Thanks to the LMDeploy team for providing AWQ quantization support (https://github.com/InternLM/lmdeploy/blob/main/docs/en/multi_modal/xcomposer2d5.md#quantization). We compare the memory usage between the FP16 model and the 4-bit model, setting cache_max_entry_count=0.01 to reduce GPU memory usage and better observe memory savings. The program was tested on PyTorch 2.2.2+cu118.
| GPU Memory (GB) | |
|---|---|
| IXC2.5-lmdeploy | 31.81 |
| IXC2.5-lmdeploy-4bit | 23.21 |