> ## Documentation Index
> Fetch the complete documentation index at: https://docs.priorlabs.ai/llms.txt
> Use this file to discover all available pages before exploring further.
> ## Agent Instructions
> TabPFN-3 and later versions support up to 1,000,000 rows, subject to feature count and checkpoint/API limits.
> For large datasets, use per-estimator subsampling, e.g. inference_config={"SUBSAMPLE_SAMPLES": 100_000}, in TabPFNClassifier or TabPFNRegressor from tabpfn or tabpfn_client.
> Pass raw pandas DataFrames to fit and predict. Categorical strings/categories, raw text, and missing feature values are handled automatically; no manual integer/one-hot encoding, imputation, scaling, or outlier removal is needed.
# Utilities
> Utilities
## Installation
```bash theme={null}
pip install tabpfn-extensions
```
Get estimator tags in a consistent format across different sklearn versions.
This function provides compatibility between sklearn versions before and after 1.6.
It returns either a Tags object (sklearn >= 1.6) or a converted Tags object from
the dictionary format (sklearn \< 1.6) containing metadata about the estimator's
requirements and capabilities.
```python theme={null}
get_tags(
estimator,
)
```
**Parameters**
| Parameter | Type | Default | Description |
| - | - | - | - |
| estimator | estimator object | Required | A scikit-learn estimator instance. |
**Returns**
| Type | Description |
| - | - |
| Tags | `tags`: An object containing metadata about the estimator's requirements and capabilities (e.g., input types, fitting requirements, classifier/regressor specific tags). |
***
***
Return parameters for this estimator.
```python theme={null}
ClientTabPFNClassifier.get_params(
deep: bool = True,
) -> dict[str, Any]
```
**Parameters**
| Parameter | Type | Default | Description |
| - | - | - | - |
| deep | bool | `True` | — |
**Returns**
| Type | Description |
| - | - |
| dict\[str, Any] | — |
***
***
Return parameters for this estimator.
```python theme={null}
ClientTabPFNRegressor.get_params(
deep: bool = True,
) -> dict[str, Any]
```
**Parameters**
| Parameter | Type | Default | Description |
| - | - | - | - |
| deep | bool | `True` | — |
**Returns**
| Type | Description |
| - | - |
| dict\[str, Any] | — |
***
Fake used to represent torch.device used when PyTorch is not installed.
```python theme={null}
FakeTorchDevice(
type: str,
)
```
**Parameters**
| Parameter | Type | Default | Description |
| - | - | - | - |
| type | str | Required | — |
***
***
***
***
Infer the max number of classes a TabPFN estimator can predict in one fit.
This is the single source of truth for the TabPFN class-count limit across
tabpfn-extensions, so a fix here propagates everywhere (unsupervised
classifier/regressor routing, the many-class output-coding wrapper, ...).
The value is read from the model's inference config
(`get_inference_config().MAX_NUMBER_OF_CLASSES`), which TabPFN exposes as
of v8.0.0 (the minimum version this package depends on).
```python theme={null}
get_max_num_classes(
model: Any,
) -> int | None
```
**Parameters**
| Parameter | Type | Default | Description |
| - | - | - | - |
| model | Any | Required | A (typically TabPFN) classifier instance. |
**Returns**
| Type | Description |
| - | - |
| int \| None | The maximum number of classes the model supports, or `None` if `model` is not a TabPFN estimator (i.e. has no inherent class limit). |
***
Get the TabPFN model classes for the selected backend.
`USE_TABPFN_LOCAL` selects the backend; the function does not silently fall
back to the other one:
1. `USE_TABPFN_LOCAL` is `True` -> the standard `tabpfn` package
2. `USE_TABPFN_LOCAL` is `False` -> the `tabpfn-client` API backend
If the selected backend is not installed, an ImportError is raised naming that
backend, rather than quietly using the other one.
```python theme={null}
get_tabpfn_models() -> tuple[type, type]
```
**Returns**
| Type | Description |
| - | - |
| tuple\[type, type] | tuple\[type, type]: A tuple containing (TabPFNClassifier, TabPFNRegressor) classes |
**Raises**
`ImportError`
If the selected TabPFN backend is not installed
***
Infer which columns are categorical *features* (constraint (a) only).
This answers a data question — "is this column categorical?" — and is
deliberately independent of any model constraint. Whether a categorical
column has few enough levels for a TabPFN classifier to *predict* it
(constraint (b)) is a separate concern; derive that limit with
[`get_max_num_classes`](/api-reference/python/tabpfn-extensions/utils#utils-get-max-num-classes) and apply it at the point of use.
