Loading...
Loading...
Found 20 Skills
Monitors and troubleshoots GKE TPU workloads, nodes, and node pools using GKE system metrics and PromQL. Use when monitoring TensorCore duty cycle, TPU memory, node readiness, multi-host TPU node pool availability, host maintenance or preemption interruptions, and calculating MTTR or MTBI metrics for GKE TPUs. Don't use for general non-TPU GKE workload monitoring or non-metric TPU debugging.
Monitors, troubleshoots, and manages GKE TPU Dynamic Slices custom resources. Use when checking TPU slice lifecycle states, troubleshooting slice provisioning failures, validating single-slice or multi-slice (JobSet) workload manifests, or safely patching stuck finalizers and disabling the slice controller. Don't use for generic GKE cluster node pool creation or standard non-TPU workload management (use gke-basics or gke-cluster-creation instead).
Monitors and troubleshoots GKE TPU workloads, nodes, and node pools using GKE system metrics and PromQL. Use when monitoring TensorCore duty cycle, TPU memory, node readiness, multi-host TPU node pool availability, host maintenance or preemption interruptions, and calculating MTTR or MTBI metrics for GKE TPUs. Don't use for general non-TPU GKE workload monitoring or non-metric TPU debugging.
Monitors, troubleshoots, and manages GKE TPU Dynamic Slices custom resources. Use when checking TPU slice lifecycle states, troubleshooting slice provisioning failures, validating single-slice or multi-slice (JobSet) workload manifests, or safely patching stuck finalizers and disabling the slice controller. Don't use for generic GKE cluster node pool creation or standard non-TPU workload management (use gke-basics or gke-cluster-creation instead).
Executes Python scripts, tests, or benchmarks on a provisioned remote cluster (GPU or TPU) using SkyPilot. Use this skill when the user asks to run code on GPU, TPU, or any "remote" cluster.
Diagnoses and prevents vbar_control_agent segfaults, out-of-memory (OOM) errors, and TPU device initialization failures on TPU v6e nodes in GKE caused by race conditions during TPU device resets or high-frequency metrics polling. Use when troubleshooting vbar_control_agent crashes, memory cgroup OOMs in serial console logs, tpu-device-plugin metrics checksum corruption errors, or custom TPU metrics collection conflicts on GKE TPU v6e nodes. Don't use for general non-TPU container OOM troubleshooting or standard GKE node lifecycle operations.
Diagnose GKE Cloud TPU multi-slice training hangs (Megascale `HANG_DETECTED` logs and hang monitored events) using the ML Diagnostics `Megascale XLA (MXLA) Hang Analyzer` (`gcloud alpha mldiagnostics monitored-events`) and 1-minute Cloud Monitoring multi-slice latency metrics (`kubernetes.io/container/multislice/*`). Distinguishes XLA compiler/HLO launch divergence and host data-input stalls from TPU chip, SparseCore, ICI, or network fabric faults. Use when multi-slice TPU training jobs freeze without progressing steps, emit `HANG_DETECTED`, or stall in collective operations. Don't use for gradual step-time throughput drops without hangs (use gke-ai-troubleshooting-tpu-performance-degradation) or pod preemption/eviction restarts (use gke-ai-troubleshooting-jobset-interruption).
Diagnose GKE Cloud TPU training throughput drops and step-time regressions (15%+ TPU duty-cycle drop) using ML Diagnostics Workload Monitoring (`gcloud alpha mldiagnostics monitored-events` / `hypercomputecluster.googleapis.com/v1alpha`) and 1-minute Cloud Monitoring system metrics (`kubernetes.io/node/accelerator/*`). Distinguishes hardware and network fabric throttling from workload resource bottlenecks (HBM capacity, host memory, or host CPU saturation). Use when TPU training throughput or duty cycle drops without crashing pods, when `PERFORMANCE_DEGRADATION` monitored events fire, or when triaging slow multi-slice training steps. Don't use for complete multi-slice XLA execution stalls with `HANG_DETECTED` logs (use gke-ai-troubleshooting-tpu-mxla-hang) or pod eviction/interruption restarts (use gke-ai-troubleshooting-jobset-interruption).
Diagnose and mitigate GKE TPU or GPU nodes stuck in NotReady / NodeStatusUnknown ("Kubelet stopped posting node status") due to host kernel panics, hardware lockups, or disabled node auto-repair. Use when nodes stop heartbeating beyond the node auto-repair threshold and pods remain stuck in Terminating. Don't use for healthy nodes, pod-only application crashes, or routine GKE upgrades.
Diagnose and resolve cluster-wide GKE TPU node pool creation or scale-up failures caused by Compute Engine reservation stockouts (GCE_STOCKOUT, ZONE_RESOURCE_POOL_EXHAUSTED) when existing unhealthy or NotReady TPU node pools in the cluster have node auto-repair disabled and hold reserved TPU capacity. Use when TPU node pool creation or scale-up fails with GCE_STOCKOUT, ZONE_RESOURCE_POOL_EXHAUSTED, or ZONE_RESOURCE_POOL_EXHAUSTED_WITH_DETAILS despite having a specific Compute Engine TPU reservation. Don't use for general pod scheduling without a TPU reservation or workload-level container crashes.
Diagnose and resolve GKE TPU Dynamic Slice creation and JobSet scheduling failures blocked by underlying Compute Engine managed instance groups (MIGs) stuck in an updating or unstable state (isStable=False). Use when TPU slice creation fails with a SliceCreationFailed status reason, when JobSets remain stuck in pending, or when MIG status reflects isStable=False or versionTarget.isReached=False during TPU slice formation. Don't use for SliceCreationFailed causes unrelated to MIG stability (use gke-ai-troubleshooting-tpu-dynamic-slices-monitoring) or general pod scheduling failures unrelated to TPU slices.
Diagnose GKE TPU VM instances stuck in Compute Engine repair (REPAIRING state) after host errors, maintenance events, or faulty host reports in single-host, multi-host, dynamic slicing, and TPU All Capacity mode node pools. Use when TPU instances in a slice node pool remain in REPAIRING, host error events appear in audit logs, or TPU slices are unschedulable due to a VM in repair. Don't use for nodes reporting Ready: Unknown when no VM is in REPAIRING (use gke-ai-troubleshooting-node-unresponsive-timeout), other NotReady nodes when no VM is in REPAIRING (use gke-node-notready), Pending Pods when all nodes are Ready (use gke-workload-troubleshooting), JobSet preemption events (use gke-ai-troubleshooting-jobset-interruption), or Slice custom resource lifecycle failures (use gke-ai-troubleshooting-tpu-dynamic-slices-monitoring).