Add PEFT checkpoint merger script

merge_peft.py

@@ -0,0 +1,71 @@
+"""
+Script for merging PEFT LoRA weights with the base model. Uses code from https://github.com/eugenepentland/landmark-attention-qlora/blob/main/llama/merge_peft.py
+Usage: python merge_peft.py [-h] [--base_model_name_or_path BASE_MODEL_NAME_OR_PATH] [--peft_model_path PEFT_MODEL_PATH] [--output_dir OUTPUT_DIR] [--device DEVICE]
+                               [--push_to_hub]
+
+"""
+import torch
+import os
+import logging
+import argparse
+from tqdm import tqdm
+from transformers import AutoModelForCausalLM, AutoTokenizer
+from peft import PeftModel
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--base_model_name_or_path", type=str)
+    parser.add_argument("--peft_model_path", type=str)
+    parser.add_argument("--output_dir", type=str)
+    parser.add_argument("--device", type=str, default="auto")
+    parser.add_argument("--push_to_hub", action="store_true")
+
+    return parser.parse_args()
+
+
+def main():
+    # Configure logging
+    logging.basicConfig(level=logging.INFO)
+    logger = logging.getLogger(__name__)
+
+    try:
+        args = get_args()
+
+        if args.device == 'auto':
+            device_arg = {'device_map': 'auto'}
+        else:
+            device_arg = {'device_map': {"": args.device}}
+
+        logger.info(f"Loading base model: {args.base_model_name_or_path}")
+        with tqdm(total=1, desc="Loading base model") as pbar:
+            base_model = AutoModelForCausalLM.from_pretrained(
+                args.base_model_name_or_path,
+                return_dict=True,
+                torch_dtype=torch.float16,
+                **device_arg
+            )
+            pbar.update(1)
+
+        logger.info(f"Loading PEFT: {args.peft_model_path}")
+        with tqdm(total=1, desc="Loading PEFT model") as pbar:
+            model = PeftModel.from_pretrained(base_model, args.peft_model_path)
+            pbar.update(1)
+
+        logger.info("Running merge_and_unload")
+        with tqdm(total=1, desc="Merge and Unload") as pbar:
+            model = model.merge_and_unload()
+            pbar.update(1)
+
+        tokenizer = AutoTokenizer.from_pretrained(args.base_model_name_or_path)
+
+        model.save_pretrained(f"{args.output_dir}")
+        tokenizer.save_pretrained(f"{args.output_dir}")
+        logger.info(f"Model saved to {args.output_dir}")
+
+    except Exception as e:
+        logger.exception("An error occurred:")
+        raise
+
+
+if __name__ == "__main__":
+    main()