Cherry pick for 0.10.1 patch (#656)

* support python 3.12 for packaging pipeline (#650)

* support python 3.12 for packaging pipeline

* turn unit test build off for python package

* update ctest flag for selected_ops build

* make onnx package to be optional. (#653)

* putting onnx package to be optional

* update the ci.yml

* add more message of missing ONNX package

---------

Co-authored-by: Wenbing Li <10278425+wenbingl@users.noreply.github.com>

Author: YUNQIUGUO

.pipelines/ci.yml

@@ -89,8 +89,7 @@ stages:
         python -m pip install --upgrade pip
         python -m pip install --upgrade setuptools
         python -m pip install onnxruntime==$(ort.version)
-        python -m pip install -r requirements.txt
-      displayName: Install requirements.txt
+      displayName: Install requirements
 
     - script: |
         CPU_NUMBER=8 python -m pip install .
@@ -283,8 +282,7 @@ stages:
         python -m pip install --upgrade setuptools
         python -m pip install --upgrade wheel
         python -m pip install onnxruntime==$(ort.version)
-        python -m pip install -r requirements.txt
-      displayName: Install requirements.txt
+      displayName: Install requirements
 
     - script: |
         python -c "import onnxruntime;print(onnxruntime.__version__)"
@@ -419,7 +417,6 @@ stages:
         call activate pyenv
         python -m pip install --upgrade pip
         python -m pip install onnxruntime==$(ort.version)
-        python -m pip install -r requirements.txt
         python -m pip install -r requirements-dev.txt
       displayName: Install requirements{-dev}.txt and cmake python modules
 
@@ -653,7 +650,6 @@ stages:
                 python3 -m pip install --upgrade pip; \
                 python3 -m pip install --upgrade setuptools; \
                 python3 -m pip install onnxruntime-gpu==$(ORT_VERSION); \
-                python3 -m pip install -r requirements.txt; \
                 python3 -m pip install -v --config-settings "ortx-user-option=use-cuda" . ; \
                 python3 -m pip install $(TORCH_VERSION) ; \
                 python3 -m pip install -r requirements-dev.txt; \

.pipelines/templates/onebranch-windows-build-stage.yml

@@ -13,7 +13,7 @@ stages:
       type: windows
 
     variables:
-      CIBW_BUILD: cp3{8,9,10,11}-*amd64
+      CIBW_BUILD: cp3{8,9,10,11,12}-*amd64
       CIBW_ARCHS: AMD64
       CIBW_ENVIRONMENT: "${{ parameters.CibwEnv }}"
       CIBW_BUILD_VERBOSITY: 1

.pipelines/wheels_linux.yml

@@ -9,7 +9,7 @@ jobs:
   timeoutInMinutes: 180
   pool: {name: 'Azure-Pipelines-EO-Ubuntu-2004-aiinfra'}
   variables:
-    CIBW_BUILD: "cp3{8,9,10,11}-*"
+    CIBW_BUILD: "cp3{8,9,10,11,12}-*"
     CIBW_ENVIRONMENT: "${{ parameters.ExtraEnv }}"
 
   steps:
@@ -29,7 +29,7 @@ jobs:
   timeoutInMinutes: 180
   pool: {name: 'Azure-Pipelines-EO-Ubuntu-2004-aiinfra'}
   variables:
-    CIBW_BUILD: "cp3{8,9,10,11}-*"
+    CIBW_BUILD: "cp3{8,9,10,11,12}-*"
     CIBW_SKIP: "*musllinux_aarch64*"
     # AzureOp doesn't support aaarch64 yet.
     # CIBW_ENVIRONMENT: "${{ parameters.ExtraEnv }}"

.pipelines/wheels_macos.yml

@@ -4,7 +4,7 @@ jobs:
   pool:
     vmImage: 'macOS-13'
   variables:
-    CIBW_BUILD: "cp3{8,9,10,11}-*"
+    CIBW_BUILD: "cp3{8,9,10,11,12}-*"
     CIBW_ARCHS_MACOS: "x86_64 universal2 arm64"
     # Skip trying to test arm64 builds on Intel Macs
     # CIBW_TEST_SKIP: "*-macosx_arm64 *-macosx_universal2:arm64"

