hf-quickstart

HF Quickstart

Simplest entry point: trains an Eagle3 draft model using the HuggingFace backend for target model inference.

Prerequisites

• 3 GPUs (1 inference + 2 training)
• Model access to Qwen/Qwen3-8B

Config

Uses configs/hf_qwen3_8b.yaml:

• Backend: HFEngine with HFTargetModel (no SGLang required)
• Training: 2 GPUs with FSDP, flex_attention
• Inference: 1 GPU

How to run

./examples/hf-quickstart/run.sh

Common customizations

# Adjust training steps
./examples/hf-quickstart/run.sh training.num_train_steps=50

# Change learning rate
./examples/hf-quickstart/run.sh training.learning_rate=5e-5