Crate cudakit 

Driver-only CUDA host stack shared by speakrs and scriptrs

cudakit loads the NVIDIA driver (libcuda) at run time through cudarc and never loads the CUDA Runtime (libcudart), so a build needs no CUDA toolkit and a process needs only the driver. With the libraries feature, and only when the consumer’s Profile allows it, cuBLAS, cuDNN and NVRTC are loaded lazily after a symbol preflight.

• Gpu owns one context and one stream, the device attributes and the optional library handles
• Runtime adds a module cache for a consumer’s KernelAreas: embedded PTX and exact-architecture cubins, loaded by exact identity (ModuleRequest)
• DeviceTensor is a shaped device buffer
• Sgemm and SgemmBatched describe row-major cuBLAS products
• Gpu::capture records stream work as a replayable Graph
• dnn plans cuDNN convolutions (feature libraries)
• identity reads the driver release for persisted tuning identities

Kernels are written with cuda-oxide in each consumer’s own kernels crate and built offline by cudakit-build; see the guide

Re-exports

pub use cudarc;

Modules

dnnlibraries

cuDNN forward convolutions: one planner and one plan type for every model

identity

Driver releases for persisted identities, such as tune files, with API-only identities kept apart so they are never saved as a release

symbols

Library entry points that cudakit checks before cudarc loads a library

Macros

embed_tier

Embeds one area variant under exactly the consumer features for which it is the best shipped PTX

Structs

AreaPtx

The PTX variants of one area embedded by this build

ArtifactHash

Content identity of bytes embedded in the running binary

BlasGuardlibraries

The shared cuBLAS handle in one math mode, from Gpu::blas

ComputeCapability

A device compute capability, ordered by major then minor version

Conv2d

An NCHW 2-D convolution on FP32 buffers (cross-correlation, as in PyTorch Conv2d)

DeviceAttributes

Device facts queried once per Gpu

DeviceAttributesBuildertest-support

Builds attributes for host-only tests in consumers

DeviceTensor

A device buffer with a row-major shape

EmbeddedCubin

One embedded binary, available only alongside its source PTX variant

EmbeddedPtx

The PTX fallback and exact-architecture binaries for one tier of one area

ExtraSymbols

Library entry points a consumer calls directly through cudarc, beyond the ones cudakit’s own wrappers reach

Gpu

One CUDA device context and stream, with optional libraries prepared on request

Graph

A recorded and instantiated sequence of work on one Gpu’s stream

LibraryVersionslibraries

Versions of the loaded numerical libraries, for tune-file identities

LoadedKernels

A kernel module loaded into a CUDA context

ModuleRequest

A complete module identity, resolved before the driver sees any bytes

Options

How to open a Runtime; consumers fill it from their own environment variables

Profile

A consumer’s build-wide CUDA policy, declared once as a constant

Runtime

A Gpu plus the kernel modules of one consumer’s areas

Sgemm

A row-major GEMM on FP32 buffers: c = alpha * op(a) · op(b) + beta * c

SgemmBatched

batch independent Sgemm products whose operands sit at fixed element strides in three buffers, such as one product per attention head

TierSet

A set of PTX tiers, usable in const context

Enums

ArtifactLoadError

Failure to load the requested artifact, without substituting another artifact

Error

Errors from cudakit

Library

A CUDA shared library that cudakit loads at run time

LibraryPolicy

Whether a consumer may load optional numerical libraries

LoadedArtifact

The exact artifact accepted by the driver for a kernel area

Math

Arithmetic precision of cuBLAS and cuDNN work on FP32 data

PtxTier

A GPU generation that committed PTX can target

Traits

KernelArea

A consumer’s kernel areas: one cuda-oxide module each, embedded per PTX tier

Functions

check_len

Checks that a buffer holds exactly the elements an operation reads or writes

element_count

Number of elements in a shape, failing on overflow

load_artifact

Load exactly the requested bytes; driver rejection never tries the other format

to_c_int

Converts a dimension for a cuBLAS or cuDNN call

unfold_windows

Writes rows of window samples cut out of a device span into target