This is an optimized version of the Llama 2 model, available from Meta under the Llama Community License Agreement found on this repository. Microsoft permits you to use, modify, redistribute and create derivatives of Microsoft's contributions to the optimized version subject to the restrictions and disclaimers of warranty and liability in the Llama Community License agreement.

The sub-modules that contain the ONNX files in this repository are access controlled. To get access permissions to the Llama 2 model, please fill out the Llama 2 access request form. If allowable, you will receive GitHub access in the next 48 hours, but usually much sooner.

Chose from the following sub-modules:

• 7B_FT_float16
• 7B_FT_float32
• 7B_float16
• 7B_float32
• 13B_FT_float16
• 13B_FT_float32
• 13B_float16
• 13B_float32

git clone https://github.com/microsoft/Llama-2-Onnx.git
cd Llama-2-Onnx
git submodule init <chosen_submodule> 
git submodule update

You can repeate the init command with a different submodule name to initialize multiple submodules. Be careful, the contained files are very large! (7B Float16 models are about 10GB)

Llama 2 is a collection of pretrained and fine-tuned generative text models. To learn more about Llama 2, review the Llama 2 model card.

Llama 2 model consists of a stack of decoder layers. Each decoder layer (or transformer block) is constructed from one self-attention layer and one feed-forward multi-layer perceptron. Llama models use different projection sizes compared with classic transformers in the feed-forward layer, for instance, both Llama 1 and Llama 2 projection use 2.7x hidden size rather than the standard 4x hidden size. A key difference between Llama 1 and Llama 2 is the architectural change of attention layer, in which Llama 2 takes advantage of Grouped Query Attention (GQA) mechanism to improve efficiency.

There are two examples provided in this repository. There is a minimum working example shown in Llama-2-Onnx/MinimumExample. This is simply a command line program that will complete some text with the chosen version of Llama 2.

Given the following input:

python MinimumExample/Example_ONNX_LlamaV2.py --onnx_file 7B_FT_float16/ONNX/LlamaV2_7B_FT_float16.onnx --embedding_file 7B_FT_float16/embeddings.pth --tokenizer_path tokenizer.model --prompt "What is the lightest element?"

Output:

The lightest element is hydrogen. Hydrogen is the lightest element on the periodic table, with an atomic mass of 1.00794 u (unified atomic mass units).

There is a more complete chat bot interface that is available in Llama-2-Onnx/ChatApp. This is a python program based on the popular Gradio web interface. It will allow you to interact with the chosen version of Llama 2 in a chat bot interface.

An example interaction can be seen here:

The fine-tuned models were trained for dialogue applications.

To get the expected features and performance for them, a specific formatting needs to be followed, including the INST tag, BOS and EOS tokens, and the whitespaces and breaklines in between (we recommend calling strip() on inputs to avoid double-spaces).

This enables models in chat mode as well as additional safeguards to reduce potentially undesirable output.

ONNX runtime execution provider might need to generate JIT binaries for the underlying hardware, typically the binary is cache and will be loaded directly in the subsequent runs to reduce the overhead.

It is possible that your device does not support native FP16 math, therefore weights will be cast to FP32 at runtime. Using the FP32 version of the model will avoid the cast overhead.

It is recommended that inputs/outputs are put on target device to avoid expensive data copies, please refer to the following document for details.

I/O Binding | onnxruntime

Users can perform temperature and top-p sampling using the model’s output logits. Please refer to Meta’s guidance for the best parameters combination; an example is located here.

In order to help developers innovate responsibly, Meta encourages you to review the Responsible Use Guide for the Llama 2 models.

Microsoft encourages you to learn more about its Responsible AI approach, including many publicly available resources and tools for developers.