Generated banner for Hyperpod Node Debugger
Generated logo for Hyperpod Node Debugger

Hyperpod Node Debugger

AWS LabsOfficialPublished by AWS Labs

  • Skill
  • Workflow
  • Cloud and DevOps
  • Apache-2.0
  • Low risk
  • Has scripts

From the author

Diagnose and remediate per-node issues on a HyperPod cluster (EKS or Slurm) — a specific node is unhealthy, unresponsive, stuck, or needs replacing. Covers on-node EFA, GPU / accelerator hardware (XID, ECC, NVLink, row-remap, DCGM), Slurm node down/drained, disk and memory pressure, per-node lifecycle-script failures, SSM agent, container runtime, kernel panics, pod networking. Read-only. Not for cluster-wide provisioning (→ hyperpod-cluster-debugger), NCCL (→ hyperpod-nccl), or MFU (→ hyperpod-mfu-debugger).
— the description by AWS Labs

AutoGineer summary

Written by AutoGineer from the skill’s own files, not by the author

The HyperPod Node Debugger provides a structured, read-only diagnostic workflow for identifying and resolving issues on individual nodes within a HyperPod cluster. It covers a wide range of failure modes, including EFA networking, GPU/accelerator hardware faults (XID, ECC, NVLink), Slurm node states, resource exhaustion, and container runtime errors. The skill emphasizes a safe, non-destructive triage process, requiring explicit customer approval for any remediation steps, and includes specific guidance for Infrastructure-as-Code (IaC) environments to prevent configuration drift.

What it does

  • Performs read-only triage of cluster nodes to identify health failures.
  • Maps detected symptoms to specific diagnostic sections for targeted troubleshooting.
  • Provides verified commands for remediation, including rebooting or replacing nodes.
  • Includes safety checks for IaC-managed resources to prevent configuration drift.
  • Supports both EKS and Slurm orchestrators for HyperPod environments.
  • Guides the user through a structured investigate-reboot-replace escalation path.

When to use it

  • When a specific node in a HyperPod cluster is unhealthy, unresponsive, or stuck.
  • When training jobs hang due to suspected hardware or networking issues on a single node.
  • When GPU/accelerator errors (e.g., XID, ECC, NVLink) are suspected.
  • When investigating node-level resource exhaustion (disk, memory, OOM).
  • When Slurm nodes are marked as down or drained unexpectedly.

Advantages

  • Standardizes the diagnostic process for complex HyperPod node failures.
  • Reduces manual effort by automating the collection of cluster and node health signals.
  • Prioritizes safety by enforcing a read-only triage phase before any remediation.
  • Provides clear guidance on when to escalate to AWS Support.
  • Integrates with existing IaC workflows to maintain infrastructure integrity.

Usage examples

“My node i-0123456789abcdef0 is stuck in a Pending state. How do I check it?”
The agent will run the triage script focused on that specific instance ID, identify the failure reason (e.g., lifecycle script or hardware), and provide the corresponding diagnostic steps.
“I suspect a GPU hardware failure on node i-0987654321fedcba0. What should I do?”
The agent will guide you to check for XID or ECC errors using the provided diagnostic commands and, if confirmed, suggest the appropriate batch-reboot or batch-replace procedure.

Install

  1. Download the skill as a ZIP:

    https://autogineer.com/agent-skills/awslabs/hyperpod-node-debugger/download
  2. Unzip it into the cross-client skills folder of your project:

    mkdir -p .agents/skills && unzip hyperpod-node-debugger.zip -d .agents/skills

Documentation by AWS Labs

HyperPod Node Debugger

Operating policy. Run read-only diagnostics yourself. Never run a command that changes cluster, node, or workload state — present each one as a Suggested command (run this yourself) block and wait for the customer. Destructive order: investigate → reboot → replace (replace destroys root + secondary volumes; not supported on Slurm controller nodes). Never discard training state, logs, or caches on speculation.

IaC note (always include with mutation commands). When you suggest any command that changes cluster, VPC, SG, subnet, or EKS configuration (e.g. authorize-security-group-*, modify-vpc-attribute, update-cluster, kubectl label/cordon/drain, create namespace, set env daemonset), ask the customer first whether the cluster / VPC / SG is managed by Infrastructure-as-Code (CloudFormation, CDK, Terraform, Pulumi). If yes, tell them: "Apply this change in your IaC source first, then deploy through the pipeline — running the command directly will drift from your template and the next stack update may overwrite it." If they need to fix the issue immediately and the IaC change will follow, flag the drift explicitly so they remember to reconcile.

Read-only triage. scripts/triage-cluster.sh (and helpers check-efa-sg.sh, check-node-reachability.sh, check-vpc-config.sh) read state and print each issue as [FAIL] ... → references/node-diagnostics-detail.md § <section>. Catalog of customer-ticket patterns: references/node-issue-catalog.md.


Workflow

  1. Collect cluster name, region, suspect instance ID, exact error string from logs.

  2. Run scripts/triage-cluster.sh (add --node <INSTANCE-ID> to focus one node).

  3. For every [FAIL] / issue entry, Read the referenced section.

  4. Present: what script detected (copy the line verbatim), root cause, exact command(s) with instance/SG IDs filled in, blast radius (e.g. "reboots i-xxx", "wipes volumes on replacement"). For any command that mutates cluster/VPC/SG/EKS state, ask whether the affected resource is IaC-managed and surface the drift warning from the operating-policy note above.

