Core ML model initialization hangs with iOS 15

Question

Florian-D OP

Created Aug ’21

Replies 1

Boosts 1

Views 1.7k

Participants 2

Hey;

I am having issues with a specific CoreML model's initialization under iOS 15

Specifically:

iOS 15 build 19A5318f on iPhone 11
App built with Xcode 13.0 (13A5212g) targeting iOS 13.0

The code basically looks like this

let configuration = MLModelConfiguration()
configuration.computeUnits = .all

let t0 = CACurrentMediaTime()
let model = try MyModelGeneratedClass(configuration: configuration).model
let t1 = CACurrentMediaTime()
print("Loading time: \(t1 - t0)")

When running that code I see a few errors in the console:

Error: Convolution configuration cannot fit in KMEM (Given: 258048b, Max: 65536b)
Error: Convolution configuration cannot fit in KMEM (Given: 258048b, Max: 65536b)
Error: Convolution configuration cannot fit in KMEM (Given: 258048b, Max: 65536b)
Error: Convolution configuration cannot fit in KMEM (Given: 258048b, Max: 65536b)
Error: Convolution configuration cannot fit in KMEM (Given: 258048b, Max: 65536b)
Error: Convolution configuration cannot fit in KMEM (Given: 258048b, Max: 65536b)
Error: Convolution configuration cannot fit in KMEM (Given: 258048b, Max: 65536b)
Error: Convolution configuration cannot fit in KMEM (Given: 258048b, Max: 65536b)
Error: Convolution configuration cannot fit in KMEM (Given: 258048b, Max: 65536b)

and then the app hangs, until finally

Loading time: 116.33786087499993"

(it always is around 2 minutes)

If I change the configuration so that

configuration.computeUnits = .cpuAndGPU

then everything works fine;

Did I miss something ? The CoreML SDK just hanging there for 2 minutes definitely looks like an iOS 15 bug

Boost

Answer 1

ppix OP

Sep ’21

I just ran into the same thing. My guess is this is a bug in the CoreML compiler or on-device scheduler, where it tries to put part of the network onto the Neural Engine, even though it contains convolutions that need more memory than the Neural Engine can handle.

0