Alelyon Compute Kit

Alelyon Compute Kit is being developed as a vendor-neutral compute SDK for training and general computation. The goal is one programming interface with backends for different device architectures, compiler targets and driver APIs. Support is determined by explicit capabilities and workload tests.

The PyPI distribution name is alelyon-ai. The Python namespace is alelyon_compute_kit. This repository uses the MIT license.

Current state

This initial standalone package contains immutable capability records and an explicit backend registry. It has no runtime dependencies and does not load a driver, import plugins, probe hardware or allocate device memory on import.

The native Vulkan runtime, kernels and experimental PyTorch integration are being prepared for a bounded SDK export from their existing development tree. They are not included in this initial package. No device backend is registered by default. This version cannot yet run training or replace an installed CUDA toolchain. A published release will identify its actual backend, architecture, operator, precision and platform coverage.

Capability matching

The following is a fabricated declaration illustrating the API; it performs no hardware discovery or computation:

from alelyon_compute_kit import (
    BackendDescriptor, BackendRegistry, DeviceCapabilities, DeviceIdentity,
    DType, OperatorCapability, OperatorRequirement, Precision,
)

registry = BackendRegistry()
registry.register(BackendDescriptor(
    backend_id="example",
    name="Example declaration",
    devices=(DeviceCapabilities(
        identity=DeviceIdentity("example", "device-0", "Example device"),
        operators=(OperatorCapability("mm", 1, DType.FLOAT32, Precision.FP32),),
    ),),
))

selection = registry.select((
    OperatorRequirement("mm", 1, DType.FLOAT32, Precision.FP32),
))
assert len(selection.matches) == 1

A match means the supplied declaration meets the supplied requirements. It does not verify hardware availability, numerical correctness or training readiness. Evidence records remain unverified references and do not change matching. Unsupported requirements return named reasons; selection neither ranks vendors nor silently changes precision or falls back to a different device.

Development

Use an isolated Python environment:

python -m pip install -e ".[dev]"
python -m pytest
python -m build

The current source targets Python 3.10 and later. A pure-Python wheel's platform tag describes these capability modules, not native accelerator support. Native packages will require separate platform and installation checks.

Architecture and extension boundaries describe the planned runtime layers and the acceptance required before backend adoption. The release guide describes the manually dispatched validation and PyPI publishing workflow. Publishing is an explicit option; ordinary pushes do not release a package.