  5. Wait for explicit customer approval. Destructive order: investigate → reboot → replace.

  6. Re-run triage to confirm. Iterate if not cleared.

Step 1: Triage

bash scripts/triage-cluster.sh --cluster <CLUSTER_NAME_OR_ARN> --region <REGION>

# Focus on one node:
bash scripts/triage-cluster.sh --cluster <CLUSTER_NAME_OR_ARN> --region <REGION> --node <INSTANCE_ID>

One pass collects: cluster status + NodeRecovery, events, per-node health (HyperPod + EKS labels, Slurm states), VPC/SG snapshot, CloudWatch availability, SSM readiness, on-node resource checks (disk, memory, /dev/shm, OOM, NVMe, time sync, SSM agent), Slurm node→instance mapping.

Tags: [PASS] passed · [FAIL] issue with a → references/... pointer · [WARN] advisory · [INFO] informational. Priorities: P0 blocks operation · P1 degraded · P2 informational.

Step 2: Match signal → section

Events (list-cluster-events) — provisioning-time:

EventSection
"EFA health checks did not run successfully" (public-doc verbatim signal)A: EFA/SG
Instance bootstrap or network-misconfiguration eventA + B: VPC
Lifecycle-script failure or timeoutD: Lifecycle
Insufficient-capacity or AZ-mismatch failure at creationC: Capacity
Hardware failure / UnschedulablePendingReplacementF: Hardware

EKS labels:

LabelSection
node-health-status: UnschedulablePendingReplacementF
node-health-status: UnschedulablePendingRebootF
deep-health-check-status: FailedG → F

Symptoms:

SymptomSection
Training hangs at NCCL init / AllReduceA → E
Slurm node down / "Node unexpectedly rebooted"H: Slurm
Jobs stuck PENDING / COMPLETINGH
Auto-repair not triggeringF
GPU not visible / XID / ECC errorsG
GPU row-remap pending/failed / silent NaNs / DCGM FailG § G.1.a/b
Disk full / OOM / "Cannot allocate memory"I: Resources
Wrong vCPU count (e.g. 96 instead of 192 on p5.48xlarge)J: Config
Container CrashLoopBackOff / runtime crashM: Container Runtime
aws-node CrashLoopBackOff / gRPC 50051 refusedO: CNI / Pod Networking
Pods stuck Pending with no IP / CNI errorO
DNS resolution / enableDnsSupportB § B.2
Public subnet / IGW misconfiguredB § B.3
Missing VPC endpoints (ECR / STS / FSx)B § B.4
EKS VPC / SG mismatch with HyperPodB § B.5
Kernel panic / watchdog / hung taskN: Kernel
Need shell on a nodeK: SSM
Collect logs for AWS SupportL: Log Collection

A: EFA / Security Group

Per the HyperPod prerequisites doc, the SG must allow all inbound and outbound to itself. scripts/check-efa-sg.sh validates self-ref rules on every cluster SG. On-node EFA check via scripts/check-node-reachability.sh over SSM. Full: § A.

B: VPC / Routing

SG/subnet VPC mismatch, missing S3 Gateway endpoint, EKS auth mode, worker→controller routing, VPC DNS support, private-subnet + NAT / VPC endpoints, EKS↔HyperPod VPC alignment. scripts/check-vpc-config.sh. Full: § B.

C: Capacity / AZ

Insufficient-capacity failure at creation, or no subnets in the AZ where capacity is available. Check AZ offerings via describe-instance-type-offerings, then change subnet AZ or use Flexible Training Plans / ODCR. Full: § C.

D: Lifecycle Scripts

Surfaced in cluster events + CloudWatch under LifecycleConfig/<group>/<instance-id>. Common: S3 connectivity, IAM gaps, CRLF line endings, infinite loops, parameter-name mismatch. Full: § D.

E: Software Versions

Delegate to hyperpod-version-checker to compare NVIDIA driver, CUDA, NCCL, EFA installer, OFI NCCL, PyTorch across nodes. Ensure job env has FI_PROVIDER=efa, FI_EFA_USE_DEVICE_RDMA=1, NCCL_SOCKET_IFNAME=^lo,docker. Full: § E.

F: Hardware / Auto-Repair

Confirm NodeRecovery=Automatic, inspect the EKS health labels + sagemaker.amazonaws.com/fault-details annotation, and read the SagemakerHealthMonitoringAgent/<group>/<instance> CloudWatch stream. HMA runs passive background checks on GPU and Neuron state and reboots the node on count mismatch (per the HMA doc: "if there's a mismatch between the expected number of GPUs … and the count returned by nvidia-smi, then HMA reboots the node"; same for neuron-ls). Manual recovery order: reboot first, replace only if reboot fails; the preferred path is the batch APIs (BatchReboot/BatchReplaceClusterNodes). Full: § F · patterns: node-issue-catalog.md.

G: GPU / Accelerator

NVIDIA (p4d/p5/g5/g6): nvidia-smi + dmesg over SSM for Xid, ECC, thermal throttling. Xid classification per NVIDIA's catalog: 13 Graphics Engine Exception (application-level), 31 GPU memory page fault (application, can be driver/HW), 63 GPU memory remapping event (HW/ECC), 71 CE4 Error (HW copy engine), 74 NVLink Error (HW), 79 GPU has fallen off the bus (PCIe bus), 109 Context Switch Timeout Error (HW). Any uncorrectable ECC → drain and replace. Row-remap state is the authoritative silent-degradation signal (§ G.1.a).

Trainium / Inferentia (trn1/trn2/inf2): Neuron SDK — neuron-ls, neuron-top, neuron-monitor. nvidia-smi does not apply.

