mergeLr.py

import os
import shutil
import torch
from transformers import AutoModelForCausalLM, AutoConfig, AutoTokenizer


def _load_model_fp32(model_dir: str):
    # transformers versions differ: some warn about torch_dtype, some prefer dtype
    try:
        return AutoModelForCausalLM.from_pretrained(
            model_dir,
            dtype=torch.float32,
            device_map="cpu",
            trust_remote_code=True,
        )
    except TypeError:
        return AutoModelForCausalLM.from_pretrained(
            model_dir,
            torch_dtype=torch.float32,
            device_map="cpu",
            trust_remote_code=True,
        )


def merge_instruction_residual(lr_dir, base_model_dir, output_dir):
    """
    Merge instruction residual into a (possibly vocab-resized) CPT model.

    If vocab was resized after the residual was computed, we add residual only
    for the overlapping token rows and keep extra rows (new tokens) unchanged.
    """

    adapter_file = os.path.join(lr_dir, "adapter_model.bin")
    if not os.path.exists(adapter_file):
        raise FileNotFoundError(f"Adapter checkpoint not found at {adapter_file}")

    print("Loading residual adapter...")
    residual_state_dict = torch.load(adapter_file, map_location="cpu")

    print(f"\nMerging residual into base model: {base_model_dir}")
    base_model = _load_model_fp32(base_model_dir)
    base_state_dict = base_model.state_dict()

    merged_state_dict = {}
    mismatched = []

    for key, base_tensor in base_state_dict.items():
        if key not in residual_state_dict:
            merged_state_dict[key] = base_tensor
            continue

        res_tensor = residual_state_dict[key]

        # Exact match → normal add
        if base_tensor.shape == res_tensor.shape:
            merged_state_dict[key] = (base_tensor + res_tensor).to(torch.float32)
            continue

        # Common case: vocab resized → dim0 differs, rest matches
        if (
            base_tensor.ndim == res_tensor.ndim
            and base_tensor.ndim >= 1
            and base_tensor.shape[1:] == res_tensor.shape[1:]
            and base_tensor.shape[0] != res_tensor.shape[0]
        ):
            n = min(base_tensor.shape[0], res_tensor.shape[0])
            out = base_tensor.clone().to(torch.float32)
            out[:n] += res_tensor[:n].to(torch.float32)
            merged_state_dict[key] = out
            mismatched.append((key, tuple(base_tensor.shape), tuple(res_tensor.shape), n))
            continue

        # Anything else is suspicious → don’t silently corrupt
        raise RuntimeError(
            f"Shape mismatch for key '{key}': base={tuple(base_tensor.shape)} "
            f"residual={tuple(res_tensor.shape)}. Not a simple vocab-resize mismatch."
        )

    if mismatched:
        print("\nHandled vocab-resize mismatches by partial add:")
        for k, bs, rs, n in mismatched[:20]:
            print(f" - {k}: base{bs} vs res{rs} → added first {n} rows, kept the rest unchanged")
        if len(mismatched) > 20:
            print(f" ... and {len(mismatched) - 20} more")

    # Load merged weights back
    base_model.load_state_dict(merged_state_dict, strict=True)

    # Save as bf16
    base_model = base_model.to(torch.bfloat16)
    os.makedirs(output_dir, exist_ok=True)
    base_model.save_pretrained(output_dir, safe_serialization=True)

    # Save config (optional; save_pretrained usually does it, but keeping your intent)
    base_config = AutoConfig.from_pretrained(base_model_dir)
    base_config.save_pretrained(output_dir)

    # Best way to keep tokenizer consistent (incl. added tokens)
    try:
        tok = AutoTokenizer.from_pretrained(base_model_dir, trust_remote_code=True)
        tok.save_pretrained(output_dir)
    except Exception:
        # fallback to your original file-copy approach
        for file_name in ["tokenizer.json", "tokenizer_config.json", "special_tokens_map.json"]:
            src_path = os.path.join(base_model_dir, file_name)
            dst_path = os.path.join(output_dir, file_name)
            if os.path.exists(src_path):
                shutil.copyfile(src_path, dst_path)

    print(f"\n✅ Merge complete.")
    print(f"🧠 fp32 math → saved bf16 at: {output_dir}")


if __name__ == "__main__":
    lr_file = "/workspace/Llama-3.2-3B-Lr/instruction_residual_adapter"
    base_model_file = "/workspace/v126rc_exp3/F_r10000/checkpoint-31"
    output_root = "/workspace/v126rc_exp3/F_r10000/checkpoint-31/residued"

    merge_instruction_residual(lr_file, base_model_file, output_root)