pub fn reduce_state_dicts(
reducer: &dyn DeltaReducer,
contributors: &[ContributorState],
) -> Result<HashMap<String, Tensor>>Expand description
Aggregate N contributors’ full state dicts into one reduced state dict.
Each contributor supplies a (state_dict, provenance, norm_ratios) where the
state dict uses the extract_state_dict key convention
("layer.module.lora_a" / "layer.module.lora_b"). The reducer runs the
stack + truncated-SVD per layer key and emits reduced factors at the fixed
rank R. Effective-rank / Muon-momentum metadata are intentionally dropped
(fresh optimizer post-aggregation, per the spike).
Keys present in only some contributors are aggregated over the subset that has them — this is still order-independent and tolerant of contributors that trained different module sets.