bug: Vulkan performance degrades after multiple restarts

[NitroOxid]

Version: 2.0.37

Describe the Bug

Performance degrades significantly after restarting Atomic Chat several times.
Normal performance is approximately 25 tokens/sec, but after several restarts it drops to approximately 11 tokens/sec.

Environment:

• GPU: AMD Radeon RX 9070 16 GB
• Backend: Vulkan
• Model: Qwen 3.6 27B IQ3_XSS
• Context: 30k–80k
• Flash Attention: Enabled
• KV cache: K Q8_0, V Q4_0

The issue occurs with both llama.cpp and llama.cpp-turboquant Vulkan backends. Reducing the context to 30k does not resolve it.

LM Studio and Jan do not have this problem. When Atomic Chat is degraded to ~11 tokens/sec, Jan can immediately run the same model at ~25 tokens/sec. Closing Jan and launching Atomic Chat again still results in ~11 tokens/sec.

A full Windows reboot restores Atomic Chat to ~25 tokens/sec.

Steps to Reproduce

1. Reboot Windows.
2. Start Atomic Chat and load Qwen 3.6 27B IQ3_XSS using Vulkan.
3. Observe approximately 25 tokens/sec.
4. Restart Atomic Chat several times and reload the model.
5. Performance eventually drops to approximately 11 tokens/sec.
6. Close Atomic Chat and run the same model in Jan — approximately 25 tokens/sec.
7. Close Jan and start Atomic Chat again — approximately 11 tokens/sec.
8. Reboot Windows — Atomic Chat returns to approximately 25 tokens/sec.

Screenshots / Logs

app.log

Operating System

• MacOS
• Windows
• Linux

[NitroOxid]

Everything is the same when using ROCm on llama.cpp without turboquant support.
If do not restart Windows, and after a while (20-80 minutes) start Atomic Chat again.. There is a small chance that the speed recover.
I would be that the problem is the lack of VRAM.. But it's not like that!

[NitroOxid]

This is bug llama.cpp, not atomic chat app.