Yuechen

> I am working on acceleration using mixed-precision training. Actually, it works fine when using a single GPU with several lines modification (about 8x faster than using 4 cards with...

哈哈 我本来也想提一个类似的issue，先感谢楼主

We adopt the 100K ShareGPT (caption) data in the SFT. I will calculate the number of image files used in SAM and I can extract them to form a shared...

Hi, the huggingface demo uses a less-powerful 13B-HD model and the demo at port 7860 uses our largest 34B-HD model. Thus the model may have some hallucinations and output wrong...

Here is a similar issue #11, hope this can help you.

I'm sorry for not getting back to you sooner. Those Ray-based losses backward are auto in the function grid.volume_render_fused(*), you may check it in svox2. As we pass the gt...

Hi, please refer to #6 if you encounter the same problem.

Please refer here: https://huggingface.co/datasets/YanweiLi/MGM-Instruction/blob/main/mgm_generation_pure_text.json