Merge branch 'main' into exclude-test-data

.azure-pipelines/Windows-CI.yml

@@ -56,13 +56,6 @@ jobs:
   - powershell: Write-Host "##vso[task.prependpath]$env:CONDA\Scripts"
     displayName: Add conda to PATH
 
-  # cmake 3.28.3 failed to find protobuf installed from conda-forge
-  # https://gitlab.kitware.com/cmake/cmake/-/issues/25704
-  - script: |
-      choco install cmake --version 3.27.9 -y
-      cmake --version
-    displayName: Install specific version of CMake and check version
-
   - script: |
       conda create --yes --quiet --name py$(python.version) python=$(python.version)
       if '$(protobuf_type)' == 'External' (
@@ -72,6 +65,9 @@ jobs:
       )
     displayName: Create Anaconda environment
 
+  - powershell: echo "##vso[task.setvariable variable=CMAKE_PREFIX_PATH]$env:CONDA/envs/py$(python.version)/Library"
+    displayName: Set CMAKE_PREFIX_PATH
+
   - script: |
       call activate py$(python.version)
       python -m pip install --upgrade pip

.github/workflows/check_urls.yml

@@ -33,7 +33,7 @@ jobs:
         subfolder: docs
         file_types: .md,.py,.rst,.ipynb,.cc,.h,.cpp
         print_all: false
-        timeout: 2
+        timeout: 10
         retry_count# : 2
         exclude_urls: https://github.com/onnx/onnx/blob/main/docs/Operators,https://github.com/onnx/onnx/pull/436
         force_pass: false

.github/workflows/reuse.yml

@@ -15,4 +15,4 @@ jobs:
     steps:
     - uses: actions/checkout@b4ffde65f46336ab88eb53be808477a3936bae11 # v4.1.1
     - name: REUSE Compliance Check
-      uses: fsfe/reuse-action@4f2804894b54004c8ed4b8a62b7c649e54a3aa4b # v2.0.0
+      uses: fsfe/reuse-action@a46482ca367aef4454a87620aa37c2be4b2f8106 # v3.0.0

VERSION_NUMBER

@@ -1 +1 @@
-1.16.0
+1.17.0

docs/Hub.md

@@ -212,7 +212,7 @@ All other fields in the `metadata` field are optional for the client but provide
 The simplest way to add a model to the official `onnx/models` version model hub is to follow
 [these guidelines](https://github.com/onnx/models/blob/main/contribute.md) to contribute your model. Once contributed,
 ensure that your model has a markdown table in its `README.md`
-([Example](https://github.com/onnx/models/tree/main/vision/classification/mobilenet)). The model hub
+([Example](https://github.com/onnx/models/tree/main/validated/vision/classification/mobilenet)). The model hub
  manifest generator will pull information from these markdown tables. To run the generator:
 
