Scikit-learn-intelex (DeepWiki)

https://deepwiki.com/uxlfoundation/scikit-learn-intelex
Admin
Scikit-learn-intelex is an open-source Intel oneDAL-backed accelerator that transparently patches...
Tokens:70,116
Snippets:514
Trust Score:9
Update:3 days ago
Show doc for...
Tokens:
Raw
### Linux venv-based environment setup

Source: https://deepwiki.com/uxlfoundation/scikit-learn-intelex/6-development-guide

Used in GitHub Actions. Creates a Python virtual environment, installs dependencies, activates components, and installs the package with setup.py.

```bash
python -m venv venv
source venv/bin/activate
pip install -r dependencies-dev
source .github/scripts/activate_components.sh
python setup.py install --single-version-externally-managed --record=record.txt
```

--------------------------------

### Development install (editable)

Source: https://deepwiki.com/uxlfoundation/scikit-learn-intelex/5.1-building-from-source

Performs an editable install using setup.py develop with GPU and distributed support disabled.

```bash
# Development install (editable)
NO_DIST=1 NO_DPC=1 python setup.py develop
```

--------------------------------

### Install scikit-learn-intelex from PyPI

Source: https://deepwiki.com/uxlfoundation/scikit-learn-intelex/1.1-quick-start

Installs the scikit-learn-intelex package from PyPI. Requires Python ≥ 3.10 and scikit-learn ≥ 1.0.

```bash
pip install scikit-learn-intelex
```

--------------------------------

### Build documentation locally with build-doc.sh

Source: https://deepwiki.com/uxlfoundation/scikit-learn-intelex/6.3-documentation-build

Run from the repository root after installing pandoc, Python dependencies, and building the library. Output is placed in doc/_build/html/.

```bash
cd doc
./build-doc.sh
```

--------------------------------

### Install via Conda (Intel channel)

Source: https://deepwiki.com/uxlfoundation/scikit-learn-intelex/1.1-quick-start

Creates a conda environment named 'sklex' and installs scikit-learn-intelex from Intel's channel, recommended for Intel® Distribution for Python users.

```bash
conda create -n sklex -c https://software.repos.intel.com/python/conda/ -c conda-forge --override-channels scikit-learn-intelex
conda activate sklex
```

--------------------------------

### Install via Conda (conda-forge channel)

Source: https://deepwiki.com/uxlfoundation/scikit-learn-intelex/1.1-quick-start

Creates a conda environment named 'sklex' and installs scikit-learn-intelex from the conda-forge channel, recommended for general use.

```bash
conda create -n sklex -c conda-forge --override-channels scikit-learn-intelex
conda activate sklex
```

--------------------------------

### Windows conda-based environment setup

Source: https://deepwiki.com/uxlfoundation/scikit-learn-intelex/6-development-guide

Matches the Azure Pipelines WindowsCondaEnv job. Creates a conda environment named CB with Python, impi-devel, and pyyaml, installs dal-devel>=2025.6.1, and builds via conda-recipe\bld.bat.

```bash
conda create -n CB -c conda-forge python=<VERSION> impi-devel pyyaml
call activate CB
pip install -r dependencies-dev
conda install -c conda-forge "dal-devel>=2025.6.1"
call conda-recipe\bld.bat
```

--------------------------------

### Linux conda-based environment setup

Source: https://deepwiki.com/uxlfoundation/scikit-learn-intelex/6-development-guide

Matches the Azure Pipelines LinuxCondaEnv job. Creates a conda environment named CB with Python, mpich, and pyyaml, installs dal-devel>=2025.6.1, and builds via conda-recipe/build.sh.

```bash
conda create -n CB -c conda-forge python=<VERSION> mpich pyyaml
conda activate CB
conda install -c conda-forge "dal-devel>=2025.6.1"   # or set DALROOT
pip install -r dependencies-dev
source /opt/intel/oneapi/compiler/latest/env/vars.sh  # if DPC++ installed
./conda-recipe/build.sh
```

--------------------------------

### Minimal CPU-only build (no GPU, no distributed)

Source: https://deepwiki.com/uxlfoundation/scikit-learn-intelex/5.1-building-from-source

Sets DALROOT to the oneDAL installation path and disables GPU and distributed support via NO_DIST=1 and NO_DPC=1 before running setup.py build.

```bash
# Minimal CPU-only build (no GPU, no distributed)
export DALROOT=/path/to/onedal
NO_DIST=1 NO_DPC=1 python setup.py build
```

--------------------------------

### Install test dependencies and run sklearnex test groups

Source: https://deepwiki.com/uxlfoundation/scikit-learn-intelex/6.2-testing-infrastructure

Install test dependencies and run all sklearnex test groups, producing JSON reports. Individual groups can be run with pytest commands.

```bash
pip install -r requirements-test.txt
```

```bash
bash conda-recipe/run_test.sh --json-report
```

```bash
pytest --pyargs sklearnex
```

```bash
pytest --pyargs onedal
```

```bash
pytest -s tests/
```

--------------------------------

### Example verbose output for DBSCAN.fit

Source: https://deepwiki.com/uxlfoundation/scikit-learn-intelex/1.2-supported-algorithms

Example output from verbose logging showing the accelerated version running on CPU and the fallback to original Scikit-learn for sklearn.cluster.DBSCAN.fit. This output appears when SKLEARNEX_VERBOSE is set to INFO.

```text
INFO:sklearnex: sklearn.cluster.DBSCAN.fit: running accelerated version on CPU
INFO:sklearnex: sklearn.cluster.DBSCAN.fit: fallback to original Scikit-learn
```

--------------------------------

### Run sklearn conformance tests with options

Source: https://deepwiki.com/uxlfoundation/scikit-learn-intelex/6.2-testing-infrastructure

Run sklearn conformance tests, optionally with GPU offloading or preview mode deselections.

```bash
bash .ci/scripts/run_sklearn_tests.sh
```

```bash
bash .ci/scripts/run_sklearn_tests.sh gpu
```

```bash
SKLEARNEX_PREVIEW=YES bash .ci/scripts/run_sklearn_tests.sh
```

--------------------------------

### Full build with GPU and distributed support (Linux)

Source: https://deepwiki.com/uxlfoundation/scikit-learn-intelex/5.1-building-from-source

