convert_survival_32b_gguf.py

# /// script
# dependencies = ["peft", "transformers", "torch", "huggingface_hub", "sentencepiece", "cmake"]
# ///

import os
import subprocess
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer
from huggingface_hub import HfApi, create_repo
import torch

# Configuration
BASE_MODEL_ID = "Qwen/Qwen2.5-32B-Instruct"
ADAPTER_ID = "sunkencity/survival-expert-qwen-32b"
OUTPUT_REPO = "sunkencity/survival-expert-qwen-32b-gguf"
MERGED_DIR = "merged_model"
GGUF_FILE = "survival-expert-qwen-32b.Q4_K_M.gguf"

print(f"Loading base model: {BASE_MODEL_ID}")
# Load in bfloat16 to save memory and match training
base_model = AutoModelForCausalLM.from_pretrained(
    BASE_MODEL_ID,
    device_map="auto",
    torch_dtype=torch.bfloat16,
    trust_remote_code=True
)

print(f"Loading adapter: {ADAPTER_ID}")
model = PeftModel.from_pretrained(base_model, ADAPTER_ID)

print("Merging model...")
model = model.merge_and_unload()

print(f"Saving merged model to {MERGED_DIR}...")
model.save_pretrained(MERGED_DIR)
tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL_ID)
tokenizer.save_pretrained(MERGED_DIR)

print("Cloning llama.cpp...")
if os.path.exists("llama.cpp"):
    subprocess.run(["rm", "-rf", "llama.cpp"])
subprocess.run(["git", "clone", "https://github.com/ggerganov/llama.cpp"], check=True)

print("Installing llama.cpp requirements...")
subprocess.run(["pip", "install", "-r", "llama.cpp/requirements.txt"], check=True)

print("Building llama-quantize with CMake...")
os.makedirs("llama.cpp/build", exist_ok=True)
subprocess.run(["cmake", "-B", "llama.cpp/build", "-S", "llama.cpp"], check=True)
subprocess.run(["cmake", "--build", "llama.cpp/build", "--config", "Release", "-j"], check=True)

print("Converting to GGUF (FP16)...")
# Convert to GGUF (BF16/FP16 preserved from model)
subprocess.run([
    "python", "llama.cpp/convert_hf_to_gguf.py",
    MERGED_DIR,
    "--outfile", "merged_model.gguf",
    "--outtype", "bf16" 
], check=True)

print("Quantizing to Q4_K_M...")
quantize_bin = "llama.cpp/build/bin/llama-quantize"
subprocess.run([
    quantize_bin,
    "merged_model.gguf",
    GGUF_FILE,
    "Q4_K_M"
], check=True)

print(f"Creating repo {OUTPUT_REPO}...")
api = HfApi()
create_repo(OUTPUT_REPO, repo_type="model", exist_ok=True)

print(f"Uploading {GGUF_FILE}...")
api.upload_file(
    path_or_fileobj=GGUF_FILE,
    path_in_repo=GGUF_FILE,
    repo_id=OUTPUT_REPO,
    repo_type="model"
)

print("Done! GGUF available at:", f"https://huggingface.co/{OUTPUT_REPO}")