App crashes and deletes history after long discussions with Dolphin3-Llama3 models on Vulkan

[Nekzuris]

Bug Report

Steps to Reproduce

1. Load a model like bartowski/Dolphin3.0-Llama3.1-8B-GGUF
2. Ask it to tell you a long story, and then ask for another one when it’s done.
3. After asking 3 or 4 times, the app crashes, and the entire conversation history is gone when reopening the app, not just the last message, all messages since the model was first loaded.

Expected Behavior

App should not crash, and if it does, the conversation history should still be saved.

Your Environment

• GPT4All version: v3.7.0
• Operating System: Windows 10
• Chat models used: (all with default settings)
  • bartowski/Dolphin3.0-Llama3.1-8B-GGUF
  • QuantFactory/Dolphin3.0-Llama3.1-8B-GGUF
  • bartowski/Dolphin3.0-Llama3.2-3B-GGUF

• bartowski/Dolphin3.0-Qwen2.5-3b-GGUF did not crash but entered a repetition loop.

log.txt and log-prev.txt don't say anything interesting.

[brynrmrz]

Models loaded from huggingface do not have guaranteed support. There may be additional configuration needed. I would not consider this an issue if I was maintainer.

Up to you if you want to leave open, just my two cents as a user who also works in development.

[Nekzuris]

I get that it's not critical or high priority, but since there's a built-in interface to download models from Hugging Face, it's not unreasonable to ask for a solution.
And I doubt that the issue with not saving chat history when the app crashes is related to Hugging Face models.
Chat history should be saved after each valid response so that if a crash occurs, only the last message is lost.

[Nekzuris]

I just realized that it's only an issue with Vulkan, CUDA works fine.