v8.5.0

Breaking Changes

• Cache the decoder keys instead of the train embeddings. get_embeddings(model, X_test, data_source="train") is not supported with cached infernce anymore. (#1189)
• tabpfn.model_loading.download_all_models() now raises an exception if one or more of the models fails to download. It will still download all possible models before raising the exception. (#1195)

Added

• fit() now recognizes a date-like string column internally, though nothing yet expands it into calendar features: it is still read as a plain category or text. We also added InferenceConfig.MIN_CARDINALITY_FOR_TEXT, to differentiate between category-vs-text and category-vs-number decisions; they default to the same value, since we are still not handling text. (#1205)

Changed

• Reduced peak host memory during preprocessing: the ensemble preprocessor no longer rebuilds the feature matrix in steps that cannot change it, taking transient RSS from 42.7 GB to 12.0 GB (-72%), and wall time with it, on a 666,667 x 2,000 float64 fit. Preprocessed outputs are unchanged. (#1186)
• Reduced peak host memory during preprocessing for tables with categorical columns: the reshape and ordinal-encoding steps no longer rebuild the feature matrix to reorder it or to encode part of it, taking transient RSS from 3.33 GB to 2.80 GB (-16%) on a 333,333 x 400 half-categorical fit. Preprocessed outputs are unchanged. (#1187)
• Speed up modality detection on large string columns. Deciding whether a column holds numbers or dates now stops at the first value that does not parse within a 1024-row prefix, instead of parsing every row first. Detection of a 1-million-row free-text column drops from roughly 14 seconds to under 20 milliseconds; the answers are unchanged. (#1208)

Fixed

• Fix an "illegal memory access" crash in the backward pass when fine-tuning on large batches: FlashAttention's backward indexes its workspace with 32-bit integers, so the batch is now chunked to keep each call inside that range. (#1184)
• Fix fit_mode="fit_with_cache" raising TypeError: forward() missing 1 required positional argument: 'task_type' for architectures whose forward pass takes a task_type: the KV cache build now forwards it, like the prediction paths already did. (#1197)
• fit() no longer crashes the interpreter outright on a table whose text column holds a hash-like value such as "8e2569614270f3d8b9e7038efac9f116". Modality detection asked pandas.to_numeric whether a column was numeric; below pandas 3.0 that function has a signed 32-bit integer overflow in its scientific-notation parser and segfaults on a string whose exponent lands in [2**31, 2**32) (pandas#63650, fixed upstream in pandas 3.0). A segfault cannot be caught with try/except, so below pandas 3.0 the check now reads one value at a time with Python's built-in float, which does not share the bug. On pandas 3.0 and later the check is unchanged. (#1203)
• Fix TabPFNRegressor rejecting a checkpoint whose config also describes a classification head: the criterion now follows the task the estimator is built for rather than being inferred from max_num_classes. Loading a regression checkpoint into TabPFNClassifier now raises instead of silently building an unused bar distribution. (#1204)