Quantize TinyLlama CoreML on Mac ARM (May 2025)

End-to-end workflow for quantizing a TinyLlama LLM CoreML model to int8 on Apple Silicon (M1/M2/M3). Script, logs, and all gotchas included.

---

🚀 Context

Most open-source CoreML LLMs are only provided in float16. This means files are huge (2GB+), RAM usage is high, and iOS/macOS on-device deployment isn’t always efficient. Official docs and forums rarely provide a working, up-to-date, step-by-step guide for post-conversion quantization (.mlpackage) on Mac ARM.

This repo gives you a tested Python script, setup steps, and real logs from a successful int8 quantization (May 2025, Mac ARM, Python 3.11, coremltools 8.3.0).

---

🛠️ Requirements

• Mac ARM (Apple Silicon: M1, M2, M3)
• Python 3.11+ (install with brew install python@3.11)
• coremltools >= 8.3.0 (pip3 install --upgrade coremltools)

---

📦 Quick Start

1. Download your CoreML LLM Example: TKDKid1000/TinyLlama-1.1B-Chat-v0.3-CoreML

   huggingface-cli download --local-dir ./Models/TinyLlama TKDKid1000/TinyLlama-1.1B-Chat-v0.3-CoreML

2. Copy quantize_coreml.py into the directory containing float16_model.mlpackage

3. Run the script:

   python3.11 quantize_coreml.py

---

📝 Script

import coremltools as ct
from coremltools.optimize.coreml import OpLinearQuantizerConfig, OptimizationConfig, linear_quantize_weights

float16_model_path = "float16_model.mlpackage"
quant8_model_path  = "quant8_model.mlpackage"

print(f"🔄 Loading float16 model: {float16_model_path}")
model = ct.models.MLModel(float16_model_path)

op_config = OpLinearQuantizerConfig(mode="linear_symmetric")
config = OptimizationConfig(global_config=op_config)

print("⚡ Quantizing to 8 bits (int8)...")
try:
    quant8_model = linear_quantize_weights(model, config=config)
    quant8_model.save(quant8_model_path)
    print(f"✅ Quantized int8 model saved: {quant8_model_path}")
except Exception as e:
    print(f"❌ Quantization failed: {e}")

---

📈 Results

• File size reduction: From ~2.2GB (float16) → much smaller (quant8)
• Works on device: iOS 17 / macOS Sonoma, Apple Silicon
• int4 quantization: Only available for iOS 18+ (will fail on current OS/devices)

---

⚠️ Notes and Troubleshooting

• Ignore warnings like “inf/-inf not supported by quantization. Skipped.”
• Quantization 4 bits (int4) is only supported on iOS 18+ and newer CoreML.
• If you get ModuleNotFoundError: No module named 'coremltools.optimize.coreml', update coremltools and use Python 3.11+ on Mac ARM.
• Make sure you run the script inside the directory containing float16_model.mlpackage (or adapt the path).

---

🤝 Community

• Feel free to open issues or PRs!
• Share this repo on Reddit, Hugging Face forums, or Discord to help others.

---

Created by GreenBull31 (Morgan CAMILLERI)