Nemo Mbridge Perf Moe Optimization Workflow

NVIDIAOfficialPublished by NVIDIA

Skill
Guidelines
Software development
Apache-2.0
Medium risk

From the author

Evidence-gated workflow for MoE performance optimization in Megatron Bridge. Covers measurement contracts, the Three Walls framework, parallel folding, profiling, matched A/B tuning, and final validation.

— the description by NVIDIA

AutoGineer summary

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

This skill provides a rigorous, iterative methodology for tuning MoE training performance. It centers on the Three Walls framework (memory, communication, compute-efficiency) to diagnose bottlenecks and mandates a strict measurement contract to ensure valid A/B testing. The workflow guides users through fitting the model, scaling parallelism, profiling bottlenecks, and applying targeted optimizations like dispatcher tuning, overlap, and precision adjustments, concluding with a validation phase to confirm performance gains.

What it does

Defines a measurement contract to ensure consistent, valid performance comparisons.
Provides a step-by-step sequence for fitting, scaling, and tuning MoE models.
Categorizes performance bottlenecks into memory, communication, host, and compute walls.
Offers guidance on Parallel Folding to decouple attention and MoE parallelism.
Details how to validate optimizations using matched A/B profiles and steady-state metrics.
Outlines best practices for dispatcher selection, overlap, and precision tuning.

When to use it

When performing a full MoE throughput tuning sweep.
When diagnosing a throughput regression after a configuration or code change.
When needing to resolve memory, communication, or compute bottlenecks in MoE training.
When preparing evidence for performance optimization claims.

Advantages

Reduces trial-and-error by enforcing a logical, evidence-based optimization order.
Prevents invalid comparisons through strict measurement contract requirements.
Provides specific strategies for complex MoE-specific challenges like Parallel Folding.
Helps distinguish between training-equivalent and benchmark-only optimizations.

Usage examples

“How do I start optimizing the throughput of my MoE model?”: The agent will guide you to first freeze your measurement contract, then ensure the model fits in memory, scale parallelism, and finally profile the dominant bottleneck.
“My MoE training run is hitting a communication wall. What should I do?”: The agent will suggest investigating dispatcher backends like DeepEP or HybridEP, and testing overlap configurations based on your specific topology.
“How do I validate that my MoE performance optimization actually worked?”: The agent will instruct you to run at least 50 steps, verify that the intended kernels were active, and compare steady-state metrics against a matched baseline.

Install

Download the skill as a ZIP:
https://autogineer.com/agent-skills/nvidia/nemo-mbridge-perf-moe-optimization-workflow/download

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

mkdir -p .agents/skills && unzip nemo-mbridge-perf-moe-optimization-workflow.zip -d .agents/skills

Documentation by NVIDIA

MoE Training Optimization Workflow

Stable docs: @docs/training/moe-optimization.md Card: @skills/nemo-mbridge-perf-moe-optimization-workflow/card.yaml Source: Scalable Training of MoE Models with Megatron Core

Quick Reference

Start with the paper's Three Walls:

memory wall
communication wall
compute-efficiency wall

For operational diagnosis, split the compute-efficiency wall into compute and host/launch bottlenecks. They need different evidence and different fixes. MoE tuning is iterative, so use this order:

freeze the measurement contract -> fit -> scale -> profile -> retune -> validate

First Answer Checklist

For MoE optimization workflow prompts, present the response in this order:

Freeze the measurement contract: record the exact model and task, hardware and topology, container and commits, data and routing semantics, precision, sequence and batch shape, parallelism, graph scopes, and the steady-state metric window. Label each candidate as training-equivalent or benchmark-only.
Fit: make the model memory-feasible first. Use the smallest model parallelism that fits, prefer selective recompute before full recompute, add offloading only after recompute and parallelism are insufficient, and use --fake-init-process-group to sanity-check large layouts.
Scale: maximize DP after the model fits, keep hot communication inside the fastest interconnect, use PP plus VPP for multi-node scaling, prefer EP over extra TP for expert layers, and add CP when long context makes attention memory dominant.
Profile: identify the dominant wall: memory, communication, host overhead, or compute.
Retune: change one variable at a time based on the profiled bottleneck. Dispatcher, overlap, lower precision, CUDA graphs, and recompute are candidates, not hardware defaults.
Validate: use short matched screens to reject candidates, then run the winner for at least 50 steps. Verify the requested backend or graph replay actually ran, time a declared post-warmup window, and report loss health, skipped/NaN iterations, memory, step time, and model TFLOPS/GPU.
Include the exact Parallel Folding meshes: Attention: TP x CP x DP x PP and MoE: ETP x EP x EDP x PP.
Use alltoall for safe bring-up, then A/B flex + deepep and flex + hybridep when their packages and target topology support them. Start from BF16 and eager execution; introduce lower precision or the narrowest useful CUDA-graph scope only after profiling justifies it.

