Fix: eviction_engine does not run nvidia-smi even on AMD configs anymore and some testing problems fixed.

[GabrielReusRodriguez]

Hi!

I found some bugs on eviction engine.

First, my config is all AMD hardware and when I launched lemonade I saw this line repeated all time:

sh 1: nvidia-smi not found.

looking at code, I saw at system_info.cpp funtion double SystemInfo::get_global_vram_usage_pct() that it was always executing the call to nvidia-smi program so if somebody had amd config, it will not find this program.

I fixed making a get_cuda_arch().empty()) call to check if you have some nvidia arch on your config. If so.. it launches the code.

Second, I tryed the server_eviction.py tests. but I found some errors.

• The Script loads a test model phi-3-mini-4k-instruct-q4 that did not exists on lemonade so it failed. I changed by equivalent Phi-4-mini-instruct-GGUF.
• The tests usually loaded 2 models but it did not changed the by default config so the server always had max_models_loaded = 1 and failed when the second model was loaded.
• Finally at def test_weight_factor_protects_model(self): test , i found that it loads the first model (with weight), sleeps 2 seconds and loads the second one. It always failed and unloaded the one with weight. Looking the code and doing some manual testing I found that when it loaded the second model, the first model was about 3 seconds at server and the seconds 3 ms. So the 1000 value of weight did not difference. The solution I thought was do another sleep after the second model load and when I executed i found it passed the test!! it unload the second one .

I hope these changes help.