Topologies

[training topology]

@beam/diloco

Synchronous DiLoCo

H local steps per round on workers with optimizer state, an outer aggregation on the lead and a broadcast of the merged model.

Static membershipBuiltinActivedistributed run

[topology identity]

Details

What version 1.0.1 declares.

Package
@beam/diloco
Membership
Static
Engines
distributed run
Tags
diloco, rounds, stateful, synthetic
Step kinds
broadcast, local, tree
Latest tag
latest -> 1.0.1
Publisher
registry
Visibility
Public
Checksum
sha256:aca3df8a00b3d93495dccd3c94a94e140c9dc836366778481d3012cb97919401

[roles]

Roles

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

lead1

The member that initializes, merges and broadcasts the global model.

workers1..512

Every eligible room member runs the inner steps.

[slots]

Slots

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

init
model?, neighbor? -> model, report

Initializes the global model (a training run without compute or failures).

merge
update* -> model, report?

The outer step; associative, commutative and closed (a tree on the lead above 15 workers).

trainstateful
model?, neighbor? -> update, report

H inner steps from the received global model; keeps its inner optimizer state.

[round]

Step kinds

The steps of one round, in order.

  1. broadcastfrom lead · to workers · object model
  2. localrole workers · slot train
  3. treefrom workers · to lead · object update · slot merge