onnx 1.16.1__cp310-cp310-win_amd64.whl → 1.17.0__cp310-cp310-win_amd64.whl

onnx/hub.py

@@ -5,14 +5,15 @@
 
 This implements the python client for the ONNX model hub.
 """
+from __future__ import annotations
+
 import hashlib
 import json
 import os
 import sys
-import tarfile
 from io import BytesIO
 from os.path import join
-from typing import IO, Any, Dict, List, Optional, Set, Tuple, cast
+from typing import IO, Any, Dict, List, cast
 from urllib.error import HTTPError
 from urllib.request import urlopen
 
@@ -39,7 +40,7 @@ class ModelInfo:
         opset: The opset version of the model.
     """
 
-    def __init__(self, raw_model_info: Dict[str, Any]) -> None:
+    def __init__(self, raw_model_info: dict[str, Any]) -> None:
         """Initializer.
 
         Args:
@@ -48,17 +49,17 @@ class ModelInfo:
         self.model = cast(str, raw_model_info["model"])
 
         self.model_path = cast(str, raw_model_info["model_path"])
-        self.metadata: Dict[str, Any] = cast(Dict[str, Any], raw_model_info["metadata"])
-        self.model_sha: Optional[str] = None
+        self.metadata: dict[str, Any] = cast(Dict[str, Any], raw_model_info["metadata"])
+        self.model_sha: str | None = None
         if "model_sha" in self.metadata:
             self.model_sha = cast(str, self.metadata["model_sha"])
 
-        self.tags: Set[str] = set()
+        self.tags: set[str] = set()
         if "tags" in self.metadata:
             self.tags = set(cast(List[str], self.metadata["tags"]))
 
         self.opset = cast(int, raw_model_info["opset_version"])
-        self.raw_model_info: Dict[str, Any] = raw_model_info
+        self.raw_model_info: dict[str, Any] = raw_model_info
 
     def __str__(self) -> str:
         return f"ModelInfo(model={self.model}, opset={self.opset}, path={self.model_path}, metadata={self.metadata})"
@@ -86,7 +87,7 @@ def get_dir() -> str:
     return _ONNX_HUB_DIR
 
 
-def _parse_repo_info(repo: str) -> Tuple[str, str, str]:
+def _parse_repo_info(repo: str) -> tuple[str, str, str]:
     """Gets the repo owner, name and ref from a repo specification string."""
     repo_owner = repo.split(":")[0].split("/")[0]
     repo_name = repo.split(":")[0].split("/")[1]
@@ -143,9 +144,9 @@ def _download_file(url: str, file_name: str) -> None:
 
 def list_models(
     repo: str = "onnx/models:main",
-    model: Optional[str] = None,
-    tags: Optional[List[str]] = None,
-) -> List[ModelInfo]:
+    model: str | None = None,
+    tags: list[str] | None = None,
+) -> list[ModelInfo]:
     """Gets the list of model info consistent with a given name and tags
 
     Args:
@@ -164,7 +165,7 @@ def list_models(
     manifest_url = base_url + "ONNX_HUB_MANIFEST.json"
     try:
         with urlopen(manifest_url) as response:
-            manifest: List[ModelInfo] = [
+            manifest: list[ModelInfo] = [
                 ModelInfo(info) for info in json.load(cast(IO[str], response))
             ]
     except HTTPError as e:
@@ -182,7 +183,7 @@ def list_models(
         return matching_models
 
     canonical_tags = {t.lower() for t in tags}
-    matching_info_list: List[ModelInfo] = []
+    matching_info_list: list[ModelInfo] = []
     for m in matching_models:
         model_tags = {t.lower() for t in m.tags}
         if len(canonical_tags.intersection(model_tags)) > 0:
@@ -191,7 +192,7 @@ def list_models(
 
 
 def get_model_info(
-    model: str, repo: str = "onnx/models:main", opset: Optional[int] = None
+    model: str, repo: str = "onnx/models:main", opset: int | None = None
 ) -> ModelInfo:
     """Gets the model info matching the given name and opset.
 
