build(deps): bump tabpfn from 9.0.0 to 9.1.0

[dependabot]

Bumps tabpfn from 9.0.0 to 9.1.0.

Release notes

Sourced from tabpfn's releases.

v9.1.0

Breaking Changes

• Remove the unused cache_trainset_representation argument from ArchitectureModule.get_architecture, load_model and load_model_criterion_config, and the fit_mode argument from initialize_tabpfn_model. (#1250)
• Replace ClassifierModelSpecs, RegressorModelSpecs, and BaseModelSpecs with the unified ModelSpecs dataclass. Update imports, constructors, and task-specific type checks; retain norm_criterion for legacy regression models without embedded borders. (#1315)

Added

• Add kv_cache_precision="adaptive". (#1299)
• Add a shared ModelSpecs dataclass for in-memory classification and regression. Regression automatically derives its distribution from model.regression_borders; legacy models without embedded borders still require norm_criterion. (#1315)

Changed

• Faster modality detection on wide inputs: a numeric array's distinct values are counted for all columns at once instead of per column, a numeric column stored as object is recognized with pd.api.types.infer_dtype instead of a value-by-value walk, and the nullable-dtype coercion is decided once per dtype. The inferred modalities are unchanged. (#1255)
• Building a model from a checkpoint no longer runs torch's random parameter initialisation, which made up most of the per-fit model construction time. (#1257)
• SAMPLE_SUBSAMPLING_METHOD="majority_downsample" now corrects the target prior shift it introduces, so predicted probabilities and regression distributions are calibrated to the training data rather than to the downsampled context. predict_proba_batched and predict_batched raise with this sampler; score datasets individually instead. (#1270)
• Relabel the README architecture and attention diagrams as TabPFN-3.5. (#1282)
• • examples/finetune_regressor.py now fine-tunes on the bigger OpenML diamonds dataset. (#1290)
• n_preprocessing_jobs > 1 now parallelises the per-estimator preprocessing over threads instead of processes, which we have found to be faster. Parallelisation is turned off by default, as previously. (#1295)
• Link the TabPFN-3.5 technical report and add its BibTeX entry to the README citation section. (#1296)
• • Batch estimators. (#1312)
• Estimators fitted with fit_mode="fit_with_cache" use less host memory and save to smaller files. Predictions are unchanged. (#1323)
• Note TabPFN-3.5-Fast alongside TabPFN-3 and TabPFN-3.5 in the README CPU sample-limit description. (#1328)

Fixed

• Fix FullSupportBarDistribution CDF, quantiles, and sampling to match its half-normal tails while preserving batch shape, device, and dtype. (#1215)
• The opt-in built-model cache (TABPFN_MODEL_CACHE_SIZE) now keys on device and inference precision, so estimators with different settings no longer share one model. (#1219)
• numpy StringDType arrays are accepted as input at fit and predict, like unicode and object arrays (numpy 2.5 or newer). (#1269)
• Regression density and likelihood evaluation reports far fewer infinite NLLs for targets in a narrow bucket or in a distribution's tail. (#1288)
• Fixed FinetunedTabPFNRegressor computing the loss against z-scored targets for ensemble members that use a target transform (such as the default safepower); their loss now uses the transformed targets they predict in. (#1290)
• MPS support and GQA attention now detect the installed torch version correctly when torch.__version__ is a plain string. (#1301)
• Raise a proper validation error when the training inputs have mismatched lengths or are not two-dimensional, so these mistakes are reported as user errors like all other input problems (#1307)
• Fixed TabPFN-3.5 jumping too far when it predicts past the range of a feature, most visibly on a time feature whose values repeat, such as a year column with one row per month. Predictions change for any table with values outside the range seen at fit. (#1310)
• TabPFNRegressor no longer fails or returns meaningless predictions when the target's mean is large relative to its spread (e.g. ID-like values around 1e10). (#1311)
• Allow fitted estimator archives to save supported dataclass-valued initialization parameters such as InferenceConfig. (#1321)
• Forced float16 squashing preprocessing no longer squashes a finite extreme value to zero. (#1322)
• Saving a fitted estimator no longer copies the model weights, so it uses less memory, and saving a fit_mode="fit_with_cache" estimator no longer briefly removes its caches while another thread may be predicting. (#1325)
• Fixed predict raising could not convert string to float when a column was constant or all-missing at fit and held a string at predict. Such a column now gets its fit-time value back at predict, so it carries no information there. (#1329)
• Avoid a runtime C compiler requirement in batched v3 and v3.5 regression target encoding on recent PyTorch versions. (#1338)

Changelog

Sourced from tabpfn's changelog.

[9.1.0] - 2026-10-02

Breaking Changes

• Remove the unused cache_trainset_representation argument from ArchitectureModule.get_architecture, load_model and load_model_criterion_config, and the fit_mode argument from initialize_tabpfn_model. (#1250)
• Replace ClassifierModelSpecs, RegressorModelSpecs, and BaseModelSpecs with the unified ModelSpecs dataclass. Update imports, constructors, and task-specific type checks; retain norm_criterion for legacy regression models without embedded borders. (#1315)

Added

• Add kv_cache_precision="adaptive". (#1299)
• Add a shared ModelSpecs dataclass for in-memory classification and regression. Regression automatically derives its distribution from model.regression_borders; legacy models without embedded borders still require norm_criterion. (#1315)