GPU / accelerator failures flow into § F for reboot / replace. Full: § G.

H: Slurm Node Management

Node down/unresponsive, unexpected reboots, stuck PENDING/COMPLETING jobs, Slurm-to-instance-ID translation. Primary access is SSM; diagnose slurmd first, fix the root cause, then start/resume the node per § H. Full: § H.

I: Resource Exhaustion

Disk full (HyperPod root volume defaults to 100 GB and is not intended to grow post-creation), OOM, os.fork() memory error, /dev/shm exhaustion, inode exhaustion. Fork-memory fix: export FI_EFA_USE_HUGE_PAGE=0. Redirect bulk data to /opt/sagemaker (secondary EBS) or /opt/dlami/nvme (instance store). Full: § I.

J: Configuration

p5.48xlarge reports 96 vCPU instead of 192 → set ThreadsPerCore=2 via update-cluster. Full: § J.

K: Node Access via SSM

No direct SSH on HyperPod. Target format sagemaker-cluster:<CLUSTER_ID>_<GROUP>-<INSTANCE_ID>. Failures: plugin missing, wrong prefix, IAM, VPC endpoints. Full: § K.

L: Log Collection

Delegate to hyperpod-issue-report for S3-stored bundles. Key CloudWatch streams: LifecycleConfig/<group>/<instance-id>, SagemakerHealthMonitoringAgent/<group>/<instance-id>. Full: § L.

M: Container Runtime

CrashLoopBackOff, OOMKilled, ImagePullBackOff, RunContainerError on EKS. kubectl describe pod + on-node crictl ps -a, journalctl -u containerd. Full: § M.

N: Kernel & System

Kernel panic, watchdog timeout, soft lockup, unexpected reboots not explained by HyperPod health monitoring. dmesg | grep -iE 'panic|watchdog|hung_task|NMI' + journalctl -b -1. nvrm-related signatures point at NVIDIA driver crashes. Full: § N.

O: CNI / Pod Networking

VPC CNI (aws-node) failures, IPAMD errors, gRPC 127.0.0.1:50051 refused, pods stuck Pending with FailedCreatePodSandBox. Script auto-checks aws-node, kube-proxy, CoreDNS. Full: § O.


Prerequisites

  • aws CLI v2, recent enough to support the HyperPod cluster commands (describe-cluster, list-cluster-nodes, batch-reboot-cluster-nodes, batch-replace-cluster-nodes)

  • python3, bash 4+ (associative arrays are required by the scripts)

  • kubectl authenticated to the EKS cluster (K8s checks skipped if absent)

  • session-manager-plugin for on-node hardware checks

  • unbuffer (from the expect package) — optional; if missing, SSM on-node probes are skipped while the rest of the triage still runs. Install via yum install expect / apt install expect.

Defaults

  • Region — required: pass --region or set $AWS_DEFAULT_REGION.

  • Target scope — all nodes; --node <ID> focuses one.

  • Event window — up to 500 most recent events (5 × 100, paginated).

  • Node list cap — up to 20,000 nodes (200 × 100); warns on cap.

  • SSM probes — 180 s per node with retry-on-throttle.

  • Colors — auto-disabled on non-TTY; --no-color to force off.

Error handling

FailureScriptTell the customer
aws sts get-caller-identity failsExit 1"Fix AWS credentials and rerun."
describe-cluster failsExit 1 after listing region's clusters"Confirm cluster name and region."
sagemaker:* / ec2:* / logs:* AccessDeniedWarn, add Missing IAM permission for <API>, continue"Grant the listed IAM action and rerun."
kubectl absent or unauthenticatedSkip K8s checks"Install/authenticate kubectl (see § K)."
session-manager-plugin absentSkip on-node probes"Install session-manager-plugin (see § K)."
SSM start-session fails or times out (180s)Mark node unreachable with → § K pointer"Rerun with --node <ID> to isolate; verify SSM agent on the node."
Cluster > 20,000 nodesFirst 20,000 paginated; warn"Use --node to target specific nodes."

Exit codes: 0 triage complete · 1 cluster not found or fatal prerequisite missing.

IAM permissions

Read-only diagnostic — covers triage-cluster.sh, check-efa-sg.sh, check-vpc-config.sh, and check-node-reachability.sh:

{
  "Action": [
    "sagemaker:DescribeCluster",
    "sagemaker:DescribeClusterNode",
    "sagemaker:ListClusterNodes",
    "sagemaker:ListClusterEvents",
    "sagemaker:ListClusters",
    "eks:DescribeCluster",
    "ec2:DescribeSecurityGroups",
    "ec2:DescribeSubnets",
    "ec2:DescribeVpcs",
    "ec2:DescribeVpcAttribute",
    "ec2:DescribeVpcEndpoints",
    "ec2:DescribeRouteTables",
    "ec2:DescribeNetworkInterfaces",
    "ec2:DescribeInstances",
    "ec2:DescribeInstanceTypeOfferings",
    "ec2:DescribeInstanceTypes",
    "logs:DescribeLogGroups",
    "logs:DescribeLogStreams",
    "logs:FilterLogEvents",
    "ssm:StartSession",
    "ssm:TerminateSession",
    "service-quotas:GetServiceQuota"
  ]
}

sts:GetCallerIdentity is implicit — it requires no IAM action. SSM on HyperPod uses start-session against sagemaker-cluster:<cluster-id>_<group>-<iid> targets — not send-command against bare instance IDs. For remediation commands, grant the matching write permission (e.g. ec2:AuthorizeSecurityGroupIngress / Egress, ec2:RevokeSecurityGroupIngress / Egress, ec2:ModifyVpcAttribute, sagemaker:UpdateCluster, sagemaker:BatchRebootClusterNodes, sagemaker:BatchReplaceClusterNodes). Not needed for the diagnostic itself.

