This directory contains provider-ready cuBLAS operations after Tensor and mdspan policy has been lowered but before the operation-tag backend adapter.
Contents
gemm.hpp: checked GEMM over provider-ready BLAS-compatible matrix descriptors and an already acquired cuBLAS execution lease.
Notes
- Keep raw cuBLAS API calls and handle ownership in `backend/cublas/`.
- Keep operation-tag acceptance and dispatch adaptation in `linalg/backends/cublas/`.
- Keep CUDA-mdspan recognition, storage-policy lowering, completion-ledger access, and blocking/coroutine resource admission in the backend adapter's private
detail/ implementation.
- Runtime layout rejection must occur before synchronized buffer access and provider submission. Once a provider call is submitted, failure is terminal rather than a clean backend decline.
Related Documentation
