eagle3-llama3.1-8b-instruct-speculators

Eagle-3 Speculator for Llama-3.1-8B-Instruct

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

Model Details

• Base Model: meta-llama/Meta-Llama-3.1-8B-Instruct
• Speculator Type: Eagle-3
• Draft Vocabulary Size: 32,000
• Target Vocabulary Size: 128,256
• Architecture: Single-layer transformer with vocabulary mapping

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 (3×4096 → 4096)
• Layer Normalization: Applied before residual connection (HF checkpoint style)

Usage

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

# Load verifier model
verifier = AutoModelForCausalLM.from_pretrained("meta-llama/Meta-Llama-3.1-8B-Instruct")

# Load Eagle-3 speculator
speculator = Eagle3Speculator.from_pretrained(
    "nm-testing/eagle3-llama3.1-8b-instruct-speculators",
    verifier=verifier
)

Configuration

This model uses the Eagle-3 architecture with:

• Hidden size: 4096
• Attention heads: 32
• Key-value heads: 8
• Intermediate size: 14336
• RMS norm epsilon: 1e-05

Citation

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

License

Please refer to the base Llama-3.1 model license.