v7.1.1

Added

• Add modular experiment logging for finetuning with experiment_logger parameter, including WandbLogger for W&B tracking and a FinetuningLogger protocol for custom integrations. (#815)
• Add three-tier authentication flow: browser-based login for graphical environments, headless interactive login with clipboard copy for SSH/cluster sessions, and clear step-by-step instructions for fully non-interactive environments. (#862)

Changed

• • Optimize regressor predict method for memory efficiency
    • Average ensemble outputs on-the-fly instead of accumulating all outputs
    • Reduces memory usage by avoiding storage of all intermediate outputs, especially beneficial for large n_estimators (#745)

Fixed

• Fix bugs where fit_mode="fit_with_cache" produced slightly incorrect predictions in v2.5 (but not v2): thinking tokens were added twice, inference_precision flag was not applied correctly. (#852)

v7.1.0

Added

• More informative Out-Of-Memory error message. (#805)
• Add multi-GPU DDP support for finetuning via torchrun (auto-detected, no code changes needed) (#812)
• Add task_type to forward. (#844)
• Exclude very recent package release in environment (#847)

Changed

• Switch from Hugging Face to Prior Labs website for model license acceptance (#798)
• "auto" device selection now uses all available CUDA GPUs instead of only the first one (#808)
• Optimize fingerprint hashing in preprocessing: round feature matrix once instead of per-row, avoid redundant SHA-256 calls. Speeds up fit by up to 2x for large datasets. (#818)

Fixed

• Fix the pdf() in FullSupportBarDistribution to actually compute the probability density. (#799)
• Fix float overflow in Yeo-Johnson inverse transform that produced inf values and silently degraded regression border resolution. (#838)
• Fix differentiable input for v2.6 (#843)

v7.0.1

Added

• Remove the n_out parameter from get_architecture. (#839)

Changed

• Make TabPFN-2.6 the default model (#840)

v7.0.0

Added

• Introduce TabPFN-2.6 model and use as default (#831)
• Added argument use_fixed_preprocessing_seed to FinetunedTabPFNClassifier and FinetunedTabPFNRegressor for improved finetuning performance.
• This PR changes the random seeds used in the preprocessing, which may cause slight differences in final outcomes compared to previous versions.
  (#771)
• More informative Out-Of-Memory error message. (#805)
• Added max_onehot_cardinality option to cap one-hot encoding expansion for high-cardinality categorical features. (#833)

Changed

• Introduces TabPFN-2.6 as the new default model for TabPFNClassifier and TabPFNRegressor (#831)
• Remove unused functions default_classifier_preprocessor_configs() and default_regressor_preprocessor_configs() (#831)
• "auto" device selection now uses all available CUDA GPUs instead of only the first one (#808)
• Optimize fingerprint hashing in preprocessing: round feature matrix once instead of per-row, avoid redundant SHA-256 calls. Speeds up fit by up to 2x for large datasets. (#818)
• Bump minimum torch version from 2.1 to 2.5 (#823)
• Cache loaded checkpoints across fit calls: skip redundant disk I/O when the same model is loaded repeatedly (e.g. cross-validation, hyperparameter search). (#832)

Fixed

• Fix the pdf() in FullSupportBarDistribution to actually compute the probability density. (#799)

v6.4.1

Changed

• Download lock is now scoped to the target file path, allowing concurrent downloads of different model files to proceed in parallel instead of serializing all downloads behind a single global lock. (#790)

v6.4.0

Added

• Introduces dedicated method for fitting with differentiable input called fit_with_differentiable_input() (#752)
• Pass through kwargs in FinetunedTabPFNClassifier and FinetunedTabPFNRegressor predict and predict_proba methods to allow additional options like output_type='full' (#772)
• Add MPS memory limiting to prevent macOS system crashes when using Apple Silicon GPUs. Memory is automatically limited to 70% of recommended max on import. Configurable via TABPFN_MPS_MEMORY_FRACTION environment variable. (#773)
• Added TabPFNCUDAOutOfMemoryError and TabPFNMPSOutOfMemoryError for GPU out-of-memory errors during prediction with large test sets, providing helpful guidance on batching predictions. (#774)

Changed

• Remove upper version limits on dependencies (#764)

• Refactored preprocessing pipeline:

  • Introduced FeatureSchema system to track column metadata through transformations, replacing raw categorical index lists.
  • Added PreprocessingPipeline and PreprocessingStep interfaces for modular transformations and updated all preprocessing steps.
  • Added TabPFNLabelEncoder for centralized label validation and metadata extraction.

