Add [rocm] extra, ROCm install docs, and HIP version diagnostics

Dockerfile.rocm

@@ -104,7 +104,7 @@ ENV BUILD_MONAI=1
 # than no image at all.
 ARG HIPCIM_INDEX_URL="https://pypi.amd.com/rocm-${ROCM_SERIES}.0/simple/"
 RUN if [ -n "${HIPCIM_INDEX_URL}" ]; then \
-      pip install --no-cache-dir --extra-index-url "${HIPCIM_INDEX_URL}" "amd-hipcim" \
+      pip install --no-cache-dir --extra-index-url "${HIPCIM_INDEX_URL}" "amd-hipcim>=26.6.0" \
       && python -c "from cucim import CuImage"; \
     else \
       echo "hipCIM not installed; whole-slide-image (cucim) backends are unavailable."; \

docs/source/installation.md

@@ -77,10 +77,10 @@ pip install monai
 MONAI supports the extras syntax such as `pip install 'monai[nibabel]'`. The options are
 
 ```text
-clearml, cucim, cupy, einops, fire, gdown, h5py, huggingface_hub, hyena, ignite, imagecodecs, itk, jsonschema, lmdb, lpips, matplotlib, metrics_reloaded, mlflow, nibabel, nni, nvimgcodec, onnx, openslide, optuna, pandas, pillow, polygraphy, psutil, pyamg, pybind11, pydicom, pynrrd, pynvml, pyyaml, requests, segment_anything, scipy, skimage, tensorboard, tensorboardX, tifffile, torchio, torchvision, tqdm, transformers, zarr
+clearml, cucim, cupy, einops, fire, gdown, h5py, huggingface_hub, hyena, ignite, imagecodecs, itk, jsonschema, lmdb, lpips, matplotlib, metrics_reloaded, mlflow, nibabel, nni, nvimgcodec, onnx, openslide, optuna, pandas, pillow, polygraphy, psutil, pyamg, pybind11, pydicom, pynrrd, pynvml, pyyaml, requests, rocm, segment_anything, scipy, skimage, tensorboard, tensorboardX, tifffile, torchio, torchvision, tqdm, transformers, zarr
 ```
 
-which correspond to the packages: `clearml`, `cucim` (`cucim-cu12` or `cucim-cu13`), `cupy-cuda13x`, `einops`, `fire`, `gdown`, `h5py`, `huggingface_hub`, `nvsubquadratic`, `omegaconf`, `pytorch-ignite`, `imagecodecs`, `itk`, `jsonschema`, `lmdb`, `lpips`, `matplotlib`, `MetricsReloaded`, `mlflow`, `nibabel`, `nni`, `filelock`, `nvidia-nvimgcodec-cu13`, `onnx`, `onnxruntime`, `onnx_graphsurgeon`, `onnxscript`, `openslide-python`, `openslide-bin`, `optuna`, `pandas`, `pillow`, `polygraphy`, `psutil`, `pyamg`, `pybind11`, `pydicom`, `pynrrd`, `nvidia-ml-py`, `pyyaml`, `requests`, `segment_anything`, `scipy`, `scikit-image`, `tensorboard`, `tensorboardX`, `tifffile`, `torchio`, `torchvision`, `tqdm`, `transformers`, `zarr`.
+which correspond to the packages: `clearml`, `cucim` (`cucim-cu12` or `cucim-cu13`), `cupy-cuda13x`, `einops`, `fire`, `gdown`, `h5py`, `huggingface_hub`, `nvsubquadratic`, `omegaconf`, `pytorch-ignite`, `imagecodecs`, `itk`, `jsonschema`, `lmdb`, `lpips`, `matplotlib`, `MetricsReloaded`, `mlflow`, `nibabel`, `nni`, `filelock`, `nvidia-nvimgcodec-cu13`, `onnx`, `onnxruntime`, `onnx_graphsurgeon`, `onnxscript`, `openslide-python`, `openslide-bin`, `optuna`, `pandas`, `pillow`, `polygraphy`, `psutil`, `pyamg`, `pybind11`, `pydicom`, `pynrrd`, `nvidia-ml-py`, `pyyaml`, `requests`, `amd-hipcim`, `segment_anything`, `scipy`, `scikit-image`, `tensorboard`, `tensorboardX`, `tifffile`, `torchio`, `torchvision`, `tqdm`, `transformers`, `zarr`.
 
