programming Language :: Python :: 3.14
+Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
+Classifier: Topic :: Utilities
+Classifier: Typing :: Typed
 Requires-Dist: einops>=0.8.2
-Requires-Dist: packaging>=26.0
 Requires-Dist: torch>=2.10.0
 Requires-Dist: typing-extensions>=4.0.0
 Maintainer: Hunter Hogan
 Maintainer-email: Hunter Hogan <HunterHogan@pm.me>
 Requires-Python: >=3.10
+Project-URL: Documentation, https://context7.com/hunterhogan/torch_einops_kit
 Project-URL: Donate, https://www.patreon.com/integrated
+Project-URL: Download, https://pypi.org/project/torch_einops_kit
 Project-URL: Homepage, https://github.com/hunterhogan/torch_einops_kit
 Project-URL: Issues, https://github.com/hunterhogan/torch_einops_kit/issues
 Project-URL: Repository, https://github.com/hunterhogan/torch_einops_kit.git
@@ -22,8 +36,8 @@ Description-Content-Type: text/markdown
 
 Typed tensor-shaping, masking, padding, device-routing, and checkpoint utilities for PyTorch and `einops`.
 
-[![pip install torch_einops_kit](https://img.shields.io/badge/pip%20install-torch_einops_kit-gray.svg?colorB=3b434b)](https://pypi.org/project/torch_einops_kit/)
-[![uv add torch_einops_kit](https://img.shields.io/badge/uv%20add-torch_einops_kit-gray.svg?colorB=3b434b)](https://pypi.org/project/torch_einops_kit/)
+[![pip install torch-einops-kit](https://img.shields.io/badge/pip_install-torch--einops--kit-gray.svg?labelColor=blue)](https://pypi.org/project/torch-einops-kit/)
+[![uv add torch-einops-kit](https://img.shields.io/badge/uv_add-torch--einops--kit-gray.svg?labelColor=blue)](https://pypi.org/project/torch-einops-kit/)
 
 This repository is a superset of [`lucidrains/torch-einops-utils`](https://github.com/lucidrains/torch-einops-utils). The upstream repository is a compact collection of small utilities that show up repeatedly in lucidrains model repositories. `torch_einops_kit` keeps that role. The main difference is emphasis. This fork adds roughly 6000 lines of tests, typing, and docstrings so the utility layer is easier to trust, easier to search, and easier to apply correctly.
 
@@ -38,7 +52,7 @@ Use `torch_einops_kit` when you want strict typing, a `py.typed` marker, focused
 - Project name: `torch_einops_kit`.
 - Import path: `torch_einops_kit`.
 - Python requirement: `>=3.10`.
-- Runtime dependencies: `torch`, `einops`, `packaging`, and `typing-extensions`.
+- Runtime dependencies: `torch`, `einops`, and `typing-extensions`.
 - Root package exports: helper functions, slicing helpers, rank-alignment helpers, mask helpers, safe concatenation helpers, padding helpers, normalization helpers, and PyTree / `einops` helpers.
 - Submodules with dedicated imports: `torch_einops_kit.device`, `torch_einops_kit.einops`, `torch_einops_kit.save_load`, and `torch_einops_kit.scaleValues`.
 - Typing status: the package ships a `py.typed` marker and the repository uses strict type checking.
@@ -46,10 +60,16 @@ Use `torch_einops_kit` when you want strict typing, a `py.typed` marker, focused
 
 ## Installation
 
-Install from this repository with `uv`:
+### `uv`
 
 ```bash
-uv add git+https://github.com/hunterhogan/torch_einops_kit.git
+uv add torch_einops_kit
+```
+
+### `pip`
+
+```bash
+pip install torch_einops_kit
 ```
 