  (#767)

• • Introduces AddSVDFeaturesStep as a dedicated preprocessing step for SVD feature generation
  • Removes SVD-related functionality from ReshapeFeatureDistributionsStep
  • Extracts utility functions to a new tabpfn/preprocessing/steps/utils.py module

  (#768)

• SVD preprocessing is now applied after categorical encoding for more robustness. Note that this may result in slight variations in final outcomes compared to previous versions. (#779)

• Remove random_state parameter from AddFingerprintFeaturesStep; fingerprint hashing is now fully deterministic and no longer uses a random salt. Predictions will differ slightly from previous versions due to the changed fingerprint values. (#780)

• Fix bug related to column ordering in ordinal encoder by introducing OrderPreservingColumnTransformer. Note that this change can cause slight differences in final outcomes compared to previous versions. (#788)

Fixed

• Fix race condition when model is downloaded simultaneously by multiple processes (#738)
• Fix infinite loop in fingerprint hashing when rows contain inf or very large floats (#780)

Deprecated

• Removes "scaler" as an option for global_transformer_name in PreprocessorConfig (#768)

v6.3.2

Added

• • Moved preprocessing-related code to dedicated modules inside src/tabpfn/preprocessing/
  • Renamed public functions:
    • validate_X_predict → ensure_compatible_predict_input_sklearn
    • validate_Xy_fit → ensure_compatible_fit_inputs_sklearn

  (#720)

• • Add new features to finetuning (metric selection, time limit, passing validation data)
    • Added eval_metric and time_limit parameters to FinetunedTabPFNClassifier and FinetunedTabPFNRegressor
    • Added X_val, y_val parameters to .fit() of FinetunedTabPFNClassifier and FinetunedTabPFNRegressor
  • Fix bug in finetuning for splitting very small datasets
  • Ensure finetuning compares to the default checkpoint and does not accept worse models after finetuning

  (#730)

• • Ensure TabPFNValidationError wraps both custom and sklearn's validate_data() errors (#732)

• Refactor of model encoder. Move imports from tabpfn.architectures.base.encoders to tabpfn.architectures.encoders (#733)

• Renamed the estimator's preprocessor_ attribute to ordinal_encoder_ (#756)

v6.0.6

What's Changed

• New memory saving mode heuristic that takes account of inference speed. by @oscarkey in #605
• Bump version to 6.0.1 to include improved memory saving auto selection. by @oscarkey in #609
• Add seed=42 embeddings, read them from file to match cpu and gpu embeddings. by @priorphil in #612
• Anonymously track model_version and model_path by @safaricd in #611
• Add a link to the gated model docs to the error message. by @oscarkey in #613

