Issue Reproducing Table 2 Results: NaN Occurrence for Low Param Ratio on Wiki Test

[hztohhhhh]

Hello,
Thank you for your inspiring work and for sharing the codebase. I have been trying to reproduce the results presented in Table 2 of your paper, but I encountered some issues when the param_ratio is set below 0.75. Specifically, on the WikiText2 test set, the evaluation outputs NaN. Below are the parameter configurations I used and the corresponding results:

Namespace(model_id='/hf_models/Llama-2-7b-hf', ppl_target=-1, param_ratio_target=0.75, act_aware=True, alpha=0.5, n_calib_samples=32, calib_dataset='wikitext2', scaling_method='abs_mean', sensitivity_metric='ppl', use_cache=True, weight_quant='none', eval_mmlu=False, eval_ppl='wikitext2', eval_tasks='', sigma_fuse='UV', seed=233, compress_kv_cache=False, kv_cache_ratio_target=-1, rank_align=1, raw_model=False, use_bos=False)  
{'wikitext2': 44.652503967285156}

Namespace(model_id='/hf_models/Llama-2-7b-hf', ppl_target=-1, param_ratio_target=0.7, act_aware=True, alpha=0.5, n_calib_samples=32, calib_dataset='wikitext2', scaling_method='abs_mean', sensitivity_metric='ppl', use_cache=True, weight_quant='none', eval_mmlu=False, eval_ppl='wikitext2', eval_tasks='', sigma_fuse='UV', seed=233, compress_kv_cache=False, kv_cache_ratio_target=-1, rank_align=1, raw_model=False, use_bos=False)  
{'wikitext2': nan}

Namespace(model_id='/hf_models/Llama-2-7b-hf', ppl_target=-1, param_ratio_target=0.65, act_aware=True, alpha=0.5, n_calib_samples=32, calib_dataset='wikitext2', scaling_method='abs_mean', sensitivity_metric='ppl', use_cache=True, weight_quant='none', eval_mmlu=False, eval_ppl='wikitext2', eval_tasks='', sigma_fuse='UV', seed=233, compress_kv_cache=False, kv_cache_ratio_target=-1, rank_align=1, raw_model=False, use_bos=False)  
{'wikitext2': nan}

It seems that the issue occurs as the param_ratio goes below 0.75. Could you please provide the correct parameter settings for reproducing the results in Table 2? Alternatively, do you have any suggestions to avoid this NaN issue?

Your guidance would be greatly appreciated. Thank you for your time and support!

Best regards,
Will