Install & Compatibility
Where this runs
tested against v? · pip install
Install × environment matrix
Each cell = how many times install + import succeeded across repeated harness runs. Partial = flaky.
glibc = Debian/Ubuntu slim · musl = Alpine Linux
muslpy 3.10–3.95 runs
build_error
Code
Verified usage
Verified import paths — ran on the pinned version, not inferred.
TabularPredictor
✓ from autogluon.tabular import TabularPredictor
While autogluon-core is the underlying package, users typically import high-level predictor classes from specific AutoGluon modules like autogluon.tabular, autogluon.timeseries, or autogluon.multimodal for machine learning tasks.
TabularDataset
✓ from autogluon.tabular import TabularDataset
Used for loading and managing tabular data within the AutoGluon framework, often in conjunction with TabularPredictor.
This quickstart demonstrates how to use AutoGluon's TabularPredictor for a classification task. It involves loading a dataset, fitting a predictor by specifying the target column, and then making predictions on new data. AutoGluon automatically handles model selection, hyperparameter tuning, and ensemble creation.
import pandas as pd
from autogluon.tabular import TabularPredictor, TabularDataset
# Create a dummy CSV for demonstration
data = {
'feature1': [1, 2, 3, 4, 5, 6, 7, 8, 9, 10],
'feature2': ['A', 'B', 'A', 'C', 'B', 'A', 'C', 'B', 'A', 'C'],
'target': [0, 1, 0, 1, 0, 1, 0, 1, 0, 1]
}
df = pd.DataFrame(data)
df.to_csv('train.csv', index=False)
# Load data using TabularDataset
train_data = TabularDataset('train.csv')
# Initialize and train a TabularPredictor
label = 'target'
predictor = TabularPredictor(label=label, path='AutogluonModels').fit(train_data)
# Make predictions (example test data)
test_data = TabularDataset(pd.DataFrame({
'feature1': [11, 12, 13],
'feature2': ['A', 'B', 'C']
}))
predictions = predictor.predict(test_data)
print("Predictions:")
print(predictions)
Debug
Known issues
breakingModels trained with an older version of AutoGluon are generally not compatible with newer versions. Users are advised to re-train models after upgrading the library.fixRe-train your models using the new AutoGluon version after upgrading the library.
affects: All major versions (e.g., v0.8.x, v1.0.x, v1.1.x, v1.2.x, v1.3.x, v1.4.x, v1.5.x)
breakingPython 3.8 support was dropped in AutoGluon v1.2.0. Users on Python 3.8 will need to upgrade their Python environment to use v1.2.0 or newer.fixUpgrade Python to version 3.9 or higher (current stable support: 3.10-3.12; 3.13 is experimental).
affects: >=1.2.0
deprecatedSeveral methods of `TabularPredictor` (e.g., `persist_models`, `get_model_names`, `get_pred_from_proba`) were deprecated in v1.0.0, started raising errors in v1.2.0, and were removed in v1.3.0.fixUse the new method names: `persist`, `model_names`, `predict_from_proba` respectively, and consult the release notes for a full list of renames.
affects: >=1.3.0 (removed), >=1.2.0 (raises error)
gotcha`autogluon-core` is a foundational package primarily containing core utilities, searchers, and schedulers for hyperparameter tuning. Most users performing machine learning tasks will interact with higher-level packages like `autogluon.tabular`, `autogluon.timeseries`, or `autogluon.multimodal` (via `pip install autogluon`).fixFor end-to-end ML, install the full `autogluon` package and import specific predictor classes from its submodules (e.g., `from autogluon.tabular import TabularPredictor`).
affects: All versions
gotchaThe `extreme` preset for `TabularPredictor`, introduced in v1.4.0, often requires a CUDA-compatible GPU (ideally with 32+ GB vRAM) for optimal performance and is most effective for datasets with at most 30,000 samples. Inference time can also be longer than with other presets.fixEnsure adequate GPU resources or consider other presets like `best_quality` for larger datasets or CPU-only environments.
affects: >=1.4.0
gotchaPython 3.13 support, introduced in v1.5.0, is currently experimental. Some features may not be available when running on Python 3.13, particularly on Windows.fixFor production or critical workloads, consider using fully supported Python versions (3.10, 3.11, 3.12). If using 3.13, be aware of potential limitations.
affects: >=1.5.0
breakingFor users previously using `"TABPFNV2"` as a model, AutoGluon v1.4.0 (and newer) strongly recommends switching to `"REALTABPFN-V2"` to avoid breaking changes related to the underlying TabPFN releases.fixUpdate model references from `"TABPFNV2"` to `"REALTABPFN-V2"` in your `hyperparameters` configurations.
affects: >=1.4.0
Errors
Common errors & fixes
ModuleNotFoundError: No module named 'autogluon.core'
This error occurs when the 'autogluon-core' package is not installed in the Python environment.
fixInstall the package using pip: 'pip install autogluon-core'.
ImportError: cannot import name 'is_directory' from 'PIL._util'
This error arises due to compatibility issues between AutoGluon and certain versions of the Pillow library.
fixUpgrade Pillow to a compatible version: 'pip install --upgrade Pillow'.
RuntimeError: Attempting to deserialize object on a CUDA device but torch.cuda.is_available() is False.
This error occurs when a model trained on a GPU is being loaded in an environment without GPU support.
fixEnsure that the environment has GPU support enabled, or modify the code to load the model on the CPU.
free(): invalid size
Aborted (core dumped)
This segmentation fault can occur due to an import order issue between AutoGluon modules, particularly when using 'autogluon.text' and 'autogluon.vision'.
fixReverse the import order: import 'autogluon.vision' before 'autogluon.text'.
OSError: [Errno 28] No space left on device
This error indicates that the disk space is insufficient, often encountered when using AutoGluon on environments with limited storage, such as SageMaker Notebooks.
fixIncrease the disk space allocation or clean up unnecessary files to free up space.
Upgrade
Version history
1.6.1latest on PyPI · released Aug 6, 2026
Audit
Dependencies
PythonrequiredRequired Python interpreter version range. (e.g., autogluon-core==1.5.0 requires >=3.10, <3.14).