TabFM Files

Fixed

• PyTorch weight loading: load model.safetensors via PyTorchModelHubMixin. The 1.0.0 loader looked for pytorch_model.bin, which the Hugging Face checkpoint no longer provides, so load() raised FileNotFoundError.
• EnsembleGenerator no longer re-transforms the full training set on every predict call (prediction cost now scales with query size, not context size).
• Query-axis mask/bias collapse in the memory-efficient (FLASH) JAX attention.
• predict on multi-device hosts no longer crashes (IndivisibleError / device mismatch).
• TabFMRegressor.predict before fit now raises NotFittedError.
• TabFMClassifier.predict no longer returns object-dtype labels.
• README regression example now loads the regression checkpoint.
• Loading a checkpoint whose type does not match the estimator now fails fast with an actionable error, instead of a cryptic squeeze error (classification weights in TabFMRegressor) or silently wrong predictions (regression weights in TabFMClassifier).
• The sklearn layer handles duplicate and non-string column names: duplicates fail fast with a clear message, and datetime columns with integer labels no longer crash.
• Fitted estimators and the PyTorch model are now picklable, which AutoGluon / TabArena require to save a fitted model. The PyTorch gelu activation is a module-level function, and the JAX estimators drop their memoized compiled predict functions on pickle and rebuild them lazily.

Changed

• PyTorch model runs in bfloat16 by default, matching the JAX compute dtype.
• Activation chunking is enabled by default to bound peak memory on large tasks.
• JAX and PyTorch models gained Hugging Face Hub support (from_pretrained / save_pretrained); weight downloads are narrowed to the requested model type.