update

examples/tutorials/by_deepspeed/step_2_train_model.py

@@ -40,6 +40,11 @@ def get_args():
         default=(project_path / "hub_datasets").as_posix(),
         type=str
     ),
+    parser.add_argument(
+        "--model_cache_dir",
+        default=(project_path / "hub_models").as_posix(),
+        type=str
+    ),
     parser.add_argument("--dataset_streaming", default=None, type=str),
     parser.add_argument("--valid_dataset_size", default=100, type=str),
     parser.add_argument("--shuffle_buffer_size", default=5000, type=str),
@@ -60,11 +65,13 @@ def main():
         pretrained_model_name_or_path=args.model_name,
         quantization_config=None,
         # device_map="auto",
-        trust_remote_code=True
+        trust_remote_code=True,
+        cache_dir=args.model_cache_dir,
     )
     tokenizer = AutoTokenizer.from_pretrained(
         pretrained_model_name_or_path=args.model_name,
-        trust_remote_code=True
+        trust_remote_code=True,
+        cache_dir=args.model_cache_dir,
     )
     print(model)
 

examples/tutorials/lora_transformers/step_2_train_model.py

@@ -66,11 +66,13 @@ def main():
         pretrained_model_name_or_path=args.model_name,
         quantization_config=bnb_config,
         device_map="auto",
-        trust_remote_code=True
+        trust_remote_code=True,
+        cache_dir=args.model_cache_dir,
     )
     tokenizer = AutoTokenizer.from_pretrained(
         pretrained_model_name_or_path=args.model_name,
-        trust_remote_code=True
+        trust_remote_code=True,
+        cache_dir=args.model_cache_dir,
     )
     peft_config = LoraConfig(
         r=32,  # LoRA秩

examples/tutorials/lora_unsloth/step_2_train_model.py

@@ -62,7 +62,8 @@ def main():
         dtype=None,
         load_in_4bit=True,
         load_in_8bit=False,
-        full_finetuning=False
+        full_finetuning=False,
+        cache_dir=args.model_cache_dir,
     )
 
     model = FastLanguageModel.get_peft_model(

examples/tutorials/lora_unsloth/step_3_inter_model.py

@@ -55,6 +55,7 @@ def main():
         device_map="auto",
         dtype=None,  # 自动选择最优精度
         load_in_4bit=True,  # 4bit量化节省70%显存
+        cache_dir=args.model_cache_dir,
     )
 
     # 2、注入lora适配器