Module multi_gpu 

Multi-GPU coordination, topology discovery, and cross-GPU messaging.

This module provides infrastructure for coordinating work across multiple GPUs, including:

• Device Selection - Load balancing strategies for kernel placement
• Topology Discovery - NVLink/PCIe detection and bandwidth estimation
• Cross-GPU K2K Router - Kernel-to-kernel messaging across GPUs
• Kernel Migration - Move kernels between GPUs with state transfer

Example

use ringkernel_core::multi_gpu::{MultiGpuBuilder, GpuTopology, CrossGpuK2KRouter};

let coordinator = MultiGpuBuilder::new()
    .load_balancing(LoadBalancingStrategy::LeastLoaded)
    .enable_p2p(true)
    .build();

// Discover topology
let topology = coordinator.discover_topology();

// Create cross-GPU router
let router = CrossGpuK2KRouter::new(coordinator.clone());
router.route_message(source_kernel, dest_kernel, envelope).await?;

Structs

CrossDeviceTransfer

Helper for cross-device data transfer.

CrossGpuK2KRouter

Routes K2K messages across GPU boundaries.

CrossGpuRouterStats

Statistics for cross-GPU K2K routing.

CrossGpuRouterStatsSnapshot

Snapshot of router statistics.

DeviceInfo

Information about a GPU device.

DeviceStatus

Status of a device in the multi-GPU coordinator.

DeviceUnregisterResult

Result of unregistering a device from the coordinator.

GpuConnection

Connection between two GPUs.

GpuTopology

GPU topology graph describing all device interconnections.

HotReloadConfig

Configuration for kernel hot reload operations.

HotReloadManager

Manager for kernel hot reload operations.

HotReloadRequest

Request to hot reload a kernel.

HotReloadResult

Result of a completed hot reload.

HotReloadStatsSnapshot

Snapshot of hot reload statistics.

KernelCodeSource

Kernel code source for hot reload.

KernelMigrationPlan

Plan for migrating a single kernel during device unregister.

KernelMigrator

Migrator that uses checkpoints for kernel state transfer between GPUs.

MigrationRequest

Request to migrate a kernel between devices.

MigrationResult

Result of a completed migration.

MigrationStats

Statistics for kernel migrations.

MigrationStatsSnapshot

Snapshot of migration statistics.

MultiGpuBuilder

Builder for multi-GPU coordinator.

MultiGpuConfig

Configuration for multi-GPU coordination.

MultiGpuCoordinator

Multi-GPU coordinator for managing kernels across devices.

MultiGpuStats

Multi-GPU coordinator statistics.

PendingK2KMessage

A pending cross-GPU K2K message.

Enums

HotReloadState

State of a hot reload operation.

InterconnectType

Type of interconnect between GPUs.

KernelCodeFormat

Kernel code format.

LoadBalancingStrategy

Strategy for balancing load across devices.

MigrationPriority

Priority for kernel migration.

MigrationState

State of a kernel migration.

RoutingDecision

Decision for how to route a K2K message.

Traits

HotReloadableKernel

Trait for kernels that support hot reload.

MigratableKernel

Trait for kernels that support live migration.