search

ATP-1010 / FederatedLLM

42 stars 5 forks source link

can i use client_selection_frac <1 ? #4

Open ningwanyi opened 1 week ago

ningwanyi commented 1 week ago

If client_selection_frac < 1, how can you keep the initialization of the clients in the next round same? I mean, how to keep them always having the same merged model parameters?

ATP-1010 commented 1 day ago

In our code you can use client_selection_frac < 1 since every round we use a new merged based model and new initialized LoRA.

ningwanyi commented 1 day ago

Does it mean we should broadcast the stacked lora to all the clients in each round, so as to ensure all of them have the new merged based model?