Open Models Almanac
Methodsoptimizationtraining parallelism

single mention · optimization

Replicated, unsharded LoRA parameters under FSDP

Each GPU retains a full copy of LoRA adapter parameters instead of sharding them under FSDP.

The sources on this page only mention it in passing. The line above is the extraction's one-line gloss. See training parallelism for the methods filed beside it.

Filed under optimization › training parallelism