fix: revert to tabpfn v2, add TABPFN_TOKEN support, and update v2 API usage

Dockerfile

@@ -37,14 +37,6 @@ RUN pip install --no-cache-dir -r webapp/requirements.txt
 # Install SAP-RPT-1 OSS directly from GitHub (needed for the real model)
 RUN pip install --no-cache-dir git+https://github.com/SAP-samples/sap-rpt-1-oss.git
 
-# Pre-accept TabPFN license at build time by writing the marker file TabPFN
-# looks for in the cache directory (covers all versions of the check).
-RUN python -c "\
-import os; \
-os.makedirs(os.path.expanduser('~/.cache/tabpfn'), exist_ok=True); \
-open(os.path.expanduser('~/.cache/tabpfn/license_accepted'), 'w').write('accepted'); \
-print('TabPFN license pre-accepted.')"
-
 # Expose port 7860 (Hugging Face Spaces default port)
 EXPOSE 7860
 

code/models/tabpfn_wrapper.py

@@ -18,8 +18,14 @@ from typing import Optional, Union
 import numpy as np
 import pandas as pd
 
-# Automatically accept the TabPFN license (non-interactive / CI environments).
-# Different TabPFN versions check different env var names — set all of them.
+# ── TabPFN non-interactive authentication ─────────────────────────────────────
+# For TabPFN v2 (PriorLabs), set TABPFN_TOKEN from the HF Space secret.
+# The user must add TABPFN_TOKEN as a secret in HF Space settings.
+_tabpfn_token = os.environ.get("TABPFN_TOKEN", "")
+if _tabpfn_token:
+    os.environ["TABPFN_TOKEN"] = _tabpfn_token  # ensure it's set for child processes
+
+# Cover all license-acceptance env var names across TabPFN versions.
 os.environ["TABPFN_ACCEPT_LICENSE"]   = "1"
 os.environ["TABPFN_LICENSE"]          = "accept"
 os.environ["TABPFN_ACCEPT_TERMS"]     = "1"
@@ -135,32 +141,34 @@ class TabPFNWrapper(BaseModelWrapper):
         try:
             from tabpfn import TabPFNClassifier
 
-            import torch
             import tabpfn
 
-            actual_device = 'cuda' if (self.device == 'auto' and torch.cuda.is_available()) else ('cpu' if self.device == 'auto' else self.device)
-
             # Reuse class-level cached classifier so weights are only loaded ONCE
             # per process, not once per CV fold.
             if TabPFNWrapper._shared_classifier is None:
                 logger.info("Creating new TabPFNClassifier and caching at class level...")
-                if hasattr(tabpfn, '__version__') and tabpfn.__version__.startswith('0.1'):
+                # TabPFN v2: no device/N_ensemble args; token read from TABPFN_TOKEN env var.
+                # TabPFN v0.1.x: needs device + N_ensemble_configurations.
+                version = getattr(tabpfn, '__version__', '0')
+                if version.startswith('0.1'):
+                    import torch
+                    actual_device = 'cuda' if (self.device == 'auto' and torch.cuda.is_available()) else 'cpu'
                     TabPFNWrapper._shared_classifier = TabPFNClassifier(
                         device=actual_device,
                         N_ensemble_configurations=self.n_ensemble
                     )
                 else:
-                    TabPFNWrapper._shared_classifier = TabPFNClassifier(device=actual_device)
+                    # v2+: just instantiate — auth is via TABPFN_TOKEN env var
+                    TabPFNWrapper._shared_classifier = TabPFNClassifier()
             else:
                 logger.info("Reusing cached TabPFN classifier (weights NOT reloaded).")
 
             self.model = TabPFNWrapper._shared_classifier
 
-            # Fit — for v0.1.x, overwrite_warning=True suppresses the data size warning
+            # Fit — v0.1.x accepts overwrite_warning=True; v2+ does not.
             try:
                 self.model.fit(X, y, overwrite_warning=True)
             except TypeError:
-                # v2+ doesn't accept overwrite_warning kwarg
                 self.model.fit(X, y)
 
             self.is_fitted = True

webapp/requirements.txt

@@ -9,5 +9,5 @@ scikit-learn>=1.3.0
 scipy>=1.10.0
 pandas>=2.0.0
 numpy>=1.24.0
-tabpfn==0.1.11
+tabpfn>=2.0.0
 huggingface_hub