hawk.artifact

Manifest + sidecar + the ABI exports + bundles.

Nothing in this package imports eagle or raptor: plan_view() projects a sidecar into the dict shape eagle.plan reads without importing it, so a consumer can hand a HAWK artifact to eagle while HAWK stays severed. plugins() assembles the whole plan-able object per kernel the same way, loading the device side through a caller-supplied loader rather than an eagle import.

Functions

arch(…)	The CUDA arch HAWK compiles a device artifact for -- the public name for the ONE resolver (hawk.compile.toolchain.resolve_arch()): device_arch= > $HAWK_CUDA_ARCH > the running GPU (probed once per process) > PTX for the toolkit's oldest architecture, each checked against what the toolkit compiles.
build(…)	Emit, compile and publish ONE kernel into directory.
build_bundle(…)	Build every kernel into directory and write ONE manifest.json over them.
exports(…)	The export block for one target's TU.
plan_view(…)	The plan-time declaration a v2 consumer reads off a HAWK sidecar.
plugins(…)	{name: plugin} for every kernel of bundle: the objects eagle.plan.plan and eagle.deploy take, assembled in one place.
reset_unit_memo()	Forget every memoised unit, the counters and the log — the door back to the disk path, for a row that wants a cold publisher.
unit_stats()	A copy of the publish counters (memo hits, stamp hits, real publishes).
write_manifest(…)	Write doc as the bundle's manifest.json.

Classes

Artifact	One kernel's published files + the sidecar it was described by.
Bundle	A deployment unit: one manifest.json over one or more artifacts.
Entry	One compiled target of one kernel.