test: remove environment-specific e2e defaults

e2e/cases/c11_qwen3_06b_mindspore.sh

@@ -10,17 +10,24 @@ set -euo pipefail
 HERE="$(cd "$(dirname "$0")" && pwd)"
 source "${HERE}/../lib.sh"
 
+require_env NPU_NAMESPACE "namespace for NPU e2e resources"
+require_env NPU_RESOURCE_NAME "extended resource name for one NPU, for example huawei.com/Ascend910B4"
 NS="${NPU_NAMESPACE}"
 JOB_NAME="c11-mindspore-cann-smoke-$(printf '%05x' $$)"
 IMAGE="${C11_IMAGE:-docker.io/alaudadockerhub/alauda-workbench-jupyter-mindspore-cann-py312-ubi9:v0.1.7}"
+IMAGE_PULL_SECRET="${C11_IMAGE_PULL_SECRET:-${E2E_IMAGE_PULL_SECRET:-}}"
+NPU_RESOURCE_VALUE="${NPU_RESOURCE_VALUE:-1}"
+NPU_MEMORY_RESOURCE_NAME="${NPU_MEMORY_RESOURCE_NAME:-}"
+NPU_MEMORY_RESOURCE_VALUE="${NPU_MEMORY_RESOURCE_VALUE:-8192}"
+NPU_RUNTIME_CLASS="${NPU_RUNTIME_CLASS:-}"
 
 cleanup() {
   npu_kc -n "${NS}" delete job "${JOB_NAME}" --ignore-not-found --wait=false || true
 }
 trap cleanup EXIT
 
 log "C11: submitting Job ${JOB_NAME} (image=${IMAGE})"
-cat <<YAML | retry_create npu_kc >/dev/null
+cat <<YAML | mirror_dockerhub "${NPU_DH_MIRROR}" | retry_create npu_kc >/dev/null
 apiVersion: batch/v1
 kind: Job
 metadata:
@@ -35,7 +42,8 @@ spec:
       labels: { e2e.alauda.io/case: c11 }
     spec:
       restartPolicy: Never
-      runtimeClassName: ascend
+$(yaml_scalar_field 6 runtimeClassName "${NPU_RUNTIME_CLASS}")
+$(yaml_image_pull_secrets 6 "${IMAGE_PULL_SECRET}")
       securityContext: { runAsNonRoot: true, runAsUser: 1001, runAsGroup: 0, fsGroup: 1000 }
       containers:
         - name: probe
@@ -51,7 +59,8 @@ spec:
             limits:
               cpu: "2"
               memory: 8Gi
-              huawei.com/Ascend910: "1"
+$(yaml_resource_limit 14 "${NPU_RESOURCE_NAME}" "${NPU_RESOURCE_VALUE}")
+$(yaml_resource_limit 14 "${NPU_MEMORY_RESOURCE_NAME}" "${NPU_MEMORY_RESOURCE_VALUE}")
           command: [bash, -lc]
           args:
             - |

e2e/cases/c12_kueue_preemption.sh

@@ -24,6 +24,7 @@ set -euo pipefail
 HERE="$(cd "$(dirname "$0")" && pwd)"
 source "${HERE}/../lib.sh"
 
+require_env GPU_NAMESPACE "namespace for GPU e2e resources"
 NS="${GPU_NAMESPACE}"
 RUN_ID="$(printf '%05x' $$)-$(date -u +%s)"
 
@@ -40,13 +41,16 @@ RUNTIME="c12-checkpoint-runtime-${RUN_ID}"
 PVC_NAME="c12-ckpt-${RUN_ID}"
 TRAIN_LABEL="e2e.alauda.io/c12-run=${RUN_ID}"
 
-# Same harbor image C5/C6 already use — guaranteed cached on the GPU node.
-IMAGE="${LF_IMAGE:-build-harbor.alauda.cn/mlops/llamafactory0.9-cu126-amd64:v0.1.0-build.20260603021903}"
+IMAGE="${LF_IMAGE:-docker.io/alaudadockerhub/llamafactory0.9-cu126-amd64:v0.1.0}"
+IMAGE_PULL_SECRET="${LF_IMAGE_PULL_SECRET:-${E2E_IMAGE_PULL_SECRET:-}}"
+RWX_STORAGE_CLASS="${C12_RWX_STORAGE_CLASS:-${E2E_RWX_STORAGE_CLASS:-}}"
+NODE_SELECTOR_KEY="${C12_NODE_SELECTOR_KEY:-${E2E_GPU_NODE_SELECTOR_KEY:-}}"
+NODE_SELECTOR_VALUE="${C12_NODE_SELECTOR_VALUE:-${E2E_GPU_NODE_SELECTOR_VALUE:-}}"
 
