Module blas_lt 

BlasLtActor — wraps [cudarc::cublaslt::CudaBlasLT] for transformer-shaped fused matmul (matmul + bias + activation + aux-store + bias-grad reduction) across the full dtype matrix cuBLASLt accepts.

See epilogue for the curated Epilogue enum, heuristic for the algorithm cache, workspace for the workspace pool, scaling for the fp8 scale-pointer wiring, and matmul for the typed MatmulRequest<T> plus its BlasLtDispatch impl.

Re-exports

pub use epilogue::Epilogue;

pub use heuristic::HeuristicCacheRef;

pub use heuristic::HeuristicEntry;

pub use heuristic::HeuristicKey;

pub use heuristic::DEFAULT_HEURISTIC_CAPACITY;

pub use matmul::MatmulRequest;

pub use scaling::ScaleSet;

pub use workspace::WorkspaceLease;

pub use workspace::WorkspacePool;

Modules

epilogue

Epilogue enum — atomr-accel’s curated mapping over cuBLASLt’s cublasLtEpilogue_t.

heuristic

Heuristic-cache for cuBLASLt matmul algorithms.

matmul

Typed MatmulRequest<T: GemmSupported> plus the BlasLtDispatch impl that routes it through the kernel envelope.

scaling

fp8 scale-pointer helpers for cuBLASLt matmul.

workspace

cuBLASLt workspace pool — recycles per-heuristic device buffers.

Structs

BlasLtActor

Enums

Activation

Available activation for kernel fusing in matmul

BlasLtMsg

Public message surface.