 ## Import map
@@ -61,9 +81,7 @@ from torch_einops_kit import (
     align_dims_left,
     and_masks,
     broadcast_cat,
-    l2norm,
     lens_to_mask,
-    masked_mean,
     maybe,
     once,
     or_masks,
@@ -108,6 +126,17 @@ from torch_einops_kit.save_load import (
 )
 ```
 
+Import checkpoint decorators from `torch_einops_kit.scaleValues`:
+
+```python
+from torch_einops_kit.scaleValues import (
+    exclusive_cumsum,
+    l2norm,
+    RMSNorm,
+    masked_mean,
+)
+```
+
 ## Quick examples
 
 ### Batch variable-length tensors and build a mask
@@ -265,13 +294,6 @@ These functions add numeric padding values along an existing tensor dimension.
 
 When `pad_lens=True` and `return_lens=True`, the second tensor contains padding widths rather than original lengths.
 
-### Normalization and masked reduction helpers
-
-| Name                                            | Contract                                                                                                                                                                                                                                                                                       |
-| ----------------------------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
-| `l2norm(t)`                                     | Normalizes each vector in `t` to unit length along the last dimension by dividing by its L2 norm. Delegates to `torch.nn.functional.normalize` with `p=2` and `dim=-1`.                                                                                                                        |
-| `masked_mean(t, mask=None, dim=None, eps=1e-5)` | Computes a masked mean. When `mask is None`, the function falls back to `t.mean(...)`. When no masked position is selected and `dim is None`, the function returns zero by summing over the empty selection. When `mask.ndim < t.ndim`, the function right-pads mask rank before broadcasting. |
-
 ### PyTree helpers
 
 | Name                              | Contract                                                                                                                                                               |
@@ -281,10 +303,11 @@ When `pad_lens=True` and `return_lens=True`, the second tensor contains padding
 
 ## `scaleValues` submodule reference
 
-The `torch_einops_kit.scaleValues` submodule contains vector normalization, masked mean computation, and the `RMSNorm` layer. `l2norm` and `masked_mean` are also re-exported from the package root.
+The `torch_einops_kit.scaleValues` submodule contains exclusive prefix sums, vector normalization, masked mean computation, and the `RMSNorm` layer.
 
 | Name                                            | Contract                                                                                                                                                                                                                                                                                       |
 | ----------------------------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| `exclusive_cumsum(t, dim=-1)`                   | Computes the exclusive prefix sum of `t` along `dim`. Each output position receives the sum of all elements that strictly precede it along `dim`. The element at index zero is always zero.                                                                                                    |
 | `l2norm(t)`                                     | Normalizes each vector in `t` to unit length along the last dimension. Delegates to `torch.nn.functional.normalize` with `p=2` and `dim=-1`.                                                                                                                                                   |
 | `masked_mean(t, mask=None, dim=None, eps=1e-5)` | Computes a masked mean. When `mask is None`, the function falls back to `t.mean(...)`. When no masked position is selected and `dim is None`, the function returns zero by summing over the empty selection. When `mask.ndim < t.ndim`, the function right-pads mask rank before broadcasting. |
 | `RMSNorm(dim)`                                  | `torch.nn.Module` that normalizes the last feature axis to unit length, multiplies by `√dim`, and applies a learned per-feature `gamma` parameter. Use as a pre-normalization layer before attention, feedforward, or linear projection sublayers in transformer-style modules.                |

{torch_einops_kit-0.1.0 → torch_einops_kit-0.1.2}/README.md

@@ -2,8 +2,8 @@
 
 Typed tensor-shaping, masking, padding, device-routing, and checkpoint utilities for PyTorch and `einops`.
 
-[![pip install torch_einops_kit](https://img.shields.io/badge/pip%20install-torch_einops_kit-gray.svg?colorB=3b434b)](https://pypi.org/project/torch_einops_kit/)
-[![uv add torch_einops_kit](https://img.shields.io/badge/uv%20add-torch_einops_kit-gray.svg?colorB=3b434b)](https://pypi.org/project/torch_einops_kit/)
+[![pip install torch-einops-kit](https://img.shields.io/badge/pip_install-torch--einops--kit-gray.svg?labelColor=blue)](https://pypi.org/project/torch-einops-kit/)
+[![uv add torch-einops-kit](https://img.shields.io/badge/uv_add-torch--einops--kit-gray.svg?labelColor=blue)](https://pypi.org/project/torch-einops-kit/)
 
 This repository is a superset of [`lucidrains/torch-einops-utils`](https://github.com/lucidrains/torch-einops-utils). The upstream repository is a compact collection of small utilities that show up repeatedly in lucidrains model repositories. `torch_einops_kit` keeps that role. The main difference is emphasis. This fork adds roughly 6000 lines of tests, typing, and docstrings so the utility layer is easier to trust, easier to search, and easier to apply correctly.
 
@@ -18,7 +18,7 @@ Use `torch_einops_kit` when you want strict typing, a `py.typed` marker, focused
 - Project name: `torch_einops_kit`.
 - Import path: `torch_einops_kit`.
 - Python requirement: `>=3.10`.
-- Runtime dependencies: `torch`, `einops`, `packaging`, and `typing-extensions`.
+- Runtime dependencies: `torch`, `einops`, and `typing-extensions`.
 - Root package exports: helper functions, slicing helpers, rank-alignment helpers, mask helpers, safe concatenation helpers, padding helpers, normalization helpers, and PyTree / `einops` helpers.
 - Submodules with dedicated imports: `torch_einops_kit.device`, `torch_einops_kit.einops`, `torch_einops_kit.save_load`, and `torch_einops_kit.scaleValues`.
 - Typing status: the package ships a `py.typed` marker and the repository uses strict type checking.
@@ -26,10 +26,16 @@ Use `torch_einops_kit` when you want strict typing, a `py.typed` marker, focused
 
 ## Installation
 
-Install from this repository with `uv`:
+### `uv`
 