 Almost all of these can be installed together with the `all` option. For development on MONAI, this should be accompanied by `testing` which will install the testing static checking packages. Cupy and `nvimgcodec` are omitted from `all` since the choice between
 Cuda 12 and 13 versions of the libraries can't be resolved when installing and must be manually installed.
@@ -89,6 +89,30 @@ The `nvimgcodec` extra installs GPU-accelerated DICOM decoding for `NvImgCodecPy
 (`pip install 'monai[nvimgcodec]'`). It is Linux-only in the extra definition; CUDA 13 is the
 default. CUDA 12 users should install matching `cupy-cuda12x` and `nvidia-nvimgcodec-cu12` wheels.
 
+The `rocm` extra provides the AMD ROCm counterparts of the CUDA-only imaging dependencies:
+[`amd-hipcim`](https://rocm.docs.amd.com/projects/hipCIM/en/latest/) (hipCIM), which pulls
+`amd-cupy` in transitively. `amd-hipcim` ships the `cucim` Python namespace, so all existing MONAI
+code that imports from `cucim` (e.g. `WSIReader` with `backend="cucim"`) works on ROCm without any
+code changes. This extra covers MONAI's optional imaging dependencies only: a ROCm build of PyTorch
+and the matching ROCm runtime must already be installed, see
+[PyTorch on ROCm](https://rocm.docs.amd.com/projects/ai-ecosystem/en/latest/frameworks/pytorch/install.html).
+Both `amd-hipcim` and `amd-cupy` are published on AMD's package index rather than PyPI, so an extra
+index URL matching your ROCm series is required:
+
+```bash
+pip install --extra-index-url https://pypi.amd.com/rocm-10.0.0/simple/ 'monai[rocm]'
+```
+
+AMD currently publishes these wheels for CPython 3.12 on x86_64 Linux only; on other interpreters or
+architectures the install fails with `No matching distribution found for amd-hipcim`.
+
+Do not use `monai[all]` on ROCm: the `all` extra includes CUDA-only sub-extras (`cucim`, `cupy`,
+`nvimgcodec`, `pynvml`, `hyena`) that pull in `cupy-cuda13x`, `cuda-toolkit`, and a cascade of
+`nvidia-*` packages that will not work on ROCm. `Dockerfile.rocm` in the repo root shows how to
+install the full optional dependency set on ROCm while filtering those packages — follow it as the
+authoritative reference for the complete environment setup (library paths, compiler flags,
+`rocm-sdk init`, and the full container recipe).
+
 The `hyena` extra pulls in [`nvsubquadratic`](https://github.com/NVIDIA-BioNeMo/nvSubquadratic),
 required by `HyenaNDUNETR` / `HyenaMixer` / `HyenaTransformerBlock` (subquadratic
 O(N log N) alternatives to windowed self-attention). Install with
@@ -333,8 +357,10 @@ MONAI itself:
 ```bash
 git clone https://github.com/Project-MONAI/MONAI.git
 cd MONAI/
-python monai/config/print_dependencies.py \* > requirements.txt
+python monai/config/print_dependencies.py all testing > requirements.txt
 pip install -r requirements.txt
+# Note: `all testing` is used rather than `*` because `*` now also includes the `rocm` extra,
+# which requires AMD's package index and is not resolvable from PyPI.
 ```
 
