Jeffser
Alpaca Flatpak crashes after second prompt when using built-in Ollama instance (CPU-only, AMD system)
Describe the bug
Alpaca (Flatpak version 8.4.3) consistently crashes after sending the second prompt in any chat when using the built-in Ollama instance. The issue happens even in CPU-only mode and even with small models (Llama 3.2 3B Q4).
This seems to originate from the embedded Ollama 0.13.0 runner, which silently dies after the first request, causing Alpaca to throw:
('Connection aborted.', RemoteDisconnected('Remote end closed connection without response'))
Expected behavior
Alpaca should not crash when sending the second prompt.
The embedded Ollama instance should stay alive.
Environment
OS: Flatpak / Linux
Alpaca Version: 8.4.3 (Flatpak)
Bundled Ollama version: 0.13.0
Hardware:
CPU: AMD Ryzen 7 9800X3D
GPU: AMD Radeon 9070 XT (Vulkan available but disabled during testing)
RAM: 64 GB
Model: Llama 3.2 3B Instruct (Q4_K_M), default settings
Steps to Reproduce
Install Alpaca from Flathub.
Start Alpaca from terminal:
flatpak run --env=GGML_VK_VISIBLE_DEVICES=-1 com.jeffser.Alpaca
(forces CPU-only mode)
Create a new chat.
Send the first prompt → works fine.
Send a second prompt → Alpaca crashes.
This happens regardless of:
model size (even 3B fails)
GPU enabled or disabled
ROCm/Vulkan settings
context length or temperature
Observed Behavior
The first prompt processes normally.
On second prompt:
The embedded Ollama runner stops responding.
The UI crashes.
The terminal shows no explicit error — but the Ollama runner stops emitting logs and dies silently.
The client receives:
RemoteDisconnected('Remote end closed connection without response')
Relevant Logs
GGML_VK_VISIBLE_DEVICES=-1 inference compute id=cpu library=cpu ... model weights device=CPU size="1.9 GiB" kv cache device=CPU size="1.8 GiB" runner ... Server listening on 127.0.0.1:42011 waiting for llama runner to start responding
No further errors are shown — the runner simply stops responding after the second request.
Please include the output of Alpaca, for this you'll need to run Alpaca from the terminal, then try to reproduce the error you want to report.
flatpak run com.jeffser.Alpaca
