Question about the shape of lm_head.weight

[Jay-zzcoder]

Question

I found that if I don't pass the BitsAndBytesConfig args in LlavaLlamaForCausalLM.from_pretrained() like this:
model = LlavaLlamaForCausalLM.from_pretrained( tokenizer_path, torch_dtype=torch.bfloat16 , #**bnb_model_from_pretrained_args )
the shape of lm_head.weight is torch.Size([32000, 5120])

But if I pass BitsAndBytesConfig args in LlavaLlamaForCausalLM.from_pretrained() like:
model = LlavaLlamaForCausalLM.from_pretrained( tokenizer_path, torch_dtype=torch.bfloat16 , **bnb_model_from_pretrained_args )
the shape of lm_head.weight will change and become torch.Size([81920000, 1])

Why the shape of lm_head.weight changes?