Skill delegation

NeedUse
Cluster creation / deployment failureshyperpod-cluster-debugger (§ A / B / C / H + --validate)
Cluster-wide SSM outagehyperpod-cluster-debugger § F
Single-node SSM failurestay here — § K
Cluster-wide EFA health-check failure at creation timehyperpod-cluster-debugger § A
Single-node EFA failure post-provisioningstay here — § A
NCCL AllReduce / collective-op timeouts (distributed)hyperpod-nccl
Silent GPU NaNs on a specific node (row-remap / DCGM)stay here — § G.1 (even if discovered by NCCL)
Post-deployment cluster-wide managementhyperpod-cluster-debugger
Shell / commands on nodeshyperpod-ssm
CUDA / NCCL / EFA version comparisonhyperpod-version-checker
Diagnostic bundle for AWS Supporthyperpod-issue-report
Training performance / MFU degradationhyperpod-mfu-debugger

Escalate to AWS Support

Escalate when:

  1. SG rules correct and reachability passes but EFA still fails.

  2. VPC correct but K8s bootstrap fails — check VPC flow logs for REJECT.

  3. Hardware failure where replacement keeps failing (bad physical host).

  4. Node replacement fails with an insufficient-capacity signal despite a valid ODCR.

Before opening the case
# 1. Cluster identity + affected node status
aws sagemaker describe-cluster --cluster-name <CLUSTER> --region <REGION>
aws sagemaker list-cluster-nodes --cluster-name <CLUSTER> --region <REGION> \
  --query "ClusterNodeSummaries[?InstanceId=='<INSTANCE_ID>']"

# 2. Triage bundle (scoped to the affected node where possible)
bash scripts/triage-cluster.sh --cluster <CLUSTER> --region <REGION> --node <INSTANCE_ID> > triage.txt

# 3. Per-node log/config bundle to S3 (delegates to hyperpod-issue-report)
#    See skills/hyperpod-issue-report/SKILL.md for the exact invocation.
Include in the case
  • Cluster name + ARN and AWS region

  • Orchestrator (EKS or Slurm)

  • Affected instance IDs / node names / instance-group names

  • Timestamp window (UTC start / end) of the failure

  • Exact error strings observed (copy verbatim from pod logs, CloudWatch, dmesg, events)

  • XID numbers / ECC counts / DCGM output where hardware is implicated

  • triage.txt from step 2 above

  • S3 URI of the hyperpod-issue-report bundle from step 3

Patterns from real customer tickets: node-issue-catalog.md.

Files

scripts/check-node-reachability.sh

#!/usr/bin/env bash
# check-node-reachability.sh
#
# Diagnose EFA reachability and inter-node communication health on a single
# HyperPod node. Run this on each node via the hyperpod-ssm skill.
#
# Usage (via ssm-exec.sh):
#   ssm-exec.sh --target <TARGET> --upload scripts/check-node-reachability.sh /tmp/check-node-reachability.sh
#   ssm-exec.sh --target <TARGET> 'bash /tmp/check-node-reachability.sh'
#
# Usage (direct on node):
#   bash check-node-reachability.sh [--json] [--no-color]
#
# Exit codes:
#   0 — all critical checks passed
#   1 — one or more critical checks failed

set -euo pipefail

# Note: this script runs ON the node (via SSM), so aws CLI may not be present.
# Only python3 is checked here; other tools are checked individually per section.

JSON_MODE=false
USE_COLOR=true

usage() {
  cat <<EOF
Usage: bash check-node-reachability.sh [--json] [--no-color]

Read-only on-node diagnostic for EFA reachability and inter-node communication
health. Must be executed on a HyperPod compute node (typically via the
hyperpod-ssm skill). Checks EFA interfaces, /dev/infiniband devices, GPU
count and Neuron device count against the expected counts for the node's
instance type.

Options:
  --json       Emit findings as JSON instead of human-readable output.
  --no-color   Disable ANSI colors.
  -h, --help   Show this message.

Exit codes:
  0  All critical checks passed.
  1  One or more critical checks failed.
EOF
}

while [[ $# -gt 0 ]]; do
  case "$1" in
    --json)     JSON_MODE=true;  shift ;;
    --no-color) USE_COLOR=false; shift ;;
    -h|--help)  usage; exit 0 ;;
    *) echo "Unknown argument: $1" >&2; usage >&2; exit 1 ;;
  esac
done

# Colors — auto-disable when stdout isn't a TTY.
if ! [ -t 1 ] || [ "${TERM:-}" = "dumb" ]; then
  USE_COLOR=false
fi
if "$USE_COLOR"; then
  RED='\033[0;31m'; GREEN='\033[0;32m'; YELLOW='\033[1;33m'
  BOLD='\033[1m';   NC='\033[0m'
else
  RED=''; GREEN=''; YELLOW=''; BOLD=''; NC=''
fi