 ```bash
-uv add git+https://github.com/hunterhogan/torch_einops_kit.git
+uv add torch_einops_kit
+```
+
+### `pip`
+
+```bash
+pip install torch_einops_kit
 ```
 
 ## Import map
@@ -41,9 +47,7 @@ from torch_einops_kit import (
     align_dims_left,
     and_masks,
     broadcast_cat,
-    l2norm,
     lens_to_mask,
-    masked_mean,
     maybe,
     once,
     or_masks,
@@ -88,6 +92,17 @@ from torch_einops_kit.save_load import (
 )
 ```
 
+Import checkpoint decorators from `torch_einops_kit.scaleValues`:
+
+```python
+from torch_einops_kit.scaleValues import (
+    exclusive_cumsum,
+    l2norm,
+    RMSNorm,
+    masked_mean,
+)
+```
+
 ## Quick examples
 
 ### Batch variable-length tensors and build a mask
@@ -245,13 +260,6 @@ These functions add numeric padding values along an existing tensor dimension.
 
 When `pad_lens=True` and `return_lens=True`, the second tensor contains padding widths rather than original lengths.
 
-### Normalization and masked reduction helpers
-
-| Name                                            | Contract                                                                                                                                                                                                                                                                                       |
-| ----------------------------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
-| `l2norm(t)`                                     | Normalizes each vector in `t` to unit length along the last dimension by dividing by its L2 norm. Delegates to `torch.nn.functional.normalize` with `p=2` and `dim=-1`.                                                                                                                        |
-| `masked_mean(t, mask=None, dim=None, eps=1e-5)` | Computes a masked mean. When `mask is None`, the function falls back to `t.mean(...)`. When no masked position is selected and `dim is None`, the function returns zero by summing over the empty selection. When `mask.ndim < t.ndim`, the function right-pads mask rank before broadcasting. |
-
 ### PyTree helpers
 
 | Name                              | Contract                                                                                                                                                               |
@@ -261,10 +269,11 @@ When `pad_lens=True` and `return_lens=True`, the second tensor contains padding
 
 ## `scaleValues` submodule reference
 
-The `torch_einops_kit.scaleValues` submodule contains vector normalization, masked mean computation, and the `RMSNorm` layer. `l2norm` and `masked_mean` are also re-exported from the package root.
+The `torch_einops_kit.scaleValues` submodule contains exclusive prefix sums, vector normalization, masked mean computation, and the `RMSNorm` layer.
 
