Crate cubecl_runtime 

The runtime API of CubeCL: what a kernel author compiles against.

A Client reaches a device, a Handle names memory on it, and a CubeKernel is what a launch hands over. The Server trait is the contract a runtime implements, and everything an implementation needs beyond the contract — memory pools, streams, drivers, the compilation pipeline — lives in cubecl-server, so that user code never depends on it.

Error handling

A failure belongs to the memory that work left unwritten and travels the dataflow from there: nothing here holds error state beside the buffers. Taint, ErrorGraph, Failures and ExecuteScope in cubecl-server are the four pieces, and ERROR_HANDLING.md at the repository root is the argument for why they are shaped that way — read it before changing any of them.

Modules

benchmark

Module for benchmark timings

client

Compute client module.

compiler

Compiler trait and related types

config

CubeCL config module.

device

The device type of each runtime. The device type of each runtime, named without naming the runtime crate.

dry_run

Running a workload for the compilation and tuning it provokes, without running the workload itself. Running a workload for the compilation and tuning it provokes, without running the workload itself.

id

Various identifier types used in CubeCL.

kernel

Kernel related traits.

launched

The kernels a workload launches, collected while it replays: what an environment shipped for that workload has to keep, and all it has to.

logging

Logging utilities to be used by a compute server.

memory_management

Memory management value types: reports, configuration and the managed handle. The pools behind them are cubecl-server’s.

runtime

Runtime trait and related types

server

Compute server module.

storage

Compute Storage contract.

throughput

Throughput related utilities.

tma

TMA-related runtime types

tune

Autotune module

Macros

local_tuner

Create a local tuner with the provided name.

storage_id_type

Create a new storage ID type.