Validated bundles

Validated bundles#

2026-09-01

3 min read time

Applies to Linux

AMD has validated five inference-optimized bundles for AMD MI355X, MI325X, and MI300X GPUs with ROCm 10.0.0, Ubuntu 24.04, Python 3.12, and MONAI 1.6.0.

AMD provides a ROCm overlay configuration, inference_rocm.json or inference_rocm.yaml, for each bundle. See ROCm overlays for merge behavior and shared keys. The overlay applies these settings on top of the unmodified upstream inference configuration.

Channels-last 3D memory format, torch.channels_last_3d.
BF16 automatic mixed precision, amp_kwargs={'dtype': torch.bfloat16}.
torch.compile() graph compilation.
Device-aware checkpoint loading, map_location=@device, where the upstream bundle doesn’t already place weights on-device.

Each bundle targets a volumetric CT segmentation task.

Bundle	Task	Architecture	Modality
`vista3d`	Multi-organ segmentation, 130+ structures, with zero-shot label prompts	VISTA-3D	CT
`swin_unetr_btcv_segmentation`	13-organ abdominal CT segmentation, BTCV Challenge	Swin UNETR	CT
`wholeBody_ct_segmentation`	104-structure whole-body CT segmentation, TotalSegmentator	SegResNet	CT
`spleen_deepedit_annotation`	Interactive spleen segmentation with positive and negative click guidance	DynUNet	CT
`pancreas_ct_dints_segmentation`	Pancreas and pancreatic tumor segmentation, NAS-discovered architecture	DiNTS	CT

vista3d#

vista3d is a class-prompted volumetric CT segmentation bundle.

Full name	VISTA-3D: Versatile Imaging SegmenTation and Annotation
Upstream version	0.5.11
Task	Multi-organ segmentation in CT with zero-shot label prompts, 130+ structures
Architecture	VISTA-3D, a custom encoder-decoder with class-prompt conditioning
Modality	CT
Patch size	128 x 128 x 128
Training data	Task09_Spleen, Medical Segmentation Decathlon, plus internal multi-organ data
ROCm overlay	`models/vista3d/configs/inference_rocm.json`
Configuration format	JSON
Input key	`input_dict`, with `image` and `label_prompt`

swin_unetr_btcv_segmentation#

swin_unetr_btcv_segmentation is a 13-organ abdominal CT segmentation bundle.

Full name	Swin UNETR BTCV Multi-organ Segmentation
Upstream version	0.5.8
Task	13-organ abdominal CT segmentation, Beyond the Cranial Vault Challenge
Architecture	Swin UNETR, a Swin Transformer encoder with a UNet decoder
Modality	CT
Patch size	96 x 96 x 96
Output channels	14, background plus 13 organs
Training data	BTCV Challenge dataset, Synapse
ROCm overlay	`models/swin_unetr_btcv_segmentation/configs/inference_rocm.json`
Configuration format	JSON
Input key	`dataset_dir`

wholeBody_ct_segmentation#

wholeBody_ct_segmentation is a 104-structure whole-body CT segmentation bundle.

Full name	Whole Body CT Segmentation
Upstream version	0.2.7
Task	104-structure whole-body CT segmentation covering major organs, bones, muscles, and vasculature
Architecture	SegResNet
Modality	CT
Patch size	96 x 96 x 96
Output channels	105, background plus 104 structures
Training data	TotalSegmentator dataset
ROCm overlay	`models/wholeBody_ct_segmentation/configs/inference_rocm.json`
Configuration format	JSON
Input key	`dataset_dir`

spleen_deepedit_annotation#

spleen_deepedit_annotation is an interactive spleen segmentation bundle.

Full name	Spleen DeepEdit Interactive Segmentation
Upstream version	0.5.8
Task	Interactive spleen segmentation with positive and negative click guidance, DeepEdit
Architecture	DynUNet
Modality	CT
Patch size	128 x 128 x 128
Output channels	2, background plus spleen
Training data	Task09_Spleen, Medical Segmentation Decathlon
ROCm overlay	`models/spleen_deepedit_annotation/configs/inference_rocm.json`
Configuration format	JSON
Input key	`dataset_dir`

pancreas_ct_dints_segmentation#

pancreas_ct_dints_segmentation is a NAS-discovered pancreas and tumor segmentation bundle.

Full name	Pancreas and Tumor DiNTS Segmentation
Upstream version	0.5.2
Task	Pancreas and pancreatic tumor segmentation with a NAS-discovered architecture
Architecture	DiNTS, Differentiable Neural Architecture Search
Modality	CT
Patch size	96 x 96 x 96
Output channels	3, background plus pancreas plus tumor
Training data	Task07_Pancreas, Medical Segmentation Decathlon
ROCm overlay	`models/pancreas_ct_dints_segmentation/configs/inference_rocm.yaml`
Configuration format	YAML
Input key	`dataset_dir`

CI validation#

AMD provides unit tests for each bundle in ci/unit_tests/test_bundle_name.py. The tests run the ConfigWorkflow inference pipeline with a synthetic input. The ROCm overlay applies on a ROCm build when torch.version.hip is not None. Bundles that guard weight loading with @load_pretrain run without pretrained weights.