 To install all optional dependencies with `conda` based on MONAI development environment settings (`environment-dev.yml`;

monai/config/check_env.py

@@ -123,7 +123,11 @@ def check_torch_cuda():
     """
     import torch
 
-    fprint("CUDA version:", torch.version.cuda)
+    # `torch.version.cuda` is None on a ROCm build, where the toolkit version is `torch.version.hip`.
+    if torch.version.hip:
+        fprint("HIP version:", torch.version.hip)
+    else:
+        fprint("CUDA version:", torch.version.cuda)
 
     try:
         dcount = torch.cuda.device_count()

monai/config/deviceconfig.py

@@ -214,7 +214,12 @@ def get_gpu_info() -> OrderedDict:
     _dict_append(output, "Has CUDA", lambda: bool(torch.cuda.is_available()))
 
     if output["Has CUDA"]:
-        _dict_append(output, "CUDA version", lambda: torch.version.cuda)
+        # On a ROCm build `torch.version.cuda` is None and `torch.version.hip` carries the toolkit
+        # version. Reporting "CUDA version: None" is misleading, so the key name switches with the build.
+        if torch.version.hip:
+            _dict_append(output, "HIP version", lambda: torch.version.hip)
+        else:
+            _dict_append(output, "CUDA version", lambda: torch.version.cuda)
     cudnn_ver = torch.backends.cudnn.version()
     _dict_append(output, "cuDNN enabled", lambda: bool(cudnn_ver))
     _dict_append(output, "NVIDIA_TF32_OVERRIDE", os.environ.get("NVIDIA_TF32_OVERRIDE"))

pyproject.toml

@@ -155,6 +155,10 @@ pynrrd = ["pynrrd"]
 pynvml = ["nvidia-ml-py"]
 pyyaml = ["pyyaml"]
 requests = ["requests"]
+# hipCIM is the ROCm counterpart of `cucim`; it pulls `amd-cupy` in transitively. Omitted from `all`:
+# the real wheels live on AMD's package index, not PyPI, so an `--extra-index-url` is needed (see
+# installation.md). The floor pins the first AMD-index release, skipping the PyPI placeholder stub.
+rocm = ["amd-hipcim>=26.6.0; platform_system == 'Linux'"]
 # per #5882 this requirement must be on one line to get removed by sed in weekly-preview.yml
 segment_anything = ["segment_anything @ git+https://github.com/facebookresearch/segment-anything.git@6fdee8f2727f4506cfbbe553e23b895e27956588"]
 scipy = ["scipy>=1.12.0"]

tests/config/test_print_info.py

@@ -11,10 +11,13 @@
 
 from __future__ import annotations
 
+import sys
 import unittest
 from io import StringIO
+from unittest.mock import patch
 
 from monai.config import print_debug_info
+from monai.config.deviceconfig import get_gpu_info
 
 
 class TestPrintInfo(unittest.TestCase):
@@ -24,6 +27,38 @@ def test_print_info(self):
         print_debug_info(file=out)
         self.assertGreater(out.tell(), 0)
 
+    @patch("torch.backends.cudnn.version", lambda: 0)
+    @patch("torch.cuda.device_count", lambda: 0)
+    @patch("torch.cuda.is_available", lambda: True)
+    @patch("torch.version.hip", "7.0.0")
+    def test_hip_version(self):
+        """A ROCm build reports the HIP version, where `torch.version.cuda` would be None."""
+        output = get_gpu_info()
+        self.assertEqual(output["HIP version"], "7.0.0")
+        self.assertNotIn("CUDA version", output)
+
+    @patch("torch.backends.cudnn.version", lambda: 0)
+    @patch("torch.cuda.device_count", lambda: 0)
+    @patch("torch.cuda.is_available", lambda: True)
+    @patch("torch.version.hip", None)
+    @patch("torch.version.cuda", "12.4")
+    def test_cuda_version(self):
+        """A CUDA build is unaffected and still reports the CUDA version."""
+        output = get_gpu_info()
+        self.assertEqual(output["CUDA version"], "12.4")
+        self.assertNotIn("HIP version", output)
+
+    @patch("torch.version.hip", "7.0.0")
+    def test_check_env_hip_version(self):
+        """check_env.check_torch_cuda() prints 'HIP version:' on a ROCm build."""
+        import monai.config.check_env as ce
+
+        buf = StringIO()
+        with patch.object(sys, "stdout", buf):
+            ce.check_torch_cuda()
+        self.assertIn("HIP version:", buf.getvalue())
+        self.assertNotIn("CUDA version:", buf.getvalue())
+
 
 if __name__ == "__main__":
     unittest.main()