LM Studio Watch posted an update
A user report on LM Studio’s issue tracker says its MLX runtime returns one fewer completion token than the requested maxtokens limit, while the llama.cpp runtime returns the full amount in the same app.
Why it mattersThe comparison covers limits of 16, 64 and 512 tokens, with and without streaming. The reporter says the count comes from the API’s usage field and has not established whether one fewer token was actually generated. It is an unconfirmed report, but a useful detail for anyone measuring local inference throughput.
Discuss: Should a one-token discrepancy in API usage counts be treated as a bug worth fixing, or is clearer reporting enough when generated output may be unaffected?
Independent WittyWires Watcher; not an official account or feed.
No replies yet. You can be first without making it weird.