.pipelines/wheels_win32.yml

@@ -9,7 +9,7 @@ jobs:
   timeoutInMinutes: 120
   pool: {name: 'onnxruntime-Win-CPU-2022'}
   variables:
-    CIBW_BUILD: "cp3{8,9,10,11}-*amd64"
+    CIBW_BUILD: "cp3{8,9,10,11,12}-*amd64"
     CIBW_ENVIRONMENT: "${{ parameters.ExtraEnv }}"
 
   steps:

README.md

@@ -31,12 +31,15 @@ python -m pip install git+https://github.com/microsoft/onnxruntime-extensions.gi
 
 ## Usage
 
-## 1. Generate the pre-/post- processing ONNX model
-With onnxruntime-extensions Python package, you can easily get the ONNX processing graph by converting them from Huggingface transformer data processing classes, check the following API for details.
+## 1. Generation of Pre-/Post-Processing ONNX Model
+The `onnxruntime-extensions` Python package provides a convenient way to generate the ONNX processing graph. This can be achieved by converting the Huggingface transformer data processing classes into the desired format. For more detailed information, please refer to the API below:
+
 ```python
 help(onnxruntime_extensions.gen_processing_models)
 ```
-### NOTE: These data processing model can be merged into other model [onnx.compose](https://onnx.ai/onnx/api/compose.html) if needed.
+### NOTE:
+The generation of model processing requires the **ONNX** package to be installed. The data processing models generated in this manner can be merged with other models using the [onnx.compose](https://onnx.ai/onnx/api/compose.html) if needed.
+
 ## 2. Using Extensions for ONNX Runtime inference
 
 ### Python

onnxruntime_extensions/__init__.py

@@ -10,37 +10,70 @@
 
 __author__ = "Microsoft"
 
-__all__ = [
-    'gen_processing_models',
-    'ort_inference',
-    'get_library_path',
-    'Opdef', 'onnx_op', 'PyCustomOpDef', 'PyOp',
-    'enable_py_op',
-    'expand_onnx_inputs',
-    'hook_model_op',
-    'default_opset_domain',
-    'OrtPyFunction', 'PyOrtFunction',
-    'optimize_model',
-    'make_onnx_model',
-    'ONNXRuntimeError',
-    'hash_64',
-    '__version__',
-]
 
 from ._version import __version__
 from ._ocos import get_library_path
 from ._ocos import Opdef, PyCustomOpDef
 from ._ocos import hash_64
 from ._ocos import enable_py_op
-from ._ocos import expand_onnx_inputs
-from ._ocos import hook_model_op
 from ._ocos import default_opset_domain
-from ._cuops import *  # noqa
-from ._ortapi2 import OrtPyFunction as PyOrtFunction  # backward compatibility
-from ._ortapi2 import OrtPyFunction, ort_inference, optimize_model, make_onnx_model
-from ._ortapi2 import ONNXRuntimeError, ONNXRuntimeException
-from .cvt import gen_processing_models
+
+
+_lib_only = False
+
+try:
+    import onnx  # noqa
+    import onnxruntime  # noqa
+except ImportError:
+    _lib_only = True
+    pass
+
+
+_offline_api = [
+    "gen_processing_models",
+    "ort_inference",
+    "OrtPyFunction",
+    "PyOrtFunction",
+    "optimize_model",
+    "make_onnx_model",
+    "ONNXRuntimeError",
+]
+
+__all__ = [
+    "get_library_path",
+    "Opdef",
+    "onnx_op",
+    "PyCustomOpDef",
+    "PyOp",
+    "enable_py_op",
+    "expand_onnx_inputs",
+    "hook_model_op",
+    "default_opset_domain",
+    "hash_64",
+    "__version__",
+]
 
