2 x A770 with Ollama Linux , inference responses slow down dramatically

[RobinJing]

Hi,
I use the latest cpp docker and the latest i915 dkms driver on Linux, inference with 2x A770 and Ollama, the model is Deepseek 32B int4. However, after 2 or 3 times we get 15~20 tokens, the rest of time we only get 6-8 tokens/s, also, the speed is very strange, it looks quite fast at the beginning, then drops dramatically after like 50-100 tokens, and sometimes it speeds up after 10-20 seconds..

Also, the gap between ollama and vllm is 5x times on the same machine, which is typically a bug performance.
BR