Generation slowing down as context fills up

[Mushoz]

I know the generation speed should slow down as the context starts to fill up, as LLMs are autoregressive. However, should the drop in speed be as severe as I am experiencing? I can't imagine running models at 32k or longer context sizes if the slowdown is already so substantial at sub 8k levels:

Tokens in context: Generation speed

17: 24.84 tokens per second
785: 22.67 tokens per second
1633: 20.75 tokens per second
2503: 18.91 tokens per second
3471: 17.24 tokens per second
4578: 15.79 tokens per second
5543: 14.62 tokens per second
6629: 13.35 tokens per second

I am running llama.cpp pulled 3 days ago on my 7900xtx through the following command:

llama-server --port 8999 -m /models/Qwen2.5-Coder-32B-Instruct-Q4_K_S.gguf -ngl 999 --ctx-size 8192 -fa

[wooooyeahhhh]

What backend? And yes generation speed should drop significantly