hooks: add hook for onnxruntime

Add hook for `onnxruntime` to ensure that provider plugins (located
in `onnxruntime/capi` directory) are collected.

_pyinstaller_hooks_contrib/stdhooks/hook-onnxruntime.py

@@ -0,0 +1,16 @@
+# ------------------------------------------------------------------
+# Copyright (c) 2024 PyInstaller Development Team.
+#
+# This file is distributed under the terms of the GNU General Public
+# License (version 2.0 or later).
+#
+# The full license is available in LICENSE, distributed with
+# this software.
+#
+# SPDX-License-Identifier: GPL-2.0-or-later
+# ------------------------------------------------------------------
+
+from PyInstaller.utils.hooks import collect_dynamic_libs
+
+# Collect provider plugins from onnxruntime/capi.
+binaries = collect_dynamic_libs("onnxruntime")

news/817.new.rst

@@ -0,0 +1,2 @@
+Add hook for ``onnxruntime`` to ensure that provider plugins are
+collected.

tests/test_deep_learning.py

@@ -369,3 +369,103 @@ def test_ultralytics_yolo(pyi_builder):
         model = YOLO("yolov8n.pt")  # Download and load pre-trained model
         results = model("https://ultralytics.com/images/bus.jpg")
     """)
+
+
+# Basic inference test with ONNX Runtime (as well as model export with ONNX + Torch).
+@importorskip('onnxruntime')
+@importorskip('onnx')
+@importorskip('torch')
+def test_onnxruntime_gpu_inference(pyi_builder, tmpdir):
+    model_file = tmpdir / "test-model.onnx"
+
+    # Build first application: model creation + export (Torch + ONNX)
+    pyi_builder.test_source("""
+        import sys
+        import torch
+
+        if len(sys.argv) != 2:
+            print(f"Usage: {sys.argv[0]} <model-filename>", file=sys.stderr)
+            sys.exit(1)
+        model_filename = sys.argv[1]
+
+        # Create model that performs addition of given inputs
+        class Model(torch.nn.Module):
+            def __init__(self):
+                super(Model, self).__init__()
+
+            def forward(self, x, y):
+                return x.add(y)
+
+        model = Model()
+
+        sample_x = torch.ones(3, dtype=torch.float32)
+        sample_y = torch.zeros(3, dtype=torch.float32)
+
+        # Export model to ONNX graph format
+        torch.onnx.export(
+            model,
+            (sample_x, sample_y),
+            model_filename,
+            input_names=["x", "y"],
+            output_names=["z"],
+            dynamic_axes={
+                "x": {0: "array_length_x"},
+                "y": {0: "array_length_y"},
+            },
+        )
+    """, app_name="model_builder", app_args=[str(model_file)])
+
+    assert model_file.isfile(), "Model file was not created!"
+
+    # Build second application: inference (ONNX Runtime) on CPU (and optionally with CUDA on Linux).
+    pyi_builder.test_source("""
+        import sys
+        import numpy as np
+        import onnxruntime
+
+        # torch is used for CUDA check; on linux, collecting PyPI-installed torch also ensures that
+        try:
+            import torch
+            cuda_available = torch.cuda.is_available()
+        except ImportError:
+            cuda_available = false
+
+        if len(sys.argv) != 2:
+            print(f"Usage: {sys.argv[0]} <model-filename>", file=sys.stderr)
+            sys.exit(1)
+        model_filename = sys.argv[1]
+
+        test_providers = ['CPUExecutionProvider']
+        if cuda_available:
+            test_providers += ['CUDAExecutionProvider']
+
+        for test_provider in test_providers:
+            print(f"Running test with provider: {test_provider}", file=sys.stderr)
+
+            # Load model into ONNX Runtime session
+            session = onnxruntime.InferenceSession(
+                model_filename,
+                providers=[test_provider],
+            )
+
+            # Check if the requested provider appears in list returned by session.get_providers(); if requested provider
+            # failed to initialize, onnxruntime seems to fall back to CPUExecutionProvider.
+            session_providers = session.get_providers()
+            print(f"session.get_providers(): {session_providers}", file=sys.stderr)
+            if test_provider not in session_providers:
+                raise RuntimeError(f"Provider {test_provider} is missing!")
+
+            # Run the model
+            x = np.float32([1.0, 2.0, 3.0])
+            y = np.float32([4.0, 5.0, 6.0])
+
+            print(f"x = {x}", file=sys.stderr)
+            print(f"y = {y}", file=sys.stderr)
+
+            z = session.run(["z"], {"x": x, "y": y})
+            z = z[0]
+
+            print(f"z = {z}", file=sys.stderr)
+
+            assert (z == x + y).all()
+    """, app_name="model_test", app_args=[str(model_file)])