 # rename the implementation with a more formal name
 onnx_op = Opdef.declare
 PyOp = PyCustomOpDef
+
+
+if _lib_only:
+
+    def _unimplemented(*args, **kwargs):
+        raise NotImplementedError("ONNX or ONNX Runtime is not installed")
+
+    gen_processing_models = _unimplemented
+    OrtPyFunction = _unimplemented
+    ort_inference = _unimplemented
+
+else:
+    __all__ += _offline_api
+
+    from ._cuops import *  # noqa
+    from ._ortapi2 import hook_model_op
+    from ._ortapi2 import expand_onnx_inputs
+    from ._ortapi2 import OrtPyFunction, ort_inference, optimize_model, make_onnx_model
+    from ._ortapi2 import OrtPyFunction as PyOrtFunction  # backward compatibility
+    from ._ortapi2 import ONNXRuntimeError, ONNXRuntimeException  # noqa
+    from .cvt import gen_processing_models

onnxruntime_extensions/_ocos.py

@@ -7,47 +7,48 @@
 """
 import os
 import sys
-import copy
 import glob
-import onnx
-from onnx import helper
 
 
 def _search_cuda_dir():
-    paths = os.getenv('PATH', '').split(os.pathsep)
+    paths = os.getenv("PATH", "").split(os.pathsep)
     for path in paths:
-        for filename in glob.glob(os.path.join(path, 'cudart64*.dll')):
+        for filename in glob.glob(os.path.join(path, "cudart64*.dll")):
             return os.path.dirname(filename)
 
     return None
 
 
-if sys.platform == 'win32':
+if sys.platform == "win32":
     from . import _version  # noqa: E402
-    if hasattr(_version, 'cuda'):
+
+    if hasattr(_version, "cuda"):
         cuda_path = _search_cuda_dir()
         if cuda_path is None:
-            raise RuntimeError(
-                "Cannot locate CUDA directory in the environment variable for GPU package")
+            raise RuntimeError("Cannot locate CUDA directory in the environment variable for GPU package")
 
         os.add_dll_directory(cuda_path)
 
 
 from ._extensions_pydll import (  # noqa
-    PyCustomOpDef, enable_py_op, add_custom_op, hash_64, default_opset_domain)
+    PyCustomOpDef,
+    enable_py_op,
+    add_custom_op,
+    hash_64,
+    default_opset_domain,
+)
 
 
 def get_library_path():
     """
     The custom operator library binary path
     :return: A string of this library path.
     """
-    mod = sys.modules['onnxruntime_extensions._extensions_pydll']
+    mod = sys.modules["onnxruntime_extensions._extensions_pydll"]
     return mod.__file__
 
 
 class Opdef:
-
     _odlist = {}
 
     def __init__(self, op_type, func):
@@ -57,14 +58,14 @@ def __init__(self, op_type, func):
 
     @staticmethod
     def declare(*args, **kwargs):
-        if len(args) > 0 and hasattr(args[0], '__call__'):
+        if len(args) > 0 and hasattr(args[0], "__call__"):
             raise RuntimeError("Unexpected arguments {}.".format(args))
             # return Opdef._create(args[0])
         return lambda f: Opdef.create(f, *args, **kwargs)
 
     @staticmethod
     def create(func, *args, **kwargs):
-        name = kwargs.get('op_type', None)
+        name = kwargs.get("op_type", None)
         op_type = name or func.__name__
         opdef = Opdef(op_type, func)
         od_id = id(opdef)
@@ -76,15 +77,15 @@ def create(func, *args, **kwargs):
         opdef._nativedef.op_type = op_type
         opdef._nativedef.obj_id = od_id
 
-        inputs = kwargs.get('inputs', None)
+        inputs = kwargs.get("inputs", None)
         if inputs is None:
             inputs = [PyCustomOpDef.dt_float]
         opdef._nativedef.input_types = inputs
-        outputs = kwargs.get('outputs', None)
+        outputs = kwargs.get("outputs", None)
         if outputs is None:
             outputs = [PyCustomOpDef.dt_float]
         opdef._nativedef.output_types = outputs
-        attrs = kwargs.get('attrs', None)
+        attrs = kwargs.get("attrs", None)
         if attrs is None:
             attrs = {}
         elif isinstance(attrs, (list, tuple)):
@@ -106,16 +107,15 @@ def cast_attributes(self, attributes):
             elif self._nativedef.attrs[k] == PyCustomOpDef.dt_string:
                 res[k] = v
             else:
-                raise RuntimeError("Unsupported attribute type {}.".format(
-                    self._nativedef.attrs[k]))
+                raise RuntimeError("Unsupported attribute type {}.".format(self._nativedef.attrs[k]))
         return res
 
