self.model_path change to self.absolute_model_path, Use GlobalConfig to access absolute_model_path (#105)

Co-authored-by: woodchenwu <woodchenwu@tencent.com>

Author: WOODchen7

angelslim/compressor/quant/ptq.py

@@ -39,7 +39,7 @@ def __init__(self, model, slim_config=None):
         self.quant_model = model
         # init ptq config of model
         self.quant_model.init_ptq(slim_config)
-        self.model_path = slim_config.get("model_path")
+        self.absolute_model_path = slim_config["global_config"].absolute_model_path
         self.quant_algo = self.quant_model.quant_config.quant_algo
         self.quant_helpers = self.quant_model.quant_config.quant_helpers
         if (
@@ -213,12 +213,15 @@ def _convert(self):
             ):
                 if sub_layer.weight.device.type == "meta":
                     with open(
-                        os.path.join(self.model_path, "model.safetensors.index.json"),
+                        os.path.join(
+                            self.absolute_model_path, "model.safetensors.index.json"
+                        ),
                         "r",
                     ) as f:
                         model_index = json.load(f)
                     orign_w_file = os.path.join(
-                        self.model_path, model_index["weight_map"][name + ".weight"]
+                        self.absolute_model_path,
+                        model_index["weight_map"][name + ".weight"],
                     )
                     orign_w = load_file(orign_w_file, device="cpu")
                     print_info(f"Load meta weight {name} from file {orign_w_file}")
@@ -228,7 +231,7 @@ def _convert(self):
                     if hasattr(sub_layer, "bias"):
                         if (name + ".bias") in model_index["weight_map"]:
                             orign_b_file = os.path.join(
-                                self.model_path,
+                                self.absolute_model_path,
                                 model_index["weight_map"][name + ".bias"],
                             )
                             orign_b = load_file(orign_b_file, device="cpu")

angelslim/engine.py

@@ -204,7 +204,6 @@ def prepare_compressor(
             slim_config = {
                 "global_config": global_config,
                 "compress_config": compress_config,
-                "model_path": self.model_path,
             }
         self.compress_type = compress_names
         self.only_inference = (

angelslim/utils/config_parser.py

@@ -19,7 +19,7 @@
 
 import yaml
 
-from .utils import get_hf_config
+from .utils import get_hf_config, get_hf_model_path
 
 
 class CompressionMethod(str, Enum):
@@ -62,6 +62,7 @@ class GlobalConfig:
     max_seq_length: int = field(default=2048)
     hidden_size: int = field(default=2048)
     model_arch_type: str = field(default=None)
+    absolute_model_path: str = field(default=None)
     deploy_backend: str = field(default="vllm")
 
     def update(self, model_path: str = None, max_seq_length: int = None):
@@ -78,6 +79,7 @@ def update(self, model_path: str = None, max_seq_length: int = None):
         if model_path:
             self.set_model_hidden_size(model_path)
             self.set_model_arch_type(model_path)
+            self.absolute_model_path = get_hf_model_path(model_path)
         if max_seq_length:
             self.set_max_seq_length(max_seq_length)
 

angelslim/utils/utils.py

@@ -148,6 +148,14 @@ def get_hf_config(model_path) -> dict:
         return json_data
 
 
+def get_hf_model_path(model_path) -> str:
+    "When model_path does not exist, fetch the model.config from cached_file."
+    if os.path.isfile(model_path):
+        return model_path
+    else:
+        return os.path.dirname(cached_file(model_path, "config.json"))
+
+
 def common_prefix(str1, str2):
     return "".join(
         x[0] for x in takewhile(lambda x: x[0] == x[1], zip(str1, str2))