 | Name                                            | Contract                                                                                                                                                                                                                                                                                       |
 | ----------------------------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| `exclusive_cumsum(t, dim=-1)`                   | Computes the exclusive prefix sum of `t` along `dim`. Each output position receives the sum of all elements that strictly precede it along `dim`. The element at index zero is always zero.                                                                                                    |
 | `l2norm(t)`                                     | Normalizes each vector in `t` to unit length along the last dimension. Delegates to `torch.nn.functional.normalize` with `p=2` and `dim=-1`.                                                                                                                                                   |
 | `masked_mean(t, mask=None, dim=None, eps=1e-5)` | Computes a masked mean. When `mask is None`, the function falls back to `t.mean(...)`. When no masked position is selected and `dim is None`, the function returns zero by summing over the empty selection. When `mask.ndim < t.ndim`, the function right-pads mask rank before broadcasting. |
 | `RMSNorm(dim)`                                  | `torch.nn.Module` that normalizes the last feature axis to unit length, multiplies by `√dim`, and applies a learned per-feature `gamma` parameter. Use as a pre-normalization layer before attention, feedforward, or linear projection sublayers in transformer-style modules.                |

{torch_einops_kit-0.1.0 → torch_einops_kit-0.1.2}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "torch-einops-kit"
-version = "0.1.0"
+version = "0.1.2"
 description = "Typed tensor-shaping, masking, padding, device-routing, and checkpoint utilities for PyTorch and `einops`. A superset of `lucidrains/torch-einops-utils` with similar utilities from other lucidrains packages. Adds strict typing, extensive tests, and comprehensive docstrings."
 readme = "README.md"
 requires-python = ">=3.10"
@@ -10,15 +10,37 @@ authors = [
   { name = "Hunter Hogan", email = "HunterHogan@pm.me" }
 ]
 maintainers = [{ name = "Hunter Hogan", email = "HunterHogan@pm.me" }]
+keywords = [
+  "artificial intelligence",
+  "einops",
+  "machine learning",
+  "pytorch",
+  "torch",
+]
+classifiers = [
+  "Development Status :: 4 - Beta",
+  "Environment :: Console",
+  "Environment :: GPU :: NVIDIA CUDA",
+  "Intended Audience :: Developers",
+  "Intended Audience :: Science/Research",
+  "Natural Language :: English",
+  "Operating System :: OS Independent",
+  "Programming Language :: Python :: 3.10",
+  "Programming Language :: Python :: 3.14",
+  "Topic :: Scientific/Engineering :: Artificial Intelligence",
+  "Topic :: Utilities",
+  "Typing :: Typed",
+]
 dependencies = [
   "einops>=0.8.2",
-  "packaging>=26.0",
   "torch>=2.10.0",
   "typing-extensions>=4.0.0",
 ]
 
 [project.urls]
+Documentation = "https://context7.com/hunterhogan/torch_einops_kit"
 Donate = "https://www.patreon.com/integrated"
+Download = "https://pypi.org/project/torch_einops_kit"
 Homepage = "https://github.com/hunterhogan/torch_einops_kit"
 Issues = "https://github.com/hunterhogan/torch_einops_kit/issues"
 Repository = "https://github.com/hunterhogan/torch_einops_kit.git"

{torch_einops_kit-0.1.0 → torch_einops_kit-0.1.2}/src/torch_einops_kit/__init__.py

@@ -92,10 +92,6 @@ pad_sequence_and_cat
 
 Utilities
 ---------
-l2norm
-	Normalize `Tensor` vectors to unit length along the last dimension.
-masked_mean
-	Compute a mean over positions selected by a boolean mask.
 tree_flatten_with_inverse
 	Flatten a PyTree and return an inverse reconstruction function.
 tree_map_tensor
@@ -110,7 +106,7 @@ einops
 save_load
 	Decorate `torch.nn.Module` subclasses with checkpoint save, load, and reconstruction helpers.
 scaleValues
-	Normalize feature vectors and compute masked means.
+	Compute exclusive prefix sums, normalize feature vectors, and compute masked means.
 """
 # isort: split
 from torch_einops_kit._semiotics import decreasing as decreasing, zeroIndexed as zeroIndexed

{torch_einops_kit-0.1.0 → torch_einops_kit-0.1.2}/src/torch_einops_kit/_types.py

@@ -192,8 +192,8 @@ class DehydratedCheckpoint(TypedDict):
 		`Module` instances with `DehydratedTorchNNModule` [9] reconstruction records.
 	version : str | None
 		An optional version string written at save time. When both the stored version and the
-		`save_load` `version` argument are set and differ under `packaging.version.parse` [10], the
-		generated load method prints a notice but still restores model state.
+		`save_load` `version` argument are different, the generated load method prints a notice but
+		still restores model state.
 
