fix: add missing scipy dep, TabPFN weight caching, better error logging

code/models/tabpfn_wrapper.py

@@ -144,10 +144,26 @@ class TabPFNWrapper(BaseModelWrapper):
             
             actual_device = 'cuda' if (self.device == 'auto' and torch.cuda.is_available()) else ('cpu' if self.device == 'auto' else self.device)
             
-            if hasattr(tabpfn, '__version__') and tabpfn.__version__.startswith('0.1'):
-                self.model = TabPFNClassifier(device=actual_device, N_ensemble_configurations=self.n_ensemble)
+            # Reuse a cached classifier if one was injected to avoid reloading weights
+            if hasattr(self, '_cached_classifier') and self._cached_classifier is not None:
+                self.model = self._cached_classifier
+                logger.info("Reusing cached TabPFN classifier (skipping weight reload)")
             else:
-                self.model = TabPFNClassifier(device=actual_device)
+                if hasattr(tabpfn, '__version__') and tabpfn.__version__.startswith('0.1'):
+                    self.model = TabPFNClassifier(device=actual_device, N_ensemble_configurations=self.n_ensemble)
+                else:
+                    self.model = TabPFNClassifier(device=actual_device)
+                
+                # Store in global cache for future reuse
+                try:
+                    import benchmark
+                    benchmark._TABPFN_CACHED_MODEL = self.model
+                except (ImportError, AttributeError):
+                    try:
+                        import webapp.benchmark as wb
+                        wb._TABPFN_CACHED_MODEL = self.model
+                    except (ImportError, AttributeError):
+                        pass
 
             # Fit model
             self.model.fit(X, y)

webapp/benchmark.py

@@ -55,7 +55,15 @@ def _tabpfn(task):
     if task != "classification":
         raise ValueError("TabPFN only supports classification tasks")
     from models.tabpfn_wrapper import TabPFNWrapper
-    return TabPFNWrapper(task_type=task, random_state=RAND)
+    wrapper = TabPFNWrapper(task_type=task, random_state=RAND)
+    # Re-use the cached TabPFNClassifier model if available to avoid
+    # reloading weights on every CV fold (saves ~2s per fold + RAM)
+    global _TABPFN_CACHED_MODEL
+    if _TABPFN_CACHED_MODEL is not None:
+        wrapper._cached_classifier = _TABPFN_CACHED_MODEL
+    return wrapper
+
+_TABPFN_CACHED_MODEL = None
 
 
 class _SAPModel:

webapp/main.py

@@ -212,6 +212,8 @@ async def benchmark(
         }
             
     except Exception as e:
+        import traceback
+        traceback.print_exc()
         raise HTTPException(500, f"Benchmarking failed: {e}")
 
     return JSONResponse(result)

webapp/requirements.txt

@@ -6,6 +6,7 @@ xgboost>=2.0.0
 lightgbm>=4.0.0
 catboost>=1.2.0
 scikit-learn>=1.3.0
+scipy>=1.10.0
 pandas>=2.0.0
 numpy>=1.24.0
 tabpfn>=7.1.1