Supported features and limitations

2026-09-02

2 min read time

Applies to Linux

The tables list MONAI capabilities supported in the AMD ROCm 26.08 release of amd-monai 1.6.0.

Supported features

Feature category	Feature	Notes
Inference	Sliding-window inference through SlidingWindowInferer	ROCm-optimized dynamic graph stabilization prevents graph recompilation across windows when torch.compile is active.
Inference	Patch-based and dense inference	Supported with GPU acceleration.
Network architecture	SwinUNETR (3D)	ROCm-optimized fused scaled dot-product attention auto-enabled in WindowAttention when HIP is detected.
Network architecture	DynUNet	AMD extension: use_gemm_transpose parameter and enable_gemm_transpose() method for GEMM-based transposed convolution on MI300X.
Network architecture	VISTA3D, SegResNet, UNETR, BasicUNet, and others	Supported without ROCm-specific modifications.
Transforms	GPU-accelerated transforms	Supported for spatial, intensity, and elastic transforms through the PyTorch HIP backend.
Data loading	NIfTI, DICOM, MHA, MHD, PNG, and JPEG	CPU-based I/O with GPU transfer through DataLoader.
Data loading	Whole-slide image reading through WSIReader	GPU-accelerated with amd-hipcim backend when backend="cuCIM".
GPU acceleration	Mixed precision (BF16 and FP16)	BF16 is preferred on MI300X and MI355X through PyTorch AMP (torch.amp.autocast).
GPU acceleration	torch.compile graph optimization	Supported. First-call compilation latency is expected.
Model Zoo	MONAI Bundle format	Supported. AMD overlay mechanism adds ROCm optimizations at runtime without modifying upstream bundles.
Model Zoo	Five validated bundles	VISTA3D, SwinUNETR BTCV, Whole Body CT, Spleen DeepEdit, and Pancreas DiNTS.
Metrics and losses	DiceLoss, DiceCELoss, FocalLoss, Hausdorff distance, MeanIoU	Supported.
Interoperability	NumPy, ITK, SimpleITK array conversions	Supported through the CPU bridge.
Interoperability	amd-cupy	Supported with amd-cupy 14.1.1 or later.
Foundation model	EXAONEPath 2.0	Computational pathology foundation model validated on AMD hardware. ViT-based WSI patch inference through Hugging Face.

Limitations

Limitation	Details
GPU direct storage through KvikIO or cuFile	Not supported on ROCm. Standard CPU-mediated I/O is used instead.
rocTX and NVTX profiling markers	MONAI NVTX-based profiling annotations are not functional on ROCm. Use rocprof or Omniperf directly.
CuPy version	Requires amd-cupy 14.1.1 or later. Standard NVIDIA CuPy packages are not compatible on ROCm.
hipCIM version	WSI support requires amd-hipcim 26.06.00 or later.
torch.compile first-call latency	Expect 30 to 120 seconds of compilation on the first call when torch.compile is enabled. Subsequent calls use the cached graph.

Network architecture support matrix

Architecture	ROCm support	AMD extensions	Tasks
SwinUNETR	Supported	Fused SDPA auto-enable	CT and MRI segmentation, whole-body
DynUNet	Supported	GEMM-based transpose conv	Segmentation, nnU-Net backbone
VISTA3D	Supported	BF16 and torch.compile overlay	Universal volumetric segmentation
SegResNet	Supported	None	Brain tumor segmentation (BraTS)
UNETR	Supported	None	Transformer-based segmentation
BasicUNet	Supported	None	General encoder-decoder
DiNTS (NAS)	Supported	None	NAS-discovered segmentation
EXAONEPath 2.0	Supported through Hugging Face	None	Computational pathology, WSI