@@ -226,10 +227,10 @@ def get_model_info(
 def load(
     model: str,
     repo: str = "onnx/models:main",
-    opset: Optional[int] = None,
+    opset: int | None = None,
     force_reload: bool = False,
     silent: bool = False,
-) -> Optional[onnx.ModelProto]:
+) -> onnx.ModelProto | None:
     """Downloads a model by name from the onnx model hub.
 
     Args:
@@ -251,9 +252,9 @@ def load(
     selected_model = get_model_info(model, repo, opset)
     local_model_path_arr = selected_model.model_path.split("/")
     if selected_model.model_sha is not None:
-        local_model_path_arr[
-            -1
-        ] = f"{selected_model.model_sha}_{local_model_path_arr[-1]}"
+        local_model_path_arr[-1] = (
+            f"{selected_model.model_sha}_{local_model_path_arr[-1]}"
+        )
     local_model_path = join(_ONNX_HUB_DIR, os.sep.join(local_model_path_arr))
 
     if force_reload or not os.path.exists(local_model_path):
@@ -291,11 +292,12 @@ def load(
 def download_model_with_test_data(
     model: str,
     repo: str = "onnx/models:main",
-    opset: Optional[int] = None,
+    opset: int | None = None,
     force_reload: bool = False,
     silent: bool = False,
-) -> Optional[str]:
+) -> str | None:
     """Downloads a model along with test data by name from the onnx model hub and returns the directory to which the files have been extracted.
+    Users are responsible for making sure the model comes from a trusted source, and the data is safe to be extracted.
 
     Args:
         model: The name of the onnx model in the manifest. This field is
@@ -322,9 +324,9 @@ def download_model_with_test_data(
     model_with_data_sha = selected_model.metadata["model_with_data_sha"]
 
     if model_with_data_sha is not None:
-        local_model_with_data_path_arr[
-            -1
-        ] = f"{model_with_data_sha}_{local_model_with_data_path_arr[-1]}"
+        local_model_with_data_path_arr[-1] = (
+            f"{model_with_data_sha}_{local_model_with_data_path_arr[-1]}"
+        )
     local_model_with_data_path = join(
         _ONNX_HUB_DIR, os.sep.join(local_model_with_data_path_arr)
     )
@@ -361,12 +363,14 @@ def download_model_with_test_data(
                 "download the model from the model hub."
             )
 
-    with tarfile.open(local_model_with_data_path) as model_with_data_zipped:
-        # FIXME: Avoid index manipulation with magic numbers
-        local_model_with_data_dir_path = local_model_with_data_path[
-            0 : len(local_model_with_data_path) - 7
-        ]
-        model_with_data_zipped.extractall(local_model_with_data_dir_path)
+    # FIXME: Avoid index manipulation with magic numbers,
+    # remove ".tar.gz"
+    local_model_with_data_dir_path = local_model_with_data_path[
+        0 : len(local_model_with_data_path) - 7
+    ]
+    onnx.utils._extract_model_safe(
+        local_model_with_data_path, local_model_with_data_dir_path
+    )
     model_with_data_path = (
         local_model_with_data_dir_path
         + "/"
@@ -381,10 +385,10 @@ def load_composite_model(
     preprocessing_model: str,
     network_repo: str = "onnx/models:main",
     preprocessing_repo: str = "onnx/models:main",
-    opset: Optional[int] = None,
+    opset: int | None = None,
     force_reload: bool = False,
     silent: bool = False,
-) -> Optional[onnx.ModelProto]:
+) -> onnx.ModelProto | None:
     """Builds a composite model including data preprocessing by downloading a network and a preprocessing model
     and combine it into a single model
 
@@ -418,9 +422,9 @@ def load_composite_model(
     if network is None:
         raise RuntimeError(f"Could not load the network model: {network_model}")
 
