Why is fine-tuning so slow?

[simplelifetime]

InternLM2-xcomposer has the same size of LLM as LLaVA-1.5 and the same CLIP. However, the fine-tuning time is almost 3.5 times slower than LLaVA-1.5. What's causing it?

[L0CUST]

Can I know how many gpu you use?

[simplelifetime]

8 x A100 80G

[yuhangzang]

Our model is designed for high-resolution and long-context inputs/outputs. You can adjust the fine-tune hyper-parameters, e.g., hd_num, max_len to accelerate.