Phase 0: Freeze The Measurement Contract

A comparison is valid only when the following stay fixed unless they are the single variable under test:

Bridge, MCore, Transformer Engine, container, CUDA, and NCCL versions
GPU count, SKU, node topology, and launcher/environment settings
model, task, data path, sequence length, MBS, GBS, and optimizer settings
TP, PP, VPP, CP, EP, ETP, and DP layout
routing semantics, precision, recompute, dispatcher, overlap, and graph scope
warmup and steady-state timing windows

Separate two acceptance classes:

Training-equivalent changes preserve the intended routing, loss, data, optimizer, and checkpoint/resume behavior.
Benchmark-only changes such as forced load balancing are useful for controlled kernel studies, but cannot establish production-training correctness or convergence.

Phase 1: Make The Run Memory-Feasible

Start with a configuration that fits reliably before chasing throughput.

Recommended order:

Use the smallest amount of model parallelism that still fits.
Turn on selective recompute before falling back to full recompute.
Add offloading only when recompute and parallelism are still insufficient.
Use --fake-init-process-group to sanity-check large parallel layouts on a single GPU before burning cluster time.

Recompute guidance

Prefer selective recompute for MoE runs:

good first choices: layernorm, core_attn, moe_act, mlp, or model-specific modules (shared_experts, mla_up_proj)
use full recompute only when the run still does not fit
revisit recompute after enabling CUDA graphs, because some graph scopes and full recompute paths do not mix well

As a rule of thumb, fine-grained recompute often recovers most of the needed memory while keeping throughput much closer to the non-recompute baseline than full-layer recompute does.

Phase 2: Choose Parallelism For Scale

Priority order:

Maximize DP once the model fits.
Keep the hot communication path inside the fast interconnect when possible.
Use PP, plus VPP if needed, for multi-node scaling.
Prefer EP over extra TP for expert layers.
Add CP for long context once sequence length makes attention memory dominant.

Parallel Folding

Parallel Folding decouples attention and MoE parallelism so you do not have to pick a single compromise layout:

Attention: TP × CP × DP × PP
MoE:       ETP × EP × EDP × PP

Key knobs:

--expert-model-parallel-size
--expert-tensor-parallel-size

Use it when attention prefers some TP or CP, but expert layers benefit from a larger EP degree than the dense layers can tolerate.

Phase 3: Profile The Dominant Bottleneck

Bottleneck	What it looks like	Primary fixes
Memory	Run fits only with aggressive full recompute or OOMs during warmup	selective recompute, FP8, offloading, better PP layout
Communication	Nsight shows large all-to-all or collective blocks	DeepEP or HybridEP, EP overlap, DP/TP overlap, better PP layout
Host overhead	GPU gaps, launch-bound traces, Python overhead	CUDA graphs, `--manual-gc`, higher MBS, CPU affinity tuning
Compute	Low SM utilization after comm and host issues are addressed	grouped GEMM, fusion work, FP8, dispatcher-specific kernel tuning

Profile overlap without misreading it

Use unprofiled steady iterations for the acceptance metric and a matched profile for causal explanation:

Change one overlap or dispatcher variable at a time; keep routing, graph scopes, parallelism, batch shape, and runtime fixed.
Build interval unions for communication kernels and compute kernels, then measure their intersection to quantify hidden communication.
Do not add kernel durations and call the result wall time. Concurrent kernels may run longer because of SM or bandwidth contention even while the exposed GPU-active union and end-to-end step time fall.
Corroborate the trace with dispatch/combine NVTX ranges, steady step time, model TFLOPS/GPU, loss finiteness, skipped/NaN counts, and peak memory.

On a controlled 16×H100 Qwen3 30B-A3B HybridEP run, plain EP overlap increased communication hidden by GEMM/attention from 0.11% to 36.55%. The unprofiled step fell from 24.7138s to 20.9920s and throughput rose from 244.039 to 287.305 model TFLOPS/GPU. delay_wgrad_compute remained disabled.

Phase 4: Retune From Evidence

Choose the smallest candidate that targets the profiled bottleneck and change one variable at a time.