-    all_domains: Set[str] = set()
-    domains_to_version_network: Dict[str, int] = {}
-    domains_to_version_preprocessing: Dict[str, int] = {}
+    all_domains: set[str] = set()
+    domains_to_version_network: dict[str, int] = {}
+    domains_to_version_preprocessing: dict[str, int] = {}
 
     for opset_import_entry in network.opset_import:
         domain = (

onnx/inliner/inliner.cc

@@ -66,7 +66,6 @@ struct OpsetMap : public OpsetMapBase {
   }
 
   bool Add(const google::protobuf::RepeatedPtrField<OperatorSetIdProto>& list) {
-    OpsetMapBase result;
     for (const auto& pair : list) {
       auto domain = NormalizeDomain(pair.domain());
       auto version = pair.version();

onnx/mapping.py

@@ -1,6 +1,7 @@
 # Copyright (c) ONNX Project Contributors
 #
 # SPDX-License-Identifier: Apache-2.0
+from __future__ import annotations
 
 import warnings
 from typing import Any, Dict, NamedTuple, Union, cast
@@ -93,7 +94,7 @@ TENSOR_TYPE_MAP = {
 class DeprecatedWarningDict(dict):  # type: ignore
     def __init__(
         self,
-        dictionary: Dict[int, Union[int, str, np.dtype]],
+        dictionary: dict[int, int | str | np.dtype],
         original_function: str,
         future_function: str = "",
     ) -> None:
@@ -109,7 +110,7 @@ class DeprecatedWarningDict(dict):  # type: ignore
             and self._future_function == other._future_function
         )
 
-    def __getitem__(self, key: Union[int, str, np.dtype]) -> Any:
+    def __getitem__(self, key: int | str | np.dtype) -> Any:
         if not self._future_function:
             warnings.warn(
                 str(

onnx/numpy_helper.py

@@ -1,24 +1,25 @@
 # Copyright (c) ONNX Project Contributors
 #
 # SPDX-License-Identifier: Apache-2.0
-
+from __future__ import annotations
 
 import sys
-from typing import Any, Dict, List, Optional, Sequence, Tuple, Union
+from typing import Any, Sequence
 
 import numpy as np
 
+import onnx._custom_element_types as custom_np_types
 from onnx import MapProto, OptionalProto, SequenceProto, TensorProto, helper, subbyte
 from onnx.external_data_helper import load_external_data_for_tensor, uses_external_data
 
 
-def combine_pairs_to_complex(fa: Sequence[int]) -> List[complex]:
+def combine_pairs_to_complex(fa: Sequence[int]) -> list[complex]:
     return [complex(fa[i * 2], fa[i * 2 + 1]) for i in range(len(fa) // 2)]
 
 
 def bfloat16_to_float32(
-    data: Union[np.int16, np.int32, np.ndarray],
-    dims: Optional[Union[int, Sequence[int]]] = None,
+    data: np.int16 | np.int32 | np.ndarray,
+    dims: int | Sequence[int] | None = None,
 ) -> np.ndarray:
     """Converts ndarray of bf16 (as uint32) to f32 (as uint32).
 
@@ -55,6 +56,7 @@ def _float8e4m3_to_float32_scalar(ival: int, fn: bool, uz: bool) -> np.float32:
         if ival == 127:  # noqa: PLR2004
             return np.float32(np.nan)
 
+    ival = np.uint32(ival)  # type: ignore[assignment]
     expo = (ival & 0x78) >> 3
     mant = ival & 0x07
     sign = ival & 0x80
@@ -86,8 +88,8 @@ _float8e4m3_to_float32 = np.vectorize(
 
 
 def float8e4m3_to_float32(
-    data: Union[np.int16, np.int32, np.ndarray],
-    dims: Optional[Union[int, Sequence[int]]] = None,
+    data: np.int16 | np.int32 | np.ndarray,
+    dims: int | Sequence[int] | None = None,
     fn: bool = True,
     uz: bool = False,
 ) -> np.ndarray:
@@ -133,6 +135,7 @@ def _float8e5m2_to_float32_scalar(ival: int, fn: bool, uz: bool) -> np.float32:
     else:
         raise NotImplementedError("fn and uz must be both False or True.")
 
