Skip to content

About

A host-side contract and evidence layer for coprocessor launches: models host/device roles, resource grades, transfer residues, launch receipts and CNO/OND claims. It records launch semantics; it does not run accelerator kernels.

Resources

Code of conduct

Contributing

Security policy

Stars

1 star

Watchers

0 watching

Forks

Latest commit

 

History

6 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

CoprocessorRuntime.jl

Code licence: MPL-2.0 Docs licence: CC-BY-SA-4.0 Rhodium Standard Repository OpenSSF Scorecard

Coprocessor support base library for Julia — session contracts, tropical resource grades, echo residues, epistemic receipts, and CNO/OND claims. Not a kernel runtime.

What this is

CoprocessorRuntime.jl is the host-side contract layer for talking about coprocessors (GPU, NPU, TPU, FPGA, QPU, DSP) without pretending to execute on them.

It answers:

Question What this package records

Which endpoint am I?

Choreographic HostRole / DeviceRole projection of a launch plan

What is the worst-case cost?

Tropical ResourceGrade (max-plus sequential / alternative)

What was lost in the transfer?

EchoResidue on a named, possibly non-injective map

What may the host claim afterwards?

LaunchReceipt with standpoint + warrant (no-smuggling)

Did this reset actually do nothing to the world / reveal nothing?

CNOClaim / ONDClaim — claims, not proofs

Can one adapter speak every vendor protocol?

No. universal_hub_allowed() throws HubCeilingRefusal

What this is NOT

  • It does not execute CUDA / ROCm / Metal / TPU kernels. That is CUDA.jl, AMDGPU.jl, Metal.jl, KernelAbstractions.jl.

  • It does not select a backend. That is AcceleratorGate.jl.

  • It does not heal a failing kernel. That is HardwareResilience.jl.

  • It does not prove CNO/OND. That is absolute-zero. Status here is :unproved or :model_checked.

  • It does not close K-CUT. That conjecture is open in choreographic-types.

LowLevel.jl currently describes itself as a universal coprocessor foundation. This package exists so that claim does not have to be duplicated — and so that hub_ceiling is the default, not an afterthought.

Installation

using Pkg
Pkg.add(url="https://github.com/hyperpolymath/CoprocessorRuntime.jl")

Usage

using CoprocessorRuntime

s = open_session(GPUKind())
plan = launch_plan(
    :matmul,
    [Endpoint(HostRole(), CPUKind()), Endpoint(DeviceRole(), GPUKind())],
    ResourceGrade(1 << 20, 10_000, 0.0),
    EchoResidue(collapsing_transfer(:f32_to_f16), "mantissa fibre"),
)
local_view = project(plan, DeviceRole())
@assert local_view.kcut === :open          # K-CUT is not claimed
receipt = issue_receipt(s; warrant=WarrantOnly, claim="planned")
@assert !host_may_claim(receipt)           # not factive
cno = reset_session!(s)                    # model-checked handle reset
close_session!(s)

Ecosystem

This package sits in the metal layer of Hyperpolymath.jl. Wiring instructions (Project.toml + re-export) live in docs/ecosystem/HYPERPOLYMATH-WIRING.adoc.

Type-informed backend notes — what was taken from the -type repos, and what was deliberately not forced — live in docs/theory/TYPE-INFORMED-BACKENDS.adoc.

Licence

Code, configuration and scripts are Mozilla Public License 2.0 (MPL-2.0); prose documentation is CC-BY-SA-4.0.

About

A host-side contract and evidence layer for coprocessor launches: models host/device roles, resource grades, transfer residues, launch receipts and CNO/OND claims. It records launch semantics; it does not run accelerator kernels.

Resources

Code of conduct

Contributing

Security policy

Stars

1 star

Watchers

0 watching

Forks

Releases

Sponsor this project

Packages

Used by

Contributors

Languages