  ```shell script

docs/docsgen/source/intro/python.md

@@ -25,7 +25,7 @@ to build the graph among the {ref}`l-onnx-make-function`:
   (an operator type), its inputs and outputs
 - `make_graph`: a function to create an ONNX graph with
   the objects created by the two previous functions
-- `make_model`: a last function with merges the graph and
+- `make_model`: a last function which merges the graph and
   additional metadata
 
 All along the creation, we need to give a name to every input,
@@ -1500,7 +1500,7 @@ object of type `ModelProto`. But it is not. According to
 [Protobuf 4, changes](https://developers.google.com/protocol-buffers/docs/news/2022-05-06),
 this is no longer possible after version 4 and it is safer to assume the
 only way to get a hold on the content is to serialize the model
-into bytes, give it the C function, then deserialize it.
+into bytes, give it to the C function, then deserialize it.
 Functions like `check_model` or
 `shape_inference` are calling `SerializeToString` then
 `ParseFromString` before checking the model with a C code.

onnx/common/ir.h

@@ -976,7 +976,7 @@ struct Graph final {
 
   void addInitializer(Tensor& initializer) {
     if (initializer.name().empty()) {
-      initializer.setName(ONNX_NAMESPACE::to_string(getNextUnique()));
+      initializer.setName(toVarName(getNextUnique()));
     }
     initializers_.push_back(initializer);
     initializer_names_.push_back(initializer.name());
@@ -1166,7 +1166,7 @@ struct Graph final {
   }
 
   Value* addInitializerAndInput(const Tensor& initializer) {
-    return addInitializerAndInput(initializer, ONNX_NAMESPACE::to_string(getNextUnique()));
+    return addInitializerAndInput(initializer, toVarName(getNextUnique()));
   }
 
   // Erases from graph initializer list, initializer names list, and as a graph input

onnx/defs/schema.cc

@@ -107,6 +107,30 @@ OpSchemaRegistry* OpSchemaRegistry::Instance() {
 
 void OpSchema::CheckInputOutputType(struct InferenceContext& ctx) const {
   std::unordered_map<std::string, std::string> type_constraints;
+  if (inputs_.empty() && ctx.getNumInputs() > 0) {
+    fail_check(
+        "Node (",
+        domain(),
+        "::",
+        Name(),
+        ":",
+        since_version(),
+        ") takes zero inputs, but got ",
+        ctx.getNumInputs(),
+        " in graph");
+  }
+  if (outputs_.empty() && ctx.getNumOutputs() > 0) {
+    fail_check(
+        "Node (",
+        domain(),
+        "::",
+        Name(),
+        ":",
+        since_version(),
+        ") yields zero outputs, but got ",
+        ctx.getNumOutputs(),
+        " in graph");
+  }
   // check all input types
   for (size_t in_idx = 0; in_idx < ctx.getNumInputs(); ++in_idx) {
     // If the last input is Variadic by definition, checker still needs to check the rest of actual input's type

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-ml.proto3

@@ -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.
   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.proto3

@@ -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.
   string name = 3;     // namespace Node
 
   // The symbolic identifier of the Operator to execute.

onnx/test/shape_inference_test.py

@@ -9875,6 +9875,33 @@ def test_compress_with_axis(self) -> None:
         )
         self._assert_inferred(graph, [make_tensor_value_info("output", TensorProto.INT64, (2, "N", 3, None))])  # type: ignore
 
+    def test_check_type_when_schema_has_empty_io(self):
+        input = """
+            <
+                ir_version: 7,
+                opset_import: ["" : 1]
+            >
+            agraph (X, Y) => (Z)
+            {
+                Z = CustomOp(X, Y)
+            }
+           """
+        model = onnx.parser.parse_model(input)
+
+        op_schema = defs.OpSchema(
+            "CustomOp",
+            "",
+            1,
+            inputs=[],
+            outputs=[],
+        )
+        onnx.defs.register_schema(op_schema)
+        with self.assertRaises(onnx.shape_inference.InferenceError):
+            onnx.shape_inference.infer_shapes(model, True)
+        onnx.defs.deregister_schema(
+            op_schema.name, op_schema.since_version, op_schema.domain
+        )
+
 
 if __name__ == "__main__":
     unittest.main()

pyproject.toml

@@ -1,7 +1,7 @@
 # SPDX-License-Identifier: Apache-2.0
 
 [build-system]
-requires = ["setuptools>=64", "protobuf>=3.20.2"]
+requires = ["setuptools>=64", "protobuf>=3.20.2", "cmake"]
 build-backend = "setuptools.build_meta"
 
 [project]

requirements-lintrunner.txt

@@ -1,10 +1,10 @@
 # This file is auto updated by dependabot
 lintrunner-adapters>=0.12.1
 # RUFF
-ruff==0.2.1
+ruff==0.3.0
 # MYPY
 mypy==1.8.0
-types-protobuf==4.24.0.4
+types-protobuf==4.24.0.20240129
 # BLACK-ISORT
 black==23.12.1
 isort==5.13.2

setup.py

@@ -214,18 +214,6 @@ def run(self):
                 raise RuntimeError(
                     "-DONNX_DISABLE_EXCEPTIONS=ON option is only available for c++ builds. Python binding require exceptions to be enabled."
                 )
-            if (
-                "PYTHONPATH" in os.environ
-                and "pip-build-env" in os.environ["PYTHONPATH"]
-            ):
-                # When the users use `pip install -e .` to install onnx and
-                # the cmake executable is a python entry script, there will be
-                # `Fix ModuleNotFoundError: No module named 'cmake'` from the cmake script.
-                # This is caused by the additional PYTHONPATH environment variable added by pip,
-                # which makes cmake python entry script not able to find correct python cmake packages.
-                # Actually, sys.path is well enough for `pip install -e .`.
-                # Therefore, we delete the PYTHONPATH variable.
-                del os.environ["PYTHONPATH"]
             subprocess.check_call(cmake_args)
 
             build_args = [CMAKE, "--build", os.curdir]