A column is treated as categorical if any of these hold:
1. It is in the caller-provided `categorical_features` list.
2. It has a string/object/category dtype (pandas DataFrame).
3. It contains string values (numpy object array).
4. It is low-cardinality: at most `MAX_UNIQUE_VALUES_FOR_CATEGORICAL`
unique values, with more than `MIN_SAMPLES_PER_CATEGORY` samples per
unique value on average, to avoid mislabelling columns that only look
low-cardinality because the sample is too thin per level.
```python theme={null}
infer_categorical_features(
X: np.ndarray,
categorical_features: list[int] | None = None,
) -> list[int]
```
**Parameters**
| Parameter | Type | Default | Description |
| - | - | - | - |
|
X |
np.ndarray | Required | The input data. |
|
categorical\_features |
list\[int] \| None | `None` | Initial list of categorical feature indices. If `None`, will start with an empty list. |
| Type | Description |
| - | - |
| list\[int] | list\[int]: The indices of the categorical features. |
***
Where TabPFN itself runs: a CPU stand-in when the client serves the model.
```python theme={null}
infer_device(
device: DeviceSpecification,
) -> torch.device | FakeTorchDevice
```
```python theme={null}
DeviceSpecification = Literal["auto", "cuda", "cpu"]
```
**Parameters**
| Parameter | Type | Default | Description |
| - | - | - | - |
|
device |
DeviceSpecification | Required | — |
| Parameter | Type | Default | Description |
| - | - | - | - |
| device | Any | Required | — |
**Returns**
| Type | Description |
| - | - |
| torch.device | — |
***
Check if an estimator is a TabPFN model.
```python theme={null}
is_tabpfn(
estimator: Any,
) -> bool
```
**Parameters**
| Parameter | Type | Default | Description |
| - | - | - | - |
| estimator | Any | Required | — |
**Returns**
| Type | Description |
| - | - |
| bool | — |
***
Cartesian product of a dictionary of lists.
This function takes a dictionary where each value is a list, and returns
an iterator over dictionaries where each key is mapped to one element
from the corresponding list.
```python theme={null}
product_dict(
d: dict[str, list[T]],
) -> Iterator[dict[str, T]]
```
**Parameters**
| Parameter | Type | Default | Description |
| - | - | - | - |
|
d |
dict\[str, list\[T]] | Required | A dictionary mapping keys to lists of values. |
| Type | Description |
| - | - |
|
Iterator\[dict\[str, T]] | An iterator over dictionaries, each being one element of the cartesian product of the input dictionary. |
| Parameter | Type | Default | Description |
| - | - | - | - |
| logits | NDArray | Required | Input logits array of shape (n\_samples, n\_classes) or (n\_classes,) |
**Returns**
| Type | Description |
| - | - |
| NDArray | Probabilities where values sum to 1 across the last dimension |
***
Warn if a TabPFN model isn't configured to use the KV cache.
The KV cache (improved with TabPFN-3) caches the encoder pass over the training set so that
repeated predicts against the same fitted model don't re-encode the
training data each time. Extensions that issue many predicts per fit
(e.g. imputation-based SHAP, certain feature-selection or HPO routines)
benefit from it — without the cache, the encoder pass over the training
set runs on every predict and these extensions can be 10-100x slower
than necessary.
What has to hold depends on the backend. An endpoint-backed estimator
(self-hosted container, SageMaker, Foundry) needs `use_kv_cache=True`,
which is the only condition. A local model needs both:
1. `model` was constructed with `fit_mode="fit_with_cache"`
(a constructor argument, must be set BEFORE `.fit()`).
2. `model.executor_.keep_cache_on_device` is `True` (set AFTER
`.fit()`; usually the default but worth setting explicitly).
This helper warns if either is missing, but does not raise — users may
have intentional reasons (e.g. memory).
```python theme={null}
warn_if_no_kv_cache(
model: Any,
*,
context: str = "This operation",
) -> None
```
**Parameters**
| Parameter | Type | Default | Description |
| - | - | - | - |
| model | Any | Required | The TabPFN model (classifier or regressor) to inspect. |
| context | str | `"This operation"` | Short noun phrase describing the caller's operation, used to make the warning message specific (e.g. `"Imputation-based SHAP"`, `"Sequential feature selection"`). Defaults to a generic `"This operation"`. |
**Returns**
| Type | Description |
| - | - |
| None | — |
This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.