 # --- Preflight: Kueue must be installed; skip otherwise. -----------------
 if ! gpu_kc api-resources --api-group=kueue.x-k8s.io 2>/dev/null | grep -q clusterqueues; then
   log "C12: kueue.x-k8s.io API group not present — install Kueue (see preemptible-trainjobs-with-kueue.mdx) and re-run"
-  exit 77
+  exit "${E2E_SKIP_RC}"
 fi
 
 cleanup() {
@@ -63,7 +67,7 @@ cleanup() {
 trap cleanup EXIT
 
 # --- Step 1: cohort + queues + priorities. -------------------------------
-# Quotas are sized for one Tesla P100 + HAMI (gpucores 50%, 4 GiB GPU mem):
+# Quotas are sized for one fractional HAMI GPU slice:
 # both inference and training Workloads each ask for that slice, so the
 # inference reclaim path empties exactly one borrowed training slot.
 log "C12: applying cohort (CQ ${CQ_INF}, ${CQ_TRAIN}; flavor ${FLAVOR})"
@@ -142,7 +146,7 @@ apiVersion: v1
 kind: PersistentVolumeClaim
 metadata: { name: ${PVC_NAME}, namespace: ${NS} }
 spec:
-  storageClassName: cephfs
+$(yaml_storage_class 2 "${RWX_STORAGE_CLASS}")
   accessModes: ["ReadWriteMany"]
   resources: { requests: { storage: 1Gi } }
 YAML
@@ -153,7 +157,7 @@ YAML
 # elsewhere. HuggingFace Trainer auto-resumes from the newest
 # checkpoint-N/ in CKPT_DIR if we pass it to .train(resume_from_checkpoint=).
 log "C12: applying TrainingRuntime ${RUNTIME}"
-cat <<YAML | retry_apply gpu_kc
+cat <<YAML | mirror_dockerhub "${GPU_DH_MIRROR}" | retry_apply gpu_kc
 apiVersion: trainer.kubeflow.org/v1alpha1
 kind: TrainingRuntime
 metadata:
@@ -176,8 +180,8 @@ spec:
               template:
                 spec:
                   terminationGracePeriodSeconds: 60
-                  nodeSelector: { kubernetes.io/hostname: 192.168.138.15 }
-                  imagePullSecrets: [{ name: harbor-mlops-regcred }]
+$(yaml_node_selector 18 "${NODE_SELECTOR_KEY}" "${NODE_SELECTOR_VALUE}")
+$(yaml_image_pull_secrets 18 "${IMAGE_PULL_SECRET}")
                   securityContext: { runAsNonRoot: true, runAsUser: 65534, runAsGroup: 65534, fsGroup: 65534 }
                   volumes:
                     - { name: workspace, emptyDir: {} }
@@ -330,7 +334,7 @@ log "C12: checkpoint visible on PVC"
 
 # --- Step 5: high-priority preemptor. -----------------------------------
 log "C12: submitting high-priority preemptor Job"
-INF_JOB=$(cat <<YAML | retry_create gpu_kc -o jsonpath='{.metadata.name}'
+INF_JOB=$(cat <<YAML | mirror_dockerhub "${GPU_DH_MIRROR}" | retry_create gpu_kc -o jsonpath='{.metadata.name}'
 apiVersion: batch/v1
 kind: Job
 metadata:
@@ -350,8 +354,8 @@ spec:
         e2e.alauda.io/c12-run: "${RUN_ID}"
     spec:
       restartPolicy: Never
-      nodeSelector: { kubernetes.io/hostname: 192.168.138.15 }
-      imagePullSecrets: [{ name: harbor-mlops-regcred }]
+$(yaml_node_selector 6 "${NODE_SELECTOR_KEY}" "${NODE_SELECTOR_VALUE}")
+$(yaml_image_pull_secrets 6 "${IMAGE_PULL_SECRET}")
       securityContext: { runAsNonRoot: true, runAsUser: 65534, runAsGroup: 65534, fsGroup: 65534 }
       containers:
         - name: serve

e2e/cases/c1_smoke_gpu.sh

@@ -7,18 +7,21 @@ set -euo pipefail
 HERE="$(cd "$(dirname "$0")" && pwd)"
 source "${HERE}/../lib.sh"
 
+require_env GPU_NAMESPACE "namespace for GPU e2e resources"
 NS="${GPU_NAMESPACE}"
 ASSETS="${E2E_ROOT}/../docs/en/training_guides/assets/training-runtimes"
 
