Reliable crash test in 2.7.5 and 2.8.0pre1
chrisbarrera opened this issue · comments
Bug Report
Repeatable crash on Apple M3 using Qwen 1.5 1.8B with this input "make a poem about gandalf" and letting it run for about 7 seconds. If it stops sooner than that without crashing then repeat the exercise until it does. Happens on both Metal and CPU under 2.8.0pre1, just takes longer under CPU.
Crashed Thread: 15 QSGRenderThread
Exception Type: EXC_BAD_ACCESS (SIGSEGV)
Exception Codes: KERN_INVALID_ADDRESS at 0x0000000000000008
Exception Codes: 0x0000000000000001, 0x0000000000000008
Thread 15 Crashed:: QSGRenderThread
0 QtGui 0x1039e98e0 QTextLine::textStart() const + 8
1 QtQuick 0x1015dd524 0x101504000 + 890148
2 QtQuick 0x1015e1a70 0x101504000 + 907888
3 QtQuick 0x1015c04c8 QQuickTextEdit::updatePaintNode(QSGNode*, QQuickItem::UpdatePaintNodeData*) + 5084
4 QtQuickTemplates2 0x1035a8ef0 QQuickTextArea::updatePaintNode(QSGNode*, QQuickItem::UpdatePaintNodeData*) + 340
5 QtQuick 0x1015f2c28 QQuickWindowPrivate::updateDirtyNode(QQuickItem*) + 2676
6 QtQuick 0x1015ed838 QQuickWindowPrivate::updateDirtyNodes() + 316
7 QtQuick 0x1015ed504 QQuickWindowPrivate::syncSceneGraph() + 360
8 QtQuick 0x10177a470 0x101504000 + 2581616
9 QtQuick 0x10177a870 0x101504000 + 2582640
10 QtQuick 0x10177b8a0 0x101504000 + 2586784
11 QtCore 0x1040338dc 0x103e58000 + 1947868
12 libsystem_pthread.dylib 0x19d99ef94 _pthread_start + 136
13 libsystem_pthread.dylib 0x19d999d34 thread_start + 8
Steps to Reproduce
- Load Qwen 1.5 1.8b
- Write "make a poem about gandalf"
- Wait 7-15 seconds (seems like 2.8.0 crashed quicker than 2.7.5) and repeat query if necessary.
Expected Behavior
Not Crash?
Your Environment
- GPT4All version: 2.7.5 and 2.8.0pre1
- Operating System: MacOS Sonoma 14.4.1
- Chat model used (if applicable): Qwen 1.5 1.8b
I'm going to try to reproduce this, but "Qwen 1.5 1.8B" is not specific enough. Please provide the exact filename of the model you are using. This includes the quantization type, as well as the precise name of the model (since there is a model called "Qwen1.5-1.8B-Chat" that is applicable to GPT4All, and a model called "Qwen1.5-1.8B" which is a model that has not been trained to understand user/assistant conversations and is intended as a base for finetuning).
It would be even better if you provided a link to where I can download the model, or at least a checksum so I can be sure I am using the same one.
Also, you'll have to be more specific than "repeat the exercise". Do I restart GPT4All, clear the chat, make a new chat, click "regenerate response", or simply type the same message into the box again and press enter?
Model file is: qwen1_5-1_8b-chat-q4_0.gguf downloaded directly from the Qwen page on huggingface.
% cksum qwen1_5-1_8b-chat-q4_0.gguf
1501852466 1120235360 qwen1_5-1_8b-chat-q4_0.gguf
Default prompt templates.
For me, just typing that query will have it write continuously until the chat crashes after 7 seconds. If it doesn't, try regenerate, but I haven't had to do that much, it was reliably generating for 7-15 seconds until crash.