[iphone12 ios17] crash on version higer than 0.11.9 and very slow on version 0.11.9

[kyleLUO-GPT]

When I use a version higher than 0.11.9 to perform text embedding with the embeddinggemma-300m model.

When I use version 0.11.9 to perform text embedding with the embeddinggemma-300m model, obtaining results is extremely slow.

[DenisovAV]

Tried to reproduce on 0.12.3 with embeddinggemma-300M (seq256) on iPhone 14 Pro, iOS 17.4 - works fine, no crashes.

To investigate further I need:

• flutter_gemma version from your pubspec.yaml
• Device model and iOS version
• Which embedding model (embeddinggemma-300M, Gecko, or custom)?
• Debug or Release build?
• Full crash stacktrace from Xcode/Flutter console
• Does it crash on first embedding call or after multiple calls?

[kyleLUO-GPT]

Runner PID 1644
0% CPU

549.7 MB Memory

Energy Impact High

Disk 23.2MB/s
....
Network
Zero KB/s
FPS
0 FPS
Thread 1 Queue: com.apple.main-thread (serial)
Thread 3
Thread 5
Thread 6 Queue: com.apple.root.user-initiated-qos (concurrerht)
0 std::__1::basic_string<char, std::_1::char_traits, std::__1::allocator>::_is_long[abi:de200100]
1 std::__1::basic_string<char, std:__1::char_traits, std::__1::allocator>::_get_pointer[abi:de200100]
2 std::__1::basic_string<char, std::__1::char_traits, std::__1::allocator>::data[abi:de200100]
3 std::__1::basic_string<char, std::1::char_traits, std::__1::allocator>::c_str[abi:de200100]
4 sentencepiece::util::Status::error_message
5 sentencepiece::util::Status::message
6 -[SentencePieceWrapper encode:]
7 SentencePieceTokenizer.encode(:)
8 EmbeddingModel.initializePrefixCache()
9 EmbeddingModel.loadModel()
10 GemmaEmbedding Wrapper.initialize()
11 closure #1 in PlatformServicelmpl.createEmbeddingModel(rmodelPath:tokenizerPath:preferredBackend:completion:)
17_pthread_wqthread
Enqueued from com.apple.main-thread (Thread 1)
0 dispatch_async
2 merged _C.OS_dispatch_queue.asyncUnsafe(group: Swift.0Optional<_C.OS_dispatch_group>, qos: Dispatch.DispatchQoS, flags: Dispatch.Dispatch...
3 PlatformServicelmpl.createEmbeddingModel(modelPath:tokeenizerPath:preferredBackend:completion:)
5 closure #12 in static PlatformServiceSetup.setUp(binaryMessenger:api:messageChannelSuffix:)
24 start
Enqueued from com.apple.main-thread (Thread 1)
0 dispatch_async
92 start
Thread 9
com.apple.uikit.eventfetch-thread
Thread 16
0 start_wqthread
io.flutter.1.raster
jo.flutter.1.jo
io.flutter.1.profiler
io.worker.1
ic worker.2

[kyleLUO-GPT]

flutter_gemma version on 0.12.3

embedding model is embeddinggemma-300M seq512

iphone12 device with ios 17.6 debug model xcode

crash on first embedding call

i think the reason is about the text string

debug text string value is : "task: search result | query: "

[kyleLUO-GPT]

Thread 1 Queue: com.apple.main-thread (serial)
Thread 2 Queue: com.apple.root.user-initiated-qos (concurren)t)
0 std::__1::basic_string<char, std:___1::char_traits, std::__1::allocator>::__is_long[abi:de200100]
1 std::__1::basic_string<char, std:___1::char_traits, std::__1::allocator>::_get_pointer[abi:de200100
2 std::___1::basic_string<char, std::__1::char_traits, std::__1::allocator>::data[abi:de200100]
3 std::__1::basic_string<char, std::_1::char_traits, std::__1::allocator>::c_str[abi:de200100]
4 sentencepiece::util::Status::error_message
5 sentencepiece::util::Status::message
6 -[SentencePieceWrapper loadModel:error:]
7 SentencePieceTokenizer.init(modelPath:)
9 EmbeddingModel.loadModel()
10 GemmaEmbeddingWrapper.initialize()
11 closure #1 in Platform Servicelmpl.createEmbeddingModel(modelPath:tokenizer Path:preferredBackend:cormpletion:)
17 _pthread_wqthread
Enqueued from com.apple.main-thread (Thread 1)
0 dispatch_async
2 merged _C.OS_dispatch_queue.asyncUnsafe(group: Swift.Optional<__C.OS_dispatch_group>, qos: Dispatch.DispatchQos, flag...
3 PlatformServicelmpl.createEmbeddingModel(modelPath:tokenizerPath:preferredBackend:completion:)
5 closure #12 in static PlatformServiceSetup.setUp(binaryMesssenger:api:messageChannelSuffix:)
24 start
Enqueued from com.apple.main-thread (Thread 1)
0 dispatch_async
92 start

[kyleLUO-GPT]

modelPath value is "/var/mobile/Containers/Data/Application/2C32C9D1-62D4-4DF9-B2B4-5BBB121D43C0/Documents/sentencepiece.model"

in file [SentencePieceWrapper.mm]

this is the _processor object memory information

if (!status.ok()) {
(sentencepiece::SentencePieceProcessor *) Ox13f138f00
model_ (std::unique_ptrsentencepiece::Modellnterface)
pointer = (sentencepiece::Modellnterface *) Ox3
normalizer_ (std::unique_ptrsentencepiece::normalizer::Normalizer
pointer = (sentencepiece::normalizer:: Normalizer *) Ox107c3cb60
denormalizer_ (std::unique_ptrsentencepiece::normalizer::Normalizer
pointer = (sentencepiece::normalizer::Normalizer *) Ox1458259c0
model_proto_ = (std::unique_ptrsentencepiece::ModelProto) nullptr
pointer = (sentencepiece::ModelProto *) NULL
encode_extra_options_ = (std::vectorsentencepiece::SentencePieceProcessor::ExtraOption) size=error: Invalid value for end of vector
decode_extra_options_ = (std::vectorsentencepiece::SentencePieceProcessor::ExtraOption) size=0
Runner
Thread 31
0 -[SentencePieceWrapper encode:]