merge_lora_adapter.py

#!/usr/bin/env python3
"""
Merge Codette LoRA Adapter with Base Model — v2
Merges HuggingFace PEFT adapter into base model using llama.cpp's export tool.

Run this AFTER training completes and the adapter is on HuggingFace.

Two paths:
  A) HuggingFace format → merged safetensors (for further conversion)
  B) GGUF base + GGUF LoRA → merged GGUF (if you have GGUF versions of both)

Usage:
  $env:HF_TOKEN = "your_token"
  python merge_lora_adapter.py
"""

import os
import sys
import subprocess
from pathlib import Path

# ── Config ─────────────────────────────────────────────────────────────────
HF_TOKEN      = os.environ.get("HF_TOKEN", "")

# Path A: Merge HuggingFace adapter (use this after training completes)
BASE_MODEL_HF  = "meta-llama/Llama-3.2-1B-Instruct"
ADAPTER_REPO   = "Raiff1982/codette-llama-adapter"
MERGED_HF_DIR  = Path("J:/TheAI/models/codette-v2-merged")

# Path B: Merge GGUF LoRA into GGUF base (use if you have GGUF-format LoRA)
BASE_GGUF      = Path("J:/TheAI/models/codette-v2-gguf/codette-v2.gguf")
LORA_GGUF      = Path("J:/TheAI/models/codette-rc-xi-lora.bin")
OUTPUT_GGUF    = Path("J:/TheAI/models/codette-v2-merged.gguf")
LLAMA_TOOL     = Path("J:/TheAI/llama.cpp/build/bin/Release/llama-export-lora.exe")

# ── Validate token ──────────────────────────────────────────────────────────
if not HF_TOKEN:
    print("[!] HF_TOKEN not set. Run:")
    print('    $env:HF_TOKEN = "your_token_here"')
    sys.exit(1)

print("=" * 80)
print("MERGE CODETTE v2 LORA ADAPTER WITH BASE MODEL")
print("=" * 80)
print()
print("Select merge path:")
print("  A) HuggingFace format (PEFT adapter + HF base → merged safetensors)")
print("  B) GGUF format (GGUF base + GGUF LoRA → merged GGUF)")
print()

choice = input("Enter A or B [default: A]: ").strip().upper() or "A"

# ── Path A: HuggingFace PEFT merge ─────────────────────────────────────────
if choice == "A":
    print()
    print("=" * 60)
    print("PATH A: HuggingFace PEFT Merge")
    print("=" * 60)

    import torch
    from transformers import AutoModelForCausalLM, AutoTokenizer
    from peft import PeftModel

    MERGED_HF_DIR.mkdir(parents=True, exist_ok=True)

    print(f"[*] Loading tokenizer: {BASE_MODEL_HF}")
    tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL_HF, token=HF_TOKEN)
    if tokenizer.pad_token is None:
        tokenizer.pad_token = tokenizer.eos_token

    print(f"[*] Loading base model: {BASE_MODEL_HF}")
    base = AutoModelForCausalLM.from_pretrained(
        BASE_MODEL_HF,
        torch_dtype=torch.float16,
        low_cpu_mem_usage=True,
        token=HF_TOKEN,
    )

    print(f"[*] Loading LoRA adapter: {ADAPTER_REPO}")
    model = PeftModel.from_pretrained(base, ADAPTER_REPO, token=HF_TOKEN)

    print("[*] Merging and unloading LoRA weights...")
    model = model.merge_and_unload()
    model = model.to(torch.float16)

    print(f"[*] Saving merged model to {MERGED_HF_DIR}")
    model.save_pretrained(MERGED_HF_DIR, safe_serialization=True)
    tokenizer.save_pretrained(MERGED_HF_DIR)

    size_gb = sum(f.stat().st_size for f in MERGED_HF_DIR.rglob("*") if f.is_file()) / (1024**3)
    print(f"[✓] Merged model saved — {size_gb:.2f} GB")
    print()
    print("[*] Next step — convert to GGUF:")
    print(f"    python J:/TheAI/llama.cpp/convert_hf_to_gguf.py {MERGED_HF_DIR} --outfile J:/TheAI/models/codette-v2-gguf/codette-v2.gguf --outtype q8_0")
    print()
    print("[*] Or run make_codette_gguf.py which does all steps automatically.")

# ── Path B: GGUF LoRA merge ─────────────────────────────────────────────────
elif choice == "B":
    print()
    print("=" * 60)
    print("PATH B: GGUF LoRA Merge")
    print("=" * 60)

    print("[*] Checking required files...")

    if not BASE_GGUF.exists():
        print(f"[!] Base GGUF not found: {BASE_GGUF}")
        print("[!] Run make_codette_gguf.py first to create the base GGUF.")
        sys.exit(1)
    print(f"[✓] Base GGUF: {BASE_GGUF.stat().st_size / (1024**3):.2f} GB")

    if not LORA_GGUF.exists():
        print(f"[!] LoRA GGUF not found: {LORA_GGUF}")
        print("[!] Note: HuggingFace PEFT adapters are not GGUF format.")
        print("[!] Use Path A to merge the HuggingFace adapter, then convert the result.")
        sys.exit(1)
    print(f"[✓] LoRA GGUF: {LORA_GGUF.stat().st_size / (1024**2):.2f} MB")

    if not LLAMA_TOOL.exists():
        print(f"[!] Merge tool not found: {LLAMA_TOOL}")
        print("[!] Build llama.cpp first:")
        print("    cd J:/TheAI/llama.cpp")
        print("    cmake -B build && cmake --build build --config Release")
        sys.exit(1)
    print(f"[✓] Merge tool found")

    OUTPUT_GGUF.parent.mkdir(parents=True, exist_ok=True)

    print()
    print(f"[*] Merging {BASE_GGUF.name} + {LORA_GGUF.name}")
    print(f"[*] Output: {OUTPUT_GGUF}")
    print()

    cmd = [
        str(LLAMA_TOOL),
        "--model", str(BASE_GGUF),
        "--lora", str(LORA_GGUF),
        "--output", str(OUTPUT_GGUF),
    ]

    result = subprocess.run(cmd, cwd="J:/TheAI")

    if result.returncode == 0 and OUTPUT_GGUF.exists():
        size_gb = OUTPUT_GGUF.stat().st_size / (1024**3)
        print(f"[✓] Merge complete: {OUTPUT_GGUF} ({size_gb:.2f} GB)")
        print()
        print("[*] Create Ollama model:")
        print(f"    ollama create codette-v2 -f J:/TheAI/models/codette-v2-gguf/Modelfile")
        print()
        print("[*] Or load directly in llama.cpp:")
        print(f"    llama-cli.exe -m {OUTPUT_GGUF} -p 'Your prompt here'")
    else:
        print("[!] Merge failed or output not created.")
        print()
        print("[*] Alternative — load LoRA separately at inference time:")
        print(f"    llama-cli.exe -m {BASE_GGUF} --lora {LORA_GGUF}")

else:
    print(f"[!] Unknown choice: {choice}")
    sys.exit(1)