Module backend 

Backend abstraction split into compute and weight loading concerns

This module separates the previous “fat” Backend trait into focused interfaces: ComputeBackend for tensor operations and WeightLoader for model weight management.

Structs

BackendCapabilities

Backend capabilities description

BackendConfig

Backend configuration

BackendMemoryConfig

Memory configuration for backends

BackendRequirements

Requirements for backend selection

BackendStatus

Backend status information

KernelHandle

Handle for loaded kernel

KernelInfo

Kernel information and metadata

TensorSpec

Tensor specification for weight loading

WeightLoaderCapabilities

Weight loader capabilities

WeightLoaderConfig

Weight loader configuration

WeightMetadata

Weight metadata information

Enums

BackendType

Backend types

KernelArg

Kernel argument types

MemoryGrowthStrategy

Memory growth strategies

QuantizationConfig

Quantization configuration for weights

ScalarValue

Scalar values for kernel arguments

TensorTransformation

Transformations that can be applied to loaded tensors

TransformationType

Transformation types

WeightFormat

Weight file formats

WeightSource

Weight source specification

WeightSourceType

Weight source types

Traits

BackendFactory

Backend factory for creating backend instances

BackendRegistry

Backend registry for managing multiple backends

ComputeBackend

Compute backend for tensor operations and kernel execution

KernelExecutor

Kernel executor for custom GPU kernels

WeightLoader

Weight loading interface for model parameter management