Squid: Long Context as a New Modality for on-device RAG

- Nexa Model Hub - ArXiv

nexa-octopus

Overview

Squid is a novel approach to accelerate language model inference by treating long context as a new modality, similar to image, audio, and video modalities in vision-language models. This innovative method incorporates a language encoder model to encode context information into embeddings, applying multimodal model concepts to enhance the efficiency of language model inference。 Below are model highlights:

  • 🧠 Context as a distinct modality
  • 🗜️ Language encoder for context compression
  • 🔗 Multimodal techniques applied to language processing
  • ⚡ Optimized for energy efficiency and on-device use
  • 📜 Specialized for long context understanding

Model Architecture

Squid employs a decoder-decoder framework with two main components:

  1. A smaller decoder (0.5B parameters) for transforming information from extensive contexts
  2. A larger decoder (7B parameters) for comprehending and generating responses to current queries
  3. The architecture also includes a projector to align embeddings between the text encoder and the main decoder.

Model Architecture

Running the Model

Method 1

download this repository and run the following commands:

git lfs install
git clone https://huggingface.co/NexaAIDev/Squid
python inference_example.py

Method 2

Install nexaai-squid package

pip install nexaai-squid

Then run the following commands:

from transformers import AutoModelForCausalLM, AutoTokenizer, AutoConfig
import torch
from squid.configuration_squid import SquidConfig
from squid.modeling_squid import SquidForCausalLM


def inference_instruct(mycontext, question, device="cuda:0"):
    import time
    MEMORY_SIZE = 32
    start_time = time.time()
    generated_token_ids = []
    prompt = f" <context>{question}"
    text_chunks = [tokenizer(chunk).input_ids for chunk in prompt.split("<context>")]
    input_ids = (
        torch.tensor(
            text_chunks[0] + [-1] * MEMORY_SIZE + text_chunks[1], dtype=torch.long
        )
        .unsqueeze(0)
        .to(device)
    )
    context_tokenized = tokenizer(
        mycontext + "".join([f"[memory_{i}]" for i in range(MEMORY_SIZE)]),
        return_tensors="pt",
    )
    context_tokenized = {k: v.to(device) for k, v in context_tokenized.items()}
    context_token_count = (context_tokenized["input_ids"]).shape[1] - MEMORY_SIZE
    for i in range(context_token_count):
        next_token = (
            model(
                input_ids,
                context_input_ids=context_tokenized["input_ids"],
                context_attention_mask=context_tokenized["attention_mask"],
            )
            .logits[:, -1]
            .argmax(-1)
        )
        if next_token.item() == 151643:
            break
        generated_token_ids.append(next_token.item())
        input_ids = torch.cat([input_ids, next_token.unsqueeze(1)], dim=-1)
    result = tokenizer.decode(generated_token_ids)
    print(f"Time taken: {time.time() - start_time}")
    return result


if __name__ == "__main__":
    device_name = "cuda:0" if torch.cuda.is_available() else "cpu"
    AutoConfig.register("squid", SquidConfig)
    AutoModelForCausalLM.register(SquidConfig, SquidForCausalLM)
    tokenizer = AutoTokenizer.from_pretrained('NexaAIDev/Squid')
    model = AutoModelForCausalLM.from_pretrained('NexaAIDev/Squid', trust_remote_code=True, torch_dtype=torch.bfloat16, device_map=device_name)
    
    # Run inference example
    mycontext = "Nexa AI is a Cupertino-based company founded in May 2023 that researches and develops models and tools for on-device AI applications. The company is founded by Alex and Zack. The company is known for its Octopus-series models, which rival large-scale language models in capabilities such as function-calling, multimodality, and action-planning, while remaining efficient and compact for edge device deployment. Nexa AI's mission is to advance on-device AI in collaboration with the global developer community. To this end, the company has created an on-device model hub for users to find, share, and collaborate on open-source AI models optimized for edge devices, as well as an SDK for developers to run and deploy AI models locally"
    question = "Who founded Nexa AI?"
    result = inference_instruct(mycontext, question, device=device_name)
    print("Result:", result)

Training Process

Squid's training involves three stages:

  1. Restoration Training: Reconstructing original context from compressed embeddings
  2. Continual Training: Generating context continuations from partial compressed contexts
  3. Instruction Fine-tuning: Generating responses to queries given compressed contexts

This multi-stage approach progressively enhances the model's ability to handle long contexts and generate appropriate responses.

Citation

If you use Squid in your research, please cite our paper:

@article{chen2024squidlongcontextnew,
      title={Squid: Long Context as a New Modality for Energy-Efficient On-Device Language Models}, 
      author={Wei Chen and Zhiyuan Li and Shuo Xin and Yihao Wang},
      year={2024},
      eprint={2408.15518},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2408.15518}, 
}

Contact

For questions or feedback, please contact us

Downloads last month
37
Safetensors
Model size
8.11B params
Tensor type
BF16
·
Inference API
Inference API (serverless) has been turned off for this model.

Model tree for NexaAIDev/Squid

Base model

Qwen/Qwen2-7B
Finetuned
(59)
this model

Collection including NexaAIDev/Squid