How to improve latency?

[Abhranta]

I have a Llama3 8B model with IQ4_NL quantization. I want tp use this mode to run inference on an edge device (ARM CPU). I am getting a speed of around 2.3 tokens per second (Tested using llama-bench). I have tried --np-mmap and tried on multiple thread. mmap doesn't change the speed much and increasing thread count only improves the speed upto a certain point. Any other arguments or configurations to increase the speed ? I only have a CPU platform and cannot use a GPU for this.