consistency changes

model_config.json

@@ -31,7 +31,6 @@
         },
         "input_dimensions": [100, 100, 40],
         "input_spacing": [3.0, 3.0, 2.5],
-        "input_physical_extent": [300.0, 300.0, 300.0],
         "center": "target"
     },
     "metadata": {
@@ -43,7 +42,7 @@
             "notes": "Dose-preserving DenseNet3D (BatchNorm, anti-aliased pooling); dose normalized by 70 Gy prescription."
         },
         "model_input": {
-            "type_order": ["dose", "ct", "mask"],
+            "type_order": ["dose", "mask", "ct"],
             "tensor_shapes": {
                 "dose": [1, 100, 100, 40],
                 "ct": [1, 100, 100, 40],

preprocessor.py

@@ -1,7 +1,7 @@
 """Outcome-guided preprocessor shipped with the model repository.
 
 Implements the outcome preprocessor contract expected by pyRadPlan's
-``OutcomeModel`` objective, on top of :class:`pyRadPlan.ml.BasePreprocessor`:
+``OutcomeCNN`` objective, on top of :class:`pyRadPlan.ai_models.BasePreprocessor`:
 
 - ``configure(ct, cst, cst_masks=None, device=None)`` resamples CT and masks
   onto the fixed model input grid once.
@@ -411,15 +411,28 @@ class OutcomeCnnPreprocessor(BasePreprocessor):
             mask = mask.amax(dim=1, keepdim=True)
         self._mask_tensor = mask
 
+        # `extract` multiplies the single-channel dose by the mask; with more than
+        # one mask channel this would broadcast the dose to C channels and break the
+        # single-dose-channel model. Fail early and clearly instead.
+        if self.dose_config.get("extract", False) and mask.shape[1] > 1:
+            raise ValueError(
+                "dose 'extract' requires a single mask channel; got "
+                f"{mask.shape[1]} channels. Set mask 'collapse': true or pass a single VOI."
+            )
+
     def _dose_to_tensor(self, dose_values) -> "torch.Tensor":
         """Convert a dose array of any namespace to a (dZ, dY, dX) tensor on device."""
         from pyRadPlan.core import xp_utils  # noqa: PLC0415 - avoid import cycle at module load
 
-        t = xp_utils.to_namespace(torch, dose_values).to(dtype=torch.float32)
+        # Single device+dtype cast: a cupy/numpy dose is brought onto the model
+        # device and to float32 in one step (dlpack keeps it zero-copy where possible).
+        t = xp_utils.to_namespace(torch, dose_values).to(
+            device=self.device, dtype=torch.float32
+        )
         if t.ndim == 1:
             dx, dy, dz = self._dose_grid["size"]
             # Fortran-order reshape to (X, Y, Z) == C-order reshape to (Z, Y, X)
             t = t.reshape(dz, dy, dx)
         else:
             t = t.permute(2, 1, 0)  # (X, Y, Z) -> (Z, Y, X)
-        return t.contiguous().to(self.device).detach()
+        return t.contiguous().detach()