about nuscenes data convert

[AMzhanghan]

The data conversion is very slow, about 6 minutes per frame. Why is it so slow？

[rb93dett]

Yes, the inference speed of LLaVA-1.6-34B is relatively slow, and each sample requires generating multiple QA responses. If you need to speed up data production, you might consider using 4-bit or 8-bit inference here. If you have sufficient GPU resources, you can also consider splitting the dataset and generating data in parallel.