HOSTNAME=$(hostname 2>/dev/null || echo "unknown")
TIMESTAMP=$(date -u +"%Y-%m-%dT%H:%M:%SZ")
CRITICAL_FAILURES=0
declare -A RESULTS   # associative array: check_name → pass|fail|warn|skip

pass()  { RESULTS["$1"]="pass";  [[ "$JSON_MODE" == false ]] && echo -e "  ${GREEN}[PASS]${NC}  $1${2:+ — $2}"; }
fail()  { RESULTS["$1"]="fail";  CRITICAL_FAILURES=$((CRITICAL_FAILURES+1)); \
           [[ "$JSON_MODE" == false ]] && echo -e "  ${RED}[FAIL]${NC}  $1${2:+ — $2}"; }
warn()  { RESULTS["$1"]="warn";  [[ "$JSON_MODE" == false ]] && echo -e "  ${YELLOW}[WARN]${NC}  $1${2:+ — $2}"; }
skip()  { RESULTS["$1"]="skip";  [[ "$JSON_MODE" == false ]] && echo -e "         [SKIP]  $1${2:+ — $2}"; }
info()  { [[ "$JSON_MODE" == false ]] && echo -e "         $1"; }

if [[ "$JSON_MODE" == false ]]; then
  echo ""
  echo -e "${BOLD}=== HyperPod Node EFA Reachability Check ===${NC}"
  echo -e "Host:      ${BOLD}${HOSTNAME}${NC}"
  echo -e "Timestamp: ${TIMESTAMP}"
  echo ""
fi

if [[ "$JSON_MODE" == false ]]; then echo -e "${BOLD}--- EFA Kernel Module ---${NC}"; fi

EFA_MODULE=$(lsmod 2>/dev/null | grep -E '^efa\b' | awk '{print $1}' || true)
if [[ -n "$EFA_MODULE" ]]; then
  EFA_MODULE_VER=$(modinfo efa 2>/dev/null | grep -E '^version:' | awk '{print $2}' || echo "unknown")
  pass "efa_kernel_module" "loaded (version: ${EFA_MODULE_VER})"
else
  # Read-only invariant: detect only, never `sudo modprobe efa` — loading kernel
  # modules mutates node state, which the hyperpod-ssm skill's approval flow owns.
  fail "efa_kernel_module" "not loaded — see references/node-diagnostics-detail.md § A (EFA / Security Group)"
fi

if [[ "$JSON_MODE" == false ]]; then echo ""; echo -e "${BOLD}--- EFA Devices ---${NC}"; fi

# shellcheck disable=SC2010  # /dev/ entries are kernel-named, safe to ls|grep
EFA_DEVICES=$(ls /dev/infiniband/ 2>/dev/null | grep -E 'rdma_cm|uverbs|efa' || true)

if [[ -n "$EFA_DEVICES" ]]; then
  pass "efa_devices_present" "found in /dev/infiniband/: $(echo "$EFA_DEVICES" | tr '\n' ' ')"
else
  fail "efa_devices_present" "/dev/infiniband/ is empty or missing — EFA hardware not detected"
fi

if [[ "$JSON_MODE" == false ]]; then echo ""; echo -e "${BOLD}--- libfabric EFA Provider ---${NC}"; fi

if command -v fi_info &>/dev/null; then
  # If the previous section found no EFA hardware, fi_info failing is expected —
  # don't emit [FAIL] on top of the hardware [FAIL], which would double-count and
  # conflate "libfabric can't see EFA" with "node has no EFA at all".
  if [[ -z "$EFA_DEVICES" ]]; then
    skip "fi_info_efa_provider" "no EFA devices detected upstream — see efa_devices_present"
  else
    FI_EXIT=0
    FI_OUTPUT=$(fi_info -p efa 2>&1) || FI_EXIT=$?
    if echo "$FI_OUTPUT" | grep -q "provider: efa"; then
      EFA_PROVIDER_COUNT=$(echo "$FI_OUTPUT" | { grep -c "provider: efa" 2>/dev/null; true; })
      pass "fi_info_efa_provider" "EFA provider found (${EFA_PROVIDER_COUNT} endpoint(s))"
      info "$(echo "$FI_OUTPUT" | grep -E 'provider:|fabric:|domain:|version:' | head -8 | sed 's/^/    /')"
    else
      fail "fi_info_efa_provider" "fi_info -p efa returned no EFA provider (exit code ${FI_EXIT}) — libfabric cannot enumerate EFA devices. See references/node-diagnostics-detail.md § A (EFA / Security Group)"
      info "fi_info output: ${FI_OUTPUT:0:200}"
    fi
  fi
else
  warn "fi_info_efa_provider" "fi_info not found — install libfabric to run this check (fi_info comes with EFA installer)"
fi

if [[ "$JSON_MODE" == false ]]; then echo ""; echo -e "${BOLD}--- EFA Network Interfaces ---${NC}"; fi

# EFA interfaces typically appear as eth0/ens* for primary + rdmaX or efa* for EFA devices
# EFA ifaces on p5/p5en use regular kernel names (ens*) — filter by driver via ethtool
# rather than by name pattern (the old 'rdma|efa' name grep misses ens* on p5).
EFA_IFACES=""
if command -v ethtool &>/dev/null; then
  while IFS= read -r iface; do
    [[ -z "$iface" ]] && continue
    DRIVER=$(ethtool -i "$iface" 2>/dev/null | awk -F': ' '/^driver:/{print $2}')
    if [[ "$DRIVER" == "efa" ]]; then
      EFA_IFACES+="${iface}"$'\n'
    fi
  done < <(ip -o link show 2>/dev/null | awk -F': ' '{print $2}' | awk -F'@' '{print $1}' | grep -v '^lo$')
fi
# Fallback to name-based detection for older kernels / containers without ethtool
if [[ -z "$EFA_IFACES" ]]; then
  EFA_IFACES=$(ip link show 2>/dev/null | grep -E 'rdma|efa' | awk -F': ' '{print $2}' | tr -d '@' || true)
fi
REGULAR_IFACES=$(ip link show 2>/dev/null | grep -E 'state UP' | awk -F': ' '{print $2}' | tr -d '@' || true)

