viable/strict/1788603844: [cuBLAS] Always eagerly allocate cuBLAS(Lt) workspaces (#194311)

viable/strict/1788603844: [cuBLAS] Always eagerly allocate cuBLAS(Lt) workspaces (#194311)

Release notes summarize a prototype that introduces eager allocation of cuBLAS(Lt) workspaces to replace or augment existing workspace caching. The accompanying benchmarks compare cached vs. eager configurations across several matrix shapes, showing modest increases in per-operation time for eager workspaces.

The notes include a discussion on observed issues such as CUDA_ERROR_ILLEGAL_ADDRESS with cached graphs and the potential need to isolate or clear cached workspaces at graph-capture boundaries. A simpler direction being considered is to retain the cache but manage its scope around graph captures.

Why it matters: If adopted, this change could alter performance characteristics for cuBLAS-backed operations and affect debugging of graph captures; it may lead to simpler lifetime management of workspaces but with potential per-operation overhead changes.

AI-assisted brief