Requires both DALROOT and I_MPI_ROOT environment variables to be set for a complete build with GPU and distributed support on Linux.

```bash
# Full build with GPU and distributed support (Linux)
export DALROOT=/path/to/onedal
export I_MPI_ROOT=/path/to/mpi
python setup.py build
```

--------------------------------

### Offload computation to GPU with target_offload

Source: https://deepwiki.com/uxlfoundation/scikit-learn-intelex

Use config_context with target_offload to move computations to a SYCL-compatible GPU. The option accepts a device name string like "gpu:0" or a dpctl.SyclQueue object.

```python
from sklearnex import config_context
with config_context(target_offload="gpu:0"):
    clustering = DBSCAN(eps=3, min_samples=2).fit(X)
```

--------------------------------

### GPU execution with config_context and target_offload (with patching)

Source: https://deepwiki.com/uxlfoundation/scikit-learn-intelex/1.1-quick-start

Use config_context with target_offload to run DBSCAN on a GPU device. Requires patching scikit-learn with patch_sklearn() first. Data is automatically moved to the specified device. A model fitted on GPU must also perform prediction/scoring on GPU.

```python
import numpy as np
from sklearnex import patch_sklearn, config_context
patch_sklearn()
 
from sklearn.cluster import DBSCAN
 
X = np.array([[1., 2.], [2., 2.], [2., 3.],
              [8., 7.], [8., 8.], [25., 80.]], dtype=np.float32)
with config_context(target_offload="gpu:0"):
    clustering = DBSCAN(eps=3, min_samples=2).fit(X)
```

--------------------------------

### Parallel build with 8 threads

Source: https://deepwiki.com/uxlfoundation/scikit-learn-intelex/5.1-building-from-source

Runs a parallel build using 8 threads via the --parallel flag, with distributed support disabled.

```bash
# Parallel build with 8 threads
NO_DIST=1 python setup.py build --parallel 8
```

--------------------------------

### sklearnex.basic_statistics.BasicStatistics

Source: https://deepwiki.com/uxlfoundation/scikit-learn-intelex/1.2-supported-algorithms

Basic statistics computation with GPU offload. Triggered via config_context(target_offload="gpu:0") or on-device array API objects.

```APIDOC
## sklearnex.basic_statistics.BasicStatistics

### Description
Basic statistics computation with GPU offload. Triggered via config_context(target_offload="gpu:0") or on-device array API objects.

### Parameters
#### Key Parameter Constraints
- All parameters

### Data Format
- Dense or CSR sparse

### Notes
- No `sample_weight` with CSR
```

--------------------------------

### Conda build script wrapper command

Source: https://deepwiki.com/uxlfoundation/scikit-learn-intelex/5.2-conda-packaging

Both build.sh and bld.bat call this command to delegate compilation to setup.py, which drives CMake, Cython, and pybind11 builds.

```bash
python setup.py install --single-version-externally-managed --record record.txt
```

--------------------------------

### Array API Support

Source: https://deepwiki.com/uxlfoundation/scikit-learn-intelex/1.2-supported-algorithms

A subset of algorithms supports array API-conforming inputs (e.g., dpnp.ndarray, torch.tensor on XPU). Array API must be enabled via config_context(array_api_dispatch=True).

```APIDOC
## Array API Support

### Description
A subset of algorithms supports array API-conforming inputs (e.g., dpnp.ndarray, torch.tensor on XPU). Array API must be enabled via config_context(array_api_dispatch=True).

### Patched Classes with Array API Support
- `sklearnex.basic_statistics.BasicStatistics`
- `sklearnex.basic_statistics.IncrementalBasicStatistics`
- `sklearn.cluster.DBSCAN`
- `sklearn.covariance.EmpiricalCovariance`
- `sklearnex.covariance.IncrementalEmpiricalCovariance`
- `sklearn.decomposition.PCA`
- `sklearn.ensemble.ExtraTreesClassifier`
- `sklearn.ensemble.ExtraTreesRegressor`
- `sklearn.ensemble.RandomForestClassifier`
- `sklearn.ensemble.RandomForestRegressor`
- `sklearn.linear_model.LinearRegression`
- `sklearn.linear_model.Ridge`
- `sklearnex.linear_model.IncrementalLinearRegression`
- `sklearnex.linear_model.IncrementalRidge`
- `sklearn.svm.NuSVC`
- `sklearn.svm.NuSVR`
- `sklearn.svm.SVC`
- `sklearn.svm.SVR`
```

--------------------------------

### GPU execution with config_context and target_offload (without patching)

Source: https://deepwiki.com/uxlfoundation/scikit-learn-intelex/1.1-quick-start

Use config_context with target_offload to run DBSCAN on a GPU device without patching scikit-learn. Import the accelerated estimator directly from sklearnex.cluster. Data is automatically moved to the specified device. A model fitted on GPU must also perform prediction/scoring on GPU.

```python
import numpy as np
from sklearnex import config_context
from sklearnex.cluster import DBSCAN
 
X = np.array([[1., 2.], [2., 2.], [2., 3.],
              [8., 7.], [8., 8.], [25., 80.]], dtype=np.float32)
with config_context(target_offload="gpu:0"):
    clustering = DBSCAN(eps=3, min_samples=2).fit(X)
```

--------------------------------

### Enable verbose logging with SKLEARNEX_VERBOSE=INFO

Source: https://deepwiki.com/uxlfoundation/scikit-learn-intelex/1.2-supported-algorithms

Set the SKLEARNEX_VERBOSE environment variable to INFO to enable verbose logging that shows whether the accelerated or fallback implementation is running. This is useful for verifying which implementation executes during a dispatched call.

```bash
export SKLEARNEX_VERBOSE=INFO
```

--------------------------------

### Command-line patching with python -m sklearnex

Source: https://deepwiki.com/uxlfoundation/scikit-learn-intelex/1.1-quick-start

Run an existing script without modifying it. Patching is applied automatically before the script runs.

```bash
python -m sklearnex my_application.py
```

--------------------------------

### Package identity and versioning in meta.yaml

Source: https://deepwiki.com/uxlfoundation/scikit-learn-intelex/5.2-conda-packaging

Defines the package name and version sourced from the GIT_DESCRIBE_TAG environment variable during repository builds. In feedstock builds, the version is static and the source block points to a GitHub archive.

```yaml
package:
  name: scikit-learn-intelex
  version: {{ environ.get('GIT_DESCRIBE_TAG') }}
```

