AutoGineer
Generated banner for Nemo Mbridge Perf Moe Vlm Training
Generated logo for Nemo Mbridge Perf Moe Vlm Training

Nemo Mbridge Perf Moe Vlm Training

NVIDIAOfficialPublished by NVIDIA

  • Skill
  • Guidelines
  • Data and AI
  • Apache-2.0
  • Medium risk
Download ZIP

From the author

Practical guidance for training MoE VLMs in Megatron Bridge. Compares FSDP and 3D-parallel approaches, using rounded lessons from Qwen3-VL, Qwen3-Next, and other multimodal experiments.
— the description by NVIDIA

AutoGineer summary

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

This skill provides a structured approach to training MoE VLMs using Megatron Bridge. It offers comparative analysis between FSDP and 3D-parallel strategies, drawing on empirical lessons from multimodal experiments like Qwen3-VL. The guide emphasizes the importance of using real data for performance profiling and provides a decision framework for selecting training configurations based on model stability and throughput requirements.

What it does

  • Compares FSDP and 3D-parallel training approaches for MoE VLMs.
  • Provides a decision guide for selecting the optimal training strategy based on model maturity.
  • Outlines key tuning knobs including vision stack freezing, MBS sweeping, and recompute strategies.
  • Identifies common pitfalls in VLM training, such as the unreliability of mock data.
  • Suggests representative configuration families for GB200-class systems.
  • Offers guidance on balancing memory usage and throughput for multimodal workloads.

When to use it

  • When bringing up a new MoE VLM for the first time.
  • When investigating training failures or OOM errors in multimodal MoE models.
  • When transitioning from initial model bring-up to steady-state throughput optimization.
  • When deciding between FSDP and 3D-parallel strategies for a specific VLM architecture.

Advantages

  • Provides actionable, experiment-backed performance patterns.
  • Helps avoid common pitfalls like misleading mock-data performance proxies.
  • Clarifies the trade-offs between ease of bring-up and peak throughput.
  • Offers specific configuration guidance for modern hardware like GB200.

Usage examples

“How should I approach the initial training bring-up for a new MoE VLM?”
The guide recommends starting with FSDP for its simplicity and memory-first tuning, moving to 3D parallel only once the model is stable and throughput headroom is required.
“Why is my VLM training performance different than expected?”
The skill suggests checking if you are using mock data, which is not a trustworthy performance proxy, and profiling the vision encoder, projector, and decoder separately.

Install

  1. Download the skill as a ZIP:

    https://mp-staging.autogineer.com/agent-skills/nvidia/nemo-mbridge-perf-moe-vlm-training/download

  2. Unzip it into the cross-client skills folder of your project:

    mkdir -p .agents/skills && unzip nemo-mbridge-perf-moe-vlm-training.zip -d .agents/skills

Documentation by NVIDIA

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.

Files

BENCHMARK.md

Evaluation Report

Evaluation of the nemo-mbridge-perf-moe-vlm-training skill before publication through NVSkills-Eval.

This benchmark summarizes 3-Tier Evaluation from NVSkills-Eval results for the skill. The goal is to document whether the skill is safe, discoverable, effective, and useful for agents before it is published for broader workflow use.

Evaluation Summary

  • Skill: nemo-mbridge-perf-moe-vlm-training

  • Evaluation date: 2026-06-02

  • NVSkills-Eval profile: external

  • Environment: local

  • Dataset: 1 evaluation tasks

  • Attempts per task: 2

  • Pass threshold: 50%

  • Overall verdict: PASS

Agents Used

  • claude-code

  • codex

Metrics Used

Reported benchmark dimensions:

  • Security: checks whether skill-assisted execution avoids unsafe behavior such as secret leakage, destructive commands, or unauthorized access.

  • Correctness: checks whether the agent follows the expected workflow and produces the correct final output.

  • Discoverability: checks whether the agent loads the skill when relevant and avoids using it when irrelevant.