+    ival = np.uint32(ival)  # type: ignore[assignment]
     expo = (ival & 0x7C) >> 2
     mant = ival & 0x03
     sign = ival & 0x80
@@ -160,8 +163,8 @@ _float8e5m2_to_float32 = np.vectorize(
 
 
 def float8e5m2_to_float32(
-    data: Union[np.int16, np.int32, np.ndarray],
-    dims: Optional[Union[int, Sequence[int]]] = None,
+    data: np.int16 | np.int32 | np.ndarray,
+    dims: int | Sequence[int] | None = None,
     fn: bool = False,
     uz: bool = False,
 ) -> np.ndarray:
@@ -186,8 +189,8 @@ def float8e5m2_to_float32(
 
 
 def unpack_int4(
-    data: Union[np.int32, np.ndarray],
-    dims: Union[int, Sequence[int]],
+    data: np.int32 | np.ndarray,
+    dims: int | Sequence[int],
     signed: bool,
 ) -> np.ndarray:
     """Converts ndarray of int4 (as packed uint8) to f32
@@ -218,7 +221,7 @@ def unpack_int4(
     return res
 
 
-def to_array(tensor: TensorProto, base_dir: str = "") -> np.ndarray:  # noqa: PLR0911
+def _to_array(tensor: TensorProto, base_dir: str = "") -> np.ndarray:  # noqa: PLR0911
     """Converts a tensor def object to a numpy array.
 
     Args:
@@ -332,7 +335,79 @@ def to_array(tensor: TensorProto, base_dir: str = "") -> np.ndarray:  # noqa: PL
     return np.asarray(data, dtype=storage_np_dtype).astype(np_dtype).reshape(dims)
 
 
-def from_array(arr: np.ndarray, name: Optional[str] = None) -> TensorProto:
+def to_array(tensor: TensorProto, base_dir: str = "") -> np.ndarray:
+    """Converts a tensor def object to a numpy array.
+    Supports types defined in :mod:`onnx._custom_element_types`.
+
+    Args:
+        tensor: a TensorProto object.
+        base_dir: if external tensor exists, base_dir can help to find the path to it
+
+    Returns:
+        arr: the converted array.
+    """
+    elem_type = tensor.data_type
+    if elem_type == TensorProto.BFLOAT16:
+        if uses_external_data(tensor):
+            load_external_data_for_tensor(tensor, base_dir)
+        data = tensor.int32_data
+        shape = tuple(tensor.dims)
+        y = np.empty(shape, dtype=custom_np_types.bfloat16).ravel()
+        for i, d in enumerate(data):
+            y[i] = d
+        return y.reshape(shape)
+
+    if elem_type in (
+        TensorProto.FLOAT8E4M3FN,
+        TensorProto.FLOAT8E4M3FNUZ,
+        TensorProto.FLOAT8E5M2,
+        TensorProto.FLOAT8E5M2FNUZ,
+    ):
+        m = {
+            TensorProto.FLOAT8E4M3FN: custom_np_types.float8e4m3fn,
+            TensorProto.FLOAT8E4M3FNUZ: custom_np_types.float8e4m3fnuz,
+            TensorProto.FLOAT8E5M2: custom_np_types.float8e5m2,
+            TensorProto.FLOAT8E5M2FNUZ: custom_np_types.float8e5m2fnuz,
+        }
+
+        if uses_external_data(tensor):
+            load_external_data_for_tensor(tensor, base_dir)
+        if tensor.HasField("raw_data"):
+            data = tensor.raw_data  # type: ignore[assignment]
+        else:
+            data = tensor.int32_data
+        shape = tuple(tensor.dims)
+        y = np.empty(shape, dtype=m[elem_type]).ravel()  # type: ignore[index]
+        for i, d in enumerate(data):
+            y[i] = d
+        return y.reshape(shape)
+
+    if elem_type in (TensorProto.UINT4, TensorProto.INT4):
+        if uses_external_data(tensor):
+            load_external_data_for_tensor(tensor, base_dir)
+        if tensor.HasField("raw_data"):
+            data = tensor.raw_data  # type: ignore[assignment]
+        else:
+            data = tensor.int32_data
+        shape = tuple(tensor.dims)
+        m = {
+            TensorProto.INT4: custom_np_types.int4,
+            TensorProto.UINT4: custom_np_types.uint4,
+        }
+        dtype = m[elem_type]  # type: ignore[index]
+        signed = elem_type == TensorProto.INT4
+        # 2 packed int4 elements must be represented as a single uint8 value.
+        # Therefore, y is np.uint8 (not the dtype to which the int4 maps)
+        y = np.empty(len(data), dtype=np.uint8).ravel()  # type: ignore[assignment]
+        for i, d in enumerate(data):
+            y[i] = d
+        unpacked_data = unpack_int4(y, dims=shape, signed=signed)
+        return unpacked_data.astype(dtype)
+
+    return _to_array(tensor, base_dir=base_dir)
+
+
+def _from_array(arr: np.ndarray, name: str | None = None) -> TensorProto:
     """Converts a numpy array to a tensor def.
 
