Llama-v2-7B-Chat: Optimized for Mobile Deployment

State-of-the-art large language model useful on a variety of language understanding and generation tasks

Llama 2 is a family of LLMs. The "Chat" at the end indicates that the model is optimized for chatbot-like dialogue. The model is quantized to 4-bit weights and 16-bit activations making it suitable for on-device deployment. For Prompt and output length specified below, the time to first token is Llama-PromptProcessor-Quantized's latency and average time per addition token is Llama-TokenGenerator-KVCache-Quantized's latency.

This model is an implementation of Llama-v2-7B-Chat found here. This repository provides scripts to run Llama-v2-7B-Chat on Qualcomm® devices. More details on model performance across various devices, can be found here.

Model Details

• Model Type: Text generation
• Model Stats:
  • Number of parameters: 7B
  • Model size: 3.6GB
  • Model-1 (Prompt Processor): Llama-PromptProcessor-Quantized
  • Max context length: 1024
  • Prompt processor input: 1024 tokens
  • Prompt processor output: 1024 output tokens + KVCache for token generator
  • Model-2 (Token Generator): Llama-TokenGenerator-KVCache-Quantized
  • Token generator input: 1 input token + past KVCache
  • Token generator output: 1 output token + KVCache for next iteration
  • Decoding length: 1024 (1 output token + 1023 from KVCache)
  • Use: Initiate conversation with prompt-processor and then token generator for subsequent iterations.
  • QNN-SDK: 2.19

Deploying Llama 2 on-device

Large Language Model (LLM) such as Llama 2 has the following complexities to deploy on-device:

1. Model size is too large to fit in device memory for inference
2. Multi-Head Attention (MHA) has large activations leading to fallback from accelerators
3. High model load and inference time

We can tackle the above constraints with the following steps:

1. Quantize weights to reduce on-disk model size, e.g., int8 or int4 weights
2. Quantize activations to reduce inference time memory pressure
3. Graph transformations to reduce inference time memory pressure, e.g., Multi-Head to Split-Head Attention (MHA -> SHA)
4. Graph transformations to convert or decompose operations into more accelerator friendly operations e.g. Linear to Conv
5. For LLM with 7B or more parameters, above steps are still not good enough on mobile, hence we go one step further and split model into sub-parts.

Here, we divide the model into 4 parts in order to

1. Make model exportable with low memory usage
2. Avoid inference time out-of-memory errors

In order to export Llama 2, please ensure

1. Host machine has >40GB memory (RAM+swap-space)
2. If you don't have enough memory, export.py will dump instructions to increase swap space accordingly

Device	Chipset	Target Runtime	Inference Time (ms)	Peak Memory Range (MB)	Precision	Primary Compute Unit	Target Model
Samsung Galaxy S23 Ultra (Android 13)	Snapdragon® 8 Gen 2	QNN Model Library	104.953 ms	316 - 4785 MB	UINT16	NPU	Llama-TokenGenerator-KVCache-Quantized
Samsung Galaxy S23 Ultra (Android 13)	Snapdragon® 8 Gen 2	QNN Model Library	1917.811 ms	0 - 1028 MB	UINT16	NPU	Llama-PromptProcessor-Quantized

Installation

This model can be installed as a Python package via pip.

pip install "qai-hub-models[llama_v2_7b_chat_quantized]"

Configure Qualcomm® AI Hub to run this model on a cloud-hosted device

Sign-in to Qualcomm® AI Hub with your Qualcomm® ID. Once signed in navigate to Account -> Settings -> API Token.

With this API token, you can configure your client to run models on the cloud hosted devices.

qai-hub configure --api_token API_TOKEN

Navigate to docs for more information.

Demo off target

The package contains a simple end-to-end demo that downloads pre-trained weights and runs this model on a sample input.

python -m qai_hub_models.models.llama_v2_7b_chat_quantized.demo

The above demo runs a reference implementation of pre-processing, model inference, and post processing.

NOTE: If you want running in a Jupyter Notebook or Google Colab like environment, please add the following to your cell (instead of the above).

%run -m qai_hub_models.models.llama_v2_7b_chat_quantized.demo

Run model on a cloud-hosted device

In addition to the demo, you can also run the model on a cloud-hosted Qualcomm® device. This script does the following:

• Performance check on-device on a cloud-hosted device
• Downloads compiled assets that can be deployed on-device for Android.
• Accuracy check between PyTorch and on-device outputs.

python -m qai_hub_models.models.llama_v2_7b_chat_quantized.export

Profile Job summary of Llama-TokenGenerator-KVCache-Quantized
--------------------------------------------------
Device: Snapdragon X Elite CRD (11)
Estimated Inference Time: 118.14 ms
Estimated Peak Memory Range: 64.97-64.97 MB
Compute Units: NPU (34842) | Total (34842)

Profile Job summary of Llama-PromptProcessor-Quantized
--------------------------------------------------
Device: Snapdragon X Elite CRD (11)
Estimated Inference Time: 2302.57 ms
Estimated Peak Memory Range: 10.29-10.29 MB
Compute Units: NPU (31766) | Total (31766)

Deploying compiled model to Android

The models can be deployed using multiple runtimes:

• TensorFlow Lite (.tflite export): This tutorial provides a guide to deploy the .tflite model in an Android application.

• QNN (.so export ): This sample app provides instructions on how to use the .so shared library in an Android application.

View on Qualcomm® AI Hub

Get more details on Llama-v2-7B-Chat's performance across various devices here. Explore all available models on Qualcomm® AI Hub

License

• The license for the original implementation of Llama-v2-7B-Chat can be found here.
• The license for the compiled assets for on-device deployment can be found here

References

• LLaMA: Open and Efficient Foundation Language Models
• Source Model Implementation

Community

• Join our AI Hub Slack community to collaborate, post questions and learn more about on-device AI.
• For questions or feedback please reach out to us.