I wrote about problem /thread/86102#256651when after use 0.6.2/3 version my mlmodel became 2x slower
Solution: remove "hint_fallback_from_metal" from deconvolution layer description.
I wrote about problem /thread/86102#256651when after use 0.6.2/3 version my mlmodel became 2x slower
Solution: remove "hint_fallback_from_metal" from deconvolution layer description.
How do you do that?
1) compile model on device
2) go to model dir and change the flag "hint_fallback_from_metal" in the file model.espresso.net
Hmm, this doesn't seem to happen when you compile the models with Xcode. At least I can't find that flag in the models in my app bundle.
On which device is this? I benchmarked my models and it seems the device needs to have device family 2 or higher in order to support accelleration with Metal.