Llama2 model parameters count is half

I just loaded the meta-llama/llama2-13b-hf model and did a counted the parameters using the numel() fn. It looks like the 13b model has ~6.7b params. I’m see a similar discrepancy with the meta-llama/llama2-7b-hf model as well.

However the model sizes were fine a week ago.

Did the models change?

1 Like

I noticed the same in 4 bit quantized models (BitsAndBytes). It shows the correct number in the “original” models.