prepare_donor.py

# prepare_donor.py
import torch
import os
import argparse
from tqdm import tqdm
from transformers import AutoConfig, AutoModelForCausalLM, AutoTokenizer

def main(source_model_id, output_path):
    """
    Loads a Qwen2.5 model, removes all '.bias' tensors, adds placeholder
    'q_norm.weight' and 'k_norm.weight' tensors, and saves the result.
    This creates an architecturally compatible donor for a Qwen3 merge.
    """
    print(f"Loading source donor model: {source_model_id}")
    # Load on CPU to save VRAM
    model = AutoModelForCausalLM.from_pretrained(
        source_model_id, 
        torch_dtype=torch.bfloat16,
        device_map="cpu",
        trust_remote_code=True
    )
    tokenizer = AutoTokenizer.from_pretrained(source_model_id, trust_remote_code=True)
    config = model.config
    
    source_state_dict = model.state_dict()
    new_state_dict = {}
    
    # --- Part 1: Remove '.bias' tensors ---
    print("Removing all '.bias' tensors...")
    for name, tensor in tqdm(source_state_dict.items(), desc="Filtering Tensors"):
        if not name.endswith(".bias"):
            new_state_dict[name] = tensor
    
    # --- Part 2: Add placeholder 'q_norm' and 'k_norm' tensors ---
    print("Adding placeholder 'q_norm' and 'k_norm' tensors...")
    # These norms are 1D vectors of size `head_dim` (128)
    # A value of 1.0 is a standard, neutral initialization for a norm weight.
    norm_dim = config.hidden_size // config.num_attention_heads # Should be 128 for this model
    placeholder_norm = torch.ones(norm_dim, dtype=torch.bfloat16)

    for i in tqdm(range(config.num_hidden_layers), desc="Adding Norm Tensors"):
        q_norm_name = f"model.layers.{i}.self_attn.q_norm.weight"
        k_norm_name = f"model.layers.{i}.self_attn.k_norm.weight"
        new_state_dict[q_norm_name] = placeholder_norm.clone()
        new_state_dict[k_norm_name] = placeholder_norm.clone()

    # The original model is a fine container, we just need to load the modified state dict.
    # strict=False is crucial because we have removed and added keys.
    print("Loading the new state dict back into the model shell...")
    model.load_state_dict(new_state_dict, strict=False, assign=True)

    print(f"Saving the architecturally aligned model to: {output_path}")
    os.makedirs(output_path, exist_ok=True)
    model.save_pretrained(output_path)
    tokenizer.save_pretrained(output_path)

    print("\nDonor preparation complete!")
    print(f"The aligned donor is ready at '{output_path}'.")

if __name__ == "__main__":
    parser = argparse.ArgumentParser(description="Prepare a Qwen2.5 donor model for merging with Qwen3.")
    parser.add_argument("--source_model", type=str, default="Qwen/Qwen2.5-72B-Instruct", help="The Hugging Face model ID of the source model.")
    parser.add_argument("--output_path", type=str, required=True, help="The local directory path to save the prepared donor model.")
    args = parser.parse_args()
    
    # Example: python prepare_donor.py --output_path ./Qwen2.5-72B-Instruct-Aligned
    main(args.source_model, args.output_path)