README_mlx.md

# Burmese Coder 4B - MLX (4-bit Quantized)

This directory contains the **Burmese Coder 4B** model converted to the native **MLX format**, optimized for Apple Silicon (M1/M2/M3/M4).

## 🚀 Performance
- **Format**: MLX (4-bit quantization)
- **Size**: ~2.2 GB
- **Inference Speed**: ~60 tokens/sec (on standard Apple Silicon)
- **Memory Usage**: ~2.3 GB Peak VRAM

## 🛠️ Prerequisites

Ensure you have `mlx-lm` installed in your Python environment (Python 3.10+ required):
```bash
pip install mlx-lm
```

## 💬 Usage (CLI Inference)

You can run the model directly from the command line:

```bash
python -m mlx_lm.generate --model burmese_coder_4b_mlx --prompt "Python မှာ list တစ်ခုကို ဘယ်လို sort လုပ်ရသလဲ?" --max-tokens 300
```

## 🐍 Python API Usage

```python
from mlx_lm import load, generate

model, tokenizer = load("burmese_coder_4b_mlx")

prompt = "Python မှာ list တစ်ခုကို ဘယ်လို sort လုပ်ရသလဲ?"
messages = [{"role": "user", "content": prompt}]
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)

response = generate(model, tokenizer, prompt=prompt, verbose=True)
print(response)
```

## 📝 Conversion Details
- **Base Model**: `unsloth/gemma-3-4b-it`
- **Adapter**: `WYNN747/burmese-coder-4b` (LoRA)
- **Quantization**: 4-bit (`--q-bits 4`)
- **Fusing**: Merged using `mlx-lm` fuse tool before quantization for maximum integrity.