Comment on Mesh LLM: distributed AI computing on irohComments−jmercouris2moI thought about this too, but the throughput over a network is incredibly slow. It’s not usable for interactive use.−stymaar2moThroughput is not a problem as you just share relatively small vectors (a few kilobytes in size), the key issue is network latency.−jmercouris2moI didn't mean the throughput of the network, I meant the throughput of the calculations. Yes, the latency is the issue.−i3862moThat isn’t true. llama RPC is incredibly slow but staged splits in skippy are orders of magnitude faster.
Comments
I thought about this too, but the throughput over a network is incredibly slow. It’s not usable for interactive use.
Throughput is not a problem as you just share relatively small vectors (a few kilobytes in size), the key issue is network latency.
I didn't mean the throughput of the network, I meant the throughput of the calculations. Yes, the latency is the issue.
That isn’t true. llama RPC is incredibly slow but staged splits in skippy are orders of magnitude faster.