Module tensor_device 

Shared tensor-backend adapters for host-staged rank and in-process collectives.

Transfers use the existing collective transports. Reductions execute through the selected backend, including its normal runtime, fusion and copy-on-write rules. This module does not provide a peer-memory or NCCL ABI implementation.

Structs

TensorBuffer

A typed, one-dimensional collective buffer on a tensor backend.

TensorDevice

Execution context shared by both collective device interfaces.

TensorReductionLaunch

Checked launch coordinates, in elements rather than bytes.

Enums

TensorDeviceError

Traits

TensorElement

Storage types supported by the shared collective device adapter.