Module cuda_host 

2026-09-25: The process-scoped CUDA context and stream (CudaHost), and release, which unloads one model’s crate::registry modules.

Owner: gpu-runtime. Invariants:

• The process has at most one CudaHost, on the ordinal of the first host call; a later call with another ordinal is an error, and a failed first creation is returned as the same error to every later call.
• release unloads nothing while another Arc to the registry is live.

Structs

CudaHost

2026-09-25: The CUDA context and stream for the process. Each model’s MetraleRegistry loads its modules into this one context.

Functions

host

2026-09-25: Get, or on the first call create, the process CUDA host on ordinal. The first call fixes the ordinal; asking for another is an error.

release

2026-09-25: Unload a model’s kernel modules. Errors, unloading nothing, while another strong Arc to the registry is live; otherwise drops every module the registry holds.