tomevault-io AI Infrastructure on Azure
Operational knowledge for Azure HPC GPU clusters running on CycleCloud Workspace for Slurm or AKS, with NVIDIA GB300 (NDv6) and H100 (NDv5) GPUs.
This single skill is the entry point for all cluster validation, diagnosis, and remediation questions. It contains a routing index that points to detailed reference files in references/. Read the relevant reference files before giving commands — do not invent thresholds or procedures.
How to Use This Skill
- Classify the request using the Intent Map below.
- Read the listed reference files before answering.
- Use only commands and thresholds that appear in the reference files. If the user's question can't be answered from the references, say so and ask for clarification.
- Ask for missing inputs explicitly: SKU (
Standard_ND128isr_GB300_v6orStandard_ND96isr_H100_v5), nodelist, cluster name, orchestrator (Slurm or AKS), failing job context. - Respond using the Response Contract at the bottom.
If the user invokes this skill with /ai-infra <request>, treat that as an explicit invocation and proceed with the same workflow.
Reference File Index
All reference files live in skills/ai-infra/references/.
Concepts and baselines (orchestrator-agnostic)
| File | Covers |
|---|---|
sku_baselines.md |
Expected NCCL busbw, GPU GFlops, thermal limits, IB ports, rack sizes for GB300 and H100. Warn and GHR thresholds. |
rack_topology.md |
MNNVL domain discovery via nvidia-smi, ClusterUUID, expected rack sizes per SKU, FabricManager troubleshooting. |
ib_validation.md |
InfiniBand port state (operstate, ibstat), partition keys, error counters, link flap detection, soft fixes. |
nccl_diagnosis.md |
Bisection algorithm, intra-rack vs inter-rack scoping, GPU vs network root cause, bandwidth pattern interpretation. |
outlier_detection.md |
Statistical methods (absolute threshold, z-score, MAD) for fleet-wide GPU and NCCL analysis. |
ghr.md |
Azure Guest Health Reporting (GHR) — used for reporting a node failure to Azure. Covers full impact category reference, IMDS/KVP data collection, REST API format, insight polling. |
Test execution (currently Slurm; AKS counterparts to be added)
| File | Covers |
|---|---|
nccl_test.md |
Run all_reduce_perf via the Slurm launcher. Per-SKU env vars (MNNVL/SHARP/GDR), output columns, quick vs full sweep. AKS path: see infrastructure_validations/aks/NCCL/. |
gpu_validation.md |
Run ubergemm GEMM benchmark via Slurm. Parse CSV output, identify underperforming GPUs. |
thermal_test.md |
Run dcgmproftester thermal stress via Slurm. Interpret pass/fail, throttle reasons, DCGMI levels. AKS path: see infrastructure_validations/aks/. |
Remediation
| File | Covers |
|---|---|
node_drain_and_replace.md |
Slurm drain/undrain/reboot, decision tree for drain vs reboot vs GHR, post-replacement validation. |
Intent Map
New cluster bring-up / full validation
Read: sku_baselines.md, rack_topology.md, nccl_test.md, gpu_validation.md, thermal_test.md
Slow training or low multi-node throughput
Read: nccl_diagnosis.md, sku_baselines.md, ib_validation.md, rack_topology.md (when topology correlation is needed)
NCCL failures or low all-reduce bandwidth
Read: nccl_test.md, nccl_diagnosis.md, ib_validation.md, outlier_detection.md (fleet-wide analysis)
GPU underperformance on one or more nodes
Read: gpu_validation.md, outlier_detection.md, sku_baselines.md
Thermal throttling or suspected cooling issues
Read: thermal_test.md, sku_baselines.md, gpu_validation.md (if thermal impact on GEMM is suspected)
InfiniBand link / pkey / errors investigation
Read: ib_validation.md, nccl_diagnosis.md, rack_topology.md (when rack-locality matters)
Identify degraded nodes across fleet
Read: outlier_detection.md, sku_baselines.md, plus gpu_validation.md and/or nccl_test.md depending on the metric source
Node remediation / drain / replace / file GHR (Guest Health Report)
Triggers: "file a GHR", "guest health report", "report a bad node to Azure", "report a node failure to Azure", "open a hardware issue with Azure", "health impact category".
Read: node_drain_and_replace.md, ghr.md
Cross-Cutting Rules
These apply to every operational answer:
- Always collect node metadata before rebooting or draining. PhysicalHostName (Hyper-V KVP pool 3) and Resource ID (IMDS) are required for any GHR. If the node goes down, you lose access to them. See
ghr.mdfor exact commands. - Use exact thresholds from
sku_baselines.md. Do not invent warn/GHR cutoffs. - Never suppress test failures — non-zero
#wrongin NCCL output, dcgmproftester non-zero exit, XID errors in dmesg are all hard signals. - For NCCL bandwidth issues, scope before bisecting. Determine intra-rack vs inter-rack from per-rack tests before isolating individual nodes.
- GHR impact category must match the symptom. See the category reference table in
ghr.md. UseHpcGenericFailureonly when nothing else fits. - Test scripts live in
infrastructure_validations/in this repo. The references describe how to run them and how to interpret output; the runnable artifacts are the scripts themselves.
Test Script Paths
infrastructure_validations/slurm/NCCL/— NCCLall_reduce_perflauncher with per-SKU configsinfrastructure_validations/slurm/gpu_test/— GPU GEMM benchmark (ubergemm)infrastructure_validations/slurm/thermal_test/— Thermal stress test (dcgmproftester)infrastructure_validations/slurm/NHC/— Node Health Checkinfrastructure_validations/aks/NCCL/— NCCL on AKS (via MPI Operator)infrastructure_validations/aks/NHC/— Node Health Check on AKSinfrastructure_validations/aks/fio/— Storage I/O test on AKS
Response Contract
For operational responses, follow this structure:
- Selected references — list the reference files you used.
- Run plan — concise ordered steps.
- Exact commands — copied from the reference files.
- Pass/fail thresholds — from
sku_baselines.mdor the relevant test reference. - Action decision — continue, isolate, drain, reboot, or file GHR with the specific impact category.
Source: Azure/ai-infrastructure-on-azure — distributed by TomeVault.