Module kernel 

Kernel-actor wrappers around CUDA library handles (§3.2).

Each library actor follows a uniform shape:

• a Real { handle, stream, completion, state, … } variant holding the cudarc handle plus the per-actor caches it needs;
• a Mock variant for GPU-free tests;
• a props(stream, allocator, completion, state) constructor that panics with "ContextPoisoned: <Lib>::new failed: …" if the handle can’t be created, so the supervisor restarts;
• a mock_props() constructor that replies Unrecoverable("…not supported in mock mode") to every variant.

The shared kernel-enqueue body lives in envelope::run_kernel — every library actor calls it instead of reimplementing the validate / enqueue / spawn-completion-await / reply / drop-keep-alive sequence.

F2 ships: BlasActor, CudnnActor, FftActor, RngActor. F3 adds: SolverActor, BlasLtActor, NvrtcActor. F4 adds: CollectiveActor (NCCL).

Re-exports

pub use dispatch::BlasLtDispatch;

pub use dispatch::BlasLtDispatchCtx;

pub use dispatch::CollectiveDispatch;

pub use dispatch::CollectiveDispatchCtx;

pub use dispatch::CudnnDispatch;

pub use dispatch::CudnnDispatchCtx;

pub use dispatch::DevSliceArg;

pub use dispatch::GemmDispatchCtx;

pub use dispatch::NvrtcDispatchCtx;

pub use dispatch::NvrtcLaunchDispatch;

pub use dispatch::RngDispatch;

pub use dispatch::ScalarArg;

pub use dispatch::FftDispatch;

pub use dispatch::FftDispatchCtx;

pub use dispatch::SendSparseHandle;

pub use dispatch::SparseDispatch;

pub use dispatch::SparseDispatchCtx;

pub use dispatch::SparseOp;

pub use dispatch::TensorDispatch;

pub use dispatch::TensorDispatchCtx;

pub use dispatch::WorkspacePool;

pub use blas::AsumRequest;

pub use blas::AxpyRequest;

pub use blas::BlasActor;

pub use blas::BlasMsg;

pub use blas::CopyRequest;

pub use blas::DotRequest;

pub use blas::GeamRequest;

pub use blas::GemmRequest;

pub use blas::GemmStridedBatchedRequest;

pub use blas::GemvRequest;

pub use blas::GerRequest;

pub use blas::IamaxRequest;

pub use blas::IaminRequest;

pub use blas::Nrm2Request;

pub use blas::RotRequest;

pub use blas::ScalRequest;

pub use blas::SwapRequest;

pub use blas::SyrkRequest;

pub use blas::TrsmRequest;

pub use cudnn::ActivationFwdRequest;

pub use cudnn::ActivationKind;

pub use cudnn::ActivationRequest;

pub use cudnn::AttentionMask;

pub use cudnn::AttentionParams;

pub use cudnn::BatchNormRequest;

pub use cudnn::ConvBwdDataRequest;

pub use cudnn::ConvBwdFilterRequest;

pub use cudnn::ConvDescParams;

pub use cudnn::ConvForwardRequest;

pub use cudnn::ConvFwdRequest;

pub use cudnn::ConvParams;

pub use cudnn::CudnnActor;

pub use cudnn::CudnnMsg;

pub use cudnn::DropoutFwdRequest;

pub use cudnn::EpilogueKind;

pub use cudnn::GroupNormRequest;

pub use cudnn::InstanceNormRequest;

pub use cudnn::LayerNormRequest;

pub use cudnn::LrnFwdRequest;

pub use cudnn::LrnParams;

pub use cudnn::MultiHeadAttnBwdRequest;

pub use cudnn::MultiHeadAttnFwdRequest;

pub use cudnn::NormBwdRequest;

pub use cudnn::NormMode;

pub use cudnn::NormPhase;

pub use cudnn::PoolBwdRequest;

pub use cudnn::PoolFwdRequest;

pub use cudnn::PoolMode;

pub use cudnn::PoolParams;

pub use cudnn::RnnBwdRequest;

pub use cudnn::RnnDirection;

pub use cudnn::RnnFwdRequest;

pub use cudnn::RnnMode;

pub use cudnn::RnnParams;

pub use cudnn::SoftmaxFwdRequest;

pub use cudnn::SoftmaxMode;

