Struct StreamShardAllocator 

pub struct StreamShardAllocator { /* private fields */ }

Streaming shard allocator.

Callers feed (key, cost) pairs; the allocator returns the target device plus a running snapshot of per-device load. Initial landing is shard_by_blake3. If the target device’s running cost exceeds the least-loaded device’s cost by more than spill_threshold, the item spills to the least-loaded device. Least-loaded selection is heap-backed and keeps stale heap entries compacted to GPU-count scale.

Implementations

impl StreamShardAllocator

pub fn new(n_gpus: u32, spill_threshold: u64) -> Result<Self, StreamShardError>

Create an allocator for n_gpus devices with an initial zero-cost load vector.

pub fn seed_load( &mut self, device: u32, cost: u64, ) -> Result<(), StreamShardError>

Inject pre-existing load, such as already-queued work.

pub fn assign( &mut self, key: &[u8], cost: u64, ) -> Result<Option<u32>, StreamShardError>

Assign one item.

Returns the chosen device index, or None when cost is zero.

pub fn load(&self) -> &[u64]

Snapshot of per-device cost. Index = device id.