PTQ-Level Efficiency, QAT-Level Accuracy | 训练后量化的效率,量化感知训练的精度
better warning for not support mllm while quant_non_text_module (#1411)
Heng Guo96cfbc9566c0
diff --git a/auto_round/compressors/mllm/compressor.py b/auto_round/compressors/mllm/compressor.pyindex b5cfd8598..905277bed 100644--- a/auto_round/compressors/mllm/compressor.py+++ b/auto_round/compressors/mllm/compressor.py@@ -183,12 +183,16 @@ class MLLMCompressor(BaseCompressor):self.model = modelquant_nontext_module = self._check_quant_nontext(layer_config, quant_nontext_module)if quant_nontext_module and iters > 0:+ logger.warning(+ "quant_nontext_module is experimental feature and not work well for all models,"+ "please remove --quant_nontext_module and retry if you meet some errors."+ )import importlib.utilmissing_libs = []- for require_lib in ["pillow", "torchvision"]:+ for require_lib in ["PIL", "torchvision"]:if importlib.util.find_spec(require_lib) is None:- missing_libs.append(require_lib)+ missing_libs.append(require_lib.replace("PIL", "pillow"))if len(missing_libs) > 0:logger.error(f"{', '.join(missing_libs)} are required for quantizing non-text modules,"@@ -213,21 +217,22 @@ class MLLMCompressor(BaseCompressor):# TODO: Remove after fixing https://github.com/huggingface/transformers/issues/43005model.config.model_type = model.config.to_dict()["model_type"]+ # try default template if model_type not in preset templatesif template is None and model.config.model_type not in TEMPLATES:- self.template = None+ self.template = "default"else:self.template = template if template is not None else model.config.model_type- if not isinstance(dataset, torch.utils.data.DataLoader):- self.template = get_template(- self.template,- model=model,- tokenizer=tokenizer,- processor=processor,- image_processor=image_processor,- use_rtn=iters == 0,- quiet=not self.quant_nontext_module,- )- dataset = self.template.default_dataset if dataset is None else dataset+ if not isinstance(dataset, torch.utils.data.DataLoader):+ self.template = get_template(+ self.template,+ model=model,+ tokenizer=tokenizer,+ processor=processor,+ image_processor=image_processor,+ use_rtn=iters == 0,+ quiet=not self.quant_nontext_module,+ )+ dataset = self.template.default_dataset if dataset is None else datasetelse:self.template = None