Full Changelog: v6.0.0...v6.0.6

v6.0.0

What's Changed

• Bump actions/checkout from 4.2.2 to 5.0.0 by @dependabot[bot] in #454
• Bump astral-sh/setup-uv from 6.4.3 to 6.5.0 by @dependabot[bot] in #456
• Bump astral-sh/ruff-action from 3.5.0 to 3.5.1 by @dependabot[bot] in #455
• Finetuning rergression debugging - Renaming bar distributions by @rosenyu304 in #410
• Fix infer cat nans by @noahho in #432
• Improve column selection in RemoveEmptyFeaturesEncoderStep by @benraha in #468
• Ensure tests parameters are in deterministic order. by @oscarkey in #472
• Rename public methods to remove "_" prefix. And fix typo in method name. by @oscarkey in #476
• Bump astral-sh/setup-uv from 6.5.0 to 6.6.0 by @dependabot[bot] in #467
• Use generated data instead of "California housing" in tests. by @oscarkey in #481
• Declare joblib dependency. by @oscarkey in #477
• Handle arbitrary CI environment variable values by @LeoGrin in #483
• Add support for uneven splits for large data by @anuragg1209 in #494
• Forward cache_trainset_representation to load_model. by @priorphil in #492
• Add note to softmax_temperature for the classifier by @brendan-priorlabs in #497
• Add consistency checks for the low_memory and fit_with_cache modes. by @oscarkey in #498
• Bump actions/setup-python from 5.6.0 to 6.0.0 by @dependabot[bot] in #489
• Bump astral-sh/setup-uv from 6.6.0 to 6.6.1 by @dependabot[bot] in #490
• Set up methods for anonymous usage analytics by @safaricd in #499
• QKV calculation improvements in attention mechanism by @benraha in #488
• Select multiple gpus for inference, if available. Don't use them yet. by @oscarkey in #496
• Evaluate the estimators in parallel. by @oscarkey in #484
• Increase versioning for TabPFN by @safaricd in #504
• refactor preprocessors by @bejaeger in #503
• Update classifier model file link in README by @anuragg1209 in #500
• Increase versioning from 2.1.4 to 2.2.0 by @safaricd in #505
• Update README to test workflow by @bejaeger in #511
• Add TELEMETRY.md file by @safaricd in #512
• Resolved issue saving & loading the predictor (Regressor/Classifier) model on different devices (cuda / cpu) by @MagnusBuehler in #486
• If device="auto" and multiple GPUs are present, only select the first. by @oscarkey in #517
• Increase tolerance for test_normalize_data_basic test by @MagnusBuehler in #515
• Fix synchronisation bug in multi-gpu inference. by @oscarkey in #524
• Run the tests on merge to main in addition to in PRs. by @oscarkey in #519
• relax type for features_per_group by @LeoGrin in #529
• Disable all telemetry in pytest by @safaricd in #527
• Workflow Image by @klemens-floege in #532
• [Resolved Merge Conflicts from PR 312] Remove shell download script and fix fallback URLs by @MagnusBuehler in #520
• Add global seed fixture for reproducible tests by @MagnusBuehler in #516
• Bump max Sklearn version to < 1.8 by @klemens-floege in #534
• Switch to uv's "--resolution" option to install minimum dependency set. by @oscarkey in #535
• Re-enable onnx tests. by @oscarkey in #526
• Manage CI dependencies in pyproject and specify numpy dependency. by @oscarkey in #541
• Use a Generator object to specify the seed for the random embeddings. by @oscarkey in #525
• Add CODEOWNERS file for review team assignment by @noahho in #540
• Bump mypy from 1.17.1 to 1.18.2 by @dependabot[bot] in #522
• Fixed a formatting issue in README.md by @azurechen97 in #542
• Fix potential race condition in InferenceEngineOnDemand. by @oscarkey in #543
• Bump astral-sh/setup-uv from 6.6.1 to 7.1.0 by @dependabot[bot] in #547
• Enable MPS devices on the CI, and fix associated bugs. by @oscarkey in #518
• Add a note on creating/linking issues before opening a PR in the PR template. by @priorphil in #548
• Update Ruff to 0.14. by @oscarkey in #549
• Introduce method that returns raw logits for individual estimators by @bejaeger in #550
• Add KV cache example for fast predictions by @noahho in #553
• Update README.md by @noahho in #552
• Explicitly disable Dynamo onnx exports to fix CI. by @oscarkey in #554
• Add an n_preprocessing_jobs option to the classifier and regressor. by @oscarkey in #555
• Fix build error on Python 3.9 by @brendan-priorlabs in #562
• Bugfix: improve logics for handling regression on constant target values by @liam-sbhoo in #564
• Bump astral-sh/setup-uv from 7.1.0 to 7.1.2 by @dependabot[bot] in #566
• Fix column reordering in process_text_na_dataframe with nan and missing values due to Columntransformer by @MagnusBuehler in #539
• Update pyobjc-framework-metal requirement from <12.0 to <13.0 by @dependabot[bot] in #567
• Check for incompatible licenses by @brendan-priorlabs in #561
• Fix cuda configuration error. by @oscarkey in #570
• Fix precision of output in temperature=1.0 case by @bejaeger in #569
• Allow to run Classifier and Regressor with multiple models by @bejaeger in #557
• Rename _resolve_model_path to resolve_model_path by @bejaeger in #573
• Loading files from UCI instead of openml now for stability by @noahho in #572
• Fix KDITransformerWithNaN.fit in case of all-nan column. by @priorphil in #574
• Add concurrency control to CI workflow by @noahho in #568
• Rename ModelInterfaceConfig to InferenceConfig. by @oscarkey in #575
• Load the inference config from the checkpoint, if present. by @oscarkey in #576
• Add unit tests for inference config. by @oscarkey in #578
• Add kdi transform to deps. by @priorphil in #580
• Update huggingface-hub requirement from <1,>=0.0.1 to >=0.0.1,<2 by @dependabot[bot] in #581
• Changes to support TabPFN 2.5, once it is released. by @oscarkey in #579
• Check that inference_config is actually set when loading. by @priorphil in #586
• Introduce option to calibrate temperature and tune decision thresholds by @bejaeger in #584
• Remove type restriction on REGRESSION_Y_PREPROCESS_TRANSFORMS. by @oscarkey in #587
• More robust performance test by @bejaeger in #588
• Update 'balance probabilities test' and 'get_embeddings' test by @bejaeger in #589
• Fix seed for random embeddings on CPU & MPS by @bejaeger in #592
• Changed OrderPreservingColumnTransformer back to ColumnTransformer by @magnus-priorlabs in #596
• Increase the default and v2.5 pretraining limits. by @oscarkey in #597
• Update the device doc string to include multi-gpu inference. by @oscarkey in #599
• Parse model version from model_path by @brendan-priorlabs in #585
• Forward HF errors as they explain gating / missing tokens nicely. by @priorphil in #600
• Set `ensure_y...

V2.1.3

What's Changed

• Bump mypy from 1.17.0 to 1.17.1 by @dependabot[bot] in #434
• fix saving and loading model including weights by @spirousschuh in #420
• raise an error if the model in TabPFNClassifier doesn't have encoder … by @LeoGrin in #439
• Fix GQA api detection on PyTorch<2.5. by @oscarkey in #438
• Adding GPU testing to our CI pipeline by @noahho in #440
• Gpu runner #2 by @noahho in #441
• Fix validation error when .env file contains settings from other apps. by @oscarkey in #446
• Ignore notebooks in the Ruff linter. by @oscarkey in #450
• Make the Gemini review bot less verbose by disabling the summary. by @oscarkey in #452
• Enable type annotation rules in the Ruff linter. by @oscarkey in #451
• Test all models by @noahho in #449
• Add additional finetuned model checkpoints to model loading by @anuragg1209 in #462
• Update CHANGELOG.md by @noahho in #463
• bump package version by @noahho in #464
• CPU attention mechanism using PyTorch implementation by @benraha in #459

New Contributors

• @spirousschuh made their first contribution in #420

Full Changelog: v2.1.1...v2.1.3