Module gpu

GPU acceleration module (currently CPU fallback) GPU acceleration for neural network operations

This module provides GPU-accelerated implementations of neural network primitives. Includes CUDA support via safe wrappers, mixed precision operations, multi-GPU training, and comprehensive GPU memory management.

Structs

CudaTensor

CUDA safe wrapper for tensor operations

DeviceInfo

GPU device information

GpuContext

GPU context for managing multiple devices

GpuMemoryHandle

GPU memory handle

MemoryStats

GPU memory statistics

MixedPrecisionConfig

Mixed precision configuration

MultiGpuTrainer

Multi-GPU training coordinator

NeuralOps

Enhanced neural operations accelerator with full GPU support

Enums

ReductionStrategy

Gradient reduction strategy for multi-GPU training

Functions

create_neural_ops

Helper function to create neural operations with automatic backend detection

create_neural_ops_with_backend

Helper function to create neural operations with preferred backend