PTQ-Level Efficiency, QAT-Level Accuracy | 训练后量化的效率,量化感知训练的精度
Refine funcs (#446)
WeiweiZhang13a1432842d91
diff --git a/auto_round/calib_dataset.py b/auto_round/calib_dataset.pyindex ad4f3f0dc..585ae3d5d 100644--- a/auto_round/calib_dataset.py+++ b/auto_round/calib_dataset.py@@ -150,7 +150,7 @@ def get_pile_val_dataset(tokenizer, seqlen, dataset_name="swift/pile-val-backup"@register_dataset("BAAI/CCI3-HQ")-def get_CCI3_HQ_dataset(tokenizer, seqlen, dataset_name="BAAI/CCI3-HQ", split=None, seed=42, apply_chat_template=False):+def get_cci3_hq_dataset(tokenizer, seqlen, dataset_name="BAAI/CCI3-HQ", split=None, seed=42, apply_chat_template=False):"""Returns a dataloader for the specified dataset and split.Args:@@ -235,18 +235,19 @@ def get_new_chinese_title_dataset(seed=42,apply_chat_template=False):- """Returns a dataloader for the specified dataset and split.+ """+ Returns a tokenized dataset for the specified parameters.Args:- tokenizer: The tokenizer to be used for tokenization.- seqlen: The maximum sequence length.- data_name: The name of the dataset.- split: The data split to be used (e.g., "train", "test").- seed: The random seed for shuffling the dataset.- apply_chat_template: Whether to apply chat template in tokenization.+ tokenizer: The tokenizer to use.+ seqlen: Maximum sequence length.+ dataset_name: Name of the dataset to load.+ split: Which split of the dataset to use.+ seed: Random seed for shuffling.+ apply_template: Whether to apply a template to the data.Returns:- A dataloader for the specified dataset and split, using the provided tokenizer and sequence length.+ A tokenized and shuffled dataset."""def get_tokenizer_function(tokenizer, seqlen, apply_chat_template=apply_chat_template):@@ -639,3 +640,4 @@ def get_dataloader(calib_dataloader = DataLoader(dataset_final, batch_size=bs, shuffle=False, collate_fn=collate_batch)return calib_dataloader+diff --git a/auto_round/export/export_to_autoround/export.py b/auto_round/export/export_to_autoround/export.pyindex 23843e0d8..5deb60c89 100644--- a/auto_round/export/export_to_autoround/export.py+++ b/auto_round/export/export_to_autoround/export.py@@ -50,7 +50,7 @@ def check_neq_config(config, data_type, bits, group_size, sym):return [key for key, expected_value in expected_config.items() if config.get(key) != expected_value]-def dynamic_import_quantLinear_for_packing(backend, bits, group_size, sym):+def dynamic_import_quant_linear_for_packing(backend, bits, group_size, sym):"""Dynamically imports and returns the appropriate QuantLinear class based on the specified backend and parameters.@@ -97,7 +97,7 @@ def pack_layer(name, model, layer_config, backend, pbar):layer = get_module(model, name)device = layer.weight.device- QuantLinear = dynamic_import_quantLinear_for_packing(backend, bits, group_size, sym)+ QuantLinear = dynamic_import_quant_linear_for_packing(backend, bits, group_size, sym)if isinstance(layer, nn.Linear):in_features = layer.in_features@@ -286,3 +286,4 @@ def save(model: nn.Module, save_dir: str, max_shard_size: str = "5GB", safe_seriif hasattr(model, "config") and hasattr(model.config, "quantization_config"):with open(os.path.join(save_dir, config_file), "w", encoding="utf-8") as f:json.dump(model.config.quantization_config, f, indent=2)+diff --git a/auto_round/low_cpu_mem/utils.py b/auto_round/low_cpu_mem/utils.pyindex 3cb481a3b..f4165b4eb 100644--- a/auto_round/low_cpu_mem/utils.py+++ b/auto_round/low_cpu_mem/utils.py@@ -78,7 +78,7 @@ def get_named_children(model, pre=[]):return module_list-def dowload_hf_model(repo_id, cache_dir=None, repo_type=None, revision=None):+def download_hf_model(repo_id, cache_dir=None, repo_type=None, revision=None):"""Download hugging face model from hf hub."""from huggingface_hub.constants import DEFAULT_REVISION, HUGGINGFACE_HUB_CACHEfrom huggingface_hub.file_download import REGEX_COMMIT_HASH, repo_folder_name@@ -116,7 +116,7 @@ def load_empty_model(pretrained_model_name_or_path, cls=AutoModelForCausalLM, saif is_local: # pragma: no coverpath = pretrained_model_name_or_pathelse:- path = dowload_hf_model(pretrained_model_name_or_path)+ path = download_hf_model(pretrained_model_name_or_path)torch_dtype = kwargs.pop("torch_dtype", None)if cls.__base__ == _BaseAutoModelClass:config = AutoConfig.from_pretrained(path, **kwargs)@@ -258,7 +258,7 @@ def _get_path(pretrained_model_name_or_path):if is_local: # pragma: no coverpath = pretrained_model_name_or_pathelse:- path = dowload_hf_model(pretrained_model_name_or_path)+ path = download_hf_model(pretrained_model_name_or_path)return path@@ -471,3 +471,4 @@ def layer_wise_load(path):d = pickle.loads(d)state_dict.update(d)return state_dict+diff --git a/auto_round/mllm/processor.py b/auto_round/mllm/processor.pyindex 302d01200..2ec127622 100644--- a/auto_round/mllm/processor.py+++ b/auto_round/mllm/processor.py@@ -20,7 +20,7 @@ from .utils import fetch_imagePROCESSORS = {}-def regist_processor(name):+def register_processor(name):def register(processor):PROCESSORS[name] = processorreturn processor@@ -28,7 +28,7 @@ def regist_processor(name):return register-@regist_processor("basic")+@register_processor("basic")class BasicProcessor:def __init__(self):pass@@ -111,7 +111,7 @@ class BasicProcessor:return ret-@regist_processor("qwen2_vl")+@register_processor("qwen2_vl")class Qwen2VLProcessor(BasicProcessor):@staticmethoddef squeeze_result(ret):@@ -122,7 +122,7 @@ class Qwen2VLProcessor(BasicProcessor):return ret-@regist_processor("cogvlm2")+@register_processor("cogvlm2")class CogVLM2Processor(BasicProcessor):def get_input(self, text, images, truncation=False,@@ -205,7 +205,7 @@ from ..utils import LazyImportllava_train = LazyImport("llava.train.train")-@regist_processor("llava")+@register_processor("llava")class LlavaProcessor(BasicProcessor):def post_init(self, model, tokenizer, image_processor=None, **kwargs):self.model = model@@ -245,3 +245,4 @@ class LlavaProcessor(BasicProcessor):def data_collator(self, batch):return self.collator_func(batch)+