output += torch.matmul(subA, state.subB) RuntimeError: mat1 and mat2 shapes cannot be multiplied (1x2 and 4x5120)

Very often when I do a sequence of questions that requires compare or calculate a number using math operations I'm receiving the error below:

File "/root/miniconda3/envs/llama2/lib/python3.10/site-packages/bitsandbytes/autograd/_functions.py", line 421, in forward output += torch.matmul(subA, state.subB) RuntimeError: mat1 and mat2 shapes cannot be multiplied (1x2 and 4x5120)

I have reinstalled h2ogpt but same error, and when this happen some crazy answers come up until I restart h2ogpt.

I'm using a fine tuned model based on llama2 (frivasplata/ALE-GPT-llama2-13B-2388-12t-advanced), the error occurs with 7B or 13B variations.