File size: 1,777 Bytes
57c904c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
# Eagle-3 Speculator for Llama-3.3-70B-Instruct

This is an Eagle-3 speculator checkpoint converted to the [speculators](https://github.com/neuralmagic/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

```python
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](https://huggingface.co/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.