if [[ -n "$EFA_IFACES" ]]; then
  pass "efa_interfaces_up" "EFA interfaces found: $(echo "$EFA_IFACES" | tr '\n' ' ')"
  while IFS= read -r iface; do
    [[ -z "$iface" ]] && continue
    IP=$(ip addr show "$iface" 2>/dev/null | grep 'inet ' | awk '{print $2}' || true)
    if [[ -n "$IP" ]]; then
      info "  $iface → $IP"
    else
      warn "efa_interface_ip_${iface}" "interface $iface has no IP address — check DHCP/subnet config"
    fi
  done <<< "$EFA_IFACES"
else
  info "No EFA interfaces detected (by driver or name)"
  if [[ -n "$REGULAR_IFACES" ]]; then
    skip "efa_interfaces_up" "no separate EFA interface — primary interfaces: $(echo "$REGULAR_IFACES" | tr '\n' ' ' | head -c 80)"
  else
    warn "efa_interfaces_up" "no UP network interfaces found"
  fi
fi

if [[ "$JSON_MODE" == false ]]; then echo ""; echo -e "${BOLD}--- EFA Installation ---${NC}"; fi

EFA_VER_FILE="/opt/amazon/efa_installed_packages"
if [[ -f "$EFA_VER_FILE" ]]; then
  # Format is "EFA installer version: 1.30.0" — grab only the version token.
  EFA_VER=$(grep -iE '^EFA installer version' "$EFA_VER_FILE" 2>/dev/null \
              | head -1 \
              | grep -oE '[0-9]+\.[0-9]+(\.[0-9]+)?' \
              | head -1 || echo "")
  if [[ -z "$EFA_VER" ]]; then
    warn "efa_installer_present" "EFA installer file present but version line not parsed"
  else
    pass "efa_installer_present" "EFA installer version: ${EFA_VER}"
  fi
else
  warn "efa_installer_present" "EFA installer marker not found at ${EFA_VER_FILE} — EFA may not be installed via standard method"
fi

if [[ "$JSON_MODE" == false ]]; then echo ""; echo -e "${BOLD}--- NCCL / OFI Configuration ---${NC}"; fi

NCCL_VARS=("FI_PROVIDER" "FI_EFA_USE_DEVICE_RDMA" "NCCL_SOCKET_IFNAME" "NCCL_ALGO" "LD_LIBRARY_PATH")
ANY_NCCL_SET=false
for var in "${NCCL_VARS[@]}"; do
  val="${!var:-}"
  if [[ -n "$val" ]]; then
    info "  ${var}=${val}"
    ANY_NCCL_SET=true
  fi
done

if "$ANY_NCCL_SET"; then
  FI_PROVIDER_VAL="${FI_PROVIDER:-}"
  if [[ -n "$FI_PROVIDER_VAL" && "$FI_PROVIDER_VAL" != "efa" ]]; then
    warn "nccl_fi_provider" "FI_PROVIDER=${FI_PROVIDER_VAL} — for EFA workloads this should be 'efa'"
  elif [[ "$FI_PROVIDER_VAL" == "efa" ]]; then
    pass "nccl_fi_provider" "FI_PROVIDER=efa"
  fi
else
  skip "nccl_env_vars" "no NCCL/OFI env vars set in current shell — may be set in job launcher environment"
fi

if [[ "$JSON_MODE" == false ]]; then echo ""; echo -e "${BOLD}--- AWS OFI NCCL Plugin ---${NC}"; fi

OFI_LIB=$(find /opt/amazon/efa /opt/aws-ofi-nccl /usr/local/lib /usr/lib \
  -name "libnccl-net.so*" -o -name "aws-ofi-nccl.so*" 2>/dev/null | head -1 || true)

if [[ -n "$OFI_LIB" ]]; then
  pass "aws_ofi_nccl_plugin" "found: ${OFI_LIB}"
else
  if [[ -f "$EFA_VER_FILE" ]] && grep -q "ofi\|OFI" "$EFA_VER_FILE" 2>/dev/null; then
    pass "aws_ofi_nccl_plugin" "referenced in ${EFA_VER_FILE}"
  else
    warn "aws_ofi_nccl_plugin" "libnccl-net.so not found — required for EFA-accelerated NCCL (distributed training)"
  fi
fi

if [[ "$JSON_MODE" == false ]]; then echo ""; echo -e "${BOLD}--- Instance Metadata Reachability ---${NC}"; fi

IMDS_TOKEN=$(curl -s -X PUT "http://169.254.169.254/latest/api/token" \
  -H "X-aws-ec2-metadata-token-ttl-seconds: 60" --connect-timeout 3 -m 5 2>/dev/null || true)

if [[ -n "$IMDS_TOKEN" ]]; then
  INSTANCE_TYPE=$(curl -s -H "X-aws-ec2-metadata-token: $IMDS_TOKEN" \
    http://169.254.169.254/latest/meta-data/instance-type --connect-timeout 3 -m 5 2>/dev/null || echo "unknown")
  LOCAL_IP=$(curl -s -H "X-aws-ec2-metadata-token: $IMDS_TOKEN" \
    http://169.254.169.254/latest/meta-data/local-ipv4 --connect-timeout 3 -m 5 2>/dev/null || echo "unknown")
  pass "imds_reachable" "instance-type=${INSTANCE_TYPE}, local-ipv4=${LOCAL_IP}"

