MONAI Model Zoo

MONAI Model Zoo#

2025-09-26

3 min read time

Applies to Linux

The MONAI Model Zoo is a hub for researchers and data scientists to share, discover, and deploy the latest models from across the biomedical imaging community. By utilizing the standardized MONAI Bundle format, you can easily get started on building workflows or integrating new models into your projects.

MONAI for AMD ROCm provides seamless compatibility with the vast majority of models in the Model Zoo, helping both researchers and clinicians to accelerate state-of-the-art AI pipelines directly on AMD Instinct GPUs. Segmentation, detection, and classification models, including 2D and 3D workflows, run out of the box with minimal setup.

EXAONEPath model (hf_exaonepath-crc-msi-predictor) on ROCm#

EXAONEPath 2.0 is a foundation model designed to deliver highly efficient, directly supervised patch-level representation learning for whole-slide images (WSIs). Except for a few other model zoo entries exclusively designed for NVIDIA, advanced models for computational pathology, such as EXAONEPath 2.0, are now supported on AMD hardware. Unlike typical patch-based self-supervised learning (SSL), EXAONE Path 2.0 leverages end-to-end slide-level supervision for powerful biomarker and molecular characteristic prediction with improved data efficiency.

Setting up EXAONEPath 2.0 on ROCm#

To run EXAONEPath 2.0 on AMD platforms using MONAI for AMD ROCm, follow these steps:

  1. Install MONAI. For installation instructions, see Installing MONAI for AMD ROCm.

  2. Clone the EXAONEPath 2.0 repo:

    git clone https://huggingface.co/LGAI-EXAONE/EXAONE-Path-2.0

  3. Install EXAONEPath-specific dependencies:

    pip install pydantic==2.10.3                   \
        openslide-bin==4.0.0.6             \
        openslide-python==1.4.1            \
        rectangle_packer==2.0.2            \
        opencv-python-headless==4.11.0.86  \
        huggingface_hub

  4. Go to the EXAONEPath repo directory:

    cd EXAONE-Path-2.0

  5. Run inference using EXAONEPath:

    (Replace YOUR_HUGGING_FACE_ACCESS_TOKEN with your actual token)

from exaonepath import EXAONEPathV20
hf_token = "YOUR_HUGGING_FACE_ACCESS_TOKEN"
model = EXAONEPathV20.from_pretrained("LGAI-EXAONE/EXAONE-Path-2.0", use_auth_token=hf_token)

svs_path = "samples/sample.svs"
patch_features = model(svs_path)[0]
print(patch_features)

Key takeaways#

  • Most MONAI Model Zoo models run out of the box on ROCm, fully utilizing the AMD Instinct GPUs.

  • EXAONEPath 2.0, which was previously exclusive to NVIDIA, is now supported on AMD platforms using MONAI for AMD ROCm. The setup instructions are provided in Setting up EXAONEPath 2.0 on ROCm.

This reinforces MONAI for AMD ROCm as a truly open, high-performance AI platform, that removes vendor lock-in and unleashes broader access to foundational pathology models.