--------------------------------

### versions.json format

Source: https://deepwiki.com/uxlfoundation/scikit-learn-intelex/6.3-documentation-build

Seed file for the version switcher, containing only the "latest" entry initially. The full file is regenerated by the release script and deployed to GitHub Pages.

```json
[
  { "name": "latest",  "version": "2025.1", "url": "/scikit-learn-intelex/latest/" },
  { "name": "2025.1",  "version": "2025.1", "url": "/scikit-learn-intelex/2025.1/" },
  { "name": "2024.7",  "version": "2024.7", "url": "/scikit-learn-intelex/2024.7/" }
]
```

--------------------------------

### IncrementalBasicStatistics constructor

Source: https://deepwiki.com/uxlfoundation/scikit-learn-intelex/3.5-basic-statistics

Constructor for the incremental estimator. Supports incremental computation via partial_fit calls. batch_size is only used by fit() and defaults to 5 * n_features.

```python
IncrementalBasicStatistics(result_options="all", batch_size=None)
```

--------------------------------

### CMake Invocation Commands for Backend Build

Source: https://deepwiki.com/uxlfoundation/scikit-learn-intelex/5.1-building-from-source

Shows the three subprocess calls run by custom_build_cmake_clib for each interface (host, dpc, spmd_dpc). The -S flag points to the scripts directory, -B to the build directory for the interface, and -j{n_threads} controls parallelism.

```bash
cmake -S scripts/ -B build/backend_{iface}/ [options...]
cmake --build build/backend_{iface}/ -j{n_threads}
cmake --install build/backend_{iface}/
```

--------------------------------

### enable_thread_pinning

Source: https://deepwiki.com/uxlfoundation/scikit-learn-intelex/4-daal4py-interface

Binds threads to physical cores. Disabled by default.

```APIDOC
## enable_thread_pinning

### Description
Binds threads to physical cores. Disabled by default.

### Parameters
- **enabled** (bool) - Optional - Whether to enable thread pinning. Default is True.

```

--------------------------------

### Enable verbose logging

Source: https://deepwiki.com/uxlfoundation/scikit-learn-intelex/1.1-quick-start

Enable verbose logging to confirm which implementation (sklearnex or stock sklearn) is executing a given call. Prints messages indicating whether each operation uses the Intel oneDAL backend or falls back to stock scikit-learn.

```python
import logging
logging.basicConfig(level=logging.INFO)
```

--------------------------------

### LogisticRegression._onedal_fit CPU and GPU paths

Source: https://deepwiki.com/uxlfoundation/scikit-learn-intelex/3-algorithm-implementations

Shows the branch inside _onedal_fit: CPU path delegates to daal4py_fit via _onedal_cpu_fit, GPU path uses the onedal estimator with a queue. The queue parameter is None for CPU and a dpctl.SyclQueue for GPU.

```python
CPU path:  self._onedal_cpu_fit(X, y, sample_weight)   # delegates to daal4py_fit
GPU path:  self._onedal_estimator.fit(X, y, queue=queue)
```

--------------------------------

### BasicStatistics

Source: https://deepwiki.com/uxlfoundation/scikit-learn-intelex/3.5-basic-statistics

Compute basic statistics in a single fit call. Accepts a result_options parameter to control which statistics are computed.

```APIDOC
## BasicStatistics

### Description
Compute basic statistics in a single fit call. Accepts a result_options parameter to control which statistics are computed.

### Constructor Parameters
- **result_options** (str or list of str) - Optional - Controls which statistics are computed. Can be "all", a single statistic name, or a list of statistic names. Default is "all".

### Available Statistics
| result_options value | Output attribute | Description |
| --- | --- | --- |
| "min" | `min_` | Per-feature minimum |
| "max" | `max_` | Per-feature maximum |
| "sum" | `sum_` | Per-feature sum |
| "mean" | `mean_` | Per-feature mean |
| "variance" | `variance_` | Per-feature variance (Bessel's correction) |
| "variation" | `variation_` | Per-feature variation (Bessel's correction) |
| "sum_squares" | `sum_squares_` | Per-feature sum of squares |
| "standard_deviation" | `standard_deviation_` | Per-feature unbiased standard deviation |
| "sum_squares_centered" | `sum_squares_centered_` | Per-feature centered sum of squares |
| "second_order_raw_moment" | `second_order_raw_moment_` | Per-feature second raw moment |
| "all" | all of the above | Computes every statistic |

### Output Attributes
Each output attribute is an `ndarray` of shape `(n_features,)`. An attribute is only present on the fitted estimator if its corresponding option was requested.

### Module
`sklearnex/basic_statistics/basic_statistics.py`

### SPMD Variant
An SPMD (distributed) variant exists at `sklearnex/spmd/basic_statistics/basic_statistics.py`.
```

--------------------------------

### PatchingConditionsChain usage in LinearRegression.fit

Source: https://deepwiki.com/uxlfoundation/scikit-learn-intelex/2.1-patching-and-dispatching

Shows how to use PatchingConditionsChain to decide at runtime whether to use the oneDAL backend or fall back to sklearn. Conditions are (bool, str) tuples; and_conditions requires all true. get_status(logs=True) logs INFO 'running accelerated version on CPU' when enabled, or DEBUG per failing condition and INFO 'fallback to original Scikit-learn' when disabled.

```python
_patching_status = PatchingConditionsChain("sklearnex.linear_model.LinearRegression.fit")
_patching_status.and_conditions([
    (dtype in [np.float32, np.float64], "X not float32 or float64"),
    (not sparse, "sparse input not supported"),
])
if _patching_status.get_status(logs=True):
    # call oneDAL backend
else:
    # call sklearn backend
```

--------------------------------

### sklearn.decomposition.PCA

Source: https://deepwiki.com/uxlfoundation/scikit-learn-intelex/1.2-supported-algorithms

Principal Component Analysis with GPU offload. Triggered via config_context(target_offload="gpu:0") or on-device array API objects.

```APIDOC
## sklearn.decomposition.PCA

### Description
Principal Component Analysis with GPU offload. Triggered via config_context(target_offload="gpu:0") or on-device array API objects.

### Parameters
#### Key Parameter Constraints
- `svd_solver` in `['full','covariance_eigh','onedal_svd']`

### Data Format
- No sparse
```