Changed

• Faster modality detection on wide inputs: a numeric array's distinct values are counted for all columns at once instead of per column, a numeric column stored as object is recognized with pd.api.types.infer_dtype instead of a value-by-value walk, and the nullable-dtype coercion is decided once per dtype. The inferred modalities are unchanged. (#1255)
• Building a model from a checkpoint no longer runs torch's random parameter initialisation, which made up most of the per-fit model construction time. (#1257)
• SAMPLE_SUBSAMPLING_METHOD="majority_downsample" now corrects the target prior shift it introduces, so predicted probabilities and regression distributions are calibrated to the training data rather than to the downsampled context. predict_proba_batched and predict_batched raise with this sampler; score datasets individually instead. (#1270)
• Relabel the README architecture and attention diagrams as TabPFN-3.5. (#1282)
• • examples/finetune_regressor.py now fine-tunes on the bigger OpenML diamonds dataset. (#1290)
• n_preprocessing_jobs > 1 now parallelises the per-estimator preprocessing over threads instead of processes, which we have found to be faster. Parallelisation is turned off by default, as previously. (#1295)
• Link the TabPFN-3.5 technical report and add its BibTeX entry to the README citation section. (#1296)
• • Batch estimators. (#1312)
• Estimators fitted with fit_mode="fit_with_cache" use less host memory and save to smaller files. Predictions are unchanged. (#1323)
• Note TabPFN-3.5-Fast alongside TabPFN-3 and TabPFN-3.5 in the README CPU sample-limit description. (#1328)

Fixed

• Fix FullSupportBarDistribution CDF, quantiles, and sampling to match its half-normal tails while preserving batch shape, device, and dtype. (#1215)
• The opt-in built-model cache (TABPFN_MODEL_CACHE_SIZE) now keys on device and inference precision, so estimators with different settings no longer share one model. (#1219)
• numpy StringDType arrays are accepted as input at fit and predict, like unicode and object arrays (numpy 2.5 or newer). (#1269)
• Regression density and likelihood evaluation reports far fewer infinite NLLs for targets in a narrow bucket or in a distribution's tail. (#1288)
• Fixed FinetunedTabPFNRegressor computing the loss against z-scored targets for ensemble members that use a target transform (such as the default safepower); their loss now uses the transformed targets they predict in. (#1290)
• MPS support and GQA attention now detect the installed torch version correctly when torch.__version__ is a plain string. (#1301)
• Raise a proper validation error when the training inputs have mismatched lengths or are not two-dimensional, so these mistakes are reported as user errors like all other input problems (#1307)
• Fixed TabPFN-3.5 jumping too far when it predicts past the range of a feature, most visibly on a time feature whose values repeat, such as a year column with one row per month. Predictions change for any table with values outside the range seen at fit. (#1310)
• TabPFNRegressor no longer fails or returns meaningless predictions when the target's mean is large relative to its spread (e.g. ID-like values around 1e10). (#1311)
• Allow fitted estimator archives to save supported dataclass-valued initialization parameters such as InferenceConfig. (#1321)
• Forced float16 squashing preprocessing no longer squashes a finite extreme value to zero. (#1322)
• Saving a fitted estimator no longer copies the model weights, so it uses less memory, and saving a fit_mode="fit_with_cache" estimator no longer briefly removes its caches while another thread may be predicting. (#1325)
• Fixed predict raising could not convert string to float when a column was constant or all-missing at fit and held a string at predict. Such a column now gets its fit-time value back at predict, so it carries no information there. (#1329)
• Avoid a runtime C compiler requirement in batched v3 and v3.5 regression target encoding on recent PyTorch versions. (#1338)

Commits

• 0b1a081 Release v9.1.0 (#1343)
• d07a55c test(preprocessing): mark the inf-mask timing guards as slow (#1337)
• 8f2053d fix(regressor): keep predictions exact for targets far from zero (#1311)
• 6e4ee78 fix: avoid runtime compilation in regression target encoding (ENG-1093) (#1338)
• f21beb2 Bump anthropics/claude-code-action from 1.0.231 to 1.0.236 (#1333)
• 8ed2398 Restore fit-time values of constant columns at predict (#1329)
• d9edde5 Bound out-of-range ECDF ranks by the train extremes' midranks (#1310)
• ba6ea31 fix(torch): prevent float16 squashing overflow (#1322)
• 09b4188 Run equal-shape ensemble members as one batched forward pass [RES-3045] (#1312)
• dbba403 [Automated PR (update-readme)]: Include TabPFN-3.5-Fast in the README CPU sam...
• Additional commits viewable in compare view

Dependabot will resolve any conflicts with this PR as long as you don't alter it yourself. You can also trigger a rebase manually by commenting @dependabot rebase.

---

Dependabot commands and options

You can trigger Dependabot actions by commenting on this PR:

• @dependabot rebase will rebase this PR
• @dependabot recreate will recreate this PR, overwriting any edits that have been made to it
• @dependabot show <dependency name> ignore conditions will show all of the ignore conditions of the specified dependency
• @dependabot ignore this major version will close this PR and stop Dependabot creating any more for this major version (unless you reopen the PR or upgrade to it yourself)
• @dependabot ignore this minor version will close this PR and stop Dependabot creating any more for this minor version (unless you reopen the PR or upgrade to it yourself)
• @dependabot ignore this dependency will close this PR and stop Dependabot creating any more for this dependency (unless you reopen the PR or upgrade to it yourself)

[amrit110]

Automated fix applied and PR merged

The agentic fix loop successfully fixed this PR and merged it.

✓ Successfully fixed merge_only failures - Modified 0 files - Executed 89 agent actions - (56 info, 15 tool_call, 1 error, 12 tool_result, 5 reasoning)

View PR history on dashboard

AI Engineering Maintenance Bot