axiolid-backend-gpu
An API-neutral seam for GPU graph compilation. GpuCompiler adapts any
GpuGraphExecutor to the MeshCompiler contract: it validates device,
precision, residency and roots before submitting one batch, and checks the
executor's results afterwards. The crate chooses no GPU API (CUDA, Metal,
Vulkan, WebGPU) and ships no executor, so default builds carry no driver
stack. Concrete executors live in their own crates, including out of tree
(ADR 0011).
- API documentation: docs.rs/axiolid-backend-gpu
- Reference page: axiolid.github.io/kernel
- Source and issues: github.com/axiolid/kernel
Design notes
- Each further GPU operation gets its own narrow executor trait and adapter, never a method on one catch-all backend.
- A GPU path is evidence only if a CPU differential test checks it. Maturing the executor is tracked in #22.