Refresh RAPIDS and RoCE examples

dholt · dholt · commit 9db02ec86c87 · 2026-06-02T17:16:30.000-06:00
diff --git a/config.example/helm/rapids-dask.yml b/config.example/helm/rapids-dask.yml
@@ -4,10 +4,8 @@
 # Specify the resources used for each worker as well as the number of workers.
 worker:
   image:
-    # repository: nvcr.io/nvidia/rapidsai/rapidsai
-    # repository: dask-rapids
-    repository: supertetelman/k8s-rapids-dask
-    tag: cuda9.2-runtime-ubuntu16.04
+    repository: nvcr.io/nvidia/rapidsai/notebooks
+    tag: 26.04-cuda12-py3.13
     env:
   replicas: 1
   resources: 
@@ -18,15 +16,15 @@ worker:
 
 scheduler:
   image:
-    repository: supertetelman/k8s-rapids-dask
-    tag: cuda9.2-runtime-ubuntu16.04
+    repository: nvcr.io/nvidia/rapidsai/notebooks
+    tag: 26.04-cuda12-py3.13
 
 # By default we should be doing all Dask works on workers using calls to distributed.Client()
 # If you would like to run/test your GPU code without using workers you may comment the resources section
 jupyter:
   image:
-    repository: supertetelman/k8s-rapids-dask
-    tag: cuda9.2-runtime-ubuntu16.04
+    repository: nvcr.io/nvidia/rapidsai/notebooks
+    tag: 26.04-cuda12-py3.13
   resources: 
     requests:
       nvidia.com/gpu: 0
diff --git a/docs/container/docker-rootless.md b/docs/container/docker-rootless.md
@@ -77,7 +77,7 @@ module load rootless-docker
 
 start_rootless_docker.sh # specify --quiet option to hide rootles docker messages
 
-docker run --gpus all -it --rm nvcr.io/nvidia/cuda:11.0-base-ubuntu18.04
+docker run --gpus all -it --rm nvcr.io/nvidia/cuda:13.0.2-base-ubuntu24.04
 
 root@445bf5cca686:/# echo NGPUS: $(nvidia-smi -L | wc -l)
 NGPUS: 1
diff --git a/docs/k8s-cluster/roce_backend.md b/docs/k8s-cluster/roce_backend.md
@@ -80,11 +80,15 @@ dev_id: 101c
 
 num_vf: 8
 
-5. Mellanox Ofed place and image name - mofed_site_place, mofed_file_name.
-
-mofed_site_place: "MLNX_OFED-4.6-1.0.1.1"
-
-mofed_file_name: "MLNX_OFED_LINUX-4.6-1.0.1.1-ubuntu18.04-x86_64.iso"
+5. NVIDIA OFED version, site place and image name - mofed_version, mofed_site_place, mofed_file_name.
+
+For new Kubernetes RDMA/RoCE deployments, prefer NVIDIA Network Operator coverage with DOCA-OFED driver management. This DeepOps role remains a legacy direct-host install path for environments that still need it, so verify the OFED package against your exact operating system and kernel before using it in production.
+
+mofed_version: "24.10-4.1.4.0"
+
+mofed_site_place: "MLNX_OFED-24.10-4.1.4.0"
+
+mofed_file_name: "MLNX_OFED_LINUX-24.10-4.1.4.0-ubuntu24.04-x86_64.iso"
 
 ## Dependencies
 
diff --git a/roles/roce_backend/README.md b/roles/roce_backend/README.md
@@ -71,13 +71,15 @@ dev_id: "101c"
 
