When I push a model that has multiple shards to a repo that originally had less or more shards, it fetches all of them, even if they're a different architecture (ie pushing a 3-shard model to what was a 1-shard model only fetches the 1-shard model, instead of overwriting it)
push to hub 1-shard model
push to hub 3-shard model
fetch from hub - only fetches model.safetensors, not model-00001-of-00003.safetensors through model-00003-of-00003.safetensors