--------------------------------

### Debug build

Source: https://deepwiki.com/uxlfoundation/scikit-learn-intelex/5.1-building-from-source

Builds with debug symbols by setting DEBUG_BUILD=1 and using the --debug flag, with distributed support disabled.

```bash
# Debug build
NO_DIST=1 DEBUG_BUILD=1 python setup.py build --debug
```

--------------------------------

### Building PatchingConditionsChain for LogisticRegression.fit

Source: https://deepwiki.com/uxlfoundation/scikit-learn-intelex/3-algorithm-implementations

Accumulates boolean conditions and reason strings for a named operation, then logs 'running accelerated version on CPU/GPU' when all conditions are True, or 'fallback to original Scikit-learn' with the first failing reason when any is False. Observable via SKLEARNEX_VERBOSE=INFO and matched by tests in sklearnex/tests/test_patching.py137-143.

```python
# Building conditions
patching_status = PatchingConditionsChain(
    "sklearn.linear_model.LogisticRegression.fit"
)
patching_status.and_conditions([
    (self.penalty == "l2",          "Only l2 penalty is supported."),
    (self.dual == False,            "dual=True is not supported."),
    (sample_weight is None,         "Sample weight is not supported."),
    (target_type == "binary",       "Only binary classification is supported"),
])
```

--------------------------------

### daalfini

Source: https://deepwiki.com/uxlfoundation/scikit-learn-intelex/4-daal4py-interface

Finalizes the MPI environment (MPI_Finalize). Required when using distributed mode without mpi4py.

```APIDOC
## daalfini

### Description
Finalises the MPI environment (MPI_Finalize). Required when using distributed mode without mpi4py.

```

--------------------------------

### sklearn.cluster.KMeans

Source: https://deepwiki.com/uxlfoundation/scikit-learn-intelex/1.2-supported-algorithms

K-Means clustering with GPU offload. Triggered via config_context(target_offload="gpu:0") or on-device array API objects.

```APIDOC
## sklearn.cluster.KMeans

### Description
K-Means clustering with GPU offload. Triggered via config_context(target_offload="gpu:0") or on-device array API objects.

### Parameters
#### Key Parameter Constraints
- `algorithm='lloyd'`; `n_clusters!=1`; uniform weights; `init='k-means++'` falls back to CPU

### Data Format
- None
```

--------------------------------

### table.__dlpack__

Source: https://deepwiki.com/uxlfoundation/scikit-learn-intelex/2.3-data-conversion

Exports the table as a DLPack capsule, allowing other frameworks (PyTorch, JAX, array_api_strict) to consume the table directly. Supports specifying stream, max version, device, and copy options.

```APIDOC
## table.__dlpack__

### Description
Exports the table as a DLPack capsule, allowing other frameworks (PyTorch, JAX, array_api_strict) to consume the table directly. Supports specifying stream, max version, device, and copy options.

### Method
__dlpack__

### Parameters
- **stream** - Optional - Stream for asynchronous operations
- **max_version** - Optional - Maximum DLPack version supported
- **dl_device** - Optional - Target device for the export (e.g., (kDLCPU, 0) to move to host)
- **copy** - Optional - Whether to copy the data

### Response
Returns a DLPack capsule representing the table data.
```

--------------------------------

### BasicStatistics

Source: https://deepwiki.com/uxlfoundation/scikit-learn-intelex/1.2-supported-algorithms

Estimator unique to sklearnex for computing basic statistics. Supports result_options parameter and array API.

```APIDOC
## BasicStatistics

### Description
Estimator unique to sklearnex for computing basic statistics. Supports result_options parameter and array API.

### Class
`sklearnex.basic_statistics.BasicStatistics`

### Parameters
- `result_options` - Supported parameter

### Supported Platforms
- CPU: Yes
- GPU: Yes
- SPMD: Yes
```

--------------------------------

### In-script patching with patch_sklearn

Source: https://deepwiki.com/uxlfoundation/scikit-learn-intelex/1.1-quick-start

Call patch_sklearn() before importing sklearn modules to route supported estimators to accelerated versions. Unsupported algorithms fall back silently to stock scikit-learn.

```python
from sklearnex import patch_sklearn
patch_sklearn()
 
from sklearn.cluster import DBSCAN   # now routes to sklearnex
import numpy as np
 
X = np.array([[1., 2.], [2., 2.], [2., 3.],
              [8., 7.], [8., 8.], [25., 80.]], dtype=np.float32)
clustering = DBSCAN(eps=3, min_samples=2).fit(X)
```

--------------------------------

### table.__dlpack_device__

Source: https://deepwiki.com/uxlfoundation/scikit-learn-intelex/2.3-data-conversion

Returns the device type and ID where the table's underlying memory resides. Returns either kDLCPU or kDLOneAPI depending on the memory location.

```APIDOC
## table.__dlpack_device__

### Description
Returns the device type and ID where the table's underlying memory resides. Returns either kDLCPU or kDLOneAPI depending on the memory location.

### Method
__dlpack_device__

### Response
Returns a tuple (device_type, device_id) where device_type is either kDLCPU or kDLOneAPI.
```

--------------------------------

### IncrementalBasicStatistics

Source: https://deepwiki.com/uxlfoundation/scikit-learn-intelex/3.5-basic-statistics

Compute statistics incrementally via partial_fit batches. Accepts the same result_options parameter as BasicStatistics.

```APIDOC
## IncrementalBasicStatistics

### Description
Compute statistics incrementally via partial_fit batches. Accepts the same result_options parameter as BasicStatistics.

### Constructor Parameters
- **result_options** (str or list of str) - Optional - Controls which statistics are computed. Can be "all", a single statistic name, or a list of statistic names. Default is "all".

### Available Statistics
| result_options value | Output attribute | Description |
| --- | --- | --- |
| "min" | `min_` | Per-feature minimum |
| "max" | `max_` | Per-feature maximum |
| "sum" | `sum_` | Per-feature sum |
| "mean" | `mean_` | Per-feature mean |
| "variance" | `variance_` | Per-feature variance (Bessel's correction) |
| "variation" | `variation_` | Per-feature variation (Bessel's correction) |
| "sum_squares" | `sum_squares_` | Per-feature sum of squares |
| "standard_deviation" | `standard_deviation_` | Per-feature unbiased standard deviation |
| "sum_squares_centered" | `sum_squares_centered_` | Per-feature centered sum of squares |
| "second_order_raw_moment" | `second_order_raw_moment_` | Per-feature second raw moment |
| "all" | all of the above | Computes every statistic |

### Output Attributes
Each output attribute is an `ndarray` of shape `(n_features,)`. An attribute is only present on the fitted estimator if its corresponding option was requested.

### Module
`sklearnex/basic_statistics/incremental_basic_statistics.py`
```

