ci(uat): remove AWS GB200 nightly UAT lane

GB200 hardware is being retired. Remove the aws-gb200 reservation row,
its GB200 EKS cluster/test configs, the dedicated cross-account janitor
job, and the broker/coverage test expectations that pinned it. The
multi-arch/arm64 build machinery in uat-aws.yaml is left fully intact for
a successor ARM instance type.

Signed-off-by: Nathan Hensley <nhensley@nvidia.com>

Author: njhensley

.github/workflows/uat-janitor.yaml

@@ -163,44 +163,6 @@ jobs:
           GH_TOKEN: ${{ github.token }}
         run: ./.github/scripts/uat-janitor.sh aws
 
-  # GB200 lives in a SEPARATE AWS account — the cross-account Capacity Block in
-  # tenancy 897722687756 (tests/uat/aws/cluster-config-gb200.yaml), not the
-  # standard 615299774277 the `aws` job above covers — so it needs its own role
-  # assumption + config, or aws-gb200 orphans are never discovered or reaped.
-  # PREREQUISITE: infra/uat-aws-account applied in 897722687756 (GitHub OIDC
-  # provider + github-actions-role-aicr whose trust permits this workflow's ref).
-  # Until then the auth step fails harmlessly and the first dry-run surfaces the gap.
-  aws-gb200:
-    if: >-
-      github.repository == 'nvidia/aicr' &&
-      github.ref == 'refs/heads/main' &&
-      (github.event_name == 'schedule' || inputs.cloud == 'all' || inputs.cloud == 'aws')
-    runs-on: ubuntu-latest
-    timeout-minutes: 60
-    permissions:
-      contents: read
-      id-token: write
-      actions: read
-    env:
-      AWS_REGION: "us-east-1"
-      JANITOR_CONFIG: "tests/uat/aws/cluster-config-gb200.yaml"
-      JANITOR_ACTUATOR_IMAGE: "ghcr.io/mchmarny/cluster/eks@sha256:8bc33d14e6e5659d242aa2cdbcd69fca389cfcdb9e94f7a7a1b82fd33176befa"
-    steps:
-      - name: Checkout
-        uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
-        with:
-          persist-credentials: false
-      - name: Authenticate to AWS (GB200 account)
-        uses: aws-actions/configure-aws-credentials@e6de054238d6b7531b4efff3b6587d9aade6a06c # v6.2.3
-        with:
-          role-to-assume: "arn:aws:iam::897722687756:role/github-actions-role-aicr"
-          aws-region: ${{ env.AWS_REGION }}
-          role-session-name: GitHubActions-UAT-Janitor
-      - name: Reconcile GB200 orphans
-        env:
-          GH_TOKEN: ${{ github.token }}
-        run: ./.github/scripts/uat-janitor.sh aws
-
   azure:
     if: >-
       github.repository == 'nvidia/aicr' &&

docs/design/017-uat-cluster-name-convention.md

@@ -105,7 +105,6 @@ reservation name is too long and too cloud-scoped to be. Launch values:
 | `aws-h100`  | `ah1` |
 | `gcp-h100`  | `gh1` |
 | `azure-h100`| `zh1` |
-| `aws-gb200` | `ag2` |
 | `kind-h100` | `kh1` |
 
 The nvkind lane (`kind-h100`) gets a slug for uniformity, though it provisions

infra/uat/reservations.yaml

@@ -147,51 +147,6 @@ reservations:
     nightly-intent-min-versions:
       inference: v0.18.0
       training: v0.18.0
-  - name: aws-gb200
-    slug: ag2
-    cloud: aws
-    # CROSS-ACCOUNT: the GB200 Capacity Block lives in a SEPARATE tenancy
-    # (dgx-cloud-runai-prod-22, account 897722687756), not the standard UAT
-    # account. The account is NOT a field here: the AWS pipeline derives the
-    # tenancy it assumes the OIDC role into from the target cluster-config's
-    # .deployment.tenancy (cluster-config-gb200.yaml sets it to 897722687756),
-    # the single source of truth. Prereq: apply infra/uat-aws-account in
-    # 897722687756 (GitHub OIDC provider + github-actions-role-aicr).
-    # GB200 reservations are Capacity Blocks (marketType: capacity-block in
-    # cluster-config-gb200.yaml), not standing on-demand CRs like the h100 rows,
-    # so a run must fall inside the block's active window or actuator apply
-    # fails to acquire capacity. This particular block (cr-0e2f3833a602809a6,
-    # CapacityBlockId cbr-0a6a5543dd5c74727) is booked long: State=active,
-    # StartDate 2025-11-15, EndDate 2029-08-22 (TotalInstanceCount 18). A 04:00
-    # UTC nightly cron therefore always falls inside the window today, which is
-    # why nightly enrollment below is safe without broker window-awareness. If
-    # the block is ever re-booked to a shorter window, add a scheduling gate to
-    # the broker (or revert to `nightly-intents: []`) before it lapses.
-    reservation-id: cr-0e2f3833a602809a6
-    accelerator: gb200
-    # 4 Blackwell GPUs per p6e-gb200.36xlarge node; the cluster-config runs 2
-    # nodes (activates the multi-node NCCL East-West fabric check).
-    gpu-count: 4
-    cluster-config-path: tests/uat/aws/cluster-config-gb200.yaml
-    test-config-dir: tests/uat/aws/tests
-    # Enrolled in the nightly batch with both intents, matching the aws-h100
-    # row. Training runs the Kubeflow TrainJob CUJ; inference stands up and
-    # validates the served inference stack (the vLLM serve run itself stays
-    # disabled repo-wide, #1644). Each (version × intent) is a full
-    # provision→CUJ→teardown cell dispatched through the shared per-reservation
-    # lease, so the two intents serialize with no contention. Safe to enroll
-    # because the Capacity Block is booked active through 2029 (see the block
-    # note above) and the GB200 EKS UAT config first shipped in v0.18.0.
-    nightly-intents: [training, inference]
-    # Gate release cells to v0.18.0: cluster-config-gb200.yaml and the GB200 EKS
-    # UAT lane first shipped in v0.18.0 (absent in v0.17.0 and earlier), so a
-    # release cell below that tag has no GB200 config to run and would fail
-    # closed. `main` always runs both intents (built from source); older
-    # releases are skipped until v0.18.0 is the oldest cell in the window. Bump
-    # if the actual first-supported tag differs.
-    nightly-intent-min-versions:
-      training: v0.18.0
-      inference: v0.18.0
   - name: kind-h100
     slug: kh1
     cloud: kind