-log "C1: applying torch2.6-cu126-amd64 TrainingRuntime to ns/${NS} (rewriting docker.io → ${GPU_DH_MIRROR})"
-sed "s/namespace: kubeflow-admin-cpaas-io/namespace: ${NS}/" \
-  "${ASSETS}/torch2.6-cu126-amd64-trainingruntime.yaml" \
+if [ -n "${GPU_DH_MIRROR}" ]; then
+  log "C1: applying torch2.6-cu126-amd64 TrainingRuntime to ns/${NS} (Docker Hub mirror=${GPU_DH_MIRROR})"
+else
+  log "C1: applying torch2.6-cu126-amd64 TrainingRuntime to ns/${NS}"
+fi
+set_metadata_namespace "${NS}" < "${ASSETS}/torch2.6-cu126-amd64-trainingruntime.yaml" \
   | mirror_dockerhub "${GPU_DH_MIRROR}" \
   | retry_apply gpu_kc
 
 log "C1: submitting TrainJob from trainjob-smoke.yaml"
-TJ_NAME=$(sed -e "s/namespace: kubeflow-admin-cpaas-io/namespace: ${NS}/" \
-              "${ASSETS}/trainjob-smoke.yaml" \
+TJ_NAME=$(set_metadata_namespace "${NS}" < "${ASSETS}/trainjob-smoke.yaml" \
           | retry_create gpu_kc -o jsonpath='{.metadata.name}')
 log "C1: trainjob=${TJ_NAME}"
 

e2e/cases/c2_kubeflow_trainer_mnist.sh

@@ -11,16 +11,17 @@ set -euo pipefail
 HERE="$(cd "$(dirname "$0")" && pwd)"
 source "${HERE}/../lib.sh"
 
+require_env GPU_NAMESPACE "namespace for GPU e2e resources"
 NS="${GPU_NAMESPACE}"
 
 log "C2: applying ClusterTrainingRuntime torch-distributed (from kubeflow-trainer-quick-start.md)"
 # Pulled directly out of docs/en/training_guides/kubeflow-trainer-quick-start.md.
-# Image override: docker-mirrors.alauda.cn proxies docker.io but the lazy-cache
-# returns EOF on the torch-distributed blobs (only torch2.6-* are cached). The
-# alauda-local registry mirror has the image pre-pulled. Doc itself stays at
-# the public alaudadockerhub/ name; only the test overrides for network reasons.
-TORCH_DIST_IMAGE="${TORCH_DIST_IMAGE:-152-231-registry.alauda.cn:60070/mlops/torch-distributed:v2.9.1-aml2}"
-cat <<'YAML' | sed "s@alaudadockerhub/torch-distributed:v2.9.1-aml2@${TORCH_DIST_IMAGE}@g" | retry_apply gpu_kc
+# Override TORCH_DIST_IMAGE only when the public Docker Hub image is mirrored locally.
+TORCH_DIST_IMAGE="${TORCH_DIST_IMAGE:-alaudadockerhub/torch-distributed:v2.9.1-aml2}"
+cat <<'YAML' \
+  | sed "s@alaudadockerhub/torch-distributed:v2.9.1-aml2@${TORCH_DIST_IMAGE}@g" \
+  | mirror_dockerhub "${GPU_DH_MIRROR}" \
+  | retry_apply gpu_kc
 apiVersion: trainer.kubeflow.org/v1alpha1
 kind: ClusterTrainingRuntime
 metadata:

e2e/cases/c3_traininghub_sft.sh

@@ -2,16 +2,17 @@
 # C3 — exercises training_hub.sft against the published
 # traininghub0.1-cu126-amd64:v0.1.0 runtime image (used by sft-comprehensive-tutorial.ipynb).
 # Tiny synthetic Qwen2-style HF checkpoint + synthetic JSONL are generated inside the
-# Pod so the case has no external dependency. nproc_per_node is forced to 1 because
-# the dev GPU node exposes only 1 whole GPU.
+# Pod so the case has no external dependency. nproc_per_node is forced to 1
+# to keep the smoke case small.
 set -euo pipefail
 HERE="$(cd "$(dirname "$0")" && pwd)"
 source "${HERE}/../lib.sh"
 