--------------------------------

### Instantiate oneDAL SVM estimator in _onedal_fit

Source: https://deepwiki.com/uxlfoundation/scikit-learn-intelex/3.4-svm

Sets the _onedal_estimator attribute by instantiating the _onedal_factory with SVM hyperparameters, then calls fit with sample_weight, class_count, and queue. Used during _onedal_fit in sklearnex/svm/_base.py.

```python
self._onedal_estimator = self._onedal_factory(**onedal_params)
self._onedal_estimator.fit(X, y, sample_weight, class_count=..., queue=queue)
```

--------------------------------

### sklearn.decomposition.IncrementalPCA

Source: https://deepwiki.com/uxlfoundation/scikit-learn-intelex/1.2-supported-algorithms

Incremental Principal Component Analysis with GPU offload. Triggered via config_context(target_offload="gpu:0") or on-device array API objects.

```APIDOC
## sklearn.decomposition.IncrementalPCA

### Description
Incremental Principal Component Analysis with GPU offload. Triggered via config_context(target_offload="gpu:0") or on-device array API objects.

### Parameters
#### Key Parameter Constraints
- `svd_solver` in `['full','covariance_eigh']`

### Data Format
- No sparse

### Notes
- Preview mode only
```

--------------------------------

### sklearn.cluster.DBSCAN

Source: https://deepwiki.com/uxlfoundation/scikit-learn-intelex/1.2-supported-algorithms

DBSCAN clustering with GPU offload. Triggered via config_context(target_offload="gpu:0") or on-device array API objects.

```APIDOC
## sklearn.cluster.DBSCAN

### Description
DBSCAN clustering with GPU offload. Triggered via config_context(target_offload="gpu:0") or on-device array API objects.

### Parameters
#### Key Parameter Constraints
- `metric='euclidean'` only; `algorithm` in `['brute','auto']`

### Data Format
- Dense only
```

--------------------------------

### sklearn.neighbors.NearestNeighbors

Source: https://deepwiki.com/uxlfoundation/scikit-learn-intelex/1.2-supported-algorithms

Nearest Neighbors search with GPU offload. Triggered via config_context(target_offload="gpu:0") or on-device array API objects.

```APIDOC
## sklearn.neighbors.NearestNeighbors

### Description
Nearest Neighbors search with GPU offload. Triggered via config_context(target_offload="gpu:0") or on-device array API objects.

### Parameters
#### Key Parameter Constraints
- `algorithm='brute'` only; no callable `weights`; `metric` in `[euclidean, manhattan, minkowski, chebyshev, cosine]`

### Data Format
- Dense only
```

--------------------------------

### sklearn.svm.SVC

Source: https://deepwiki.com/uxlfoundation/scikit-learn-intelex/1.2-supported-algorithms

Support Vector Classification with GPU offload. Triggered via config_context(target_offload="gpu:0") or on-device array API objects.

```APIDOC
## sklearn.svm.SVC

### Description
Support Vector Classification with GPU offload. Triggered via config_context(target_offload="gpu:0") or on-device array API objects.

### Parameters
#### Key Parameter Constraints
- `kernel` in `["linear","rbf"]`
- no `class_weight`
- `probability=False` recommended

### Data Format
- Dense only; no negative weights

### Notes
- Binary only; `probability=True` runs initial fit on GPU, calibration on CPU
```

--------------------------------

### num_threads

Source: https://deepwiki.com/uxlfoundation/scikit-learn-intelex/4-daal4py-interface

Returns the current configured thread count.

```APIDOC
## num_threads

### Description
Returns the current configured thread count.

### Returns
- **int** - Current thread count.

```

--------------------------------

### sklearn.linear_model.LinearRegression

Source: https://deepwiki.com/uxlfoundation/scikit-learn-intelex/1.2-supported-algorithms

Linear Regression with GPU offload. Triggered via config_context(target_offload="gpu:0") or on-device array API objects.

```APIDOC
## sklearn.linear_model.LinearRegression

### Description
Linear Regression with GPU offload. Triggered via config_context(target_offload="gpu:0") or on-device array API objects.

### Parameters
#### Key Parameter Constraints
- No `sample_weight`, no `positive=True`

### Data Format
- Dense only
```

--------------------------------

### Direct import from sklearnex submodules

Source: https://deepwiki.com/uxlfoundation/scikit-learn-intelex/1.1-quick-start

Import accelerated estimators directly from sklearnex submodules without patching. This keeps accelerated and stock scikit-learn classes independent, giving explicit control over which estimators are accelerated.

```python
from sklearnex.cluster import DBSCAN as exDBSCAN
from sklearn.cluster import DBSCAN as stockDBSCAN
```

--------------------------------

### C++ generated code structure for PCA algorithm

Source: https://deepwiki.com/uxlfoundation/scikit-learn-intelex/4.1-code-generation-system

Shows the C++ side of the generated code for a given algorithm namespace (e.g., algorithms::pca). Includes typemap, interface struct, and manager struct templates.

```cpp
// --- typemap (Result) ---
typedef daal::algorithms::pca::ResultPtr algorithms_pca_result;
extern "C" NumericTablePtr* get_algorithms_pca_result_eigenvalues(algorithms_pca_result* obj_);
// ... (implementation in .cpp)

// --- interface struct ---
struct pca__iface__ : public algo_manager__iface__ {
    virtual algorithms_pca_result* compute(data_or_file& data, ..., bool setup_only) = 0;
};

// --- manager struct (template) ---
template<typename fptype, pca::Method method>
struct pca_manager<fptype, method> : public pca__iface__ {
    // member storage for all parameters
    // _algob: SharedPtr to Batch<fptype, method>
    // batch(), distributed(), compute() methods
};
```

