nemo-mbridge-perf-moe-vlm-training

tarafından nvidia

Megatron Bridge'te MoE VLM'lerin eğitimi için pratik rehber. FSDP ve 3D-paralel yaklaşımlarını karşılaştırır, Qwen3-VL, Qwen3-Next ve…'den yuvarlanmış dersler kullanır.

npx skills add https://github.com/nvidia/skills --skill nemo-mbridge-perf-moe-vlm-training

MoE VLM Training

Stable docs: @docs/training/moe-optimization.md Card: @skills/nemo-mbridge-perf-moe-vlm-training/card.yaml

FSDP vs 3D Parallel

ApproachStrengthBest fit
FSDPSimplest path to a working multimodal runfirst bring-up, memory-first tuning, awkward PP boundaries
3D parallelHigher ceiling after tuningstable models with a clean PP layout and time for deeper sweeps

For MoE VLMs, the practical workflow is usually:

  1. get the first reliable run with FSDP
  2. stabilize real-data input, recompute, and memory behavior
  3. move to 3D parallel only if the throughput headroom is worth the extra work

Rounded Findings From Recent VLM Runs

Qwen3-VL class models

The main patterns were consistent across the tracker:

  • FSDP on GB200-class systems can already reach healthy high-teens utilization with a comparatively simple setup
  • B200 FSDP runs are viable, but more sensitive to recompute choice and frozen vision settings
  • 3D parallel can recover to a similar or better operating point, but only after tuning MBS, recompute, and the real vision path together

Real data vs mock data

Mock-data VLM runs are not trustworthy performance proxies. In the experiments, image-free mock runs looked closer to "roughly twice as fast" than "slightly optimistic" when compared with real multimodal input.

Use real or realistic image payloads before drawing any conclusion about VLM throughput.

Smaller multimodal MoE runs

The smaller Qwen3.5-style multimodal experiments reinforce the same lessons:

  • HybridEP is a solid default on GB200
  • TE-scoped CUDA graphs help once the training loop is stable
  • larger MBS can pay off, but only if the vision encoder does not become the next bottleneck

Decision Guide

Choose FSDP when

  • you are bringing up a new VLM for the first time
  • the model has awkward stage boundaries across embedding, vision, and decoder
  • memory fit matters more than absolute throughput
  • you may freeze the vision stack during decoder-focused tuning

Choose 3D parallel when

  • the model is already stable under FSDP
  • the PP layout is clear and repeatable
  • you can sweep MBS, recompute, and CUDA-graph scope together
  • the goal is best steady-state throughput, not easiest bring-up

Key Tuning Knobs

  1. Freeze the vision stack when appropriate: if the work is decoder-focused, freezing the vision side often gives a small but real throughput gain and reduces memory pressure.

  2. Sweep MBS aggressively: VLMs are more MBS-sensitive than text-only MoE runs because the vision path changes the compute-to-overhead balance.

  3. Prefer selective recompute once the model fits: full recompute is a useful bring-up tool, but selective recompute is usually the better steady state.

  4. Match CUDA-graph scope to the workload: attn moe_router moe_preprocess is the safer MoE default, while narrower scopes can still be useful for controlled experiments.

  5. Use ETP only when EP alone is insufficient: it can unlock a layout, but it also introduces more communication and more tuning surface.

Representative Config Families

FSDP-first GB200 path

TP=1  CP=1  PP=1
EP sized to the expert topology, often large
Dispatcher: HybridEP on GB200-class systems
Recompute: start with full, then relax toward selective recompute

3D-parallel GB200 path

TP=1  CP=1  PP=1 or modest PP
EP and ETP sized to the expert topology
Dispatcher: HybridEP
CUDA Graph: start narrow, then widen only after the real-data path is stable

Compatibility

FeatureFSDP3D parallel
HybridEP on GB200strong defaultstrong default once topology is stable
CUDA graphsuseful after bring-upuseful, but more scope-sensitive
Freeze visionnatural fitpossible, but less often used as the headline perf path
Selective recomputerecommendedrecommended

Pitfalls

  1. Mock multimodal data is misleading: it can make the decoder look much healthier than the real end-to-end VLM path.

  2. The vision encoder can dominate unexpectedly: profile encoder, projector, and decoder separately before attributing everything to the dispatcher.

  3. Do not compare FSDP and 3D-parallel runs with different effective work: normalize by useful tokens and workload shape, not only by step time.

  4. ETP is not free: use it as a fit or topology tool, not as the default.

  5. Recompute and CUDA-graph choices are coupled: the setting that gets the model to fit is often not the setting that gives the best steady-state speed.

nvidia tarafından daha fazla skill

compileiq-debug
nvidia
Bir şeyler yanlış olduğunda kullanın: Search() takılıyor, tüm değerlendirmeler INVALID_SCORE döndürüyor, puanlar iyileşmiyor, her yapılandırma aynı sayıyı döndürüyor, ptxas hataları…
create-github-pr
nvidia
gh CLI kullanarak GitHub pull request'leri oluşturun. Kullanıcı yeni bir PR oluşturmak, kodu incelemeye göndermek veya bir pull request açmak istediğinde kullanın. Tetikleyici anahtar kelimeler -…
nemoclaw-maintainer-cross-issue-sweep
nvidia
Diğer açık sorunları tarayarak, belirli bir PR'ın da düzeltebileceği veya yanlışlıkla bozabileceği sorunları bulur. Dosya:satır… ile bitişik düzeltme fırsatlarını ve çelişki risklerini çıktı olarak verir.
fhir-basics
nvidia
Ajanlara FHIR R4 API'lerinin nasıl çalıştığını, hangi kaynakların mevcut olduğunu, arama parametreleriyle nasıl sorgulanacağını ve tüm yanıt formatlarının nasıl doğru şekilde ayrıştırılacağını öğretir…
compileiq-validate-result
nvidia
Bir arama tamamlandıktan SONRA ve herhangi bir hızlandırma talep etmeden veya bir ACF göndermeden ÖNCE kullanın. dump_results CSV dosyasını yükler, en iyi K adayı (tek amaçlı) çıkarır…
changelog-audit
nvidia
Bir sürüm öncesinde Warp CHANGELOG.md dosyasını denetle: kayıp girdileri kurtar, kullanıcı etkisine göre sırala, girdi dilini iyileştir, satır kaydırma yap ve (sürüm dalı modunda) karşılaştırmayı artır…
maintain-dynamic-plugins
nvidia
NeMo Relay dinamik eklenti yükleyicilerini, manifestolarını, Rust yerel SDK'larını, gRPC işçi protokolünü, Python işçi SDK'sını, dokümantasyonu, testleri ve sürüm iş akışı kapsamını korur
dgx-diagnose
nvidia
Yaygın DGX Station GB300 sorunlarını teşhis edin — CUDA çökmeleri, yanlış GPU hedefleme, vLLM/SGLang konteyner hataları, MIG durumu sorunları, NVLink/Fabric Manager hataları,…