     Args:
@@ -352,7 +427,7 @@ def from_array(arr: np.ndarray, name: Optional[str] = None) -> TensorProto:
     if name:
         tensor.name = name
 
-    if arr.dtype == object:
+    if arr.dtype == object or np.issubdtype(arr.dtype, np.str_):
         # Special care for strings.
         tensor.data_type = helper.np_dtype_to_tensor_dtype(arr.dtype)
         # TODO: Introduce full string support.
@@ -397,7 +472,68 @@ def from_array(arr: np.ndarray, name: Optional[str] = None) -> TensorProto:
     return tensor
 
 
-def to_list(sequence: SequenceProto) -> List[Any]:
+def from_array(tensor: np.ndarray, name: str | None = None) -> TensorProto:
+    """Converts an array into a TensorProto including
+    supported type defined in :mod:`onnx._custom_element_types`.
+
+    Args:
+        tensor: a numpy array.
+        name: (optional) the name of the tensor.
+
+    Returns:
+        TensorProto: the converted tensor def.
+    """
+    if not isinstance(tensor, np.ndarray):
+        return _from_array(tensor, name)
+
+    dt = tensor.dtype
+    if dt == custom_np_types.float8e4m3fn and dt.descr[0][0] == "e4m3fn":
+        to = TensorProto.FLOAT8E4M3FN
+        dt_to = np.uint8  # type: ignore[assignment]
+    elif dt == custom_np_types.float8e4m3fnuz and dt.descr[0][0] == "e4m3fnuz":
+        to = TensorProto.FLOAT8E4M3FNUZ
+        dt_to = np.uint8  # type: ignore[assignment]
+    elif dt == custom_np_types.float8e5m2 and dt.descr[0][0] == "e5m2":
+        to = TensorProto.FLOAT8E5M2
+        dt_to = np.uint8  # type: ignore[assignment]
+    elif dt == custom_np_types.float8e5m2fnuz and dt.descr[0][0] == "e5m2fnuz":
+        to = TensorProto.FLOAT8E5M2FNUZ
+        dt_to = np.uint8  # type: ignore[assignment]
+    elif dt == custom_np_types.bfloat16 and dt.descr[0][0] == "bfloat16":
+        to = TensorProto.BFLOAT16
+        dt_to = np.uint16  # type: ignore[assignment]
+    elif dt == custom_np_types.int4 and dt.descr[0][0] == "int4":
+        to = TensorProto.INT4
+        dt_to = np.int8  # type: ignore[assignment]
+    elif dt == custom_np_types.uint4 and dt.descr[0][0] == "uint4":
+        to = TensorProto.UINT4
+        dt_to = np.uint8  # type: ignore[assignment]
+    else:
+        return _from_array(tensor, name)
+
+    if to in (TensorProto.UINT4, TensorProto.INT4):
+        value = tensor.astype(dt_to).ravel()
+        if value.size % 2 == 1:
+            raise ValueError(
+                f"The conversion of a tensor of INT4 or UINT4 requires an even size "
+                f"(shape={tensor.shape}). Every byte stores two INT4 or two UINT4."
+            )
+        buffer = (value[::2] & 0x0F) + (value[1::2] << 4)
+
+        pb = TensorProto()
+        pb.dims.extend(tensor.shape)
+        if name:
+            pb.name = name
+        pb.raw_data = buffer.tobytes()
+        pb.data_type = to
+        return pb
+
+    t = _from_array(tensor.astype(dt_to), name)
+    t.data_type = to
+    return t
+
+
+def to_list(sequence: SequenceProto) -> list[Any]:
     """Converts a sequence def to a Python list.
 
