Speculative-Decoding-DraftToken-Analysis

#MLX #AppleSilicon #SpeculativeDecoding This project analyzes the performance and quality trade-offs in speculative decoding using draft tokens and different quantized model configurations. It evaluates:

• ⚡ Generation speed (tokens/sec)
• 🔍 Semantic similarity (Accuracy metric) (cosine similarity via sentence embeddings)
• 📝 Text quality (Accuracy metric) (ROUGE-L score)

---

▶️ Install the pip package

pip install speculative-decoding-metrics

---

Customize the Speculative decoding analysis

After installing the package, create a demo script with a default prompt and model:

demo/run_example.py

from speculative_decoding_metrics.main import run_evaluation

run_evaluation(
    base_model="phi-3-mini-4k-instruct",  #Use a model that'll run on your local
    main_quant="8bit",          #use q8 instead of "8bit" based on HuggingFace Repo name 
    draft_quant="4bit",         #use q4 instead of "4bit" based on HuggingFace Repo name
    prompt="How do LLMs work?",
    max_tokens=64,              #Tweak max tokens per output
    num_draft_tokens_list=[0, 1, 2, 3, 4]
)

▶️ Run the Demo

python demo/run_example.py

---

📌 What Is Speculative Decoding?

Speculative decoding speeds up language generation by using a smaller "draft" model to propose tokens, which are then verified by a larger "main" model.

This repo benchmarks speculative decoding using:

• Main model: Quantized to 8bit → mlx-community/<model>-8bit
• Draft model: Quantized to 4bit → mlx-community/<model>-4bit

---

📊 Visualized Metrics

Based on these results, the user can decide which speculative decoding they want to run for the best results! This package generates plots comparing output quality and speed across draft token counts:

1. Tokens/sec – Speed boost with draft tokens
2. Cosine Similarity – Semantic match with baseline (no draft)
3. ROUGE-L – Text overlap quality score

---

🛠 Customization Tips

• 🔧 Change the prompt – Modify the prompt in demo/run_example.py
• 🧠 Try other models – Swap the base_model string (e.g., Mistral, TinyLlama)
• 🎛️ Adjust draft token range – Modify num_draft_tokens_list for finer control
• 📏 Set max output length – Use max_tokens to limit generation length

---

Acknowledgments

• Apple MLX – Lightweight ML framework
• HuggingFace – Transformers + SentenceTransformers
• Google Research – ROUGE scoring tools