 
 def _on_pyop_invocation(k_id, feed, attributes):
     if k_id not in Opdef._odlist:
         raise RuntimeError(
-            "Unable to find function id={}. "
-            "Did you decorate the operator with @onnx_op?.".format(k_id))
+            "Unable to find function id={}. " "Did you decorate the operator with @onnx_op?.".format(k_id)
+        )
     op_ = Opdef._odlist[k_id]
     rv = op_.body(*feed, **op_.cast_attributes(attributes))
     if isinstance(rv, tuple):
@@ -127,86 +127,7 @@ def _on_pyop_invocation(k_id, feed, attributes):
         res = tuple(res)
     else:
         res = (rv.shape, rv.flatten().tolist())
-    return (k_id, ) + res
-
-
-def _ensure_opset_domain(model):
-    op_domain_name = default_opset_domain()
-    domain_missing = True
-    for oi_ in model.opset_import:
-        if oi_.domain == op_domain_name:
-            domain_missing = False
-
-    if domain_missing:
-        model.opset_import.extend(
-            [helper.make_operatorsetid(op_domain_name, 1)])
-
-    return model
-
-
-def expand_onnx_inputs(model, target_input, extra_nodes, new_inputs):
-    """
-    Replace the existing inputs of a model with the new inputs, plus some extra nodes
-    :param model: The ONNX model loaded as ModelProto
-    :param target_input: The input name to be replaced
-    :param extra_nodes: The extra nodes to be added
-    :param new_inputs: The new input (type: ValueInfoProto) sequence
-    :return: The ONNX model after modification
-    """
-    graph = model.graph
-    new_inputs = [n for n in graph.input if n.name !=
-                  target_input] + new_inputs
-    new_nodes = list(model.graph.node) + extra_nodes
-    new_graph = helper.make_graph(
-        new_nodes, graph.name, new_inputs, list(graph.output), list(graph.initializer))
-
-    new_model = copy.deepcopy(model)
-    new_model.graph.CopyFrom(new_graph)
-
-    return _ensure_opset_domain(new_model)
-
-
-def hook_model_op(model, node_name, hook_func, input_types):
-    """
-    Add a hook function node in the ONNX Model, which could be used for the model diagnosis.
-    :param model: The ONNX model loaded as ModelProto
-    :param node_name: The node name where the hook will be installed
-    :param hook_func: The hook function, callback on the model inference
-    :param input_types: The input types as a list
-    :return: The ONNX model with the hook installed
-    """
-
-    # onnx.shape_inference is very unstable, useless.
-    # hkd_model = shape_inference.infer_shapes(model)
-    hkd_model = model
-
-    n_idx = 0
-    hnode, nnode = (None, None)
-    nodes = list(hkd_model.graph.node)
-    brkpt_name = node_name + '_hkd'
-    optype_name = "op_{}_{}".format(hook_func.__name__, node_name)
-    for n_ in nodes:
-        if n_.name == node_name:
-            input_names = list(n_.input)
-            brk_output_name = [i_ + '_hkd' for i_ in input_names]
-            hnode = onnx.helper.make_node(
-                optype_name, n_.input, brk_output_name, name=brkpt_name, domain=default_opset_domain())
-            nnode = n_
-            del nnode.input[:]
-            nnode.input.extend(brk_output_name)
-            break
-        n_idx += 1
-
-    if hnode is None:
-        raise ValueError("{} is not an operator node name".format(node_name))
-
-    repacked = nodes[:n_idx] + [hnode, nnode] + nodes[n_idx+1:]
-    del hkd_model.graph.node[:]
-    hkd_model.graph.node.extend(repacked)
-
-    Opdef.create(hook_func, op_type=optype_name,
-                 inputs=input_types, outputs=input_types)
-    return _ensure_opset_domain(hkd_model)
+    return (k_id,) + res
 
 
 PyCustomOpDef.install_hooker(_on_pyop_invocation)