fix: normalize CUDA device requests

CHANGELOG.md

@@ -2,6 +2,13 @@
 
 All notable changes to `ml4t-models` will be documented in this file.
 
+## 0.1.0a5
+
+- Added Apple Silicon MPS device resolution and seeding support for Torch-backed models.
+- Added a non-skipping macOS arm64 CI smoke test that proves PyTorch MPS tensor execution.
+- Updated the NumPy requirement to allow NumPy 2.5 releases.
+- Fixed CUDA device normalization for whitespace or case variants such as `CUDA:0`.
+
 ## 0.1.0a4
 
 - Aligned conditional autoencoder training with shuffled mini-batches, BatchNorm hidden layers,

src/ml4t/models/__init__.py

@@ -4,7 +4,7 @@
 
 from importlib import import_module
 
-__version__ = "0.1.0a4"
+__version__ = "0.1.0a5"
 
 from ml4t.models.api import (
     AssetMapper,

src/ml4t/models/_internal/torch_runtime.py

@@ -18,7 +18,7 @@ def import_torch() -> Any:
 def resolve_device(torch: Any, requested: str) -> Any:
     requested_device = requested.strip().lower()
     if requested_device.startswith("cuda") and torch.cuda.is_available():
-        return torch.device(requested)
+        return torch.device(requested_device)
     mps_backend = getattr(torch.backends, "mps", None)
     if requested_device == "mps" and mps_backend is not None and mps_backend.is_available():
         return torch.device("mps")

tests/test_torch_runtime.py

@@ -46,6 +46,8 @@ def __init__(self, *, cuda_available: bool = False, mps_available: bool = False)
         self.seeds: list[int] = []
 
     def device(self, requested: str) -> _FakeDevice:
+        if requested != requested.strip().lower():
+            raise ValueError("fake torch expects normalized device strings")
         return _FakeDevice(requested.split(":", maxsplit=1)[0])
 
     def manual_seed(self, seed: int) -> None:
@@ -64,6 +66,12 @@ def test_resolve_device_cuda_when_available() -> None:
     assert resolve_device(torch, "cuda:0").type == "cuda"
 
 
+def test_resolve_device_normalizes_cuda_request() -> None:
+    torch = _FakeTorch(cuda_available=True)
+
+    assert resolve_device(torch, " CUDA:0 ").type == "cuda"
+
+
 def test_resolve_device_cuda_falls_back_to_cpu() -> None:
     torch = _FakeTorch(cuda_available=False)