Llama 4 Day 0 Support on AMD Instinct GPUs

In today’s fast-evolving AI landscape, innovation is driven by powerful collaborations and cutting-edge hardware. AMD is proud to announce Day 0 support for Meta’s latest breakthrough — the Llama 4 Maverick and Scout models on our AMD Instinct™ MI300X and MI325X GPU accelerators.

A Collaborative Milestone

AMD’s deep-rooted partnerships have always been about pushing the limits of what’s possible. By working closely with Meta, vLLM, and Hugging Face, we’ve ensured that Llama 4 can run seamlessly on our GPUs from Day 0 with Pytorch and vLLM. This unified effort not only accelerates innovation but also empowers developers to harness the full potential of open-source AI with optimal performance and efficiency.

Inside Llama 4

Llama 4 represents a leap forward in AI technology. This model harnesses a mixture-of-experts (MoE) architecture, combining 17 billion activated parameters with up to a staggering 400 billion parameters. Designed for multimodal text and image processing, it supports extended context lengths—enabling applications that require deep, nuanced understanding of vast amounts of data.

The AMD Advantage

The AMD Instinct MI300X and MI325X GPU accelerators are engineered to meet the challenges of next-generation AI models. Key highlights include:

• Both MI300X and MI325X are capable of running the massive 400B-parameter Llama 4 Maverick in BF16 datatype on a single node, streamlining deployment and reducing infrastructure complexity.
• With their large HBM capacities, these GPUs effortlessly support Llama4’s extended context lengths, ensuring smooth performance even under the most demanding scenarios.
• With optimized Triton and AITER kernels, both MI300X and MI325X can achieve best-in-class performance and TCO for Llama 4 in production deployment.

Looking Ahead

This milestone is not just a technical achievement—it’s a commitment to driving AI innovation forward. By leveraging the strengths of AMD’s GPU accelerators and the pioneering work behind Llama 4, developers and enterprises are now empowered to explore new frontiers in AI applications.

Stay tuned for more updates and in-depth technical insights on how AMD and our partners support and optimize the exciting Llama 4.