Change default minimum_deployment_target to None and support CoreML models with no inputs (#13053)

This PR changes the default minimum_deployment_target in CoreML from the
fixed iOS15 to None. This will automatically select the minimum
deployment target required to export the model.

A warning occurs after export telling the user the deploymnet target
selected, and directing them to specify a specific target if that is
what they need.

This is more in line with how CoreML standalone works.

In addition, this PR allows running CoreML models with no user inputs,
which requires CoreML deployment target iOS18 or higher.

This addresses:
* #11719
* #12906

Author: metascroy

backends/apple/coreml/compiler/coreml_preprocess.py

@@ -126,15 +126,18 @@ def model_compute_precision_from_compile_specs(
 
     @staticmethod
     def generate_minimum_deployment_target_compile_spec(
-        min_deployment_target: ct.target,
+        min_deployment_target: Optional[ct.target],
     ) -> CompileSpec:
         """
         Returns the compile spec representing the minimum deployment target on which the model can run,
         for additional details please refer to the documentation for ``coremltools.target``.
         """
+        value = str("").encode("utf-8")
+        if min_deployment_target is not None:
+            value = str(min_deployment_target.value).encode("utf-8")
         return CompileSpec(
             COMPILE_SPEC_KEYS.MIN_DEPLOYMENT_TARGET.value,
-            str(min_deployment_target.value).encode("utf-8"),
+            value,
         )
 
     @staticmethod
@@ -146,10 +149,13 @@ def min_deployment_target_from_compile_specs(
         """
         for compile_spec in compile_specs:
             if compile_spec.key == COMPILE_SPEC_KEYS.MIN_DEPLOYMENT_TARGET.value:
-                compile_spec_value: int = int(compile_spec.value.decode("utf-8"))
+                value = compile_spec.value.decode("utf-8")
+                if value == "":
+                    return None
+                compile_spec_value: int = int(value)
                 return ct.target(compile_spec_value)
 
-        return ct.target.iOS15
+        return None
 
     @staticmethod
     def compute_unit_from_compile_specs(
@@ -211,7 +217,7 @@ def op_linear_quantizer_config_from_compile_specs(
     @staticmethod
     def generate_compile_specs(
         compute_unit: ct.ComputeUnit = ct.ComputeUnit.ALL,
-        minimum_deployment_target: ct.target = ct.target.iOS15,
+        minimum_deployment_target: Optional[ct.target] = None,
         compute_precision: ct.precision = ct.precision.FLOAT16,
         model_type: MODEL_TYPE = MODEL_TYPE.MODEL,
         op_linear_quantizer_config: Optional[Dict] = None,
@@ -248,6 +254,13 @@ def model_metadata_from_spec(
         input_names: List[str] = [input.name for input in model_spec.description.input]
         output_names = [output.name for output in model_spec.description.output]
 
+        if len(output_names) == 0:
+            raise ValueError("Cannot lower a model with no outputs in CoreML.")
+        if len(input_names) == 0:
+            assert (
+                model_spec.specificationVersion >= 9
+            ), "Deploying a model with no inputs in CoreML requires you set minimum_deployment_target to iOS18 or later in the CoreMLPartitioner."
+
         return ModelMetadata(
             inputNames=input_names, outputNames=output_names, identifier=identifier
         )
@@ -352,6 +365,12 @@ def preprocess_model(
         dir_path: Path = Path("tmp") / identifier
         model_dir_path: Path = dir_path / "lowered_module"
         model_spec: ct.proto.Model_pb2 = mlmodel.get_spec()
+        logger.warning(
+            f"The model with identifier {identifier} was exported with CoreML specification version {model_spec.specificationVersion}, and it will not run on all version of iOS/macOS."
+            " See https://apple.github.io/coremltools/mlmodel/Format/Model.html#model for information on what OS versions are compatible with this specifcation version."
+            " If you want to control the deployment target, please set the minimum_deployment_target compile spec in the CoreMLPartitioner."
+        )
+
         model_metadata: ModelMetadata = CoreMLBackend.model_metadata_from_spec(
             model_spec=model_spec,
             identifier=identifier,
@@ -418,7 +437,7 @@ def preprocess(
         model_compute_precision: ct.precision = (
             CoreMLBackend.model_compute_precision_from_compile_specs(compile_specs)
         )
-        minimum_deployment_target: ct.target = (
+        minimum_deployment_target: Optional[ct.target] = (
             CoreMLBackend.min_deployment_target_from_compile_specs(compile_specs)
         )
         compute_units: ct.ComputeUnit = CoreMLBackend.compute_unit_from_compile_specs(

backends/apple/coreml/runtime/delegate/model_metadata.h

@@ -29,9 +29,7 @@ struct ModelMetadata {
     inline ModelMetadata() noexcept { }
 
     /// Returns `true` if the metadata is valid otherwise `false`.
-    inline bool is_valid() const noexcept {
-        return !identifier.empty() && !input_names.empty() && !output_names.empty();
-    }
+    inline bool is_valid() const noexcept { return !identifier.empty() && !output_names.empty(); }
 
     inline std::string to_json_string() const noexcept { return executorchcoreml::serde::json::to_json_string(*this); }