--------------------------------

### Verifying patch status with is_patched_instance

Source: https://deepwiki.com/uxlfoundation/scikit-learn-intelex/1.1-quick-start

Check whether a given object uses the accelerated implementation. Returns False before patching and True after patching and re-importing the estimator.

```python
from sklearnex import patch_sklearn, is_patched_instance
from sklearn.linear_model import LinearRegression
 
is_patched_instance(LinearRegression())  # False before patching
 
patch_sklearn()
from sklearn.linear_model import LinearRegression  # re-import after patch
is_patched_instance(LinearRegression())  # True
```

--------------------------------

### sklearn.neighbors.KNeighborsRegressor

Source: https://deepwiki.com/uxlfoundation/scikit-learn-intelex/1.2-supported-algorithms

K-Nearest Neighbors Regressor with GPU offload. Triggered via config_context(target_offload="gpu:0") or on-device array API objects.

```APIDOC
## sklearn.neighbors.KNeighborsRegressor

### Description
K-Nearest Neighbors Regressor with GPU offload. Triggered via config_context(target_offload="gpu:0") or on-device array API objects.

### Parameters
#### Key Parameter Constraints
- `algorithm='brute'` only; no callable `weights`; `metric` euclidean/minkowski(p=2)

### Data Format
- Dense only
```

--------------------------------

### Cython generated code structure for PCA algorithm

Source: https://deepwiki.com/uxlfoundation/scikit-learn-intelex/4.1-code-generation-system

Shows the Cython side (.pyx) of the generated code, including cdef classes for result and manager, and the compute method.

```cython
cdef class algorithms_pca_result:
    cdef algorithms_pca_result* c_ptr
    @property
    def eigenvalues(self): ...     # calls get_algorithms_pca_result_eigenvalues
cdef extern from "daal4py.h":
    cdef cppclass c_pca_manager__iface__:
        algorithms_pca_result* compute(data_or_file data, ..., bool setup_only) except +
cdef class pca_manager__iface__:
    cdef c_pca_manager__iface__sptr c_ptr
cdef class pca_manager(pca_manager__iface__):
    def __cinit__(self, fptype="double", method="defaultDense", ...):
        self.c_ptr = mk_pca(...)
    def compute(self, data, ...):
        ...
```

--------------------------------

### Run SPMD distributed mode with MPI

Source: https://deepwiki.com/uxlfoundation/scikit-learn-intelex

Launch a multi-GPU SPMD execution using mpirun with mpi4py. This mode is supported on Linux for selected estimators in sklearnex.spmd.

```bash
mpirun -n 4 python linear_regression_spmd.py
```

--------------------------------

### sklearn.ensemble.ExtraTreesRegressor

Source: https://deepwiki.com/uxlfoundation/scikit-learn-intelex/1.2-supported-algorithms

Extra Trees Regressor with GPU offload. Triggered via config_context(target_offload="gpu:0") or on-device array API objects.

```APIDOC
## sklearn.ensemble.ExtraTreesRegressor

### Description
Extra Trees Regressor with GPU offload. Triggered via config_context(target_offload="gpu:0") or on-device array API objects.

### Parameters
#### Key Parameter Constraints
- Same as `RandomForestRegressor` (GPU)

### Data Format
- No multi-output; no sparse
```

--------------------------------

### Assemble onedal_params dict in BaseSVC._onedal_fit

Source: https://deepwiki.com/uxlfoundation/scikit-learn-intelex/3.4-svm

Builds the parameter dictionary passed to the oneDAL SVM estimator from the scikit-learn estimator's hyperparameters. Includes C, nu, kernel, degree, gamma, coef0, tol, shrinking, max_iter, and cache_size.

```python
onedal_params = {
    "C": self.C,
    "nu": self.nu,
    "kernel": self.kernel,
    "degree": self.degree,
    "gamma": self._compute_gamma_sigma(X),
    "coef0": self.coef0,
    "tol": self.tol,
    "shrinking": self.shrinking,
    "max_iter": self.max_iter,
    "cache_size": self.cache_size,
}
```

--------------------------------

### Using GPU arrays with array API dispatch

Source: https://deepwiki.com/uxlfoundation/scikit-learn-intelex/1.1-quick-start

Use array API dispatch to avoid redundant data transfers when data is already on a GPU device as a dpnp.ndarray or torch.tensor. Requires enabling array API support in scikit-learn via SCIPY_ARRAY_API environment variable and setting array_api_dispatch=True in config_context. Not all estimators support array API inputs.

```python
import os
os.environ["SCIPY_ARRAY_API"] = "1"
 
import torch
from sklearnex import config_context
from sklearnex.cluster import DBSCAN
 
X = torch.tensor(..., device="xpu")  # already on Intel GPU
with config_context(array_api_dispatch=True):
    clustering = DBSCAN(eps=3, min_samples=2).fit(X)
```

--------------------------------

### sklearn.linear_model.LogisticRegression

Source: https://deepwiki.com/uxlfoundation/scikit-learn-intelex/1.2-supported-algorithms

Logistic Regression with GPU offload. Triggered via config_context(target_offload="gpu:0") or on-device array API objects.

```APIDOC
## sklearn.linear_model.LogisticRegression

### Description
Logistic Regression with GPU offload. Triggered via config_context(target_offload="gpu:0") or on-device array API objects.

### Parameters
#### Key Parameter Constraints
- `solver='newton-cg'` only; no `class_weight`, `sample_weight`, `dual`, `intercept_scaling!=1`, `warm_start`, `l1_ratio`

### Data Format
- None

### Notes
- Binary only
```

--------------------------------

### table.__sycl_usm_array_interface__

Source: https://deepwiki.com/uxlfoundation/scikit-learn-intelex/2.3-data-conversion

Exposes the SYCL USM array interface property. This is only available when compiled with the DPC backend (ONEDAL_DATA_PARALLEL).

```APIDOC
## table.__sycl_usm_array_interface__

### Description
Exposes the SYCL USM array interface property. This is only available when compiled with the DPC backend (ONEDAL_DATA_PARALLEL).

### Method
__sycl_usm_array_interface__

### Response
Returns a dictionary representing the SYCL USM array interface.
```

--------------------------------

### IncrementalBasicStatistics

Source: https://deepwiki.com/uxlfoundation/scikit-learn-intelex/1.2-supported-algorithms