  # Static list of EFA-capable families; unknown types fall through to the
  # EC2 API check. aws CLI may not be present on-node, so the static path
  # covers the common case.
  case "$INSTANCE_TYPE" in
    p4de*|p4d*|p5en*|p5e*|p5*|p6*|trn1*|trn2*|inf2*|g5.48xlarge|g6e.48xlarge|g6.48xlarge|hpc6a*|hpc6id*|hpc7a*|hpc7g*|dl1*|dl2q*)
      pass "efa_capable_instance" "${INSTANCE_TYPE} supports EFA" ;;
    *)
      if command -v aws &>/dev/null; then
        EFA_CHECK=$(aws ec2 describe-instance-types \
          --instance-types "${INSTANCE_TYPE}" \
          --query 'InstanceTypes[0].NetworkInfo.EfaSupported' \
          --output text 2>/dev/null || echo "unknown")
        if [[ "$EFA_CHECK" == "True" ]]; then
          pass "efa_capable_instance" "${INSTANCE_TYPE} supports EFA (verified via API)"
        elif [[ "$EFA_CHECK" == "False" ]]; then
          warn "efa_capable_instance" "${INSTANCE_TYPE} does NOT support EFA"
        else
          warn "efa_capable_instance" "${INSTANCE_TYPE} — could not verify EFA support"
        fi
      else
        warn "efa_capable_instance" "${INSTANCE_TYPE} — not in known EFA list; verify with: aws ec2 describe-instance-types --instance-types ${INSTANCE_TYPE} --query 'InstanceTypes[0].NetworkInfo.EfaSupported'"
      fi
      ;;
  esac
  # Multi-EFA validation — counts per EC2 instance-type documentation.
  # NOTE: EFA counts vary between instance families (p5en has fewer than p5/p5e).
  EXPECTED_EFA=0
  case "$INSTANCE_TYPE" in
    p5.48xlarge|p5e.48xlarge)   EXPECTED_EFA=32 ;;
    p5en.48xlarge)              EXPECTED_EFA=16 ;;
    p4d.24xlarge|p4de.24xlarge) EXPECTED_EFA=4 ;;
    trn1.32xlarge)              EXPECTED_EFA=8 ;;
    trn2.48xlarge)              EXPECTED_EFA=16 ;;
    # p6 family and newer: don't hardcode counts; discover via ethtool to avoid false FAILs.
  esac

  if [[ "$EXPECTED_EFA" -gt 0 ]]; then
    # Count actual EFA devices — avoid grep -c pattern that returns "0\n0" fallthrough.
    ACTUAL_EFA=$(find /dev/infiniband -maxdepth 1 -name 'uverbs*' 2>/dev/null | wc -l)
    [[ -z "$ACTUAL_EFA" ]] && ACTUAL_EFA=0
    if [[ "$ACTUAL_EFA" -ge "$EXPECTED_EFA" ]]; then
      pass "multi_efa_interfaces" "${ACTUAL_EFA}/${EXPECTED_EFA} EFA interfaces present for ${INSTANCE_TYPE}"
    elif [[ "$ACTUAL_EFA" -gt 0 ]]; then
      warn "multi_efa_interfaces" "only ${ACTUAL_EFA}/${EXPECTED_EFA} EFA interfaces — some may not be attached or driver issue"
    else
      fail "multi_efa_interfaces" "0/${EXPECTED_EFA} EFA interfaces on ${INSTANCE_TYPE} — EFA driver or attachment issue"
    fi
  fi
else
  warn "imds_reachable" "IMDS not reachable. If running inside a container: check IMDSv2 HttpPutResponseHopLimit on the instance (default 1 is often too low for container networking — set to 2 or higher). Otherwise: verify the instance metadata service is enabled (HttpEndpoint != disabled) and that no local iptables / nftables rules block 169.254.169.254. Note: SGs do not filter link-local addresses."
fi

if [[ "$JSON_MODE" == false ]]; then echo ""; echo -e "${BOLD}--- Network Interface Statistics ---${NC}"; fi

if command -v ip &>/dev/null; then
  IFACE_ERRORS=$(ip -s link show 2>/dev/null | awk '
    BEGIN { rx_err=0; tx_err=0; iface="" }
    /^[0-9]+:/ {
      if (iface != "" && (rx_err > 0 || tx_err > 0))
        print "  " iface ": RX errors=" rx_err " TX errors=" tx_err
      iface=$2; gsub(/:$/, "", iface)
      rx_err=0; tx_err=0
    }
    /RX:/ { getline; rx_err=$3+0 }
    /TX:/ { getline; tx_err=$3+0 }
    END {
      if (iface != "" && (rx_err > 0 || tx_err > 0))
        print "  " iface ": RX errors=" rx_err " TX errors=" tx_err
    }
  ' || true)

  if [[ -n "$IFACE_ERRORS" ]]; then
    warn "network_interface_errors" "interfaces with errors detected:"
    info "$IFACE_ERRORS"
  else
    pass "network_interface_errors" "no RX/TX errors on active interfaces"
  fi
else
  skip "network_interface_errors" "ip command not available"
fi

if [[ "$JSON_MODE" == false ]]; then echo ""; echo -e "${BOLD}--- Neuron Devices (Trainium/Inferentia) ---${NC}"; fi

