I am running a CoreML model (vision transformer, fixed input size) with hardware units cpuAndNeuralEngine on both iPhone 15 and iPhone SE 2. The memory usage is normal on iPhone 15. On iPhone SE 2 [A13], it is more than double. The model loading on SE2 takes forever, and puts a few of these warnings in the console: [Espresso::handle_ex_plan] exception=ANECF error: failed to load ANE model [...] FAILED: err=( CompilationFailure ) doUnloadModel:options:qos:error:: nil _ANEModel ANECF error: Unable to build plan. @ PrepareOpForEncode E5RT: Unable to build plan. (2) Before seemingly falling back on CPU because inference completes successfully. But it looks as though the failed attempt at Espresso compilation leaves behind a ~800MB memory leak. What's interesting is, if I set CoreML to use cpuOnly and not cpuAndNeuralEngine, the memory usage is lower by 800MB to around ~100MB, less than what I would expect.

Facing the same issue. You can get around the queryable index requirement as follows: Change the predicate from NSPredicate(value: true) to a predicate w/ condition. If you really want to get all the records, you could make a dummy field which is always the same value and check that. before (DOES NOT WORK): let query = CKQuery(recordType: "Snapshot", predicate: NSPredicate(value: true)) database.perform(query, inZoneWith: zoneID) { records, err in ... Instead:         let query = CKQuery(recordType: "Snapshot", predicate: NSPredicate(format: "version = %d", argumentArray: [version])) now it doesn't complain. Will test this with CKQueryOperation as well. Especially useful if you're writing a library and using JIT schema - no need to open up the web UI at all.