[ipex-llm[cpp]][ollama] low performance and gpu usage when running minicpm3-4B model

[jianjungu]

I'm trying to running miniCPM3-4B:Q4_K_M model with ollama 0.5.1-ipex-llm-20250107 on intel MTL iGpu and Arc A770.

• When running the model on MTL igpu, there is no GPU compute usage with very high VRAM occupation

before load the model into VRAM, the VRAM usage is 2.4GB

when running the model, the compute usage is almost 0 and VRAM usage is 8.2GB(this 4B model is using VRAM higher than GLM4:9B)

• Runing the model on Arc A770

before load the model , the VRAM usage is 1.1GB

when runing the modek ,the compute usage is around 43% and VRAM usage is 9.5GB
running with "ollama run gfunsai/minicpm3-4b:q4_k_m" command

running with curl command

The output token per second performance is significant dropped when doing inference on both GPUs.

Please help fix this issue :)

[sgwhat]

Hi @jianjungu, we have reproduced this issue and are working on fixing it :)