torch_to_nnef.utils

Classes:

Name	Description
ReactiveNamedItemDict	Named items ordered Dict data structure.
SemanticVersion	SemVer 2.0 compatible version class.
T2NExtra	Special extra names used in T2N framework.

Functions:

Name	Description
blank_from_init	Create a blank instance of cls by bypassing init.
cache	LRU cache helper that avoid pylint complains.
cd	Context manager for changing the current working directory.
check_torch_ecosystem	Check that torch, torchaudio and torchvision versions are compatible.
dedup_list	Remove duplicates from list while preserving order.
ensure_tuple_io	Normalize inputs/outputs into a tuple.
flatten_dict	Flatten a nested dictionary.
flatten_dict_tuple_or_list	Flatten dict/list/tuple recursively, return types, indexes and values.
fullname	Full class name with module path from an object.
init_empty_weights	A context manager under which models init with meta device.
init_on_device	Context manager under which models are init on the specified device.
normalize_cli_list_option	Normalize repeated/CSV CLI options into a list of unique strings.
require_extra	Lazily import an optional dependency and raise a error if missing.
torch_safe_load	torch.load that avoids arbitrary code execution by default.
torch_version	Semantic version for torch.

ReactiveNamedItemDict

ReactiveNamedItemDict(items: Optional[Iterable[NamedItem]] = None)

Named items ordered Dict data structure.

Ensure that 'NO' 2 items are inserted with same 'name' attribute and maintains fast name update and with some additive colision protections.

Warning! only aimed at NamedItem subclass.

Expose a 'list' like interface. (with limited index access)

Example

from dataclasses import dataclass @dataclass ... class DummyItem(NamedItem): ... name: str ... namespace = ReactiveNamedItemDict() item = DummyItem("hello") for i in "abc": ... namespace.append(DummyItem(i)) namespace.append(item) try: ... namespace.append(DummyItem("a")) ... assert False ... except T2NErrorDataNodeValue: ... pass item.name = "world" namespace.append(DummyItem("hello"))

Methods:

Name	Description
append	Append item to ordered set.

append

append(item: NamedItem)

Append item to ordered set.

WARNING: This is crucial that all added items use this function as it set the hook to listen to name changes

SemanticVersion

SemanticVersion(major, minor, patch, prerelease=None, build=None)

SemVer 2.0 compatible version class.

Supports

1.2.3 1.2.3-alpha 1.2.3-alpha.1 1.2.3-rc.1 1.2.3+build.5 (build metadata ignored in ordering)

Allows symmetric comparison with strings

"1.2.0" < SemanticVersion.from_str("1.3.0")

Example

version = SemanticVersion.from_str("1.2.13") "1.2.12" < version < "1.2.14" True "1.3.12" < version False version == "1.2.13" True

Methods:

Name	Description
to_str	Return canonical SemVer string.

to_str

to_str() -> str

Return canonical SemVer string.

T2NExtra

Bases: str, Enum

Special extra names used in T2N framework.

blank_from_init

blank_from_init(cls)

Create a blank instance of cls by bypassing init.

cache

cache(func: Callable[..., C]) -> C

LRU cache helper that avoid pylint complains.

cd

cd(path)

Context manager for changing the current working directory.

check_torch_ecosystem

check_torch_ecosystem()

Check that torch, torchaudio and torchvision versions are compatible.

This is a common source of runtime errors, so we proactively check and raise a clear error message with instructions if we detect a mismatch.

(avoid cryptic symbol not found errors that can occur missmatched versions)

dedup_list

dedup_list(lst: List[Any]) -> T.List[T.Any]

Remove duplicates from list while preserving order.

ensure_tuple_io

ensure_tuple_io(value: Any) -> T.Tuple[T.Any, ...]

Normalize inputs/outputs into a tuple.

