Emine

That is fair. The gain is real and it is a one-off discovery, not a recurring reward.

Build options mattered more than anything the frontend exposed. A build compiled for the machine it runs on rather than a generic package was a meaningful speedup on identical hardware and identical weights.

Second thing: check what it reports at startup about which acceleration it found. Half the "why is this slow" posts are a build that silently fell back to plain CPU.