convert_to_gguf_q8.py

#!/usr/bin/env python3
# /// script
# dependencies = [
#     "transformers>=4.36.0",
#     "peft>=0.7.0",
#     "torch>=2.0.0",
#     "accelerate>=0.24.0",
#     "huggingface_hub>=0.20.0",
#     "sentencepiece>=0.1.99",
#     "protobuf>=3.20.0",
#     "numpy",
#     "gguf",
# ]
# ///

"""GGUF Conversion - Q8_0 Quantization"""

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

print("🔄 GGUF Conversion Script - Q8_0")
print("=" * 60)

ADAPTER_MODEL = "chaddy81/qwen3-0.6b-multicode-sft"
BASE_MODEL = "Qwen/Qwen3-0.6B"
OUTPUT_REPO = "chaddy81/qwen3-0.6b-multicode-sft-gguf"

print(f"\n📦 Configuration:")
print(f"   Base model: {BASE_MODEL}")
print(f"   Adapter model: {ADAPTER_MODEL}")
print(f"   Output repo: {OUTPUT_REPO}")
print(f"   Quantization: Q8_0 (8-bit)")

# Step 1: Load and merge
print("\n🔧 Step 1: Loading base model and LoRA adapter...")
base_model = AutoModelForCausalLM.from_pretrained(
    BASE_MODEL, torch_dtype=torch.float16, device_map="auto", trust_remote_code=True,
)
print("   ✅ Base model loaded")

model = PeftModel.from_pretrained(base_model, ADAPTER_MODEL)
print("   ✅ Adapter loaded")

merged_model = model.merge_and_unload()
print("   ✅ Models merged!")

tokenizer = AutoTokenizer.from_pretrained(ADAPTER_MODEL, trust_remote_code=True)
print("   ✅ Tokenizer loaded")

# Step 2: Save merged model
print("\n💾 Step 2: Saving merged model...")
merged_dir = "/tmp/merged_model"
merged_model.save_pretrained(merged_dir, safe_serialization=True)
tokenizer.save_pretrained(merged_dir)
print(f"   ✅ Merged model saved")

# Step 3: Setup llama.cpp
print("\n📥 Step 3: Setting up llama.cpp...")
subprocess.run(["apt-get", "update", "-qq"], check=True, capture_output=True)
subprocess.run(["apt-get", "install", "-y", "-qq", "build-essential", "cmake"], check=True, capture_output=True)
print("   ✅ Build tools installed")

subprocess.run(["git", "clone", "--depth", "1", "https://github.com/ggerganov/llama.cpp.git", "/tmp/llama.cpp"], check=True, capture_output=True)
print("   ✅ llama.cpp cloned")

subprocess.run(["pip", "install", "-r", "/tmp/llama.cpp/requirements.txt"], check=True, capture_output=True)
subprocess.run(["pip", "install", "sentencepiece", "protobuf"], check=True, capture_output=True)
print("   ✅ Dependencies installed")

# Step 4: Convert to GGUF (FP16)
print("\n🔄 Step 4: Converting to GGUF format...")
gguf_output_dir = "/tmp/gguf_output"
os.makedirs(gguf_output_dir, exist_ok=True)

model_name = "qwen3-0.6b-multicode-sft"
gguf_file = f"{gguf_output_dir}/{model_name}-f16.gguf"

result = subprocess.run(
    ["python", "/tmp/llama.cpp/convert_hf_to_gguf.py", merged_dir, "--outfile", gguf_file, "--outtype", "f16"],
    check=True, capture_output=True, text=True
)
print(f"   ✅ FP16 GGUF created")

# Step 5: Build quantize and create Q8_0
print("\n⚙️  Step 5: Creating Q8_0 quantization...")
os.makedirs("/tmp/llama.cpp/build", exist_ok=True)
subprocess.run(
    ["cmake", "-B", "/tmp/llama.cpp/build", "-S", "/tmp/llama.cpp", "-DGGML_CUDA=OFF"],
    check=True, capture_output=True, text=True
)
subprocess.run(
    ["cmake", "--build", "/tmp/llama.cpp/build", "--target", "llama-quantize", "-j", "4"],
    check=True, capture_output=True, text=True
)
print("   ✅ Quantize tool built")

quant_file = f"{gguf_output_dir}/{model_name}-q8_0.gguf"
result = subprocess.run(
    ["/tmp/llama.cpp/build/bin/llama-quantize", gguf_file, quant_file, "Q8_0"],
    check=True, capture_output=True, text=True
)
size_mb = os.path.getsize(quant_file) / (1024 * 1024)
print(f"   ✅ Q8_0: {size_mb:.1f} MB")

# Step 6: Upload
print("\n☁️  Step 6: Uploading to Hub...")
api = HfApi()
api.create_repo(repo_id=OUTPUT_REPO, repo_type="model", exist_ok=True)
print("   ✅ Repository ready")

api.upload_file(path_or_fileobj=quant_file, path_in_repo=f"{model_name}-q8_0.gguf", repo_id=OUTPUT_REPO)
print("   ✅ Q8_0 uploaded")

readme = f"""---
base_model: {BASE_MODEL}
tags:
- gguf
- llama.cpp
- quantized
- trl
- sft
- qwen3
- code
---

# {model_name} GGUF

GGUF conversion of [{ADAPTER_MODEL}](https://huggingface.co/{ADAPTER_MODEL}).

## Model Details

- **Base Model:** {BASE_MODEL}
- **Fine-tuned Model:** {ADAPTER_MODEL}
- **Training:** SFT on code datasets (Codeforces, Golang, Vue/Nuxt, React)
- **Quantization:** Q8_0 (8-bit, high quality)

## Usage

### With Ollama

```bash
huggingface-cli download {OUTPUT_REPO} {model_name}-q8_0.gguf
echo "FROM ./{model_name}-q8_0.gguf" > Modelfile
ollama create qwen3-multicode -f Modelfile
ollama run qwen3-multicode
```

### With llama.cpp

```bash
huggingface-cli download {OUTPUT_REPO} {model_name}-q8_0.gguf
./llama-cli -m {model_name}-q8_0.gguf -p "Write a React component"
```

### With LM Studio

Download the `.gguf` file and import into LM Studio.
"""

api.upload_file(path_or_fileobj=readme.encode(), path_in_repo="README.md", repo_id=OUTPUT_REPO)
print("   ✅ README uploaded")

print("\n" + "=" * 60)
print("✅ GGUF Conversion Complete!")
print(f"📦 https://huggingface.co/{OUTPUT_REPO}")
print(f"📥 huggingface-cli download {OUTPUT_REPO} {model_name}-q8_0.gguf")
print("=" * 60)