rag

How to run RAG LLM on AMD Ryzen AI NPU

To leverage AMD Ryzen AI NPU for RAG LLM application involves setting up your environment correctly. The following steps will guide you through preparing your environment, quantizing the model for efficient execution on the NPU, and integrating it into the RAG framework.

Note: This example is intended solely for demonstrating the integration of the Ryzen-AI LLM flow with LlamaIndex for the Retrieval-Augmented Generation (RAG) application. The context passed with RAG could have a prompt length greater than 2048 tokens, for which it is not tuned for performance optimization yet.

1. Clone Ryzen AI Transformers Repository

• Clone the Ryzen AI SW repository. For the initial environment setup, follow the detailed steps provided in the RyzenAI/example/transformers/models/llm folder.

2. Quantize the model

• Download the original model weights and place it in the RyzenAI-SW/example/transformers/models/llm directory. Now to generate the quantized model to run on the NPU, follow these steps:

  cd RyzenAI-SW/example/transformers/models/llm

  python run_awq.py --model_name llama-2-7b-chat --task quantize

If you have followed these instructions, you would now have a quantized model file, such as quantized_llama-2-7b-chat_w4_g128_awq.pth in the RyzenAI-SW/example/transformers/models/llm/quantized_models folder.

3. Prepare the RAG LLM Framework

Once you have completed the setup and quantization using the Ryzen AI transformers guide, you can integrate the quantized model into the RAG Framework.

• Within the same conda environment, install additional dependencies required for the RAG Framework:

  cd RyzenAI-SW/example/transformers/models/rag

  pip install -r requirements.txt

• Copy the original llama-2-7b-chat model from step 1, into models/rag/llama-2-7b-chat

• Copy the quantized model from step 1, into models/rag/quantized_models

4. Configure and run the RAG Application

Configure the RAG Application to use the quantized model, enabling the optional features like speculative decoding (assisted generation)

python run.py --model_name llama-2-7b-chat --target aie --no-direct_llm --quantized --assisted_generation

Note: fast_attention optimization is currently only supported for input prompt/token length <=2048, and is turned off in this RAG example for 1.2 release.