Can I broadcast different models for each of the clients in federated_tensorflow?

Within the TFF language, there is not currently a way to express this. The contract of tff.federated_broadcast is that the value that arrives at each client is the same. There isn't an alternative communication intrinsic (yet) for other ways of moving tensors from server→clients.

That said, its possible to simulate clients that receive different weights. Setup the clients state outside TFF in the Python execution environment, then pass in a list (one element for each client). Example:

@tff.tf_computation(MODEL_WEIGHT_TYPE):
def client_work(model_weights):
  # Do local training, or any other logic.


@tff.federated_computation(tff.FederatedType(MODEL_WEIGHT_TYPE, tff.CLIENTS))
def run_one_round(model):
   return tff.federated_map(client_work, model)


client_models = [...]  # a list of each client, possibly different models.
run_one_round(client_models)