Fix multi-input ONNX binding and add OWLv2 CPU fp32 recipe

examples/recipes/google_owlv2-base-patch16/cpu/cpu/zero-shot-object-detection_fp32_config.json

@@ -0,0 +1,76 @@
+{
+  "export": {
+    "opset_version": 17,
+    "batch_size": 1,
+    "export_params": true,
+    "do_constant_folding": true,
+    "verbose": false,
+    "dynamo": false,
+    "enable_hierarchy_tags": true,
+    "clean_onnx": false,
+    "hierarchy_tag_format": "full",
+    "input_tensors": [
+      {
+        "name": "input_ids",
+        "dtype": "int32",
+        "shape": [
+          1,
+          16
+        ],
+        "value_range": [
+          0,
+          49408
+        ]
+      },
+      {
+        "name": "pixel_values",
+        "dtype": "float32",
+        "shape": [
+          1,
+          3,
+          960,
+          960
+        ],
+        "value_range": [
+          0,
+          1
+        ]
+      },
+      {
+        "name": "attention_mask",
+        "dtype": "int32",
+        "shape": [
+          1,
+          16
+        ],
+        "value_range": [
+          0,
+          2
+        ]
+      }
+    ],
+    "output_tensors": [
+      {
+        "name": "logits"
+      },
+      {
+        "name": "pred_boxes"
+      },
+      {
+        "name": "text_embeds"
+      },
+      {
+        "name": "image_embeds"
+      }
+    ]
+  },
+  "optim": {
+    "clamp_constant_values": true
+  },
+  "quant": null,
+  "loader": {
+    "task": "zero-shot-object-detection",
+    "model_class": "AutoModelForZeroShotObjectDetection",
+    "model_type": "owlv2"
+  }
+}

src/winml/modelkit/export/htp/exporter.py

@@ -20,6 +20,7 @@
 from __future__ import annotations
 
 import contextlib
+import inspect
 import logging
 import sys
 import time
@@ -445,9 +446,25 @@ def _convert_model_to_onnx(
         output_path = str(Path(output_path).resolve())
         Path(output_path).parent.mkdir(parents=True, exist_ok=True)
 
-        # Input names from config, fallback to inputs dict keys
+        # Input names from config, fallback to inputs dict keys.
         input_names = export_config.get_input_names() or list(inputs.keys())
 
+        if not hasattr(model, "get_export_args"):
+            # torch.onnx.export binds kwargs by name when invoking forward(), but
+            # applies input_names positionally to the resulting graph inputs in
+            # forward-signature order. Keep kwargs for safe invocation and align
+            # the ONNX names independently so recipe order cannot swap bindings.
+            try:
+                parameters = inspect.signature(model.forward).parameters
+                forward_input_names = [name for name in parameters if name in input_names]
+                if len(forward_input_names) == len(input_names):
+                    input_names = forward_input_names
+            except (TypeError, ValueError):
+                logger.debug(
+                    "Could not inspect %s.forward; preserving configured input order.",
+                    type(model).__name__,
+                )
+
         # Output names: infer from traced hierarchy, validate against config
         traced_outputs = self._hierarchy_builder.get_outputs() if self._hierarchy_builder else None
         inferred_names = infer_output_names(traced_outputs) if traced_outputs is not None else []

tests/unit/export/test_pytorch_export.py

@@ -350,12 +350,12 @@ def test_normalize_false_skips_normalization(self, tmp_path) -> None:
         assert not _all_value_info_have_shape(onnx_model)
 
     def test_mismatched_input_order_exports_successfully(self, tmp_path) -> None:
-        """Export succeeds when InputTensorSpec order differs from forward() param order.
+        """ONNX names bind correctly when specs differ from forward() param order.
 
         Regression test for CLIP bug: OnnxConfig listed pixel_values before input_ids,
-        but CLIPModel.forward() expected input_ids first. With positional args this
-        caused 'not enough values to unpack (expected 4, got 2)'. The fix uses
-        kwargs= in torch.onnx.export so name-based binding makes order irrelevant.
+        but CLIPModel.forward() expected input_ids first. kwargs make model invocation
+        order-independent, while input_names must separately follow forward order
+        because torch.onnx.export applies those names positionally to graph inputs.
         """
         model = MismatchedInputOrderModel()
 
@@ -374,9 +374,14 @@ def test_mismatched_input_order_exports_successfully(self, tmp_path) -> None:
 
         onnx_model = onnx.load(str(tmp_path / "model.onnx"))
         onnx.checker.check_model(onnx_model)
-        input_names = {i.name for i in onnx_model.graph.input}
-        assert "pixel_values" in input_names
-        assert "text_input" in input_names
+        input_shapes = {
+            tensor.name: tuple(dim.dim_value for dim in tensor.type.tensor_type.shape.dim)
+            for tensor in onnx_model.graph.input
+        }
+        assert input_shapes == {
+            "text_input": (1, 16),
+            "pixel_values": (1, 3, 8, 8),
+        }
 
 
 class TestStaleExternalDataCleanup: