ci(uat): remove AWS GB200 nightly UAT reservation entry

GB200 hardware is being retired. Drop the aws-gb200 row from the UAT
reservation registry so the nightly broker no longer dispatches it, and
update TestCommittedRegistryValid to match the reduced set (it now locks
the exact reservation set so a re-add fails closed).

Also scrub the retired Capacity Block CR ID from the GB200 cluster-config
(target: "" + note), since the reservation no longer exists.

Scoped intentionally: the GB200 cluster/test configs, the janitor job,
and the multi-arch build machinery in uat-aws.yaml are left in place for
a successor ARM instance type.

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

Author: njhensley

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 {
@@ -829,6 +828,17 @@ func TestCommittedRegistryValid(t *testing.T) {
 		}
 	}
 
+	// The registry must carry EXACTLY these reservations. The positive lookup
+	// loop above would still pass if the retired aws-gb200 row (or any other)
+	// were re-added, so lock the full set here — a re-add or a stray row fails
+	// closed rather than sliding in unnoticed.
+	gotNames := append([]string(nil), reg.Names()...)
+	slices.Sort(gotNames)
+	wantNames := []string{"aws-h100", "azure-h100", "gcp-h100", "kind-h100"}
+	if !slices.Equal(gotNames, wantNames) {
+		t.Errorf("committed registry reservations = %v, want exactly %v", gotNames, wantNames)
+	}
+
 	// The committed daytime-name discovery slugs (ADR-017). Locked here so a
 	// future edit cannot silently rename or collide a slug — the daytime
 	// cluster name (aicr-uat-day-<slug>-<slot>-<run_id>) and its guard/teardown
@@ -837,7 +847,6 @@ 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 +899,6 @@ 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 +922,10 @@ 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)

tests/uat/aws/cluster-config-gb200.yaml

@@ -146,7 +146,10 @@ compute:
             desired: 2
             reservation:
               preference: capacity-reservations-only
-              target: cr-0e2f3833a602809a6
+              # Capacity Block CR ID removed: the GB200 block is retired and this
+              # lane is no longer in the reservation registry. Re-fill with the
+              # successor reservation's id before re-enabling.
+              target: ""
               marketType: capacity-block    # REQUIRED — GB200 CRs are Capacity Blocks
           labels:
             nodeGroup: gpu-worker