This directory contains private CUDA-storage lowering for CublasBackend.

gemm.hpp recognizes supported CUDA mdspan accessors and layouts, validates buffer bounds, aliasing, and device placement, and creates the provider-ready GEMM plan shared by direct and coroutine dispatch. Direct dispatch acquires an execution lease with the blocking pool interface. The backend's optional try_make_kernel_task hook returns a CudaTask that awaits the same resource.

Provider-ready matrix calls remain in `linalg/cublas/`, and raw cuBLAS API wrappers remain in `backend/cublas/`.

See CUDA kernel dispatch and provider scheduling for the direct and coroutine execution paths.

Return to the cuBLAS backend adapter.