if command -v neuron-ls &>/dev/null; then
  NEURON_OUTPUT=$(neuron-ls 2>&1 || true)
  NEURON_DEVICE_COUNT=$(echo "$NEURON_OUTPUT" | { grep -c "neuron_device" 2>/dev/null; true; })
  if [[ "$NEURON_DEVICE_COUNT" -gt 0 ]]; then
    pass "neuron_devices" "${NEURON_DEVICE_COUNT} Neuron device(s) detected"
    info "$(echo "$NEURON_OUTPUT" | head -10 | sed 's/^/    /')"
  else
    NEURON_MOD=$(lsmod 2>/dev/null | grep -E '^neuron' || true)
    if [[ -n "$NEURON_MOD" ]]; then
      warn "neuron_devices" "Neuron driver loaded but neuron-ls shows 0 devices → references/node-diagnostics-detail.md § G.2 (Trainium/Inferentia)"
    else
      fail "neuron_devices" "Neuron driver not loaded → references/node-diagnostics-detail.md § G.2 (Trainium/Inferentia)"
    fi
  fi
elif ls /dev/neuron* &>/dev/null 2>&1; then
  NEURON_DEV_COUNT=$(find /dev -maxdepth 1 -name 'neuron*' 2>/dev/null | wc -l)
  NEURON_DEV_COUNT=${NEURON_DEV_COUNT:-0}
  warn "neuron_devices" "${NEURON_DEV_COUNT} /dev/neuron* device(s) found but neuron-ls not installed → references/node-diagnostics-detail.md § G.2 (Trainium/Inferentia)"
else
  skip "neuron_devices" "not a Trainium/Inferentia instance (no Neuron devices)"
fi

if [[ "$JSON_MODE" == false ]]; then
  echo ""
  echo -e "${BOLD}--- Summary ---${NC}"
  TOTAL=${#RESULTS[@]}
  PASSED=$(printf '%s\n' "${RESULTS[@]}" | { grep -c "^pass$" 2>/dev/null; true; })
  WARNED=$(printf '%s\n' "${RESULTS[@]}" | { grep -c "^warn$" 2>/dev/null; true; })
  FAILED=$(printf '%s\n' "${RESULTS[@]}" | { grep -c "^fail$" 2>/dev/null; true; })
  SKIPPED=$(printf '%s\n' "${RESULTS[@]}" | { grep -c "^skip$" 2>/dev/null; true; })
  echo -e "  Host: ${HOSTNAME}"
  echo -e "  Checks: ${TOTAL} total | ${GREEN}${PASSED} passed${NC} | ${YELLOW}${WARNED} warnings${NC} | ${RED}${FAILED} failed${NC} | ${SKIPPED} skipped"

  if [[ $CRITICAL_FAILURES -eq 0 ]]; then
    echo -e "\n  ${GREEN}${BOLD}Node EFA reachability checks PASSED.${NC}"
    echo "  If inter-node communication still fails, verify security group rules with check-efa-sg.sh"
    echo "  and compare EFA versions across nodes with the hyperpod-version-checker skill."
  else
    echo -e "\n  ${RED}${BOLD}Node EFA reachability checks FAILED (${CRITICAL_FAILURES} critical issue(s)).${NC}"
    echo "  See [FAIL] items above. Each finding ends with a pointer of the form"
    echo "  '→ references/node-diagnostics-detail.md § <section>' — open that section"
    echo "  for root cause and remediation. Remediation lives in references, not in scripts."
  fi
  echo ""
else
  CHECKS_JSON=""
  for key in "${!RESULTS[@]}"; do
    val="${RESULTS[$key]}"
    CHECKS_JSON+="\"${key}\": \"${val}\","
  done
  CHECKS_JSON="${CHECKS_JSON%,}"  # remove trailing comma

  cat <<EOF
{
  "hostname": "${HOSTNAME}",
  "timestamp": "${TIMESTAMP}",
  "critical_failures": ${CRITICAL_FAILURES},
  "overall_pass": $([ $CRITICAL_FAILURES -eq 0 ] && echo true || echo false),
  "checks": { ${CHECKS_JSON} }
}
EOF
fi

exit "$([[ $CRITICAL_FAILURES -eq 0 ]] && echo 0 || echo 1)"

Requirements and compatibility

Works with: Claude Code, Claude apps, OpenAI Codex, GitHub Copilot, Cursor, Gemini CLI, Kiro, OpenCode, Goose, Amp, Roo Code, JetBrains Junie, Factory Droid, Devin (ex-Windsurf), Mistral Vibe, OpenHands, Letta

“Works with” means the tool reads the open Agent Skills format; a skill written for one tool may use features others lack.

Needs: Needs a command-line tool, Needs Python, Needs Node.js

Security

What our scanner noticed· Low risk

  • Scripts that your agent may run

This skill includes scripts that your agent may run. Read them before you use it.

A static scan is not a guarantee. Skills are third-party content: review a skill before you let an agent use it.

Attribution and licence

Published by
AWS Labs
Publisher
AWS Labs (Official)
Source
awslabs/agent-plugins · plugins/sagemaker-ai/skills/hyperpod-node-debugger · commit fbefb37
Licence
Apache License 2.0 (redistribution allowed)
  • LICENSE (added by AutoGineer from LICENSE)
  • NOTICE (added by AutoGineer from NOTICE)
  • ATTRIBUTION.md (written by AutoGineer)
Synced

AutoGineer is not affiliated with or endorsed by AWS Labs or AWS Labs.

Download ZIP

The ZIP holds the skill’s files, its licence files and an ATTRIBUTION.md. Download hyperpod-node-debugger.zip