# AMD releases xDiT configurations for scaling diffusion inference on RDNA 4 GPUs

> **Key takeaway:** AMD provides optimized Docker containers using the open-source xDiT engine to scale large diffusion models across multiple RDNA 4 GPUs via ROCm.

**Published:** 2026-11-02T22:03:26.979903+00:00 
**Source:** AMD ROCm Official Blog 
**Category:** ai-local-edge 
**Canonical URL:** https://fosswire.org/news/amd-releases-xdit-configurations-for-scaling-diffusion-infer.html 

## Summary
AMD has published technical guidance and optimized software configurations for running large-scale diffusion transformer (DiT) workloads on RDNA 4 hardware. By utilizing the open-source xDiT inference engine, users can distribute heavy image and video generation models, such as FLUX.2 and Stable Diffusion 3.5, across one, two, or four AMD Radeon AI PRO R9700S GPUs.

The release focuses on overcoming memory limitations and PCIe communication bottlenecks inherent in multi-GPU setups. AMD provides a specialized Docker container, amdsiloai/pytorch-xdit-custom, which includes tuned configurations for RDNA 4 to enable efficient distributed inference through techniques like FP8 execution and parallel VAE decoding.

## Technical implementation of xDiT on RDNA 4
- **Distributed Inference Engine:** Uses the open-source xDiT engine to manage model weights and communication across multiple GPUs, supporting scaling via the ULYSSES_DEGREE parameter.
- **Memory and Bandwidth Optimization:** Employs memory-efficient replicated loading to prevent host-memory spikes and utilizes FP8 quantization for both GEMM and communications to reduce data transfer overhead.
- **Pipeline Acceleration:** Implements native FP8 execution and the aiter_flydsl_fp8 attention backend to accelerate the transformer stages of the diffusion process.
- **VAE Scaling:** Supports parallel VAE decoding and tiling (with configurable height, width, and overlap) to mitigate memory bottlenecks during the final pixel decoding stage.
- **Software Stack:** Requires ROCm 10.0 and the amdgpu-dkms driver, packaged within a containerized environment for consistent deployment on Linux.

## Why it matters
This development provides a concrete path for developers to deploy high-end generative AI workloads on RDNA 4 hardware without requiring massive single-GPU memory capacities. By leveraging open-source tools like xDiT and Docker, it reduces vendor lock-in and allows for more flexible, scalable AI inference nodes in Linux-based edge or local environments.