Although I have one question @apeskov.
I have successfully tuned iOS model, but when I try to test it, I’m getting an error trying to execute function “run”. And I have noticed that some of my tuning tries have this issue.
Could you please give me an advice where to dig in order to fix that problem?
Hi,
It depends. What particular problem do you faced? Could you provide some error message example?
Unfortunately metal backend is not mature as would like. There is known problem with improper detecting of maxTotalThreadsPerThreadgroup. It may be good enough for tuning period of time, but wrong while using via graph_executor api.
Other issues are possible as well…
You may set breakpoint into metal_module.mm and define a real reason of failure.
That’s the problem, that run function doesn’t give me any text feedback about why it has crashed. I’ll dig more into that and will do my best to fix it and provide pull requests
I may be experiencing a similar issue. Did you ever find a solution to the error?