+require_env GPU_NAMESPACE "namespace for GPU e2e resources"
 NS="${GPU_NAMESPACE}"
 JOB_NAME="c3-traininghub-sft-$(printf '%05x' $$)"
-# Bundled-training_hub image built from kubeflow-plugin/training-runtimes/traininghub0.1-cu126-amd64.
-IMAGE="${C3_IMAGE:-build-harbor.alauda.cn/mlops/traininghub0.1-cu126-amd64:v0.1.0-build.20260609030710}"
+IMAGE="${C3_IMAGE:-docker.io/alaudadockerhub/traininghub0.1-cu126-amd64:v0.1.0}"
+IMAGE_PULL_SECRET="${C3_IMAGE_PULL_SECRET:-${E2E_IMAGE_PULL_SECRET:-}}"
 
 cleanup() {
   gpu_kc -n "${NS}" delete job "${JOB_NAME}" --ignore-not-found --wait=false || true
@@ -24,7 +25,7 @@ log "C3: submitting Job ${JOB_NAME} (image=${IMAGE})"
 # single_gpu_dev preset; we do the same here. Liger is disabled because the
 # image's CUDA *runtime* lacks nvcc (see training-runtimes.mdx — same caveat
 # applies to JIT op compilation).
-cat <<YAML | gpu_kc -n "${NS}" create -f -
+cat <<YAML | mirror_dockerhub "${GPU_DH_MIRROR}" | gpu_kc -n "${NS}" create -f -
 apiVersion: batch/v1
 kind: Job
 metadata:
@@ -44,8 +45,7 @@ spec:
         runAsUser: 1000
         runAsGroup: 1000
         fsGroup: 1000
-      imagePullSecrets:
-        - name: harbor-mlops-regcred
+$(yaml_image_pull_secrets 6 "${IMAGE_PULL_SECRET}")
       volumes:
         - name: workspace
           emptyDir: {}

e2e/cases/c4_traininghub_osft.sh

@@ -6,9 +6,11 @@ set -euo pipefail
 HERE="$(cd "$(dirname "$0")" && pwd)"
 source "${HERE}/../lib.sh"
 
+require_env GPU_NAMESPACE "namespace for GPU e2e resources"
 NS="${GPU_NAMESPACE}"
 JOB_NAME="c4-traininghub-osft-$(printf '%05x' $$)"
-IMAGE="${C4_IMAGE:-build-harbor.alauda.cn/mlops/traininghub0.1-cu126-amd64:v0.1.0-build.20260609030710}"
+IMAGE="${C4_IMAGE:-docker.io/alaudadockerhub/traininghub0.1-cu126-amd64:v0.1.0}"
+IMAGE_PULL_SECRET="${C4_IMAGE_PULL_SECRET:-${E2E_IMAGE_PULL_SECRET:-}}"
 
 cleanup() {
   gpu_kc -n "${NS}" delete job "${JOB_NAME}" --ignore-not-found --wait=false || true
@@ -17,7 +19,7 @@ trap cleanup EXIT
 
 log "C4: submitting Job ${JOB_NAME} (image=${IMAGE})"
 
-cat <<YAML | gpu_kc -n "${NS}" create -f -
+cat <<YAML | mirror_dockerhub "${GPU_DH_MIRROR}" | gpu_kc -n "${NS}" create -f -
 apiVersion: batch/v1
 kind: Job
 metadata:
@@ -37,8 +39,7 @@ spec:
         runAsUser: 1000
         runAsGroup: 1000
         fsGroup: 1000
-      imagePullSecrets:
-        - name: harbor-mlops-regcred
+$(yaml_image_pull_secrets 6 "${IMAGE_PULL_SECRET}")
       volumes:
         - name: workspace
           emptyDir: {}
@@ -80,7 +81,7 @@ spec:
               cd /workspace
               # training_hub is bundled in the runtime image. mini_trainer (the
               # OSFT backend) unconditionally `import flash_attn`, which only
-              # supports sm_75+; the dev GPU node is Tesla P100 (sm_60).
+              # supports sm_75+; older GPU architectures need this stub.
               # Stub the module so the import succeeds; transformers falls
               # through to torch SDPA, which is what training_hub.sft uses too.
               mkdir -p /workspace/stubs/flash_attn

e2e/cases/c5_trainer_v2_llamafactory.sh

@@ -12,34 +12,34 @@ set -euo pipefail
 HERE="$(cd "$(dirname "$0")" && pwd)"
 source "${HERE}/../lib.sh"
 
+require_env GPU_NAMESPACE "namespace for GPU e2e resources"
 NS="${GPU_NAMESPACE}"
 # Per-run suffix so re-runs / concurrent runs don't share PVC contents or fight
 # over the same TrainingRuntime. The trap below tears down both.
 RUN_ID="$(printf '%05x' $$)-$(date -u +%s)"
 RUNTIME="c5-llamafactory-finetune-runtime-${RUN_ID}"
 PVC_NAME="c5-models-${RUN_ID}"
-# The doc notebook says "Use `alaudadockerhub/fine_tune_with_llamafactory:v0.1.11`,
-# or build your own". The Dockerhub copy is firewalled (mirror blob EOF on large
-# images), the harbor rebuild at the same tag lacks torch, and the catalog runtime
-# image `llamafactory0.9-cu126-amd64:v0.1.0` has the same blob-EOF issue. Use the
-# build-suffixed copy of the catalog image already cached on the dev cluster.
-IMAGE="${LF_IMAGE:-build-harbor.alauda.cn/mlops/llamafactory0.9-cu126-amd64:v0.1.0-build.20260603021903}"
+IMAGE="${LF_IMAGE:-docker.io/alaudadockerhub/llamafactory0.9-cu126-amd64:v0.1.0}"
+IMAGE_PULL_SECRET="${LF_IMAGE_PULL_SECRET:-${E2E_IMAGE_PULL_SECRET:-}}"
+RWX_STORAGE_CLASS="${C5_RWX_STORAGE_CLASS:-${E2E_RWX_STORAGE_CLASS:-}}"
+NODE_SELECTOR_KEY="${C5_NODE_SELECTOR_KEY:-${E2E_GPU_NODE_SELECTOR_KEY:-}}"
+NODE_SELECTOR_VALUE="${C5_NODE_SELECTOR_VALUE:-${E2E_GPU_NODE_SELECTOR_VALUE:-}}"
 
-log "C5: ensuring shared PVC ${PVC_NAME} exists (RWX on cephfs so any node can mount)"
+log "C5: ensuring shared RWX PVC ${PVC_NAME} exists"
 cat <<YAML | retry_apply gpu_kc
 apiVersion: v1
 kind: PersistentVolumeClaim
 metadata:
   name: ${PVC_NAME}
   namespace: ${NS}
 spec:
-  storageClassName: cephfs
+$(yaml_storage_class 2 "${RWX_STORAGE_CLASS}")
   accessModes: ["ReadWriteMany"]
   resources: { requests: { storage: 4Gi } }
 YAML
 
 log "C5: applying TrainingRuntime ${RUNTIME} to ns/${NS} (image=${IMAGE})"
-cat <<YAML | retry_apply gpu_kc
+cat <<YAML | mirror_dockerhub "${GPU_DH_MIRROR}" | retry_apply gpu_kc
 apiVersion: trainer.kubeflow.org/v1alpha1
 kind: TrainingRuntime
 metadata:
@@ -64,10 +64,8 @@ spec:
               template:
                 spec:
                   securityContext: { runAsNonRoot: true, runAsUser: 65534, runAsGroup: 65534, fsGroup: 65534 }
-                  nodeSelector:
-                    kubernetes.io/hostname: 192.168.138.15
-                  imagePullSecrets:
-                    - name: harbor-mlops-regcred
+$(yaml_node_selector 18 "${NODE_SELECTOR_KEY}" "${NODE_SELECTOR_VALUE}")
+$(yaml_image_pull_secrets 18 "${IMAGE_PULL_SECRET}")
                   volumes:
                     - name: models
                       persistentVolumeClaim: { claimName: ${PVC_NAME} }
@@ -123,10 +121,8 @@ spec:
               template:
                 spec:
                   securityContext: { runAsNonRoot: true, runAsUser: 65534, runAsGroup: 65534, fsGroup: 65534 }
-                  nodeSelector:
-                    kubernetes.io/hostname: 192.168.138.15
-                  imagePullSecrets:
-                    - name: harbor-mlops-regcred
+$(yaml_node_selector 18 "${NODE_SELECTOR_KEY}" "${NODE_SELECTOR_VALUE}")
+$(yaml_image_pull_secrets 18 "${IMAGE_PULL_SECRET}")
                   volumes:
                     - name: models
                       persistentVolumeClaim: { claimName: ${PVC_NAME} }
@@ -192,10 +188,8 @@ spec:
               template:
                 spec:
                   securityContext: { runAsNonRoot: true, runAsUser: 65534, runAsGroup: 65534, fsGroup: 65534 }
-                  nodeSelector:
-                    kubernetes.io/hostname: 192.168.138.15
-                  imagePullSecrets:
-                    - name: harbor-mlops-regcred
+$(yaml_node_selector 18 "${NODE_SELECTOR_KEY}" "${NODE_SELECTOR_VALUE}")
+$(yaml_image_pull_secrets 18 "${IMAGE_PULL_SECRET}")
                   volumes:
                     - { name: workspace, emptyDir: {} }
                     - name: models