TabPFN Files

Breaking Changes

• Major release: TabPFN-3 is now the default model. New users and existing users who do not pin a model will automatically get TabPFN-3 going forward. To use a previous model version, use the create_default_for_version() classmethod on TabPFNClassifier / TabPFNRegressor, or pass an explicit model_path to the estimator constructor to pin a specific model file. (#948)

Added

• Add opt-in feature subsampling strategies across ensemble members when the number of features exceeds max_features_per_estimator. Set FEATURE_SUBSAMPLING_METHOD in the inference config to one of "random" (default), "balanced", or "constant_and_balanced". (#851)
• Add enable_torch_compile to PerformanceOptions. (#879)
• Add GPU preprocessing pipeline that runs feature transformations (quantile normalization, SVD) directly on the GPU as part of the model forward pass. (#884)
• Add get_inference_config() method to TabPFNClassifier and TabPFNRegressor. This method loads the model checkpoint if needed and returns the active InferenceConfig, allowing inspection of preprocessing and inference settings before calling fit(). (#890)
• Add an optional show_progress_bar flag to TabPFN classifier and regressor inference, defaulting to False. (#899)
• Add a nightly workflow that reproduces every example notebook's pip-install sequence in a fresh venv and asserts tabpfn resolves to the latest PyPI release. (#901)
• Add gini_feature_importance and gini_feature_importance_lightgbm as new FEATURE_SUBSAMPLING_METHOD options. Both rank features by importance and always include the top-K most predictive features per estimator when the dataset exceeds max_features_per_estimator. LightGBM is an optional dependency (pip install tabpfn[lightgbm]). (#908)
• Add TabPFN v3 support: TabPFNClassifier and TabPFNRegressor now support ModelVersion.V3, including create_default_for_version(ModelVersion.V3) and explicit v3 model paths. (#909)
• Add auto as a new FEATURE_SUBSAMPLING_METHOD option. When selected, it automatically uses gini_feature_importance (LightGBM-based) for datasets with more than 100k samples where feature subsampling is needed, and falls back to balanced otherwise. LightGBM is now a required dependency (previously optional via pip install tabpfn[lightgbm]). (#913)
• Add embedding_dim abstract property to the Architecture interface, exposing the output embedding dimension for all architecture implementations. (#924)
• Stratified row subsampling for the classifier: when SUBSAMPLE_SAMPLES is set, each ensemble member now draws rows that preserve the original class proportions, using a balanced round-robin pool per class to ensure uniform row coverage across estimators. (#928)
• Add opt-in FlashAttention-3 backend selector for v3 (PerformanceOptions.attention_backend). On Hopper GPUs, "auto" routes to FA3 once the sequence length amortises FA3's dispatch overhead; otherwise falls back to PyTorch SDPA. (#935)
• Auto-scale n_estimators at fit time so every feature is covered by at least one ensemble member. The effective count is exposed as n_estimators_; a UserWarning is emitted when scaling triggers. (#937)
• Add TorchSquashingScaler and TorchSquashingScalerStep — a torch implementation of SquashingScaler mirroring the CPU version. (#938)
• Run SVD on GPU when enable_gpu_preprocessing=True by pre-warming PyTorch's LAPACK lazy wrapper on the main thread before parallel dispatch to avoid a multi-GPU race in torch.svd_lowrank -> torch.linalg.qr. (#941)
• Schedule the squashing scaler on GPU when the configuration is eligible. This makes the preprocessing significantly faster. (#944)

Changed

• Introduces balanced subsampling of features for improved performance for datasets with large number of features. Results may vary slightly because of different seeds. (#851)
• Model checkpoint caching now automatically invalidates when the file on disk changes (detected via mtime and size), so replaced checkpoints (e.g. during finetuning) are always reloaded. (#863)
• Row subsampling across ensemble members now uses round-robin balanced sampling. This replaces the previous random sampling approach. (#886)
• Remove unused v2.6 defaults from InferenceConfig.get_default(). V2.6 checkpoints always embed their own InferenceConfig, so these defaults were never used at inference time. The v2.6 preprocessor config factories are also removed from tabpfn.preprocessing. (#890)
• Renamed InferenceConfig.CONSTANT_FEATURE_COUNT to FEATURE_SUBSAMPLING_CONSTANT_FEATURE_COUNT to better reflect its purpose. Old checkpoints that store the previous key name are migrated transparently on load. (#900)
• Updated copyright year to 2026 and consolidated the authors field in pyproject.toml to a single Prior Labs entry. (#916)
• Speed up ReshapeFeatureDistributionsStep ~2x on large numerical workloads (~1670 ms → ~870 ms on 100k×100): inline SquashingScaler's robust/minmax branches into a single nanpercentile pass, and call ColumnTransformer.fit_transform once instead of fit + transform (sklearn's fit already runs the transform internally). Behavior unchanged. (#938)
• Keep the inference cache on the GPU by default when fit_mode="fit_with_cache", avoiding host/device transfers on each predict call. The per-estimator KV caches are reachable via model.executor_.kv_caches. (#942)
• Clean up README and inline references to removed/deprecated tabpfn-extensions modules (rf_pfn, post_hoc_ensembles, hpo) and the retired large_datasets example. Drops the now-stale workflow mermaid diagram, updates the OOM error message to link to the Models page, and removes the unused AutoTabPFNClassifier import from the Colab demo notebook. (#945)

Fixed

• Fix inference precision to respect force_inference_dtype in KV cache engine and skip thinking tokens during cache-building. (#802)
• Reduce TabPFNRegressor peak GPU memory at large test-set sizes by chunking the row dimension inside translate_probs_across_borders. Output is unchanged; peak drops ~60% at n_test=250k (57.6 GB → 22.8 GB on an H100). (#882)
• Fix v2.6 producing near-random outputs on Apple Silicon (MPS). F.scaled_dot_product_attention on MPS silently returns wrong values for non-contiguous q/k/v (upstream: pytorch/pytorch#181133); we now force contiguity before the call. Iris multiclass accuracy on MPS: 0.48 → 0.98. (#888)
• Fix FinetunedTabPFNClassifier/FinetunedTabPFNRegressor dropping pandas feature names from the final inference model. The raw training inputs are now retained so the fitted inference estimator records feature_names_in_, and calling predict_proba/predict with a DataFrame no longer triggers spurious sklearn feature-name warnings. (#892)
• Adapt recompute_layer flag in FinetunedTabPFNClassifier/FinetunedTabPFNRegressor to new PerformanceOptions interface. (#917)
• Fix save_tabpfn_model not setting architecture_name="tabpfn_v3" for v3 configs and not persisting inference_config_, which broke resuming v3 finetuning from a saved checkpoint. (#930)
• Reduce KV cache GPU memory in fit_with_cache by materialising only the kept KV head(s) at cache-build time. Output is unchanged. (#933)
• Fix RuntimeError: No available kernel on v3 inference for GPUs where none of FlashAttention / EfficientAttention / CuDNN-Attention are eligible (e.g. Turing-class cards like the T4) by adding SDPBackend.MATH as a final fallback in _SDPA_BACKENDS. (#947)