Request update of llama3.2 1b and 3b experimental results

[shifeiwen]

I recently tried to save the model after R @ Weight obtained by SpinQuant, and found that the quantization effect was not improved much after hadamard_utils.get_hadK did not match the appropriate size k=1. Is there any experimental record about small models, such as llama3.2 1B or 3B?
thanks

[lonleyodd]

I recently tried to save the model after R @ Weight obtained by SpinQuant, and found that the quantization effect was not improved much after hadamard_utils.get_hadK did not match the appropriate size k=1. Is there any experimental record about small models, such as llama3.2 1B or 3B? thanks

same error, apply rotation on llama3.2 1b is not correct，have you solved it?

[lonleyodd]

@shifeiwen hi, i found a note in the code:

I think this may the key cause rotation useless in llama3.2， or you can refer llama3.2 demo file post by official :https://colab.research.google.com/gist/zxdmike/abbb2c9b0d1fd1f4ed8cdae8c02180f4#scrollTo=4yQuaPzv3yle