  • Effectiveness: checks whether the agent performs measurably better with the skill than without it.

  • Efficiency: checks whether the agent uses fewer tokens and avoids redundant work.

Underlying evaluation signals used in this run:

  • security (Security): checks for unsafe operations, secret leakage, and unauthorized access.

  • skill_execution (Skill Execution): verifies that the agent loaded the expected skill and workflow.

  • skill_efficiency (Efficiency): checks routing quality, decoy avoidance, and redundant tool usage.

  • accuracy (Accuracy): grades final-answer correctness against the reference answer.

  • goal_accuracy (Goal Accuracy): checks whether the overall user task completed successfully.

  • behavior_check (Behavior Check): verifies expected behavior steps, including safety expectations.

  • token_efficiency (Token Efficiency): compares token usage with and without the skill.

Test Tasks

The benchmark dataset contained 1 evaluation tasks:

  • Positive tasks: 1 tasks where the skill was expected to activate.

  • Negative tasks: 0 tasks where no skill was expected.

  • Unlabeled tasks: 0 tasks where positive/negative intent could not be inferred.

Task composition is derived from the evaluation dataset when possible. Entries with expected_skill set are treated as positive skill-activation cases, while entries with expected_skill: null are treated as negative activation cases.

Results

DimensionNumclaude-codecodex
Security2100% (+0%)100% (+0%)
Correctness2100% (+0%)88% (+0%)
Discoverability2100% (+0%)62% (+0%)
Effectiveness290% (+1%)86% (-1%)
Efficiency293% (-0%)60% (-0%)

Score values show skill-assisted performance. Values in parentheses show uplift versus the no-skill baseline when baseline data is available.

Tier 1: Static Validation Summary

Tier 1 validation passed with observations. NVSkills-Eval ran 9 checks and found 10 total findings.

Top findings:

  • MEDIUM QUALITY/quality_correctness: SKILL_SPEC recommended field missing: 'metadata.author' (skills/nemo-mbridge-perf-moe-vlm-training/SKILL.md)

  • MEDIUM QUALITY/quality_correctness: SKILL_SPEC recommended field missing: 'metadata.tags' (skills/nemo-mbridge-perf-moe-vlm-training/SKILL.md)

  • MEDIUM SCHEMA/body_recommended_section: Missing recommended section: '## Instructions' (skills/nemo-mbridge-perf-moe-vlm-training/SKILL.md)

  • MEDIUM SCHEMA/body_recommended_section: Missing recommended section: '## Examples' (skills/nemo-mbridge-perf-moe-vlm-training/SKILL.md)

  • MEDIUM SCHEMA/author_missing: Author not specified in metadata (skills/nemo-mbridge-perf-moe-vlm-training/SKILL.md)

Tier 2: Deduplication Summary

Tier 2 validation passed. NVSkills-Eval ran 2 checks and found 0 total findings.

Notable observations:

  • Context Deduplication: Collected 1 file(s)

  • Inter-Skill Deduplication: Parsed skill 'nemo-mbridge-perf-moe-vlm-training': 185 char description

Publication Recommendation

The skill is suitable to proceed toward NVSkills-Eval publication based on this benchmark. Skill owners should keep this file with the skill and refresh it when the evaluation dataset, skill behavior, or target agents materially change.

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.

Security

What our scanner noticed· Medium risk

  • Large base64-encoded blobs

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
NVIDIA
Publisher
NVIDIA (Official)
Source
NVIDIA/skills · skills/nemo-mbridge-perf-moe-vlm-training · commit fce3cca
Licence
Apache License 2.0 (redistribution allowed)
  • LICENSE (added by AutoGineer from LICENSE-APACHE)
  • ATTRIBUTION.md (written by AutoGineer)
Synced

AutoGineer is not affiliated with or endorsed by NVIDIA.

Download ZIP

The ZIP holds the skill’s files, its licence files and an ATTRIBUTION.md. Download nemo-mbridge-perf-moe-vlm-training.zip