Add cli merge lora (#168)

Author: Jintao-Huang

README.md

@@ -245,6 +245,10 @@ CUDA_VISIBLE_DEVICES=0 swift infer --model_id_or_path qwen/Qwen-7B-Chat --datase
 
 # Fine-tuned Model
 CUDA_VISIBLE_DEVICES=0 swift infer --ckpt_dir 'xxx/vx_xxx/checkpoint-xxx'
+
+# Merge LoRA incremental weights and perform inference
+swift merge-lora --ckpt_dir 'xxx/vx_xxx/checkpoint-xxx'
+CUDA_VISIBLE_DEVICES=0 swift infer --ckpt_dir 'xxx/vx_xxx/checkpoint-xxx-merged'
 ```
 
 **Web-UI**:
@@ -254,6 +258,10 @@ CUDA_VISIBLE_DEVICES=0 swift web-ui --model_id_or_path qwen/Qwen-7B-Chat
 
 # Fine-tuned Model
 CUDA_VISIBLE_DEVICES=0 swift web-ui --ckpt_dir 'xxx/vx_xxx/checkpoint-xxx'
+
+# Merge LoRA incremental weights and use web UI
+swift merge-lora --ckpt_dir 'xxx/vx_xxx/checkpoint-xxx'
+CUDA_VISIBLE_DEVICES=0 swift web-ui --ckpt_dir 'xxx/vx_xxx/checkpoint-xxx-merged'
 ```
 
 

README_CN.md

@@ -242,6 +242,10 @@ CUDA_VISIBLE_DEVICES=0 swift infer --model_id_or_path qwen/Qwen-7B-Chat --datase
 
 # 微调后的模型
 CUDA_VISIBLE_DEVICES=0 swift infer --ckpt_dir 'xxx/vx_xxx/checkpoint-xxx'
+
+# merge LoRA增量权重并推理
+swift merge-lora --ckpt_dir 'xxx/vx_xxx/checkpoint-xxx'
+CUDA_VISIBLE_DEVICES=0 swift infer --ckpt_dir 'xxx/vx_xxx/checkpoint-xxx-merged'
 ```
 
 **Web-UI**:
@@ -251,6 +255,10 @@ CUDA_VISIBLE_DEVICES=0 swift web-ui --model_id_or_path qwen/Qwen-7B-Chat
 
 # 微调后的模型
 CUDA_VISIBLE_DEVICES=0 swift web-ui --ckpt_dir 'xxx/vx_xxx/checkpoint-xxx'
+
+# merge LoRA增量权重并使用web-ui
+swift merge-lora --ckpt_dir 'xxx/vx_xxx/checkpoint-xxx'
+CUDA_VISIBLE_DEVICES=0 swift web-ui --ckpt_dir 'xxx/vx_xxx/checkpoint-xxx-merged'
 ```
 
 

examples/pytorch/llm/README.md

@@ -208,6 +208,10 @@ CUDA_VISIBLE_DEVICES=0 swift infer --model_id_or_path qwen/Qwen-7B-Chat --datase
 
 # Fine-tuned Model
 CUDA_VISIBLE_DEVICES=0 swift infer --ckpt_dir 'xxx/vx_xxx/checkpoint-xxx'
+
+# Merge LoRA incremental weights and perform inference
+swift merge-lora --ckpt_dir 'xxx/vx_xxx/checkpoint-xxx'
+CUDA_VISIBLE_DEVICES=0 swift infer --ckpt_dir 'xxx/vx_xxx/checkpoint-xxx-merged'
 ```
 
 **Web-UI**:
@@ -217,6 +221,10 @@ CUDA_VISIBLE_DEVICES=0 swift web-ui --model_id_or_path qwen/Qwen-7B-Chat
 
 # Fine-tuned Model
 CUDA_VISIBLE_DEVICES=0 swift web-ui --ckpt_dir 'xxx/vx_xxx/checkpoint-xxx'
+
+# Merge LoRA incremental weights and use web UI
+swift merge-lora --ckpt_dir 'xxx/vx_xxx/checkpoint-xxx'
+CUDA_VISIBLE_DEVICES=0 swift web-ui --ckpt_dir 'xxx/vx_xxx/checkpoint-xxx-merged'
 ```
 
