Mobilint NPU Python

Shared runtime support for applications that run MXQ models on Mobilint NPUs or ONNX models through ONNX Runtime. mblt-npu-python provides the common backend, device-selection rules, Hugging Face artifact resolution, and model-detail logging used by Mobilint Python packages. It is a library dependency, rather than an end-user model catalog.

Version 0.0.0 is the initial standalone release.

logging ships here rather than with its only caller because the two are mutually dependent — npu_backend imports log_model_details, and log_model_details reads a MobilintNPUBackend's fields.

Installation

pip install mblt-npu-python

This package requires a supported Linux environment with mobilint-qb-runtime available and Python 3.10 through 3.12.

Public API

Import the backend from mblt_npu:

from mblt_npu import MobilintNPUBackend

backend = MobilintNPUBackend(
    mxq_path="model.mxq",
    core_mode="single",
)
backend.create()
try:
    backend.launch()
    outputs = backend.mxq_model.infer([input_tensor])
finally:
    backend.dispose()

MobilintNPUBackend selects the appropriate implementation from target_device (default: "aries-rb"). "aries-rb" selects MobilintAriesBackend; "regulus-ra", "regulus-rb", "regulus-ra-usb", and "regulus-rb-usb" select MobilintRegulusBackend. The former generic values "aries" and "regulus" remain accepted when loading older configurations. The board name is forwarded to qbruntime.Accelerator so the runtime opens the matching device (requires mobilint-qb-runtime>=1.4.0). backend_class_for() and BACKEND_CLASSES are available for integrations that need to inspect the supported targets.

Multi-slot MXQ execution

max_batch_size is aggregate capacity. At create(), the backend probes the compiled per-model capacity K and loads ceil(max_batch_size / K) model slots. Slots are distributed round-robin over the devices named by canonical target strings and reuse one accelerator per device. mxq_model and acc continue to refer to slot zero for compatibility; concurrent callers can use infer_slot(slot_index, inputs). Allocation failures dispose all created slots and raise MobilintBackendAllocError with the failed slot and device.

Hub-backed configurations retain name_or_path, revision, and commit_hash through to_dict() / from_dict(). Artifact lookup never substitutes an unpinned revision or an unrelated cached MXQ.

For ONNX inference, install the optional runtime extra and use ONNXBackend:

pip install "mblt-npu-python[onnxruntime]"

from mblt_npu import ONNXBackend

backend = ONNXBackend("model.onnx")
backend.create()
outputs = backend({"images": input_array})
backend.dispose()

ONNXBackend imports onnxruntime only when it creates a session.

Most users should access the backend through a model package such as mblt-vision-python, which owns model configuration, preprocessing, and postprocessing.

Testing helpers

The optional test extra provides a shared pytest plugin with NPU options and the npu_params fixture used by Mobilint package test suites:

pip install "mblt-npu-python[test]"

Import mblt_npu.pytest_plugin from a repository's root tests/conftest.py to register its options. The plugin is intentionally not auto-registered, so projects control when those command-line options are exposed.

Support and issues

For installation, runtime, or integration support, visit the Mobilint forum. Report reproducible package issues in the mblt-npu-python issue tracker.

License

Distributed under the BSD 3-Clause License.