CAMEL

Introduction

CAMEL(Context-Aware Modifier for Efficient Language model) is a speculative decoding method inspired by EAGLE. It compresses former input hidden states according to window size and then make speculations.

Installation

pip install modifier

Quick Start

CAMEL only supports meta-llama/Llama-2-7b-chat-hf currently.

import torch
from camel import CamelModel

prompt = "What is artificial intelligence?"
model = CamelModel.from_pretrained(
    base_model_path="meta-llama/Llama-2-7b-chat-hf",
    modifier_path="0xWe11es/camel-llama2-h1024-w1",
    torch_dtype=torch.float16,
    device_map="auto"
)
tokenizer = model.get_tokenizer()
input_ids = tokenizer(prompt).input_ids
output_ids = model.generate(input_ids)
output = tokenizer.decode(output_ids)
print(output)

CAMEL has the following modifier based on Llama2 (h stands for hidden size, w stands for window size):

• 0xWe11es/camel-llama2-h256-w1
• 0xWe11es/camel-llama2-h256-w4
• 0xWe11es/camel-llama2-h256-w16
• 0xWe11es/camel-llama2-h256-w64
• 0xWe11es/camel-llama2-h1024-w1
• 0xWe11es/camel-llama2-h1024-w4
• 0xWe11es/camel-llama2-h1024-w16
• 0xWe11es/camel-llama2-h1024-w64

Performance

We test modifier 0xWe11es/camel-llama2-h1024-w4 on several datasets, and get the following results compared to vanilla model (hf version).

Dataset	Model	Temperature	Speed(Token/s)	Speedup
MT-Bench	LlaMa2 7B	0.0	71.85	1.92x
MT-Bench	LlaMa2 7B	1.0	57.54	1.62x
GSM8K	LlaMa2 7B	0.0	73.51	2.20x
GSM8K	LlaMa2 7B	1.0	57.15	1.77x
Alpaca	LlaMa2 7B	0.0	68.92	1.88x
Alpaca	LlaMa2 7B	1.0	55.38	1.56x

Reference

• Medusa

• EAGLE