Dispatcher And Overlap Guidance

Use dispatcher choice as a bottleneck fix, not as a hardware lookup table.

moe_token_dispatcher_type="alltoall": safest bring-up path, fine for smaller EP sizes
moe_token_dispatcher_type="flex" + moe_flex_dispatcher_backend="deepep": candidate when DeepEP is installed and communication is exposed
moe_token_dispatcher_type="flex" + moe_flex_dispatcher_backend="hybridep": topology-sensitive candidate on both NVL8 and NVL72 systems when HybridEP is installed

HybridEP plus plain EP overlap is the current measured winner for the canonical 16×H100 Qwen3 30B-A3B shape, while the canonical 256×H100 Qwen3 235B recipe uses standard alltoall plus overlap. Benchmark backend compatibility and throughput in the target container; neither GPU name nor EP degree determines the winner by itself.

If the all-to-all path is visible in profiles, combine dispatcher tuning with:

--overlap-moe-expert-parallel-comm
--overlap-grad-reduce
--tp-comm-overlap

Test plain EP overlap, shared-expert overlap, and delayed weight-gradient compute as separate candidates first. A combination can regress even when one component helped on another model.

Lower-Precision Candidate Matrix

Start with a verified BF16 baseline. Hardware capability only determines which lower-precision candidates are legal; it does not guarantee a speedup.

Platform	Candidate after BF16 is stable
Hopper	per-tensor, current-scaling, or blockwise FP8 supported by the target stack
Blackwell	MXFP8 or another supported FP8 recipe
Blackwell, speed-first exploration	NVFP4 after the BF16/FP8 path is stable

Keep the router in FP32. The largest wins usually come from expert GEMMs and other heavy matrix math, not from trying to quantize every small MoE component. Require logs or traces showing that the intended kernels ran, and judge the candidate by end-to-end steady step time rather than theoretical peak FLOPS.

CUDA Graphs For MoE

Use CUDA graphs only after a profile shows meaningful host/launch gaps. For dropless MoE, start with the narrowest partial TE-scoped graph candidate:

moe_router
moe_preprocess

Add attn only if it is supported for the model and improves the same matched stack. A successful capture is not evidence of a speedup, and a graph win can disappear after dispatcher, overlap, or precision changes.

This path keeps dynamic expert work outside the graph. Budget extra memory, verify that shapes remain static, confirm replay rather than capture alone, and time only post-capture iterations.

Use full-iteration graphs only for graph-friendly workloads such as drop-and-pad or tightly controlled static-shape experiments.

Related references:

@skills/nemo-mbridge-perf-cuda-graphs/SKILL.md
@docs/training/cuda-graphs.md
@docs/training/activation-recomputation.md

Phase 5: Validate And Package Evidence

Use 6–12 post-warmup iterations for inexpensive screening when the workload allows it. For the selected candidate, run at least 50 steps and report a fixed steady window such as steps 41–50. The final evidence bundle should contain:

exact command/config diff, commits, container, hardware, and topology
declared routing/data semantics and training-equivalent vs benchmark-only label
proof that the intended dispatcher, precision kernels, overlap, and graph replay were active
step time and model TFLOPS/GPU from the same unprofiled steady window
finite loss, skipped/NaN counts, peak memory, and checkpoint/optimizer-state validation when the production path requires it
matched A/B profile evidence for the claimed causal mechanism

Do not attribute the total gain of a final multi-change winner to one earlier A/B. For example, the Qwen3 overlap experiment isolated a rise from 244.039 to 287.305 TFLOPS/GPU; the later canonical recipe reached 299.352 after additional HybridEP tuning. They answer different questions.

Pitfalls

Do not optimize in the wrong order: fitting the model and selecting sane parallelism matter more than micro-optimizations.
Platform changes the limiting wall: H100-class runs often feel more communication-bound, while GB200 or GB300 runs often expose CPU or launch overhead earlier.
FP8 MFU can look misleadingly low: compare absolute throughput as well as MFU when switching precision modes.
CUDA graphs and recompute interact: TE-scoped graphs are usually paired with selective recompute, not blanket full recompute.
Parallel Folding is not optional at large scale: once attention and expert layers want clearly different layouts, a single shared TP or EP plan becomes a tax on both.
Summed kernel time is not exposed time: use interval unions and communication/compute intersection when validating overlap.
Benchmark-only semantics are not production acceptance: forced routing, synthetic data, or disabled optimizer/checkpoint paths must be disclosed and validated separately from training-equivalent results.
Feature activation needs evidence: a config dump is insufficient when a backend can fall back, a graph can capture without helping, or a lower- precision recipe can miss the intended kernels.

Last signature refresh: 2026-08-03.

