update: support new Llama API + assess OpenGVLab/OmniQuant#113

[Tfloow]

See #113 for issue

Also support new Llama API. Work in Progress, Need some code review to make sure I don't break any other things in Omniquant.

[Tfloow]

Also, I updated the weight_only=False for torch to be able to load properly the testloader/dataloader

[Tfloow]

Benchmark results:

python main.py --model meta-llama/Llama-3.2-1B --epochs 0 --output_dir ./log --eval_ppl --wbits 4 --abits 16 --group_size 128 --lwc 

On Llama-3.2-1B with Wikitext2 epochs 0 lwc

• Previous version: PPL = 11.66
• My proposition: PPL = 11.56

Roughly 1% PPL reduction by removing unnecessary quantization of activation when --abits 16