Bug: gemma-2-9b-it inference speed very slow 1.73 tokens per second

[ninth99]

What happened?

System Info
Device: Ascend 910B3
OS: Ubuntu 20.04.6 LTS
Arch: aarch64

command:
./build/bin/llama-cli -m ./models/gemma-2-9b-it.gguf -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 33 -sm none -mg 0

log:
llama_perf_sampler_print: sampling time = 152.66 ms / 414 runs ( 0.37 ms per token, 2711.82 tokens per second)
llama_perf_context_print: load time = 7152.40 ms
llama_perf_context_print: prompt eval time = 619.28 ms / 14 tokens ( 44.23 ms per token, 22.61 tokens per second)
llama_perf_context_print: eval time = 230735.63 ms / 399 runs ( 578.28 ms per token, 1.73 tokens per second)
llama_perf_context_print: total time = 231911.05 ms / 413 tokens

Name and Version

./build/bin/llama-cli --version
version: 3923 (becfd38)
built with cc (Ubuntu 9.4.0-1ubuntu1~20.04.2) 9.4.0 for aarch64-linux-gnu

What operating system are you seeing the problem on?

Linux

Relevant log output

The CPU usage is very high, while the NPU usage is low, suggesting that the NPU is not being utilized during inference.