Support differentiable_input on TabPFNRegressor

changelog/923.fixed.md

@@ -0,0 +1 @@
+Add `TabPFNRegressor.fit_with_differentiable_input(X, y)` so gradients can flow from a downstream loss back through the regressor into upstream torch modules feeding `X` (and `y`, when it carries grads). Mirrors the existing classifier-side path — previously `TabPFNRegressor.fit` raised `ValueError("Differentiable input is not supported for regressors yet.")` and there was no differentiable counterpart.

src/tabpfn/base.py

@@ -370,6 +370,24 @@ def create_inference_engine(  # noqa: PLR0913
     raise ValueError(f"Invalid fit_mode: {fit_mode}")
 
 
+def reject_categoricals_for_differentiable_input(
+    categorical_features_indices: Sequence[int] | None,
+) -> None:
+    """Reject categorical features in the differentiable-input fit path.
+
+    The differentiable path uses an identity preprocessor (no
+    ordinal-encoding step), so categorical columns have no valid handling
+    and would corrupt the prompt-tuning signal.
+    """
+    if (
+        categorical_features_indices is not None
+        and len(categorical_features_indices) > 0
+    ):
+        raise ValueError(
+            "Categorical features are not supported for differentiable input."
+        )
+
+
 def initialize_model_variables_helper(
     calling_instance: TabPFNRegressor | TabPFNClassifier,
     model_type: Literal["regressor", "classifier"],

src/tabpfn/classifier.py

@@ -41,6 +41,7 @@
     get_embeddings,
     initialize_model_variables_helper,
     initialize_telemetry,
+    reject_categoricals_for_differentiable_input,
 )
 from tabpfn.constants import (
     PROBABILITY_EPSILON_ROUND_ZERO,
@@ -635,13 +636,7 @@ def _initialize_for_differentiable_input(
         )
 
         # Minimal preprocessing for prompt tuning
-        if (
-            self.categorical_features_indices is not None
-            and len(self.categorical_features_indices) > 0
-        ):
-            raise ValueError(
-                "Categorical features are not supported for differentiable input."
-            )
+        reject_categoricals_for_differentiable_input(self.categorical_features_indices)
         n_features = X.shape[1]
         features = [Feature(name=None, modality=FeatureModality.NUMERICAL)] * n_features
         self.inferred_feature_schema_ = FeatureSchema(features=features)