PTQ-Level Efficiency, QAT-Level Accuracy | 训练后量化的效率,量化感知训练的精度
fix vlm hf config loading issue (#624)
Weiwei514939043a88
diff --git a/auto_round/utils.py b/auto_round/utils.pyindex 48ef3d987..326529201 100644--- a/auto_round/utils.py+++ b/auto_round/utils.py@@ -1334,13 +1334,26 @@ def mllm_load_model(import jsonimport transformersfrom transformers import AutoProcessor, AutoTokenizer, AutoModelForCausalLM- from huggingface_hub import HfApi, HfFileSystem+ from huggingface_hub import HfApi, hf_hub_download, HfFileSystemif os.path.isdir(pretrained_model_name_or_path):config = json.load(open(os.path.join(pretrained_model_name_or_path, "config.json")))else:- hf_file = HfFileSystem()- config = json.load(hf_file.open(pretrained_model_name_or_path + "/config.json"))+ from huggingface_hub import hf_hub_download, list_repo_files+ file_list = list_repo_files(pretrained_model_name_or_path)+ if "config.json" in file_list:+ # Load plain JSON+ config_path = hf_hub_download(pretrained_model_name_or_path, "config.json")+ with open(config_path, "r", encoding="utf-8") as f:+ config = json.load(f)+ elif "config.json.gz" in file_list:+ # Load gzipped JSON+ import gzip+ config_path = hf_hub_download(pretrained_model_name_or_path, "config.json.gz")+ with gzip.open(config_path, "rt", encoding="utf-8") as f:+ config = json.load(f)+ else:+ raise FileNotFoundError(f"No config.json or config.json.gz found for {pretrained_model_name_or_path}")if "model_type" in config:model_type = config["model_type"]@@ -1830,3 +1843,4 @@ def get_gguf_qtype_by_layer_config(layer_config):if bits == 8 and sym and group_size == 32:return gguf.GGMLQuantizationType.Q8_0raise ValueError(f"Unknown layer config")+