[training topology]
@beam/subnetSubnet (miners and validators)
Shaped like a Bittensor training subnet; miners pull data shards and send gradients to the model holder; validators send their scores of the miners to a consensus holder that merges them into weights each round (the tempo). Members register and leave freely.
[topology identity]
Details
What version 1.0.3 declares.
- Package
@beam/subnet- Membership
- Elastic
- Engines
- distributed run
- Tags
- bittensor, data-grants, elastic, live-session, subnet
- Step kinds
- local, gather, broadcast
- Latest tag
- latest -> 1.0.3
- Publisher
- registry
- Visibility
- Public
- Checksum
sha256:a55f35c3761617846cb44b4559117c520b04f723f8d8b80a2c599654bbab6e6e
[roles]
Roles
Who takes part, and how many members each role holds.
Hands out data shards and holds the model.
Leads the run and holds the weights.
Registers and leaves freely; pulls data shards.
Registers and leaves freely; scores the miners.
[slots]
Slots
The actions a run binds: what each slot consumes and produces.
Initializes the model the miners train from and the validators score.
Merges the miners' gradients into the next model version.
Evaluates the miners and sends its scores.
Trains on pulled data shards and sends a gradient.
Merges the validators' scores into weights (the Yuma consensus step).
[round]
Step kinds
The steps of one round, in order.
- localrole miner · slot train
- localrole validator · slot score
- gatherfrom miner · to lead · object gradients · slot merge
- gatherfrom validator · to chain · object scores · slot consensus
- broadcastfrom lead · to miner · object model