EAGLE3-LLaMA3.3-Instruct-70B-speculators

Eagle-3 Speculator for Llama-3.3-70B-Instruct

This is an Eagle-3 speculator checkpoint converted to the speculators format.

Model Details

• Base Model: meta-llama/Llama-3.3-70B-Instruct
• Speculator Type: Eagle-3
• Draft Vocabulary Size: 32,000
• Target Vocabulary Size: 128,256
• Architecture: Single-layer transformer with vocabulary mapping
• Target Model Hidden Size: 8,192
• Draft Model Hidden Size: 6,144

Key Features

• Vocabulary Mapping: Maps between draft (32K) and target (128K) vocabularies
• Custom Attention: Modified attention layer accepting 2×hidden_size input
• Fusion Layer: Processes 3 verifier layers from target model (3×8192 → 6144)
• Optimized for 70B Models: Specifically configured for Llama-3.3-70B architecture

Usage

from speculators.models.eagle3 import Eagle3Speculator, Eagle3SpeculatorConfig
from transformers import AutoModelForCausalLM

# Load verifier model
verifier = AutoModelForCausalLM.from_pretrained("meta-llama/Llama-3.3-70B-Instruct")

# Load Eagle-3 speculator
speculator = Eagle3Speculator.from_pretrained(
    "nm-testing/EAGLE3-LLaMA3.3-Instruct-70B-speculators",
    verifier=verifier
)

Configuration

This model uses the Eagle-3 architecture with:

• Hidden size: 6,144 (draft model)
• Target hidden size: 8,192 (70B Llama model)
• Attention heads: 48
• Key-value heads: 8
• Intermediate size: 16,384
• RMS norm epsilon: 1e-05

Original Model

Converted from: yuhuili/EAGLE3-LLaMA3.3-Instruct-70B

Citation

Based on the Eagle-3 paper: https://arxiv.org/abs/2503.01840

License

Please refer to the base Llama-3.3 model license.