Module blas 

BlasActor — full cuBLAS surface (Phase 1 cuBLAS slice).

Wraps a [cudarc::cublas::CudaBlas] handle, performs cuBLAS L1/L2/ L3 ops on its assigned stream, and returns completion via the configured CompletionStrategy (§3.2 stateless-handle archetype + §5.10 callback wiring).

Sub-modules:

• gemm — typed Gemm<T> for f32/f64/f16/bf16 (cudarc safe layer) and the legacy SgemmRequest adapter that routes through Gemm<f32> for back-compat.
• gemm_strided_batched — strided-batched gemm via cudarc’s safe layer for f32/f64/f16/bf16; can drop to crate::sys::cublas::gemm_strided_batched_ex if more dtypes are needed in a follow-up.
• l1 — axpy / dot / nrm2 / scal / asum / iamax / iamin / copy / swap / rot via the cuBLAS ex-suffix entry points.
• l2 — gemv / ger via cudarc’s Gemv<T> and the local cublasGemv_v2 / cublasGer_v2 wrappers.
• l3 — geam / syrk / trsm via the local cublasSgeam / cublasSsyrk_v2 / cublasStrsm_v2 wrappers (and dgeam/dsyrk/ dtrsm).
• scaling — fp8 scaling-factor helpers (per-tensor / per-row), stubbed under the cublas-fp8 feature for use by cublasGemmEx on Hopper+.

The mailbox is freed immediately after the kernel is enqueued — the actor never blocks on the GPU (§5.2). Reply delivery happens on the Tokio task spawned by crate::kernel::envelope::run_kernel.

Re-exports

pub use gemm::GemmRequest;

pub use gemm_strided_batched::GemmStridedBatchedRequest;

pub use l1::AsumRequest;

pub use l1::AxpyRequest;

pub use l1::CopyRequest;

pub use l1::DotRequest;

pub use l1::IamaxRequest;

pub use l1::IaminRequest;

pub use l1::Nrm2Request;

pub use l1::RotRequest;

pub use l1::ScalRequest;

pub use l1::SwapRequest;

pub use l2::GemvRequest;

pub use l2::GerRequest;

pub use l3::GeamRequest;

pub use l3::SyrkRequest;

pub use l3::TrsmRequest;

Modules

gemm

Typed GemmRequest<T> + GemmDispatch impls.

gemm_strided_batched

Typed GemmStridedBatchedRequest<T> + GemmStridedBatchedDispatch impls.

l1

Typed L1 ops: axpy, dot, nrm2, scal, asum, iamax, iamin, copy, swap, rot.

l2

Typed L2 ops: gemv, ger.

l3

Typed L3 ops other than gemm: geam (matrix add/scale), syrk (symmetric rank-k update), trsm (triangular solve).

scaling

fp8 scaling-factor helpers.

Structs

BlasActor

Two-track construction: a real cuBLAS-backed actor (props), and a mock variant used by examples/echo_no_gpu and unit tests where no GPU is present.

Enums

BlasMsg

Public messages for BlasActor. Each variant boxes a typed dispatcher trait object so the dtype dimension travels through the box without forcing an N-fold mailbox explosion.