MONAILabel on ROCm documentation

MONAILabel on ROCm documentation#

2026-09-30

2 min read time

Applies to Linux

MONAILabel is an open-source framework that provides a server for AI-assisted medical image annotation. It connects 3D Slicer, OHIF, and QuPath clients to MONAI-powered deep learning models, providing a means for radiologists and researchers to build labeled datasets interactively through active-learning workflows.

MONAILabel on ROCm is the AMD ROCm-enabled release of MONAILabel validated on AMD GPUs. It brings interactive segmentation, auto-segmentation, and model fine-tuning to AMD GPUs through PyTorch-ROCm.

The AMD ROCm port runs MONAILabel on AMD GPUs without requiring changes to application code. All GPU compute runs through PyTorch-ROCm. torch.cuda.* APIs map to HIP on AMD hardware. There are no new GPU kernels and no change to the public Python API.

ROCm ships a PyTorch build where HIP presents itself as CUDA through the torch.cuda namespace. MONAILabel uses torch.cuda.is_available(), tensor.cuda(), and model.to("cuda"). Those calls work on AMD hardware without application changes. AMD systems query VRAM with rocm-smi instead of nvidia-smi.

The port is validated on AMD Instinct MI300X, MI325X, and MI355X. It is compatible with amd-monai 1.6.0, Python 3.12, and PyTorch for ROCm 10.0.0.

The AMD ROCm port makes these changes to upstream MONAILabel.

File	Change
`monailabel/utils/others/generic.py`	`gpu_memory_map()` adds a `rocm-smi` fallback for AMD GPU VRAM queries when `nvidia-smi` is absent. The function returns free VRAM in MB per device. MONAILabel automatic image-size tuning uses that value.
`monailabel/endpoints/logs.py`	The `/gpu` admin endpoint dispatches to `nvidia-smi` on NVIDIA systems or `rocm-smi` on AMD systems, whichever is present.
`Dockerfile`	Documents an AMD ROCm 10.0.0 base image and the AMD PyPI wheel index as an alternative to the default CUDA-based installation.

The code is open and hosted at AMD-Ecosystem/MONAILabel.

Install

Installing MONAILabel on ROCm

To contribute to MONAILabel on ROCm, see Contributing to MONAILabel.

Licensing information is on the Licensing page.