[training topologies]
Training topologies
How members train together: roles, slots and the steps of a round. Studio compiles a topology into a run; a template binds one by exact version.
Synchronous DiLoCo
@beam/diloco
H local steps per round on workers with optimizer state, an outer aggregation on the lead and a broadcast of the merged model.
DiLoCo (elastic)
@beam/elastic-diloco
One role holds run.lead, data.grant and the model; members join and leave and pull data units (the CR-2 star).
Federated averaging
@beam/fedavg
Rounds of local training on every worker, a tree all-reduce merge of their models and a stop on the training loss.
Federated averaging on bucket data
@beam/fedavg-data
Rounds of training on a room bucket prefix; each worker trains on its own share of the files (kept across rounds), then a tree all-reduce merge and a stop on the training loss.
FedAvg (live session)
@beam/fedavg-live
A server holds run.lead and the model; a fixed cohort of clients trains each round and the server averages every client's update (quorum = all clients).
Gossip ring
@beam/gossip-ring
Each worker keeps its model and mixes in its ring predecessor's every iteration, then all workers merge the final models.
Scored subnet
@beam/scored-subnet
Miners pull data shards and send gradients on a broadcast lane; validators observe every gradient, score it and send signed scores to the model holder, which merges the round weighted by the median of their scores. A round nobody scored is weighted by units. Validators need the runs.observe and runs.score room permissions and join by invite only.
Split authorities
@beam/split-authorities
The run lead is not the aggregator: two stars share the trainers.
DiLoCo (fixed cohort)
@beam/static-diloco
One role holds run.lead and the model; a fixed cohort contributes without data grants (the CR-1 star).
Subnet (miners and validators)
@beam/subnet
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.
Subnet with observing validators
@beam/subnet-observed
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).
Hyperparameter sweep
@beam/sweep
A fan-out over sweep points from a common baseline, one merged model per point, then a final merge.
Two models
@beam/two-models
Two lanes from the same trainers to two state holders, one with data grants and one without.