     Args:
@@ -419,7 +555,7 @@ def to_list(sequence: SequenceProto) -> List[Any]:
 
 
 def from_list(
-    lst: List[Any], name: Optional[str] = None, dtype: Optional[int] = None
+    lst: list[Any], name: str | None = None, dtype: int | None = None
 ) -> SequenceProto:
     """Converts a list into a sequence def.
 
@@ -475,7 +611,7 @@ def from_list(
     return sequence
 
 
-def to_dict(map_proto: MapProto) -> Dict[Any, Any]:
+def to_dict(map_proto: MapProto) -> dict[Any, Any]:
     """Converts a map def to a Python dictionary.
 
     Args:
@@ -484,7 +620,7 @@ def to_dict(map_proto: MapProto) -> Dict[Any, Any]:
     Returns:
         The converted dictionary.
     """
-    key_list: List[Any] = []
+    key_list: list[Any] = []
     if map_proto.key_type == TensorProto.STRING:
         key_list = list(map_proto.string_keys)
     else:
@@ -501,7 +637,7 @@ def to_dict(map_proto: MapProto) -> Dict[Any, Any]:
     return dictionary
 
 
-def from_dict(dict_: Dict[Any, Any], name: Optional[str] = None) -> MapProto:
+def from_dict(dict_: dict[Any, Any], name: str | None = None) -> MapProto:
     """Converts a Python dictionary into a map def.
 
     Args:
@@ -559,7 +695,7 @@ def from_dict(dict_: Dict[Any, Any], name: Optional[str] = None) -> MapProto:
     return map_proto
 
 
-def to_optional(optional: OptionalProto) -> Optional[Any]:
+def to_optional(optional: OptionalProto) -> Any | None:
     """Converts an optional def to a Python optional.
 
     Args:
@@ -585,7 +721,7 @@ def to_optional(optional: OptionalProto) -> Optional[Any]:
 
 
 def from_optional(
-    opt: Optional[Any], name: Optional[str] = None, dtype: Optional[int] = None
+    opt: Any | None, name: str | None = None, dtype: int | None = None
 ) -> OptionalProto:
     """Converts an optional value into a Optional def.
 
@@ -637,7 +773,7 @@ def from_optional(
 
 
 def convert_endian(tensor: TensorProto) -> None:
-    """Call to convert endianess of raw data in tensor.
+    """Call to convert endianness of raw data in tensor.
 
     Args:
         tensor: TensorProto to be converted.
@@ -650,7 +786,7 @@ def convert_endian(tensor: TensorProto) -> None:
 
 
 def create_random_int(
-    input_shape: Tuple[int], dtype: np.dtype, seed: int = 1
+    input_shape: tuple[int], dtype: np.dtype, seed: int = 1
 ) -> np.ndarray:
     """Create random integer array for backend/test/case/node.
 

onnx/onnx-ml.proto

@@ -212,7 +212,7 @@ message NodeProto {
   repeated string output = 2;   // namespace Value
 
   // An optional identifier for this node in a graph.
-  // This field MAY be absent in ths version of the IR.
+  // This field MAY be absent in this version of the IR.
   optional string name = 3;     // namespace Node
 
   // The symbolic identifier of the Operator to execute.

onnx/onnx.in.proto

@@ -209,7 +209,7 @@ message NodeProto {
   repeated string output = 2;   // namespace Value
 
   // An optional identifier for this node in a graph.
-  // This field MAY be absent in ths version of the IR.
+  // This field MAY be absent in this version of the IR.
   optional string name = 3;     // namespace Node
 
   // The symbolic identifier of the Operator to execute.

onnx/onnx.proto

@@ -210,7 +210,7 @@ message NodeProto {
   repeated string output = 2;   // namespace Value
 
