Announcing preview support for Llama 2 in DirectML

Posted by

This post has been republished via RSS; it originally appeared at: Windows Blog.

At Inspire this year we talked about how developers will be able to run Llama 2 on Windows with DirectML and the ONNX Runtime and we've been hard at work to make this a reality. We now have a sample showing our progress with Llama 2 7B! See This sample relies on first doing an optimization pass on the model with Olive, a powerful optimization tool for ONNX models. Olive utilizes powerful graph fusion optimizations from ONNX Runtime and a model architecture optimized for DirectML to speed up inference times by up to 10X! After this optimization pass, Llama 2 7B runs fast enough that you can have a conversation in real time on multiple vendors’ hardware! We’ve also built a little UI to make it easy to see the optimized model in action. Thank you to our hardware partners who helped make this happen. For more on how Llama 2 lights up on our partners’ hardware with DirectML see: We're excited about this milestone, but this is only a first peek - stay tuned for future enhancements to support even larger models, fine-tuning and lower-precision data types.

Getting started

Requesting Llama 2 access

To run our Olive optimization pass in our sample you should first request access to the Llama 2 weights from Meta.


We recommend upgrading to the latest drivers for the best performance.
  • AMD has released optimized graphics drivers supporting AMD RDNA™ 3 devices including AMD Radeon™ RX 7900 Series graphics cards. Download Adrenalin Edition™ 23.11.1 or newer (
  • Intel has released optimized graphics drivers supporting Intel Arc A-Series graphics cards. Download the latest drivers.
  • NVIDIA: Users of NVIDIA GeForce RTX 20, 30 and 40 Series GPUs, can see these improvements first hand, in GeForce Game Ready Driver 546.01.

Leave a Reply

Your email address will not be published. Required fields are marked *

This site uses Akismet to reduce spam. Learn how your comment data is processed.