 num_vf: 8
 
-5. Mellanox Ofed version, site place and image name - mofed_version, mofed_site_place, mofed_file_name.
-```
-#Mellanox OFED parameters
-mofed_version: "4.7-3.2.9.0"
-mofed_site_place: "MLNX_OFED-4.7-3.2.9.0"
-mofed_file_name: "MLNX_OFED_LINUX-4.7-3.2.9.0-ubuntu18.04-x86_64.iso"
-```
+5. NVIDIA OFED version, site place and image name - mofed_version, mofed_site_place, mofed_file_name.
+
+For new Kubernetes RDMA/RoCE deployments, prefer NVIDIA Network Operator coverage with DOCA-OFED driver management. This DeepOps role remains a legacy direct-host install path for environments that still need it, so verify the OFED package against your exact operating system and kernel before using it in production.
+```
+# NVIDIA OFED parameters
+mofed_version: "24.10-4.1.4.0"
+mofed_site_place: "MLNX_OFED-24.10-4.1.4.0"
+mofed_file_name: "MLNX_OFED_LINUX-24.10-4.1.4.0-ubuntu24.04-x86_64.iso"
+```
 
 
 Dependencies
diff --git a/roles/roce_backend/tasks/mofed-install.yaml b/roles/roce_backend/tasks/mofed-install.yaml
@@ -22,7 +22,7 @@
       shell: |
         cd /tmp
         rm -f /tmp/{{ mofed_file_name }}
-        wget http://content.mellanox.com/ofed/{{ mofed_site_place }}/{{ mofed_file_name }}
+        wget https://content.mellanox.com/ofed/{{ mofed_site_place }}/{{ mofed_file_name }}
         mkdir -p /mnt/iso
         mount -o loop /tmp/{{ mofed_file_name }} /mnt/iso
         /mnt/iso/mlnxofedinstall --all -q
diff --git a/roles/roce_backend/vars/main.yml b/roles/roce_backend/vars/main.yml
@@ -31,18 +31,12 @@ vendor: "15b3"
 dev_id: "101c"
 num_vf: 8
 
-#Mellanox OFED parameters
-mofed_version: "4.7-3.2.9.0"
-mofed_site_place: "MLNX_OFED-4.7-3.2.9.0"
-mofed_file_name: "MLNX_OFED_LINUX-4.7-3.2.9.0-ubuntu18.04-x86_64.iso"
-
-# before K8s 1.16
-multus_ds: "https://raw.githubusercontent.com/intel/multus-cni/master/images/multus-daemonset-pre-1.16.yml"
-sriov_dp_ds: "https://raw.githubusercontent.com/intel/sriov-network-device-plugin/master/deployments/k8s-v1.10-v1.15/sriovdp-daemonset.yaml"
-sriov_cni_ds: "https://raw.githubusercontent.com/intel/sriov-cni/master/images/k8s-v1.10-v1.15/sriov-cni-daemonset.yaml"
-
-# from K8s 1.16
-# multus_ds: "https://raw.githubusercontent.com/intel/multus-cni/master/images/multus-daemonset.yml"
-# sriov_dp_ds: "https://raw.githubusercontent.com/intel/sriov-network-device-plugin/master/deployments/k8s-v1.16/sriovdp-daemonset.yaml"
-# sriov_cni_ds: "https://raw.githubusercontent.com/intel/sriov-cni/master/images/k8s-v1.16/sriov-cni-daemonset.yaml"
+# NVIDIA OFED parameters. Prefer NVIDIA Network Operator with DOCA-OFED for
+# new Kubernetes RDMA/RoCE deployments; this role is a legacy direct-host path.
+mofed_version: "24.10-4.1.4.0"
+mofed_site_place: "MLNX_OFED-24.10-4.1.4.0"
+mofed_file_name: "MLNX_OFED_LINUX-24.10-4.1.4.0-ubuntu24.04-x86_64.iso"
 
+multus_ds: "https://raw.githubusercontent.com/k8snetworkplumbingwg/multus-cni/master/deployments/multus-daemonset.yml"
+sriov_dp_ds: "https://raw.githubusercontent.com/k8snetworkplumbingwg/sriov-network-device-plugin/master/deployments/sriovdp-daemonset.yaml"
+sriov_cni_ds: "https://raw.githubusercontent.com/k8snetworkplumbingwg/sriov-cni/master/images/sriov-cni-daemonset.yaml"
diff --git a/workloads/examples/k8s/dask-rapids/README.md b/workloads/examples/k8s/dask-rapids/README.md
@@ -34,6 +34,8 @@ DeepOps includes support for running a local registry, but configuration of that
 In my workflow below, I am pushing the image to [Docker Hub](https://hub.docker.com).
 If you haven't used Docker Hub before, the [quickstart documentation](https://docs.docker.com/docker-hub/) provides a good tutorial.
 
