Yanwei

yanwei-li.com [email protected]

The Chinese University of Hong Kong Hong Kong, China

Results 27 comments of


Yanwei

Which deepspeed version is it

Hi, it seems the CUDA version and compiled PyTorch version are not the same one. Please try to align them to 11.7 or 12.2.

Zero-3 offload support

Hi, thanks for this suggestion! Because we do not have enough resources to validate the results of Zero3-offload at this time, we will try to support it later.

RuntimeError: Expected all tensors to be on the same device, but found at least two devices, cuda:0 and cuda:1! (when checking argument for argument weight in method wrapper_CUDA__native_layer_norm)

Hi, could you please provide more details like which script and your settings.

AttributeError: 'NoneType' object has no attribute 'is_loaded'

Hi, it seems this issue could casued by [this line](https://github.com/dvlab-research/LLaMA-VID/blob/d1074f3662a772d1b3c723416af59314ba593f67/llamavid/model/builder.py#L91), please check this and make sure the `CLIPVisionTower` in `llamavid/model/multimodal_encoder/clip_encoder.py` is normally loaded.

An error occurs during the stage 2 fine-tuning

Hi, it seems the data is not well-loaded. Please make sure all the downloaded datasets are well organized as in README.

请问为什么在训练llama的脚本中，预训练和微调所使用的conv不一样

Hi, for the LLaMA 7B and 13B, we follow the instruction format in [LLaVA](https://github.com/haotian-liu/LLaVA/blob/main/scripts/v1_5/pretrain.sh). In the pretraining stage, the main focus is image caption. So, it works well with `plain`...

请问panopticfcn的网络支持转onnx吗？

Hi, it is not officially supported. But you can try it. The whole process should be easy.