the GPU is bieng used even though the model finished his text generation , OLLAMA

[gitnohubz]

ipex-llm[cpp]==2.2.0b20250204
intel-oneapi-basekit 2024.0.0.49564-3

iam using ollama on iGPU , some models chat the first time then they go crazy like if they are drunk , garbage output as i can say

[qiuxin2012]

Please give us more details about how to reproduce your error.
Your cpu type, OS info, model type, input&output and etc.

[gitnohubz]

@qiuxin2012
iGPU under a linux system , this is an easy issue to reproduce , use any model , take llama3.2 & chat with it & ask it , switch between models if your using openwebui , the same happens yo deepseek all qwen models , just any damn model , give me a single model that can work with ipex without a garbage output , or the gpu still bieng used even though you ended chating with the model , its not an obscure issue , its clear , just use it to come across it , iam not the only one facing that i saw multiple people mention this ollama garbage output & igpu utilization even though you ended chating.

I dont know about the development procces of the lib , but i think its a good idea to have a stable branch & a testing one.

I also tried ollama in the co tainer its worse there , its barely can say hello , dont have default behaviour of lunching start-ollama script automatically , every single boot you need to run the docker command to enter it & lunch it , it hurts my automated services.

[qiuxin2012]

We are looking into it.