 	See Also
 	--------
@@ -220,9 +220,6 @@ class DehydratedCheckpoint(TypedDict):
 	[8] torch_einops_kit.save_load.dehydrate_config
 
 	[9] torch_einops_kit.DehydratedTorchNNModule
-
-	[10] packaging.version - packaging documentation
-		https://packaging.pypa.io/en/stable/version.html
 	"""
 	model: dict[str, Tensor]
 	config: bytes

{torch_einops_kit-0.1.0 → torch_einops_kit-0.1.2}/src/torch_einops_kit/save_load.py

@@ -20,7 +20,6 @@ from __future__ import annotations
 
 from collections.abc import Callable
 from functools import wraps
-from packaging import version as packaging_version
 from pathlib import Path
 from torch.nn import Module
 from torch_einops_kit import (
@@ -299,13 +298,12 @@ def save_load(
 	-----
 	The generated save method writes a dictionary that is compatible with `torch.save` [5], and the
 	generated load paths read the dictionary with `torch.load` [6] on CPU before restoring state.
-	When `version` and the stored checkpoint version both exist and differ under
-	`packaging.version.parse` [7], the generated load method prints a notice but still restores the
-	stored model state.
+	When `version` and the stored checkpoint version are different, the generated load method prints
+	a notice but still restores the stored model state.
 
 	Examples
 	--------
-	From `tests.test_save_load.test_init_and_load` [8]:
+	From `tests.test_save_load.test_init_and_load` [7]:
 
 		```python
 		from pathlib import Path
@@ -327,7 +325,7 @@ def save_load(
 		restored_model = SimpleNet.init_and_load(str(path))
 		```
 
-	From `tests.test_save_load_extended` [9]:
+	From `tests.test_save_load_extended` [8]:
 
