llama4-scout-17b-eagle3-dummy-drafter

Llama4 Scout 17B Eagle3 Dummy Drafter

This is a dummy/test drafter model for testing the Eagle3 speculative decoding implementation with Llama4 Scout 17B Instruct models in vLLM.

⚠️ WARNING: This is not a real model and should not be used for actual inference. It contains random weights and is only for testing purposes.

Model Details

• Architecture: Llama4ForCausalLM (Eagle3 drafter variant)
• Target Model: Llama4 Scout 17B Instruct (specifically RedHatAI/Llama-4-Scout-17B-16E-Instruct-quantized.w4a16)
• Base Model: Based on the Instruct version of Llama4 17B Scout model
• Hidden Size: 2048
• Layers: 1 (single decoder layer as per Eagle3 design)
• Vocabulary: 128256 tokens
• Parameters: ~322M

Configuration

This drafter model is specifically designed for the Instruct version of Llama4 Scout 17B and uses:

• Eagle3 speculative decoding architecture
• Single-layer transformer with auxiliary hidden state combination
• Llama4 layer structure with RoPE (Rotary Position Embedding)
• SGLang-compatible weight naming (midlayer.*)
• Vocabulary mappings (t2d/d2t) for draft-to-target token conversion

Usage

This model is designed specifically for testing the vLLM Eagle3 implementation:

# Use with vLLM for testing Eagle3 speculative decoding with Llama4 Scout
vllm serve RedHatAI/Llama-4-Scout-17B-16E-Instruct-quantized.w4a16 \
    --speculative-config '{"method": "eagle3", "model": "nm-testing/llama4-scout-17b-eagle3-dummy-drafter", ...}'

Testing Purpose Only

This model:

• Contains random weights
• Is not trained on any data
• Should not be used for actual inference
• Is only for vLLM development and testing

Related

• vLLM: https://github.com/vllm-project/vllm
• Eagle3: Speculative decoding method