Module multi_gpu 

Mockable multi-GPU work partitioning. Multi-GPU adapter probing, backend acquisition, and work partitioning.

The pure partitioner stays separately testable, but production callers can now derive device loads from live wgpu adapters and acquire one backend per selected adapter instead of stopping at scheduling math.

Two allocation modes

1. Batch + cost-aware (partition_work_stealing): caller knows every work item + its cost up front; LPT greedy assigns the heaviest item to the least-loaded device.
2. Stream + content-addressed (shard_by_blake3 + StreamShardAllocator): caller yields (key, cost) pairs one at a time from a walker. The initial device is blake3(key)[0] % n_gpus for deterministic affinity - files with the same path always land on the same GPU across runs, which enables cache-warm re-scans. Overflow (queue on the target GPU is already loaded above threshold) spills to the least-loaded neighbor to keep tail latency bounded.

Structs

BorrowedGpuWorkItem

Borrowed multi-GPU work packet for hot scan paths.

DeviceLoad

Current device load snapshot.

GpuWorkItem

One executable multi-GPU work packet.

GpuWorkOutput

Output from one dispatched GpuWorkItem.

LiveGpu

A live GPU selected for multi-device scheduling.

MultiGpuExecutor

Real multi-GPU executor backed by one crate::WgpuBackend per adapter.

Partition

Work assigned to one device.

StreamShardAllocator

Streaming shard allocator.

WeightedWorkItem

One pending unit of GPU work.

Functions

live_gpu_loads

Enumerate live wgpu GPU adapters as zero-load scheduling targets.

partition_work_stealing

Partition work by repeatedly assigning the largest remaining item to the least-loaded device.

shard_by_blake3

Deterministic content-addressed device pick.