Estimator unique to sklearnex for incremental basic statistics. Uses partial_fit and finalize_fit workflow.

```APIDOC
## IncrementalBasicStatistics

### Description
Estimator unique to sklearnex for incremental basic statistics. Uses partial_fit and finalize_fit workflow.

### Class
`sklearnex.basic_statistics.IncrementalBasicStatistics`

### Methods
- `partial_fit` - Incremental fitting
- `finalize_fit` - Finalize the fitting process

### Supported Platforms
- CPU: Yes
- GPU: Yes
- SPMD: Yes
```

--------------------------------

### Patch scikit-learn with sklearnex

Source: https://deepwiki.com/uxlfoundation/scikit-learn-intelex

Use patch_sklearn() to replace standard scikit-learn algorithms with accelerated versions. After patching, imports from sklearn return the optimized implementations; unsupported algorithms fall back silently.

```python
from sklearnex import patch_sklearn
patch_sklearn()
 
from sklearn.cluster import DBSCAN   # now backed by sklearnex
```

--------------------------------

### deselected_tests.yaml entry format

Source: https://deepwiki.com/uxlfoundation/scikit-learn-intelex/6.2-testing-infrastructure

Each entry is a pytest node ID, optionally followed by version condition(s) and a platform string.

```yaml
- node/path/test_file.py::test_name              # always deselect
```

```yaml
- node/path/test_file.py::test_name >=1.5        # deselect for sklearn >= 1.5
```

```yaml
- node/path/test_file.py::test_name >=1.5,<1.7   # deselect for sklearn 1.5.x and 1.6.x
```

```yaml
- node/path/test_file.py::test_name >=1.2 win32  # deselect on Windows, sklearn >= 1.2
```

--------------------------------

### TreeState struct definition (src/tree_visitor.h)

Source: https://deepwiki.com/uxlfoundation/scikit-learn-intelex/4.3-tree-model-export-and-distributed-transceiver

Groups the two heap-allocated arrays (node_ar and value_ar) with metadata scalars. The caller takes ownership of node_ar and value_ar.

```cpp
struct TreeState {
    skl_tree_node * node_ar;
    double        * value_ar;
    size_t          max_depth;
    size_t          node_count;
    size_t          leaf_count;
    size_t          class_count;
};
```

--------------------------------

### sklearn.neighbors.KNeighborsClassifier

Source: https://deepwiki.com/uxlfoundation/scikit-learn-intelex/1.2-supported-algorithms

K-Nearest Neighbors Classifier with GPU offload. Triggered via config_context(target_offload="gpu:0") or on-device array API objects.

```APIDOC
## sklearn.neighbors.KNeighborsClassifier

### Description
K-Nearest Neighbors Classifier with GPU offload. Triggered via config_context(target_offload="gpu:0") or on-device array API objects.

### Parameters
#### Key Parameter Constraints
- `algorithm='brute'` only; no callable `weights`; `metric` in `[euclidean, manhattan, minkowski, chebyshev, cosine]`

### Data Format
- Dense only

### Notes
- At least 2 classes
```

--------------------------------

### Coverage Reporting: pytest invocation with --cov flags

Source: https://deepwiki.com/uxlfoundation/scikit-learn-intelex/6.2-testing-infrastructure

Used in CI to collect coverage for the sklearnex and onedal test suites. The --cov-report= (empty) defers report generation; a separate Codecov upload step processes the combined .coverage.* files.

```bash
--cov=onedal --cov=sklearnex --cov-config=<COVERAGE_RCFILE> --cov-append --cov-branch --cov-report=
```

--------------------------------

### __getattr__ lazy finalization

Source: https://deepwiki.com/uxlfoundation/scikit-learn-intelex/3.5-basic-statistics

Triggers finalization on demand when accessing opt_ attributes like mean_. Calls _onedal_finalize_fit() if _need_to_finalize is True.

```python
def __getattr__(self, attr):
    sattr = attr.removesuffix("_")
    is_statistic_attr = (
        sattr in self._onedal_estimator.options
        if "_onedal_estimator" in self.__dict__
        else False
    )
    if is_statistic_attr:
        if self._need_to_finalize:
            self._onedal_finalize_fit()
        return getattr(self._onedal_estimator, attr)
    return self.__getattribute__(attr)
```

--------------------------------

### Apply class weights to sample_weight

Source: https://deepwiki.com/uxlfoundation/scikit-learn-intelex/3.4-svm

Multiplies class weights into per-sample weights before passing to the backend, because oneDAL only accepts per-sample weights. Iterates over class_weight_ entries and scales sample_weight for each class.

```python
for i, v in enumerate(self.class_weight_):
    sample_weight[y == i] *= v
```

--------------------------------

### sklearn.covariance.EmpiricalCovariance

Source: https://deepwiki.com/uxlfoundation/scikit-learn-intelex/1.2-supported-algorithms

Empirical covariance estimation with GPU offload. Triggered via config_context(target_offload="gpu:0") or on-device array API objects.

```APIDOC
## sklearn.covariance.EmpiricalCovariance

### Description
Empirical covariance estimation with GPU offload. Triggered via config_context(target_offload="gpu:0") or on-device array API objects.

### Parameters
#### Key Parameter Constraints
- All parameters

### Data Format
- Dense only

### Notes
- Preview mode only
```

--------------------------------

### sklearn.linear_model.Ridge

Source: https://deepwiki.com/uxlfoundation/scikit-learn-intelex/1.2-supported-algorithms

Ridge regression with GPU offload. Triggered via config_context(target_offload="gpu:0") or on-device array API objects.

```APIDOC
## sklearn.linear_model.Ridge

### Description
Ridge regression with GPU offload. Triggered via config_context(target_offload="gpu:0") or on-device array API objects.

### Parameters
#### Key Parameter Constraints
- `solver='auto'`; no `sample_weight`; no `positive=True`; `alpha` scalar

### Data Format
- Dense only
```

--------------------------------

### sklearn.ensemble.ExtraTreesClassifier

Source: https://deepwiki.com/uxlfoundation/scikit-learn-intelex/1.2-supported-algorithms

