PTQ-Level Efficiency, QAT-Level Accuracy | 训练后量化的效率,量化感知训练的精度
align auto_quantizer with main branch in Transformers (#437)
WeiweiZhang17d27f670284b
diff --git a/auto_round/auto_quantizer.py b/auto_round/auto_quantizer.pyindex 4856c253f..305f64ab3 100644--- a/auto_round/auto_quantizer.py+++ b/auto_round/auto_quantizer.py@@ -41,7 +41,6 @@ from transformers.pytorch_utils import Conv1Dfrom transformers.quantizers import AutoQuantizationConfig, HfQuantizerfrom transformers.quantizers.auto import AUTO_QUANTIZER_MAPPINGfrom transformers.utils.quantization_config import AwqConfig, GPTQConfig, QuantizationConfigMixin, QuantizationMethod-from auto_round.utils import (get_module, set_module, is_hpu_supported, get_block_names,get_multimodal_block_names, find_matching_blocks)@@ -191,6 +190,30 @@ class AutoHfQuantizer:warnings.warn(warning_msg)return quantization_config++ @staticmethod+ def supports_quant_method(quantization_config_dict):+ from transformers.quantizers.auto import AUTO_QUANTIZATION_CONFIG_MAPPING+ AUTO_QUANTIZATION_CONFIG_MAPPING['intel/auto-round'] = AutoRoundConfig+ AUTO_QUANTIZATION_CONFIG_MAPPING['intel/auto_round'] = AutoRoundConfig+ quant_method = quantization_config_dict.get("quant_method", None)+ if quantization_config_dict.get("load_in_8bit", False) or quantization_config_dict.get("load_in_4bit", False):+ suffix = "_4bit" if quantization_config_dict.get("load_in_4bit", False) else "_8bit"+ quant_method = QuantizationMethod.BITS_AND_BYTES + suffix+ elif quant_method is None:+ raise ValueError(+ "The model's quantization config from the arguments has no `quant_method` attribute."\+ "Make sure that the model has been correctly quantized"+ )++ if quant_method not in AUTO_QUANTIZATION_CONFIG_MAPPING.keys():+ logger.warning(+ f"Unknown quantization type, got {quant_method} - supported types are:"+ f" {list(AUTO_QUANTIZER_MAPPING.keys())}. Hence, we will skip the quantization. "+ "To remove the warning, you can delete the quantization_config attribute in config.json"+ )+ return False+ return Trueclass AutoRoundQuantizationMethod(str, Enum):@@ -758,3 +781,4 @@ if version.parse(transformers.__version__) < version.parse("4.38.0"):transformers.quantizers.auto.AutoHfQuantizer = AutoHfQuantizertransformers.modeling_utils.AutoHfQuantizer = AutoHfQuantizer+diff --git a/auto_round/script/llm.py b/auto_round/script/llm.pyindex e44cd9f0d..445599166 100644--- a/auto_round/script/llm.py+++ b/auto_round/script/llm.py@@ -463,7 +463,7 @@ def tune(args):if args.quant_lm_head:layer_config[lm_head_layer_name] = {"bits": args.bits}for format in formats:- if "auto_round" not in format:+ if "auto_round" not in format and "fake" not in format:auto_round_formats = [s for s in supported_formats if s.startswith("auto_round")]raise ValueError(f"{format} is not supported for lm-head quantization, please change to {auto_round_formats}")diff --git a/auto_round/script/mllm.py b/auto_round/script/mllm.pyindex 884ac9201..f45eab288 100644--- a/auto_round/script/mllm.py+++ b/auto_round/script/mllm.py@@ -418,7 +418,7 @@ def tune(args):if args.quant_lm_head:layer_config[lm_head_layer_name] = {"bits": args.bits}for format in formats:- if "auto_round" not in format:+ if "auto_round" not in format and "fake" not in format:auto_round_formats = [s for s in supported_formats if s.startswith("auto_round")]raise ValueError(f"{format} is not supported for lm-head quantization, please change to {auto_round_formats}")@@ -587,3 +587,4 @@ def lmms_eval(args):apply_chat_template=False,)return results+