+The included Dockerfile extends the current NGC RAPIDS notebooks image, `nvcr.io/nvidia/rapidsai/notebooks:26.04-cuda12-py3.13`, which provides RAPIDS, Dask CUDA, and JupyterLab on Ubuntu 24.04. If you move to a newer RAPIDS release, update both the Dockerfile and Helm values together.
+
 ### Editing the deployment scripts
 
 We'll deploy the RAPIDS/Dask container in one step using the `deploy.sh` script.
@@ -140,9 +142,8 @@ Feel free to adjust the number of CPU cores or GPUs used and the parameters for
 
 ## Experimenting further
 
-The base container we used for this benchmark contains more examples using RAPIDS in the `cuml/` directory,
-as well as an end-to-end workflow example based on a Fannie Mae mortgage dataset in the `mortgage/` directory.
-Both directories can be accessed easily via JupyterLab.
+The base container used for this benchmark includes RAPIDS libraries, Dask CUDA, and JupyterLab.
+For more examples and current image tags, use the RAPIDS documentation and NGC container catalog.
 
 You can also experiment with the custom container by making changes to the `Dockerfile` used to create it,
 in `examples/k8s/dask-rapids/docker`.
diff --git a/workloads/examples/k8s/dask-rapids/docker/Dockerfile b/workloads/examples/k8s/dask-rapids/docker/Dockerfile
@@ -1,24 +1,13 @@
-# Base our new image on the CUDA 9.2 RAPIDS image from upstream
-FROM nvcr.io/nvidia/rapidsai/rapidsai:cuda9.2-runtime-ubuntu16.04
+# Base the example on the current RAPIDS notebooks image from NGC.
+FROM nvcr.io/nvidia/rapidsai/notebooks:26.04-cuda12-py3.13
+
+USER root
 
-# Fix font-manager package
 RUN apt-get update && \
-    apt-get install -y --fix-missing font-manager && \
+    apt-get install -y --no-install-recommends font-manager && \
     rm -rf /var/lib/apt/lists/*
 
-# The name of the Anaconda Python environment we'll use (from upstream)
-ENV CONDA_ENV rapids
-
-# Install additional Python packages into the environment
-# (If you want to install more packages, add them here!)
-RUN source activate $CONDA_ENV && \
-    conda install -y unzip python-graphviz && \
-    pip install ipyvolume dask-kubernetes matplotlib cupy-cuda92
+USER rapids
 
 # Copy the parallel sum notebook in
-COPY ParallelSum.ipynb /rapids/notebooks/ParallelSum.ipynb
-
-# Set up image to be run
-COPY prepare.sh /usr/bin/prepare.sh
-CMD ["jupyter", "lab", "--ip=0.0.0.0", "--allow-root", "--no-browser", "--NotebookApp.token='dask'"]
-ENTRYPOINT ["tini", "--", "/usr/bin/prepare.sh"]
+COPY --chown=rapids:conda ParallelSum.ipynb /home/rapids/notebooks/ParallelSum.ipynb
diff --git a/workloads/examples/k8s/dask-rapids/docker/prepare.sh b/workloads/examples/k8s/dask-rapids/docker/prepare.sh
@@ -27,11 +27,9 @@ if [ "$EXTRA_PIP_PACKAGES" ]; then
 fi
 
 # Activate the specified or default conda environment
-#conda_env="rapids"
-#if [ "$CONDA_ENV" ]; then
-#    conda_env="$CONDA_ENV"
-#fi
-source activate $CONDA_ENV
+CONDA_ENV="${CONDA_ENV:-base}"
+source /opt/conda/etc/profile.d/conda.sh
+conda activate "$CONDA_ENV"
 
 # Run pre-commands
 if [ "$PRE_RUN_HOOK" ]; then