Behavior: - If already a tuple, return as-is. - If a list or other finite sequence, return tuple(value). - If a single Tensor, number, bool, or mapping-like (has items and getitem), wrap into a 1-tuple. - Otherwise, return (value,).

flatten_dict

flatten_dict(d: MutableMapping, parent_key: str = '', sep: str = '.') -> MutableMapping

Flatten a nested dictionary.

flatten_dict_tuple_or_list

flatten_dict_tuple_or_list(obj: Any, collected_types: Optional[List[Type]] = None, collected_idxes: Optional[List[int]] = None, current_idx: int = 0) -> T.Tuple[T.Tuple[T.Tuple[T.Type, ...], T.Tuple[T.Union[int, str], ...], T.Any], ...]

Flatten dict/list/tuple recursively, return types, indexes and values.

Flatten happen in depth first search order

Parameters:

Name	Type	Description	Default
obj	Any	dict/tuple/list or anything else (structure can be arbitrary deep) this contains N number of element non dict/list/tuple	required
collected_types	Optional[List[Type]]	do not set	None
collected_idxes	Optional[List[int]]	do not set	None
current_idx	int	do not set	0

Returns:

Type	Description
Tuple[Tuple[Tuple[Type, ...], Tuple[Union[int, str], ...], Any], ...]	tuple of N tuples each containing a tuple of: types, indexes and the element

Example

If initial obj=[{"a": 1, "b": 3}] it will output: ( ((list, dict), (0, "a"), 1), ((list, dict), (0, "b"), 3), )

fullname

fullname(o) -> str

Full class name with module path from an object.

init_empty_weights

init_empty_weights(include_buffers: Optional[bool] = None) -> T.Iterator[None]

A context manager under which models init with meta device.

Borrowed from accelerate

A context manager under which models are initialized with all parameters on the meta device, therefore creating an empty model. Useful when just initializing the model would blow the available RAM.

Parameters:

Name	Type	Description	Default
include_buffers	Optional[bool]	Whether or not to also put all buffers on the meta device while initializing.	None

Returns:

Type	Description
Iterator[None]	(None) Just a context manager

Example:

import torch.nn as nn
from  import init_empty_weights

# Initialize a model with 100 billions parameters in no time and
# without using any RAM.
with init_empty_weights():
    tst = nn.Sequential(*[nn.Linear(10000, 10000) for _ in range(1000)])

Any model created under this context manager has no weights. As such you can't do something like model.to(some_device) with it. To load weights inside an empty model, see [load_checkpoint_and_dispatch]. Make sure to overwrite the default device_map param for [load_checkpoint_and_dispatch], otherwise dispatch is not called.

init_on_device

init_on_device(device: device, include_buffers: Optional[bool] = None) -> T.Iterator[None]

Context manager under which models are init on the specified device.

Borrowed from accelerate

Parameters:

Name	Type	Description	Default
device	device	Device to initialize all parameters on.	required
include_buffers	Optional[bool]	Whether or not to also put all buffers on the meta device while initializing.	None

Example:

import torch.nn as nn
from accelerate import init_on_device

with init_on_device(device=torch.device("cuda")):
    tst = nn.Linear(100, 100)  # on `cuda` device

normalize_cli_list_option

normalize_cli_list_option(values: Optional[Iterable[Any]]) -> T.Optional[T.List[str]]

Normalize repeated/CSV CLI options into a list of unique strings.

Accepts values from argparse patterns like action="append" and also tolerates a single string. Splits on commas, strips whitespace, removes empty entries, and de-duplicates while preserving order. Returns None if the input is falsy.

require_extra

require_extra(*, extra: T2NExtra, module: str) -> T.Any

Lazily import an optional dependency and raise a error if missing.

torch_safe_load

torch_safe_load(f, **kwargs)

torch.load that avoids arbitrary code execution by default.

Prefers weights_only=True (loads tensors and primitives only). A file that genuinely needs full unpickling (a pickled nn.Module or a custom tensor subclass) falls back to weights_only=False with a warning, since that path executes pickled code: only load files you trust. On torch < 1.13 (no weights_only support) it behaves like a plain torch.load.

torch_version

torch_version() -> SemanticVersion

Semantic version for torch.