 		```python
 		import torch
@@ -367,11 +365,9 @@ def save_load(
 		https://pytorch.org/docs/stable/generated/torch.save.html
 	[6] torch.load - PyTorch documentation
 		https://pytorch.org/docs/stable/generated/torch.load.html
-	[7] packaging.version - packaging documentation
-		https://packaging.pypa.io/en/stable/version.html
-	[8] tests.test_save_load.test_init_and_load
+	[7] tests.test_save_load.test_init_and_load
 
-	[9] tests.test_save_load_extended.test_save_load_supports_custom_method_names_and_config_storage
+	[8] tests.test_save_load_extended.test_save_load_supports_custom_method_names_and_config_storage
 	"""
 	def _save_load(klass: type[TorchNNModule]) -> type[TorchNNModule]:
 		if not issubclass(klass, Module):
@@ -435,10 +431,8 @@ def save_load(
 			"""Restore model state from a checkpoint file.
 
 			You can use this method to load parameter values into an already-constructed decorated
-			module instance. The method reads the checkpoint via `torch.load` [1] on CPU, emits a
-			`UserWarning` when the stored version and the decoration-time version both exist and
-			differ under `packaging.version.parse` [2], and then restores parameter values via
-			`load_state_dict`.
+			module instance. The method reads the checkpoint via `torch.load` [1] on CPU, then
+			restores parameter values via `load_state_dict`.
 
 			Parameters
 			----------
@@ -460,16 +454,13 @@ def save_load(
 			Warns
 			-----
 			UserWarning
-				Emitted when the checkpoint's stored version and the decoration-time version both
-				exist and differ under `packaging.version.parse` [2].
+				Emitted when the checkpoint's stored version and the decoration-time version are
+				different [2].
 
 			References
 			----------
 			[1] torch.load - PyTorch documentation
 				https://pytorch.org/docs/stable/generated/torch.load.html
-
-			[2] packaging.version - packaging documentation
-				https://packaging.pypa.io/en/stable/version.html
 			"""
 			path = Path(path)
 			if not path.exists():
@@ -478,7 +469,7 @@ def save_load(
 
 			pkg: DehydratedCheckpoint = torch.load(str(path), map_location = 'cpu')
 
-			if exists(version) and exists(pkg['version']) and packaging_version.parse(version) != packaging_version.parse(pkg['version']):
+			if exists(version) and exists(pkg['version']) and (version != pkg['version']):
 				message: str = f'I received a checkpoint saved at version `{pkg["version"]}`, but the current package version is `{version}`.'
 				warnings.warn(message, UserWarning, stacklevel=2)
 
@@ -537,7 +528,7 @@ def save_load(
 		# set decorated init as well as save, load, and init_and_load
 
 		klass.__init__ = __init__  # ty:ignore[invalid-assignment]
-# TODO figure out how to use something like `wraps` to get the signature and docstring are public.
+# TODO figure out how to use something like `wraps` so the signature and docstring are public.
 		setattr(klass, save_method_name, _save)
 		setattr(klass, load_method_name, _load)
 		setattr(klass, init_and_load_classmethod_name, _init_and_load_from)

{torch_einops_kit-0.1.0 → torch_einops_kit-0.1.2}/src/torch_einops_kit/scaleValues.py

@@ -1,12 +1,14 @@
-"""Provide vector normalization, masked mean computation, and learned RMS normalization.
+"""Provide exclusive prefix sums, vector normalization, masked mean computation, and learned RMS normalization.
 
-You can use this module to normalize feature vectors to unit length, compute masked mean
-reductions over selected tensor positions, and apply learned root-mean-square normalization
-to transformer and neural network feature channels.
+You can use this module to compute exclusive prefix sums of tensors, normalize feature vectors
+to unit length, compute masked mean reductions over selected tensor positions, and apply learned
+root-mean-square normalization to transformer and neural network feature channels.
 
 Contents
 --------
 Functions
+	exclusive_cumsum
+		Compute the exclusive prefix sum of a `Tensor` along a dimension.
 	l2norm
 		Normalize `Tensor` vectors to unit length along the last dimension.
 	masked_mean
@@ -24,6 +26,47 @@ from torch_einops_kit import exists, pad_right_ndim
 import torch
 import torch.nn.functional as F
 
+def exclusive_cumsum(t: Tensor, dim: int = -1) -> Tensor:
+	"""Compute the exclusive prefix sum of `Tensor` `t` along its dimension `dim`.
+
+	You can use `exclusive_cumsum` to produce a shifted cumulative sum where each output
+	position accumulates only the elements that strictly precede it along `dim`. The element
+	at index zero is always zero. This operation is useful for converting absolute segment
+	lengths into starting offsets.
+
+	Parameters
+	----------
+	t : Tensor
+		Input `Tensor` to reduce.
+	dim : int = -1
+		Dimension along which to compute the exclusive prefix sum. Negative values index
+		from the last axis.
+
+	Returns
+	-------
+	exclusivePrefixSum : Tensor
+		`Tensor` with the same shape as `t` where `exclusivePrefixSum[..., i, ...]` equals
+		the sum of all elements of `t` with index strictly less than `i` along `dim`.
+
+	Mathematical Basis
+	------------------
+	Let t = [t₀, t₁, ..., tₙ₋₁] be the elements of `t` along `dim`. The exclusive prefix
+	sum S is defined element-wise as:
+
+		S[i] = Σⱼ₌₀^(i−1) tⱼ,  with S[0] = 0.
+
+	Equivalently, S equals the inclusive cumulative sum minus `t`:
+
+		S = cumsum(t, dim) − t.
+
+	References
+	----------
+	[1] torch.Tensor.cumsum - PyTorch documentation
+		https://pytorch.org/docs/stable/generated/torch.Tensor.cumsum.html
+	"""
+	return t.cumsum(dim = dim) - t
+
+
 def l2norm(t: Tensor) -> Tensor:
 	"""Normalize `Tensor` vectors to unit length.