Files

ATTRIBUTION.mdAdded by AutoGineer885 B
BENCHMARK.md4.8 KiB
LICENSELicence file added by AutoGineer11.2 KiB
SKILL.md12.7 KiB
card.yaml7.6 KiB
evals/evals.json3.6 KiB
skill-card.md4.3 KiB
skill.oms.sig4.7 KiB

LICENSE

Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.


                                 Apache License
                           Version 2.0, January 2004
                        http://www.apache.org/licenses/

   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION

   1. Definitions.

      "License" shall mean the terms and conditions for use, reproduction,
      and distribution as defined by Sections 1 through 9 of this document.

      "Licensor" shall mean the copyright owner or entity authorized by
      the copyright owner that is granting the License.

      "Legal Entity" shall mean the union of the acting entity and all
      other entities that control, are controlled by, or are under common
      control with that entity. For the purposes of this definition,
      "control" means (i) the power, direct or indirect, to cause the
      direction or management of such entity, whether by contract or
      otherwise, or (ii) ownership of fifty percent (50%) or more of the
      outstanding shares, or (iii) beneficial ownership of such entity.

      "You" (or "Your") shall mean an individual or Legal Entity
      exercising permissions granted by this License.

      "Source" form shall mean the preferred form for making modifications,
      including but not limited to software source code, documentation
      source, and configuration files.

      "Object" form shall mean any form resulting from mechanical
      transformation or translation of a Source form, including but
      not limited to compiled object code, generated documentation,
      and conversions to other media types.

      "Work" shall mean the work of authorship, whether in Source or
      Object form, made available under the License, as indicated by a
      copyright notice that is included in or attached to the work
      (an example is provided in the Appendix below).

      "Derivative Works" shall mean any work, whether in Source or Object
      form, that is based on (or derived from) the Work and for which the
      editorial revisions, annotations, elaborations, or other modifications
      represent, as a whole, an original work of authorship. For the purposes
      of this License, Derivative Works shall not include works that remain
      separable from, or merely link (or bind by name) to the interfaces of,
      the Work and Derivative Works thereof.

      "Contribution" shall mean any work of authorship, including
      the original version of the Work and any modifications or additions
      to that Work or Derivative Works thereof, that is intentionally
      submitted to Licensor for inclusion in the Work by the copyright owner
      or by an individual or Legal Entity authorized to submit on behalf of
      the copyright owner. For the purposes of this definition, "submitted"
      means any form of electronic, verbal, or written communication sent
      to the Licensor or its representatives, including but not limited to
      communication on electronic mailing lists, source code control systems,
      and issue tracking systems that are managed by, or on behalf of, the
      Licensor for the purpose of discussing and improving the Work, but
      excluding communication that is conspicuously marked or otherwise
      designated in writing by the copyright owner as "Not a Contribution."

      "Contributor" shall mean Licensor and any individual or Legal Entity
      on behalf of whom a Contribution has been received by Licensor and
      subsequently incorporated within the Work.

   2. Grant of Copyright License. Subject to the terms and conditions of
      this License, each Contributor hereby grants to You a perpetual,
      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
      copyright license to reproduce, prepare Derivative Works of,
      publicly display, publicly perform, sublicense, and distribute the
      Work and such Derivative Works in Source or Object form.

   3. Grant of Patent License. Subject to the terms and conditions of
      this License, each Contributor hereby grants to You a perpetual,
      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
      (except as stated in this section) patent license to make, have made,
      use, offer to sell, sell, import, and otherwise transfer the Work,
      where such license applies only to those patent claims licensable
      by such Contributor that are necessarily infringed by their
      Contribution(s) alone or by combination of their Contribution(s)
      with the Work to which such Contribution(s) was submitted. If You
      institute patent litigation against any entity (including a
      cross-claim or counterclaim in a lawsuit) alleging that the Work
      or a Contribution incorporated within the Work constitutes direct
      or contributory patent infringement, then any patent licenses
      granted to You under this License for that Work shall terminate
      as of the date such litigation is filed.

