MoE Hardware Configuration Reference
Plan MoE training layouts and throughput expectations across NVIDIA GPU platforms.
The material is primarily a read-only configuration reference, with no credential handling, covert exfiltration, malware, or destructive defaults. It includes warnings about containers, routing, precision, and cargo-culting tracker rows. Deductions apply because it lacks explicit user confirmation, execution isolation, rollback guidance, permission boundaries, and data-flow disclosure; environment variables and training settings can affect external training jobs.
SKILL.md is organized and provides platform tables, configuration families, environment variables, and pitfalls, but it has no executable key path, dependency checks, reproduction procedure, or diagnostic failure feedback. The instruction that Qwen3 235B on GB200 has VPP=unspecified conflicts materially with card.yaml's VPP12 entry. Static calibration caps this score below 10.
Triggers, audience, hardware platforms, model families, and expected guidance are reasonably clear, and the skill explicitly prohibits extrapolating VPP. Deductions apply because non-fit boundaries, input requirements, and exclusion triggers are incomplete, with no evidence of Chinese-language support. The core function does not entirely depend on overseas online services, but local environment and documentation reachability are unverified.
The files provide a name, description, Apache-2.0 license, owner, version information, known limitations, validation date, and change/update signals. Deductions apply for missing recommended metadata.author, metadata.tags, Instructions, and Examples sections, plus limited installation/dependency notes, FAQs, troubleshooting, and explicit maintenance/update ownership. The benchmark also reports a phone-number PII finding in card.yaml.
The skill offers directly useful guidance for selecting MoE configurations across H100, B200, GB200, and GB300, including representative layouts, tuning directions, performance bands, and an explicit prohibition against treating unmeasured VPP12 as authoritative. Deductions apply because outputs still require validation against the target software stack, container, routing mode, and measurements; performance claims are point-in-time, and one evaluation task does not establish broad effectiveness. Static calibration keeps this at or below 7.
The files name internal performance trackers, provide validation dates and measured-status fields, and include one evaluation report, giving limited auditability. Deductions apply because there is no committed reproducible experiment or test suite for this skill, only one positive task and no negative task, and the VPP conflict prevents independent confirmation of key claims. Static calibration keeps this at or below 5.
- card.yaml lists VPP12 for Qwen3 235B on GB200, while SKILL.md requires VPP=unspecified; resolve this conflict against measured data for the target revision before use.
- Performance bands and uplift figures are point-in-time and stack-dependent; do not treat them as promises or production configurations without checking model, precision, container, Megatron/TE versions, routing, and PP/VPP balance.
- The benchmark contains only one positive task, and this static review executed no training, configuration, or test path; it is not independent reproduction evidence.
- The evaluation report flags an international phone-number PII finding in card.yaml; review and remove unnecessary personal information before publication.
What does this skill do, and when should you use it?
This skill is for engineers planning MoE training on H100, B200, GB200, or GB300 systems. It organizes dispatcher choices, tensor, expert, and pipeline parallelism layouts, precision, recomputation, and CUDA Graph guidance by platform and model family. It also gives rounded throughput and MFU bands for planning rather than exact promises. The guidance highlights communication overlap, pipeline balance, CPU tuning, container quality, and memory headroom.
Provides representative MoE configurations for specified hardware and workloads; lists DeepEP or HybridEP, TP, EP, PP, VPP, precision, and recomputation combinations; gives planning bands for DSV3, Qwen3 235B, Qwen3 30B, and Qwen3-Next 80B; documents common CUDA Graph scopes and settings for CUDA_DEVICE_MAX_CONNECTIONS, PYTORCH_CUDA_ALLOC_CONF, and NCCL_GRAPH_REGISTER; and summarizes CPU-side tuning priorities and configuration pitfalls for GB200 and GB300. It supplies reference guidance and does not run training or benchmarks.
- An engineer training DSV3 on H100 needs a starting layout using DeepEP, TP=2, EP=64, PP=8, and VPP=4.
- A B200 team needs to assess MXFP8, DeepEP, pipeline layout, and communication tuning for DSV3.
- An engineer running DSV3 on GB200 or GB300 needs a HybridEP plan with CPU tuning and graph-friendly static shapes.
- A team planning Qwen3 235B training on H100 or GB200 needs to compare parallelism layouts and expected throughput bands.
- A training engineer facing MoE memory pressure needs a starting point for selective recomputation such as moe_act, mlp, or norm.
What are this skill's strengths and limitations?
- Covers H100, B200, GB200, and GB300 with concrete starting parallelism layouts.
- Combines configuration families, throughput planning bands, environment variables, and CPU-side tuning guidance.
- Explicitly warns against copying tracker rows without accounting for routing, containers, and pipeline layout.
- States the VPP uncertainty for Qwen3 235B on GB200 clearly.
- Throughput bands are rounded and cannot replace measurements on the target environment.
- Coverage is limited to the listed platforms and model families.
- The Qwen3 235B GB200 playbook leaves VPP unspecified, so it is not a complete layout.
- It is a reference playbook, not a set of training launch, deployment, or automated tuning commands.
How do you install this skill?
Install the specific skill with the skills CLI command documented in the repository README:
npx skills add nvidia/skills --skill nemo-mbridge-perf-moe-hardware-configs --yes
The README does not specify the exact installation directory; the CLI handles the destination.
How do you use this skill?
After installation, ask the relevant agent a hardware- or model-specific question, for example: “Give me an MoE hardware playbook for Qwen3 235B on GB200, including parallelism, CUDA Graph scopes, environment variables, and throughput caveats.” For Qwen3 235B on GB200, require VPP=unspecified unless measured data is available.