Extra Trees Classifier with GPU offload. Triggered via config_context(target_offload="gpu:0") or on-device array API objects.

```APIDOC
## sklearn.ensemble.ExtraTreesClassifier

### Description
Extra Trees Classifier with GPU offload. Triggered via config_context(target_offload="gpu:0") or on-device array API objects.

### Parameters
#### Key Parameter Constraints
- Same as `RandomForestClassifier` (GPU)

### Data Format
- No multi-output; no sparse

### Notes
- At least 2 classes
```

--------------------------------

### table Properties

Source: https://deepwiki.com/uxlfoundation/scikit-learn-intelex/2.3-data-conversion

The table object exposes properties for inspecting its data: has_data, shape, row_count, column_count, kind, and dtype.

```APIDOC
## table Properties

### Description
The table object exposes properties for inspecting its data: has_data, shape, row_count, column_count, kind, and dtype.

### Properties
- **has_data** (bool) - Whether the table holds any data
- **shape** (tuple(int, int)) - (row_count, column_count)
- **row_count** (int) - Number of rows
- **column_count** (int) - Number of columns
- **kind** (str) - "homogen", "csr", or "empty"
- **dtype** (numpy.dtype) - Element data type
```

--------------------------------

### sklearn.ensemble.RandomForestRegressor

Source: https://deepwiki.com/uxlfoundation/scikit-learn-intelex/1.2-supported-algorithms

Random Forest Regressor with GPU offload. Triggered via config_context(target_offload="gpu:0") or on-device array API objects.

```APIDOC
## sklearn.ensemble.RandomForestRegressor

### Description
Random Forest Regressor with GPU offload. Triggered via config_context(target_offload="gpu:0") or on-device array API objects.

### Parameters
#### Key Parameter Constraints
- No `warm_start`, `ccp_alpha!=0`, `criterion!='squared_error'`, `oob_score=True`, `sample_weight!=None`, `n_estimators>6024`

### Data Format
- No multi-output; no sparse
```

--------------------------------

### SPMD BasicStatistics subclass overriding _onedal_basic_statistics

Source: https://deepwiki.com/uxlfoundation/scikit-learn-intelex/3.5-basic-statistics

Defines a BasicStatistics subclass for SPMD distributed computation. It inherits all logic from the batch sklearnex BasicStatistics and only swaps the backend class to the SPMD-capable onedal_BasicStatistics, enabling multi-process distributed statistics computation.

```python
class BasicStatistics(BasicStatistics_Batch):
    __doc__ = BasicStatistics_Batch.__doc__
    _onedal_basic_statistics = staticmethod(onedal_BasicStatistics)
```

--------------------------------

### IncrementalRidge

Source: https://deepwiki.com/uxlfoundation/scikit-learn-intelex/1.2-supported-algorithms

Estimator unique to sklearnex for incremental ridge regression. Uses partial_fit and finalize_fit workflow.

```APIDOC
## IncrementalRidge

### Description
Estimator unique to sklearnex for incremental ridge regression. Uses partial_fit and finalize_fit workflow.

### Class
`sklearnex.linear_model.IncrementalRidge`

### Methods
- `partial_fit` - Incremental fitting
- `finalize_fit` - Finalize the fitting process

### Supported Platforms
- CPU: Yes
- GPU: Yes
- SPMD: Yes
```

--------------------------------

### dispatch() usage in LogisticRegression.fit

Source: https://deepwiki.com/uxlfoundation/scikit-learn-intelex/3-algorithm-implementations

Typical usage of dispatch() inside a public method of a sklearnex estimator. Passes a dictionary with 'onedal' and 'sklearn' keys mapping to the respective callable implementations. dispatch() determines the active SYCL queue, checks GPU/CPU support, and routes to the oneDAL or sklearn fallback.

```python
# sklearnex/linear_model/logistic_regression.py  (line ~153)
def fit(self, X, y, sample_weight=None):
    dispatch(
        self,
        "fit",
        {
            "onedal": self.__class__._onedal_fit,
            "sklearn": _sklearn_LogisticRegression.fit,
        },
        X, y, sample_weight,
    )
```

--------------------------------

### my_procid

Source: https://deepwiki.com/uxlfoundation/scikit-learn-intelex/4-daal4py-interface

Returns this process's MPI rank (MPI_Comm_rank). Returns 0 outside MPI.

```APIDOC
## my_procid

### Description
Returns this process's MPI rank (MPI_Comm_rank). Returns 0 outside MPI.

### Returns
- **int** - MPI rank of the current process.

```

--------------------------------

### IncrementalLinearRegression

Source: https://deepwiki.com/uxlfoundation/scikit-learn-intelex/1.2-supported-algorithms

Estimator unique to sklearnex for incremental linear regression. Uses partial_fit and finalize_fit workflow.

```APIDOC
## IncrementalLinearRegression

### Description
Estimator unique to sklearnex for incremental linear regression. Uses partial_fit and finalize_fit workflow.

### Class
`sklearnex.linear_model.IncrementalLinearRegression`

### Methods
- `partial_fit` - Incremental fitting
- `finalize_fit` - Finalize the fitting process

### Supported Platforms
- CPU: Yes
- GPU: Yes
- SPMD: Yes
```

--------------------------------

### sklearn.ensemble.RandomForestClassifier

Source: https://deepwiki.com/uxlfoundation/scikit-learn-intelex/1.2-supported-algorithms

Random Forest Classifier with GPU offload. Triggered via config_context(target_offload="gpu:0") or on-device array API objects.

```APIDOC
## sklearn.ensemble.RandomForestClassifier

### Description
Random Forest Classifier with GPU offload. Triggered via config_context(target_offload="gpu:0") or on-device array API objects.

### Parameters
#### Key Parameter Constraints
- No `warm_start`, `ccp_alpha!=0`, `criterion!='gini'`, `oob_score=True`, `sample_weight!=None`, `n_estimators>6024`

### Data Format
- No multi-output; no sparse

### Notes
- At least 2 classes
```

--------------------------------

### num_procs

Source: https://deepwiki.com/uxlfoundation/scikit-learn-intelex/4-daal4py-interface

Returns the MPI communicator size (MPI_Comm_size). Returns 1 outside MPI.

```APIDOC
## num_procs

### Description
Returns the MPI communicator size (MPI_Comm_size). Returns 1 outside MPI.

### Returns
- **int** - MPI communicator size.

```
Tokens:
Raw