   4. Redistribution. You may reproduce and distribute copies of the
      Work or Derivative Works thereof in any medium, with or without
      modifications, and in Source or Object form, provided that You
      meet the following conditions:

      (a) You must give any other recipients of the Work or
          Derivative Works a copy of this License; and

      (b) You must cause any modified files to carry prominent notices
          stating that You changed the files; and

      (c) You must retain, in the Source form of any Derivative Works
          that You distribute, all copyright, patent, trademark, and
          attribution notices from the Source form of the Work,
          excluding those notices that do not pertain to any part of
          the Derivative Works; and

      (d) If the Work includes a "NOTICE" text file as part of its
          distribution, then any Derivative Works that You distribute must
          include a readable copy of the attribution notices contained
          within such NOTICE file, excluding those notices that do not
          pertain to any part of the Derivative Works, in at least one
          of the following places: within a NOTICE text file distributed
          as part of the Derivative Works; within the Source form or
          documentation, if provided along with the Derivative Works; or,
          within a display generated by the Derivative Works, if and
          wherever such third-party notices normally appear. The contents
          of the NOTICE file are for informational purposes only and
          do not modify the License. You may add Your own attribution
          notices within Derivative Works that You distribute, alongside
          or as an addendum to the NOTICE text from the Work, provided
          that such additional attribution notices cannot be construed
          as modifying the License.

      You may add Your own copyright statement to Your modifications and
      may provide additional or different license terms and conditions
      for use, reproduction, or distribution of Your modifications, or
      for any such Derivative Works as a whole, provided Your use,
      reproduction, and distribution of the Work otherwise complies with
      the conditions stated in this License.

   5. Submission of Contributions. Unless You explicitly state otherwise,
      any Contribution intentionally submitted for inclusion in the Work
      by You to the Licensor shall be under the terms and conditions of
      this License, without any additional terms or conditions.
      Notwithstanding the above, nothing herein shall supersede or modify
      the terms of any separate license agreement you may have executed
      with Licensor regarding such Contributions.

   6. Trademarks. This License does not grant permission to use the trade
      names, trademarks, service marks, or product names of the Licensor,
      except as required for reasonable and customary use in describing the
      origin of the Work and reproducing the content of the NOTICE file.

   7. Disclaimer of Warranty. Unless required by applicable law or
      agreed to in writing, Licensor provides the Work (and each
      Contributor provides its Contributions) on an "AS IS" BASIS,
      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
      implied, including, without limitation, any warranties or conditions
      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
      PARTICULAR PURPOSE. You are solely responsible for determining the
      appropriateness of using or redistributing the Work and assume any
      risks associated with Your exercise of permissions under this License.

   8. Limitation of Liability. In no event and under no legal theory,
      whether in tort (including negligence), contract, or otherwise,
      unless required by applicable law (such as deliberate and grossly
      negligent acts) or agreed to in writing, shall any Contributor be
      liable to You for damages, including any direct, indirect, special,
      incidental, or consequential damages of any character arising as a
      result of this License or out of the use or inability to use the
      Work (including but not limited to damages for loss of goodwill,
      work stoppage, computer failure or malfunction, or any and all
      other commercial damages or losses), even if such Contributor
      has been advised of the possibility of such damages.

   9. Accepting Warranty or Additional Liability. While redistributing
      the Work or Derivative Works thereof, You may choose to offer,
      and charge a fee for, acceptance of support, warranty, indemnity,
      or other liability obligations and/or rights consistent with this
      License. However, in accepting such obligations, You may act only
      on Your own behalf and on Your sole responsibility, not on behalf
      of any other Contributor, and only if You agree to indemnify,
      defend, and hold each Contributor harmless for any liability
      incurred by, or claims asserted against, such Contributor by reason
      of your accepting any such warranty or additional liability.

   END OF TERMS AND CONDITIONS

   APPENDIX: How to apply the Apache License to your work.

      To apply the Apache License to your work, attach the following
      boilerplate notice, with the fields enclosed by brackets "[]"
      replaced with your own identifying information. (Don't include
      the brackets!)  The text should be enclosed in the appropriate
      comment syntax for the file format. We also recommend that a
      file or class name and description of purpose be included on the
      same "printed page" as the copyright notice for easier
      identification within third-party archives.

   Copyright [yyyy] [name of copyright owner]

   Licensed under the Apache License, Version 2.0 (the "License");
   you may not use this file except in compliance with the License.
   You may obtain a copy of the License at

       http://www.apache.org/licenses/LICENSE-2.0

   Unless required by applicable law or agreed to in writing, software
   distributed under the License is distributed on an "AS IS" BASIS,
   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
   See the License for the specific language governing permissions and
   limitations under the License.

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 Node.js

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-optimization-workflow · commit fce3cca

Licence

Apache License 2.0 (redistribution allowed)

LICENSE (added by AutoGineer from LICENSE-APACHE)
ATTRIBUTION.md (written by AutoGineer)

Synced

2026-11-09

AutoGineer is not affiliated with or endorsed by NVIDIA or NVIDIA.

Download ZIP

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