Perf should be fairly straightforward to ballpark. You'll need to transfer roughly 2 . hidden_size . num_shards bytes over the network per token during autoregressive decoding. And divide that number by chunk size during prefill.