Warp Integration¶

The soromox.execution.warp namespace contains allocation-free, Warp-native building blocks for integrations that need to embed SoRoMoX mechanics in a larger Warp or Newton solver.

Not needed for ordinary SoRoMoX simulation

Normal users should select backend="warp" or leave backend="auto" on a supported system. Direct kernel launches require caller-owned buffers, exact FP64 shapes, the documented launch ordering, and explicit CUDA-graph lifetime management.

The public child namespaces are:

soromox.execution.warp ¶

Warp-native building blocks for accelerated system execution.

The family subpackages separate reusable mechanics kernels from Soromox's JAX-facing dispatch machinery. The :mod:.gvs and :mod:.pcs packages expose runtime operands, workspace-shape contracts, raw kernels, and direct launch functions that external Warp-native integrators can use with preallocated buffers. Public kernel symbols are loaded lazily and require the optional warp-lang package only when requested.

Loader and JAX transformation modules are implementation details and are not part of the stable integration API.