Implementation:Huggingface Optimum Is Backend Available

Overview

Implementation of backend availability detection functions that check whether hardware acceleration backends (ONNX Runtime, OpenVINO, IPEX) are installed in the current Python environment.

Source

File: optimum/utils/import_utils.py

Repository: optimum

APIs

Primary Availability Functions

Function	Lines	Description
`is_onnxruntime_available() -> bool`	L182-183	Checks if any of 17 ONNX Runtime distribution variants are installed. Returns cached `_onnxruntime_available` flag.
`is_openvino_available() -> bool`	L241-242	Checks if OpenVINO is installed. Returns cached `_openvino_available` flag.
`is_ipex_available() -> bool`	L237-238	Checks if Intel Extension for PyTorch is installed. Returns cached `_ipex_available` flag.

Helper Function

def _is_package_available(
    pkg_name: str,
    return_version: bool = False,
    pkg_distributions: Optional[List[str]] = None,
) -> Union[Tuple[bool, str], bool]:

Lines: L40-79

Description: Core detection function that performs a two-phase check:

Uses importlib.util.find_spec(pkg_name) to check if the package is importable
Iterates over pkg_distributions using importlib.metadata.version() to verify the package is properly installed

Module-Level Caching

The availability results are computed once at module load time and stored in module-level variables:

_openvino_available = _is_package_available("openvino")                    # L87
_ipex_available = _is_package_available("intel_extension_for_pytorch")     # L92
_onnxruntime_available = _is_package_available(                            # L97-118
    "onnxruntime",
    pkg_distributions=[
        "onnxruntime-gpu",
        "onnxruntime-rocm",
        "onnxruntime-training",
        "onnxruntime-training-rocm",
        "onnxruntime-training-cpu",
        "onnxruntime-openvino",
        "onnxruntime-vitisai",
        "onnxruntime-armnn",
        "onnxruntime-cann",
        "onnxruntime-dnnl",
        "onnxruntime-acl",
        "onnxruntime-tvm",
        "onnxruntime-qnn",
        "onnxruntime-migraphx",
        "ort-migraphx-nightly",
        "ort-rocm-nightly",
    ],
)

Import

from optimum.utils import is_onnxruntime_available, is_openvino_available, is_ipex_available

I/O

Direction	Type	Description
Input	None	Reads cached module-level boolean flags set at import time
Output	`bool`	`True` if the corresponding backend package is installed and available, `False` otherwise

Usage Example

from optimum.utils import is_onnxruntime_available, is_openvino_available, is_ipex_available

# Check individual backends
print(f"ONNX Runtime available: {is_onnxruntime_available()}")
print(f"OpenVINO available: {is_openvino_available()}")
print(f"IPEX available: {is_ipex_available()}")

# Conditional backend selection
if is_openvino_available():
    from optimum.intel import OVModelForSequenceClassification
elif is_onnxruntime_available():
    from optimum.onnxruntime import ORTModelForSequenceClassification

Internal Details

The _is_package_available helper function at L40-79 follows this logic:

Call importlib.util.find_spec(pkg_name) to check if the module is importable
If pkg_distributions is provided, append pkg_name itself to the list
Iterate through all distribution names, calling importlib.metadata.version(pkg) on each
Return True on the first successful version retrieval; False if none succeed
If return_version=True, return a tuple of (bool, str) with the version string

Page Connections

Double-click a node to navigate. Hold to expand connections.

Principle

Implementation

Heuristic

Environment