pub use cudnn::SoftmaxRequest;

pub use cudnn::TensorLayout;

pub use fft::FftActor;

pub use fft::FftCallbackKind;

pub use fft::FftDirection;

pub use fft::FftKind;

pub use fft::FftMsg;

pub use fft::FftPlan;

pub use fft::FftPlanMany;

pub use fft::FftRequest;

pub use fft::PlanKey;

pub use rng::Distribution;

pub use rng::FillRequest;

pub use rng::RngActor;

pub use rng::RngGeneratorKind;

pub use rng::RngMsg;

pub use solver::CholeskyRequest;

pub use solver::GesvdjBatchedRequest;

pub use solver::GetrfBatchedRequest;

pub use solver::HegvdRequest;

pub use solver::LuRequest;

pub use solver::LuSolveRequest;

pub use solver::PotrfBatchedRequest;

pub use solver::QrRequest;

pub use solver::SolverActor;

pub use solver::SolverDispatch;

pub use solver::SolverMsg;

pub use solver::SvdRequest;

pub use solver::SyevdRequest;

pub use solver::SygvdRequest;

pub use solver::Uplo;

pub use blas_lt::BlasLtActor;

pub use blas_lt::BlasLtMsg;

pub use blas_lt::Epilogue;

pub use blas_lt::HeuristicCacheRef;

pub use blas_lt::MatmulRequest;

pub use blas_lt::ScaleSet;

pub use blas_lt::WorkspacePool as BlasLtWorkspacePool;

pub use nvrtc::KernelArg;

pub use nvrtc::KernelHandle;

pub use nvrtc::NvrtcActor;

pub use nvrtc::NvrtcMsg;

pub use nvrtc::NvrtcOpts;

pub use collective::AllGatherRequest;

pub use collective::AllReduceRequest;

pub use collective::AllToAllRequest;

pub use collective::AllToAllvRequest;

pub use collective::BroadcastRequest;

pub use collective::CollectiveActor;

pub use collective::CollectiveMsg;

pub use collective::GroupGuard;

pub use collective::NcclCapabilities;

pub use collective::NcclReduceSupported;

pub use collective::PreMulSumOp;

pub use collective::RecvRequest;

pub use collective::ReduceRequest;

pub use collective::ReduceScatterRequest;

pub use collective::SendRequest;

pub use tensor::ComputeDesc;

pub use tensor::ContractRequest;

pub use tensor::ElementwiseBinaryRequest;

pub use tensor::ElementwiseTrinaryRequest;

pub use tensor::OperandSpec;

pub use tensor::PermutationRequest;

pub use tensor::ReductionRequest;

pub use tensor::TensorActor;

pub use tensor::TensorMsg;

pub use tensor::TensorSpec;

Modules

blas

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

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.

collective

CollectiveActor — wraps an [cudarc::nccl::Comm] for one rank within an NcclWorldActor group.

cudnn

CudnnActor — Phase 2 cuDNN slice. Wraps a [cudarc::cudnn::Cudnn] handle and exposes the v9 frontend graph API plus legacy ConvForward / Activation / Softmax shims for back-compat.

dispatch

Per-actor *Dispatch traits + their *DispatchCtx bundles (Phase 0.3).

envelope

kernel::envelope — shared kernel-actor body factored out of BlasActor::enqueue_sgemm.

fft

FftActor — wraps [cudarc::cufft::CudaFft] with an LRU cache of plans keyed by shape + transform kind + dtype + batch.

nvrtc

NvrtcActor — JIT-compile and launch user-supplied CUDA C++ kernels at runtime.

record

Capture-mode contract.

rng

RngActor — wraps a cuRAND curandGenerator_t handle and fills device buffers with the full distribution matrix.

solver

SolverActor — wraps a [cudarc::cusolver::DnHandle] for dense linear algebra and a [cudarc::cusolver::SpHandle] for sparse solves (gated cusolver-sp).

tensor

TensorActor — wraps cuTENSOR for contractions, reductions, permutations, and binary/trinary elementwise ops.

Structs

CsrMatrix

Legacy CSR sparse matrix in device memory — kept for back-compat with callers built against F-9. Prefer [SparseMatrix] for new code.

SparseActor

Enums

Activation

Available activation for kernel fusing in matmul

ReduceOp

SparseMsg

Public messages for SparseActor.