Shisa-v2-Mistral-Nemo-12B-Lorablated

import argparse
import os
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

def main():
    parser = argparse.ArgumentParser(description="Merge a LoRA into a large language model.")
    parser.add_argument("--model_name", type=str, required=True, help="Name or path of the base model.")
    parser.add_argument("--lora_name", type=str, required=True, help="Name or path of the LoRA adapter.")
    parser.add_argument("--output_model_name", type=str, required=True, help="Output directory for the merged model.")
    args = parser.parse_args()

    # Load the base model in bfloat16
    print("Loading base model...")
    base_model = AutoModelForCausalLM.from_pretrained(
        args.model_name,
        torch_dtype="bfloat16"
    )

    # Load the LoRA adapter
    print("Loading LoRA adapter...")
    peft_model = PeftModel.from_pretrained(base_model, args.lora_name)

    # Merge and unload the LoRA weights into the base model
    print("Merging LoRA into base model...")
    merged_model = peft_model.merge_and_unload()

    # Save the merged model
    print(f"Saving merged model to {args.output_model_name}...")
    merged_model.save_pretrained(args.output_model_name)

    # Save the tokenizer from the base model
    tokenizer = AutoTokenizer.from_pretrained(args.model_name)
    tokenizer.save_pretrained(args.output_model_name)

    # Create README.md file
    print("Creating README.md...")
    readme_content = f"""# Merged Model

This model is a combination of:

- **Base Model**: `{args.model_name}`
- **LoRA Adapter**: `{args.lora_name}`

The model is saved in `bfloat16` format and is ready for deployment or fine-tuning.
"""

    readme_path = os.path.join(args.output_model_name, "README.md")
    with open(readme_path, "w") as f:
        f.write(readme_content)

    print("✅ Merging complete. Model saved to:", args.output_model_name)

if __name__ == "__main__":
    main()

import argparse
import os
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

def main():
    parser = argparse.ArgumentParser(description="Merge a LoRA into a large language model.")
    parser.add_argument("--model_name", type=str, required=True, help="Name or path of the base model.")
    parser.add_argument("--lora_name", type=str, required=True, help="Name or path of the LoRA adapter.")
    parser.add_argument("--output_model_name", type=str, required=True, help="Output directory for the merged model.")
    args = parser.parse_args()

    # Load the base model in bfloat16
    print("Loading base model...")
    base_model = AutoModelForCausalLM.from_pretrained(
        args.model_name,
        torch_dtype="bfloat16"
    )

    # Load the LoRA adapter
    print("Loading LoRA adapter...")
    peft_model = PeftModel.from_pretrained(base_model, args.lora_name)

    # Merge and unload the LoRA weights into the base model
    print("Merging LoRA into base model...")
    merged_model = peft_model.merge_and_unload()

    # Save the merged model
    print(f"Saving merged model to {args.output_model_name}...")
    merged_model.save_pretrained(args.output_model_name)

    # Save the tokenizer from the base model
    tokenizer = AutoTokenizer.from_pretrained(args.model_name)
    tokenizer.save_pretrained(args.output_model_name)

    # Create README.md file
    print("Creating README.md...")
    readme_content = f"""# Merged Model

This model is a combination of:

- **Base Model**: `{args.model_name}`
- **LoRA Adapter**: `{args.lora_name}`

The model is saved in `bfloat16` format and is ready for deployment or fine-tuning.
"""

    readme_path = os.path.join(args.output_model_name, "README.md")
    with open(readme_path, "w") as f:
        f.write(readme_content)

    print("✅ Merging complete. Model saved to:", args.output_model_name)

if __name__ == "__main__":
    main()

import argparse
import os
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

def main():
    parser = argparse.ArgumentParser(description="Merge a LoRA into a large language model.")
    parser.add_argument("--model_name", type=str, required=True, help="Name or path of the base model.")
    parser.add_argument("--lora_name", type=str, required=True, help="Name or path of the LoRA adapter.")
    parser.add_argument("--output_model_name", type=str, required=True, help="Output directory for the merged model.")
    args = parser.parse_args()

    # Load the base model in bfloat16
    print("Loading base model...")
    base_model = AutoModelForCausalLM.from_pretrained(
        args.model_name,
        torch_dtype="bfloat16"
    )

    # Load the LoRA adapter
    print("Loading LoRA adapter...")
    peft_model = PeftModel.from_pretrained(base_model, args.lora_name)

    # Merge and unload the LoRA weights into the base model
    print("Merging LoRA into base model...")
    merged_model = peft_model.merge_and_unload()

    # Save the merged model
    print(f"Saving merged model to {args.output_model_name}...")
    merged_model.save_pretrained(args.output_model_name)

    # Save the tokenizer from the base model
    tokenizer = AutoTokenizer.from_pretrained(args.model_name)
    tokenizer.save_pretrained(args.output_model_name)

    # Create README.md file
    print("Creating README.md...")
    readme_content = f"""# Merged Model

This model is a combination of:

- **Base Model**: `{args.model_name}`
- **LoRA Adapter**: `{args.lora_name}`

The model is saved in `bfloat16` format and is ready for deployment or fine-tuning.
"""

    readme_path = os.path.join(args.output_model_name, "README.md")
    with open(readme_path, "w") as f:
        f.write(readme_content)

    print("✅ Merging complete. Model saved to:", args.output_model_name)

if __name__ == "__main__":
    main()