Prefer CUDA over DirectML when both are installed

tests/test_torch_device_utils.py

@@ -0,0 +1,57 @@
+import sys
+import types
+
+
+def _make_fake_torch(*, cuda_available: bool):
+    torch = types.ModuleType("torch")
+    torch.__path__ = []  # allow importing torch.backends
+
+    class _FakeCUDA:
+        @staticmethod
+        def is_available():
+            return cuda_available
+
+    torch.cuda = _FakeCUDA()
+    torch.cpu = object()
+
+    backends = types.ModuleType("torch.backends")
+    backends.__path__ = []
+    torch.backends = backends
+
+    return torch, backends
+
+
+def test_get_installed_torch_platform_prefers_cuda_over_directml(monkeypatch):
+    # Regression test for Windows environments where users have both a working CUDA/ROCm torch
+    # backend AND torch-directml installed. In that scenario we should prefer torch.cuda over
+    # DirectML to avoid mis-detecting the active backend.
+    fake_torch, fake_backends = _make_fake_torch(cuda_available=True)
+
+    fake_torch_directml = types.ModuleType("torch_directml")
+    fake_torch_directml.is_available = lambda: True
+
+    monkeypatch.setitem(sys.modules, "torch", fake_torch)
+    monkeypatch.setitem(sys.modules, "torch.backends", fake_backends)
+    monkeypatch.setitem(sys.modules, "torch_directml", fake_torch_directml)
+
+    from torchruntime.utils.torch_device_utils import get_installed_torch_platform
+
+    torch_platform_name, torch_platform = get_installed_torch_platform()
+    assert torch_platform_name == "cuda"
+    assert torch_platform is fake_torch.cuda
+
+
+def test_get_installed_torch_platform_uses_directml_when_cuda_unavailable(monkeypatch):
+    fake_torch, fake_backends = _make_fake_torch(cuda_available=False)
+
+    fake_torch_directml = types.ModuleType("torch_directml")
+    fake_torch_directml.is_available = lambda: True
+
+    monkeypatch.setitem(sys.modules, "torch", fake_torch)
+    monkeypatch.setitem(sys.modules, "torch.backends", fake_backends)
+    monkeypatch.setitem(sys.modules, "torch_directml", fake_torch_directml)
+
+    from torchruntime.utils.torch_device_utils import get_installed_torch_platform
+
+    torch_platform_name, _ = get_installed_torch_platform()
+    assert torch_platform_name == "directml"

torchruntime/utils/torch_device_utils.py

@@ -42,13 +42,15 @@ def get_installed_torch_platform():
     import torch.backends
     from platform import system as os_name
 
-    if _is_directml_platform_available():
-        return DIRECTML, torch.directml
-
     if torch.cuda.is_available():
         return CUDA, torch.cuda
     if hasattr(torch, XPU) and torch.xpu.is_available():
         return XPU, torch.xpu
+
+    # DirectML is a useful fallback on Windows, but users can have torch-directml installed
+    # alongside a working CUDA/ROCm torch build. Prefer the native torch backend when available.
+    if _is_directml_platform_available():
+        return DIRECTML, torch.directml
     if os_name() == "Darwin":
         if hasattr(torch, MPS):
             return MPS, torch.mps