pkg/uatbroker/registry_test.go

@@ -809,15 +809,14 @@ func TestDaytimeAssignmentsNone(t *testing.T) {
 }
 
 // TestCommittedRegistryValid guards the actual checked-in registry: it must
-// parse, validate, and carry the launch reservations plus the aws-gb200
-// bring-up row. A bad data edit fails here before it can break the broker
-// workflows.
+// parse, validate, and carry the launch reservations. A bad data edit fails
+// here before it can break the broker workflows.
 func TestCommittedRegistryValid(t *testing.T) {
 	reg, err := LoadRegistryFile(filepath.Join("..", "..", "infra", "uat", "reservations.yaml"))
 	if err != nil {
 		t.Fatalf("committed reservations.yaml invalid: %v", err)
 	}
-	want := map[string]string{"aws-h100": CloudAWS, "gcp-h100": CloudGCP, "azure-h100": CloudAzure, "aws-gb200": CloudAWS, "kind-h100": CloudKind}
+	want := map[string]string{"aws-h100": CloudAWS, "gcp-h100": CloudGCP, "azure-h100": CloudAzure, "kind-h100": CloudKind}
 	for name, cloud := range want {
 		res, err := reg.Lookup(name)
 		if err != nil {
@@ -837,7 +836,6 @@ func TestCommittedRegistryValid(t *testing.T) {
 		"aws-h100":   "ah1",
 		"gcp-h100":   "gh1",
 		"azure-h100": "zh1",
-		"aws-gb200":  "ag2",
 		"kind-h100":  "kh1",
 	}
 	for name, slug := range wantSlug {
@@ -890,13 +888,6 @@ func TestCommittedRegistryValid(t *testing.T) {
 		// acceptance run (29125390442); inference joined after a green
 		// manual intent=inference dispatch.
 		"azure-h100": {IntentTraining, IntentInference},
-		// aws-gb200 enrolled with both intents once the GB200 EKS UAT lane
-		// shipped (v0.18.0) and the Capacity Block was confirmed booked active
-		// through 2029 (cr-0e2f3833a602809a6). Release cells are gated to
-		// v0.18.0 via nightly-intent-min-versions (the config first shipped
-		// there), so only `main` runs GB200 nightly until v0.18.0 ages into the
-		// window.
-		"aws-gb200": {IntentTraining, IntentInference},
 		// kind-h100 (nvkind real-silicon lane, DC5 #1278) enrolled with both
 		// intents after green manual H100 acceptance runs (training 29954092703,
 		// inference 29965464868). Release cells are gated to v0.18.0 via
@@ -920,12 +911,10 @@ func TestCommittedRegistryValid(t *testing.T) {
 	// a future edit cannot silently drop a gate — which would let a release cell
 	// run a pre-fix aicr and emit failing/unusable evidence. azure-h100 gates its
 	// AKS perf + driver-only fixes; kind-h100 gates the uat-kind lane + the
-	// os-agnostic coordinate fix (#1851); aws-gb200 gates the GB200 EKS UAT
-	// config, which first shipped in v0.18.0. All landed post-v0.17.0.
+	// os-agnostic coordinate fix (#1851). All landed post-v0.17.0.
 	wantMinVersions := map[string]map[string]string{
 		"azure-h100": {IntentTraining: "v0.18.0", IntentInference: "v0.18.0"},
 		"kind-h100":  {IntentTraining: "v0.18.0", IntentInference: "v0.18.0"},
-		"aws-gb200":  {IntentTraining: "v0.18.0", IntentInference: "v0.18.0"},
 	}
 	for name, want := range wantMinVersions {
 		res, lookupErr := reg.Lookup(name)