| import os | |
| import torch | |
| from qmllm.methods.smoothquant.quantize.smooth import smooth_lm, smooth_vit | |
| from qmllm.methods.smoothquant.quantize.quantizer import quantize_model, pseudo_quantize_model_weight_act | |
| from qmllm.methods.smoothquant.quantize.gen_act_scales import get_act_scales | |
| def smoothquant_entry(model, prompt_inputs, prompt_kwargs, run_sq_process: bool, pseudo_quant: bool, scale_path: str=None, w_bit: int=4, a_bit: int=8, alpha: float=0.5): | |
| ''' | |
| model: here the model is the LLM, you have to extract the LLM first! | |
| prompt_tokens: the prompt tokens | |
| prompt_mask: the prompt mask, mask the answer language tokens | |
| ''' | |
| assert scale_path is not None | |
| if run_sq_process: | |
| act_scales = get_act_scales(model, prompt_inputs, prompt_kwargs) | |
| dirpath = os.path.dirname(scale_path) | |
| os.makedirs(dirpath, exist_ok=True) | |
| torch.save(act_scales, scale_path) | |
| print("SmoothQuant results saved at", scale_path) | |
| else: | |
| act_scales = torch.load(scale_path) | |
| if pseudo_quant: | |
| smooth_lm(model.model, act_scales, alpha) | |
| pseudo_quantize_model_weight_act(model.model, w_bit=w_bit, a_bit=a_bit) | |
| return model | |