 

examples/pytorch/llm/README_CN.md

@@ -207,6 +207,10 @@ CUDA_VISIBLE_DEVICES=0 swift infer --model_id_or_path qwen/Qwen-7B-Chat --datase
 
 # 微调后的模型
 CUDA_VISIBLE_DEVICES=0 swift infer --ckpt_dir 'xxx/vx_xxx/checkpoint-xxx'
+
+# merge LoRA增量权重并推理
+swift merge-lora --ckpt_dir 'xxx/vx_xxx/checkpoint-xxx'
+CUDA_VISIBLE_DEVICES=0 swift infer --ckpt_dir 'xxx/vx_xxx/checkpoint-xxx-merged'
 ```
 
 **Web-UI**:
@@ -216,6 +220,10 @@ CUDA_VISIBLE_DEVICES=0 swift web-ui --model_id_or_path qwen/Qwen-7B-Chat
 
 # 微调后的模型
 CUDA_VISIBLE_DEVICES=0 swift web-ui --ckpt_dir 'xxx/vx_xxx/checkpoint-xxx'
+
+# merge LoRA增量权重并使用web-ui
+swift merge-lora --ckpt_dir 'xxx/vx_xxx/checkpoint-xxx'
+CUDA_VISIBLE_DEVICES=0 swift web-ui --ckpt_dir 'xxx/vx_xxx/checkpoint-xxx-merged'
 ```
 
 

swift/cli/main.py

@@ -3,13 +3,15 @@
 import sys
 from typing import Dict, List, Optional
 
-from swift.cli import infer, sft, web_ui
+from swift.cli import infer, merge_lora, sft, web_ui
 
 ROUTE_MAPPING: Dict[str, str] = {
     'sft': sft.__file__,
     'infer': infer.__file__,
-    'web-ui': web_ui.__file__
+    'web-ui': web_ui.__file__,
+    'merge-lora': merge_lora.__file__
 }
+
 ROUTE_MAPPING.update(
     {k.replace('-', '_'): v
      for k, v in ROUTE_MAPPING.items()})

swift/cli/merge_lora.py

@@ -0,0 +1,6 @@
+from swift.llm import InferArguments, merge_lora
+from swift.utils import parse_args
+
+if __name__ == '__main__':
+    args, remaining_argv = parse_args(InferArguments)
+    merge_lora(args, replace_if_exists=True)

swift/llm/infer.py

@@ -19,6 +19,7 @@
 
 
 def merge_lora(args: InferArguments, replace_if_exists=False) -> None:
+    logger.info(f'replace_if_exists: {replace_if_exists}')
     assert args.ckpt_dir is not None
     assert args.sft_type == 'lora'
     assert 'int4' not in args.model_type, 'int4 model is not supported'
@@ -65,10 +66,21 @@ def merge_lora(args: InferArguments, replace_if_exists=False) -> None:
             res.pop('adapter_cfg', None)
             with open(new_configuration_path, 'w') as f:
                 json.dump(res, f, ensure_ascii=False, indent=4)
-        logger.info('Successfully merged LoRA.')
+        # sft_args
+        sft_args_fname = 'sft_args.json'
+        old_sft_args_path = os.path.join(old_ckpt_dir, sft_args_fname)
+        new_sft_args_path = os.path.join(args.ckpt_dir, sft_args_fname)
+        if os.path.exists(old_sft_args_path):
+            with open(old_sft_args_path, 'r') as f:
+                res = json.load(f)
+            res['sft_type'] = 'full'
+            with open(new_sft_args_path, 'w') as f:
+                json.dump(res, f, ensure_ascii=False, indent=2)
+        logger.info(f'Successfully merged LoRA and saved in {args.ckpt_dir}.')
     else:
-        logger.info('The weight directory for the merged LoRA already exists, '
-                    'skipping the saving process.')
+        logger.info(
+            f'The weight directory for the merged LoRA already exists in {args.ckpt_dir}, '
+            'skipping the saving process.')
 
 
 def prepare_model_template(