   // An optional identifier for this node in a graph.
-  // This field MAY be absent in ths version of the IR.
+  // This field MAY be absent in this version of the IR.
   optional string name = 3;     // namespace Node
 
   // The symbolic identifier of the Operator to execute.

onnx/onnx_cpp2py_export/defs.pyi

@@ -1,6 +1,4 @@
 """Submodule containing all the ONNX schema definitions."""
-from __future__ import annotations
-
 from typing import Sequence, overload
 
 from onnx import AttributeProto, FunctionProto

onnx/onnx_cpp2py_export/inliner.pyi

@@ -2,13 +2,10 @@
 #
 # SPDX-License-Identifier: Apache-2.0
 
-from __future__ import annotations
-
 def inline_local_functions(model: bytes, convert_version: bool) -> bytes:
     """Inlines calls to model-local function in input model and returns it.
     Both input and output are serialized ModelProtos.
     """
-    ...
 
 def inline_selected_functions(model: bytes, function_ids: list[tuple[str,str]], exclude: bool) -> bytes:
     """Inlines calls to selected model-local functions in input model and returns it.
@@ -16,4 +13,3 @@ def inline_selected_functions(model: bytes, function_ids: list[tuple[str,str]],
     case it inlines all functions except those specified in function_ids.
     Both input and output are serialized ModelProtos.
     """
-    ...

onnx/onnx_cpp2py_export/parser.pyi

@@ -5,7 +5,6 @@ def parse_model(model: str) -> tuple[bool, bytes, bytes]:
     If success-flag is true, then serialized-proto contains the parsed ModelProto.
     Otherwise, error-message contains a string describing the parse error.
     """
-    ...
 
 def parse_graph(graph: str) -> tuple[bool, bytes, bytes]:
     """Returns (success-flag, error-message, serialized-proto).
@@ -13,7 +12,6 @@ def parse_graph(graph: str) -> tuple[bool, bytes, bytes]:
     If success-flag is true, then serialized-proto contains the parsed GraphProto.
     Otherwise, error-message contains a string describing the parse error.
     """
-    ...
 
 def parse_function(function: str) -> tuple[bool, bytes, bytes]:
     """Returns (success-flag, error-message, serialized-proto).
@@ -21,7 +19,6 @@ def parse_function(function: str) -> tuple[bool, bytes, bytes]:
     If success-flag is true, then serialized-proto contains the parsed FunctionProto.
     Otherwise, error-message contains a string describing the parse error.
     """
-    ...
 
 def parse_node(node: str) -> tuple[bool, bytes, bytes]:
     """Returns (success-flag, error-message, serialized-proto).
@@ -29,4 +26,3 @@ def parse_node(node: str) -> tuple[bool, bytes, bytes]:
     If success-flag is true, then serialized-proto contains the parsed NodeProto.
     Otherwise, error-message contains a string describing the parse error.
     """
-    ...

onnx/parser.py

@@ -1,6 +1,7 @@
 # Copyright (c) ONNX Project Contributors
 #
 # SPDX-License-Identifier: Apache-2.0
+from __future__ import annotations
 
 import onnx
 import onnx.onnx_cpp2py_export.parser as C  # noqa: N812

onnx/printer.py

@@ -1,14 +1,13 @@
 # Copyright (c) ONNX Project Contributors
 #
 # SPDX-License-Identifier: Apache-2.0
-
-from typing import Union
+from __future__ import annotations
 
 import onnx
 import onnx.onnx_cpp2py_export.printer as C  # noqa: N812
 
 
-def to_text(proto: Union[onnx.ModelProto, onnx.FunctionProto, onnx.GraphProto]) -> str:
+def to_text(proto: onnx.ModelProto | onnx.FunctionProto | onnx.GraphProto) -> str:
     if isinstance(proto, onnx.ModelProto):
         return C.model_to_text(proto.SerializeToString())
     if isinstance(proto, onnx.FunctionProto):

onnx/reference/__init__.py

@@ -1,6 +1,7 @@
 # Copyright (c) ONNX Project Contributors
 #
 # SPDX-License-Identifier: Apache-2.0
+from __future__ import annotations
 
 __all__ = ["ReferenceEvaluator"]