Topologies

[training topology]

@beam/elastic-diloco

DiLoCo (elastic)

One role holds run.lead, data.grant and the model; members join and leave and pull data units (the CR-2 star).

Elastic membershipBuiltinActivedistributed run

[topology identity]

Details

What version 1.0.2 declares.

Package
@beam/elastic-diloco
Membership
Elastic
Engines
distributed run
Tags
data-grants, diloco, elastic, live-session, star
Step kinds
local, gather, broadcast
Latest tag
latest -> 1.0.2
Publisher
registry
Visibility
Public
Checksum
sha256:224b700ee94c463d36ccceef8a7e12a7cfcf6ec29336cd102ec0ad9094b3d053

[roles]

Roles

Who takes part, and how many members each role holds.

sync1

Leads the run, hands out data units and holds the model.

learner1..8

Members that join and leave.

[slots]

Slots

The actions a run binds: what each slot consumes and produces.

merge
update* -> model

Merges the updates of a round into the next model version.

train
model, shards -> update

Trains on pulled data units from the latest model version.

[round]

Step kinds

The steps of one round, in order.

  1. localrole learner · slot train
  2. gatherfrom learner · to sync · object update · slot merge
  3. broadcastfrom sync · to learner · object model