[WIP] ci(gcp-nvkind): run on Kubernetes 1.37 with kind v0.33.0 and nvkind main

hack/ci/gcp-nvkind/README.md

@@ -46,7 +46,12 @@ the full env. `TESTINFRA_DIR` points `e2e-test.sh` at the shared lib under
 - `registry.k8s.io/dra-driver-nvidia/dra-driver-nvidia-gpu:v0.4.0-dev` is
   not published; `install-dra-driver.sh` builds from source and
   `kind load docker-image`s the result.
-- `kindest/node:v1.34.3` (DRA GA). `v1.34.1` is not published on Docker Hub.
+- `kindest/node:v1.37.0`. Device health status (`ResourceHealthStatus`) is
+  beta and on by default since 1.36, so no extra kubelet feature gate is
+  needed to exercise it.
+- kind `v0.33.0`, and nvkind rebuilt against it: Kubernetes 1.37 requires the
+  kubeadm v1beta4 config that kind v0.32.0+ generates. See
+  `lib/setup-nvkind-node.sh`.
 - GPU Operator `v26.3.1` in minimal mode: `driver`, `toolkit`, and
   `devicePlugin` disabled; `cdi.enabled=true`, `nfd.enabled=true`.
 - DLVM ships the NVIDIA driver and toolkit but not Docker/Go/kind/helm/kubectl;

hack/ci/gcp-nvkind/e2e-test.sh

@@ -50,7 +50,7 @@ source "${LIB_DIR}/gce.sh"
 : "${BOSKOS_HOST:=http://boskos.test-pods.svc.cluster.local}"
 : "${BOSKOS_RESOURCE_TYPE:=gpu-project}"
 : "${GCE_ZONE:=us-central1-b}"
-: "${K8S_VERSION:=v1.34.3}"
+: "${K8S_VERSION:=v1.37.0}"
 : "${GPU_OPERATOR_VERSION:=v26.3.1}"
 
 mkdir -p "${ARTIFACTS}"
@@ -88,7 +88,9 @@ gce::wait_for_driver
 
 # 2. Ship repo to VM. Vendor is included (Dockerfile uses -mod=vendor).
 WORK=$(mktemp -d)
-tar --exclude='.git' --exclude='_output' --exclude='dist' --exclude='.claude' --exclude='site' \
+# COPYFILE_DISABLE and the ._* exclude keep macOS tar from adding AppleDouble
+# files, which Helm then fails to parse as CRDs on local runs from a Mac.
+COPYFILE_DISABLE=1 tar --exclude='._*' --exclude='.git' --exclude='_output' --exclude='dist' --exclude='.claude' --exclude='site' \
   -czf "${WORK}/dra-src.tgz" -C "${REPO_ROOT}" .
 gce::scp_to "${WORK}/dra-src.tgz" "/tmp/"
 gce::ssh 'rm -rf /tmp/dra-src && mkdir -p /tmp/dra-src && tar -xzf /tmp/dra-src.tgz -C /tmp/dra-src'

hack/ci/gcp-nvkind/lib/setup-nvkind-node.sh

@@ -105,11 +105,27 @@ if ! command -v helm >/dev/null 2>&1; then
   curl -fsSL https://raw.githubusercontent.com/helm/helm/main/scripts/get-helm-3 | bash
 fi
 
-# kind (pinned to the version nvkind vendors) and nvkind. nvkind has no
-# tags; pin a commit SHA. Bump with a reviewable diff after re-validating.
-: "${NVKIND_SHA:=9a3061c75e59ac7e4f29e001f1c5875d46a7cc54}"
-go install sigs.k8s.io/kind@v0.31.0
-go install "github.com/NVIDIA/nvkind/cmd/nvkind@${NVKIND_SHA}"
+# kind and nvkind. nvkind has no tags; pin a commit SHA. Bump with a
+# reviewable diff after re-validating.
+#
+# kind v0.32.0+ is required for Kubernetes 1.37, which dropped the kubeadm
+# v1beta3 config that older kind generates. nvkind still vendors kind v0.31.0
+# (NVIDIA/nvkind#75 bumps it), so nvkind is rebuilt against KIND_VERSION; the
+# pinned nvkind commit includes NVIDIA/nvkind#76, without which the containerd
+# in newer kindest/node images fails to restart after nvidia-ctk configures it.
+: "${KIND_VERSION:=v0.33.0}"
+: "${NVKIND_SHA:=c57050497cffee36f10c8b918d2727c4f26d886e}"
+go install "sigs.k8s.io/kind@${KIND_VERSION}"
+rm -rf /tmp/nvkind-src
+git clone -q https://github.com/NVIDIA/nvkind /tmp/nvkind-src
+git -C /tmp/nvkind-src checkout -q "${NVKIND_SHA}"
+(
+  cd /tmp/nvkind-src
+  GOFLAGS=-mod=mod go get "sigs.k8s.io/kind@${KIND_VERSION}"
+  GOFLAGS=-mod=mod go mod tidy
+  go mod vendor
+  go install ./cmd/nvkind
+)
 
 # `sg docker -c` avoids re-login after usermod -aG docker.
 if sg docker -c "kind get clusters" | grep -qx "${NVKIND_CLUSTER_NAME}"; then