Questions about finetunning internlm_xcomposer2_7b_chat.

[piqiuni]

Hello,

1. we want to finetune the model with our own dataset(pics + txt), and with 2 RTX4090, the following setting leads to Error as follows, does that mean the model doesn't support it: setting os.environ['CUDA_VISIBLE_DEVICES'] = '0, 1' os.environ['NPROC_PER_NODE'] = '2' Error: 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 tensors in method wrapper_CUDA_cat)

2. Then we tried to run with one card, and it led to OOM, also we tried the deepspeed but it met error blow, which seems caused by the version of transformers. AttributeError: 'DummyOptim' object has no attribute 'step'

3. And we also tried the qlora but seems it doesn't support? TypeError: Linear4bit.forward() takes 2 positional arguments but 3 were given

We are wondering if the OOM caused by our own dataset(with demo it works), and what can we do to avoid the OOM ?