Skip to main content

Module elastic

Module elastic 

Source

Structs§

ElasticConfig
Configuration of an elastic training run.
ElasticCoordinator
Drives an elastic training group: a join/leave state machine over a versioned rendezvous, with balanced re-sharding and linear learning-rate scaling.
EpochSnapshot
A point-in-time snapshot of the elastic group after an accepted event.
RendezvousState
The agreed rendezvous view shared by every worker at a given epoch.
ShardRange
A contiguous half-open range [start, end) of shard indices owned by a worker.
WorkerShards
The (generally non-contiguous) set of shards a worker owns under a hashed (rendezvous / HRW) assignment.

Enums§

MembershipEvent
A membership change requested against the elastic group.
WorkerState
Lifecycle state of a single worker in the elastic group.

Functions§

block_partition
Split the shard index space 0..dataset_shards into one contiguous ShardRange per member, balanced so the sizes differ by at most one.
hashed_assignment
Assign 0..dataset_shards to members by capacity-bounded rendezvous (HRW) hashing: each shard is placed on the highest-weight member that still has spare capacity, with capacities chosen so worker loads differ by at most one.