[training topology]
@beam/subnet-observedSubnet with observing validators
Miners pull data shards and send gradients to the model holder on a broadcast lane; validators receive every gradient (and every model version) to score the miners. Validators need the runs.observe room permission and join by invite only (CR-3).
[topology identity]
Details
What version 1.0.2 declares.
- Package
@beam/subnet-observed- Membership
- Elastic
- Engines
- distributed run
- Tags
- bittensor, data-grants, elastic, live-session, observers, subnet
- Step kinds
- local, broadcast, gather
- Latest tag
- latest -> 1.0.2
- Publisher
- registry
- Visibility
- Public
- Checksum
sha256:fddb30c0db6c86225503aeb6091376d8e7306a7ca367c5d406628974860f65ef
[roles]
Roles
Who takes part, and how many members each role holds.
Leads the run, hands out data shards and holds the model.
Registers and leaves freely; pulls data shards.
Observes every gradient and model version (runs.observe room permission).
[slots]
Slots
The actions a run binds: what each slot consumes and produces.
Merges the miners' gradients into the next model version.
Trains on pulled data shards and sends a gradient.
Scores the miners from every gradient it observes.
[round]
Step kinds
The steps of one round, in order.
- localrole miner · slot train
- broadcastfrom miner · to validator · object gradients
- gatherfrom miner · to lead · object gradients · slot merge
- broadcastfrom lead · to miner · object model
- broadcastfrom lead · to validator · object model
- localrole validator · slot observe