Nemo Mbridge Perf Parallelism Strategies
NVIDIAOfficialPublished by NVIDIA
- Skill
- Guidelines
- Software development
- Apache-2.0
- Medium risk
AutoGineer summary
Written by AutoGineer from the skill’s own files, not by the author
This skill provides a structured approach to selecting and combining parallelism strategies (TP, DP, PP, CP, EP) within the Megatron Bridge framework. It covers sizing heuristics based on model size, hardware topology, and sequence length, while clarifying the critical distinction between dense and MoE parallelism meshes. It includes essential formulas for calculating minimum GPU requirements and data parallel sizes, alongside common pitfalls to avoid during configuration.
What it does
- Provides recommended parallelism configurations for dense and MoE models based on parameter count.
- Defines hardware topology constraints for TP, PP, and DP scaling.
- Offers guidance on scaling strategies for long-sequence models using CP and SP.
- Corrects common misconceptions regarding minimum GPU count calculations for MoE models.
- Details the interaction between dense and expert parallelism meshes.
- Lists configuration pitfalls and verification steps for combined parallelism.
When to use it
- When determining the initial parallelism strategy for a new model training run.
- When troubleshooting OOM errors or performance regressions related to parallelism settings.
- When sizing infrastructure for large-scale dense or MoE model training.
- When optimizing communication overhead by aligning parallelism with hardware topology.
Advantages
- Prevents common over-allocation errors in GPU sizing.
- Provides clear, actionable heuristics for complex MoE configurations.
- Emphasizes performance-critical constraints like NVLink domain boundaries.
- Includes specific formulas for calculating DP and EDP sizes.
Usage examples
- “How do I calculate the minimum number of GPUs for a 355B MoE model with PP=8 and EP=16?”
- The agent will apply the formula PP * max(TP * CP, EP * ETP) to provide the correct minimum GPU count, explicitly warning against the incorrect product-based calculation.
- “What is the recommended parallelism strategy for a 70B dense model on 64 GPUs?”
- The agent will suggest a starting configuration of TP=8, PP=4, and DP=2, while advising to profile the first iteration for memory and communication performance.
Install
Download the skill as a ZIP:
https://autogineer.com/agent-skills/nvidia/nemo-mbridge-perf-parallelism-strategies/downloadUnzip it into the cross-client skills folder of your project:
mkdir -p .agents/skills && unzip nemo-mbridge-perf-parallelism-strategies.zip -d .agents/skills
Documentation by NVIDIA
Parallelism Strategy Selection Skill
For stable background on each parallelism type, see:
@docs/parallelisms.md
@skills/nemo-mbridge-perf-parallelism-strategies/card.yaml
Decision by Model Size
Dense models
| Model size | GPUs | Recommended starting point |
|---|---|---|
| < 1B | 1-8 | DP only |
| 1-10B | 8-16 | TP=2-4 + DP |
| 10-70B | 16-64 | TP=4-8 + PP=2-4 + DP |
| 70-175B | 64-256 | TP=8 + PP=4-8 + DP |
| 175-500B | 256-1024 | TP=8 + PP=8-16 + CP=2 + DP |
MoE models
MoE parallelism differs from dense models. Because only a fraction of parameters are active per token, TP can often stay at 1 or 2 — the active parameter shard already fits on a single GPU. EP is the primary scaling dimension, with PP handling cross-node layer distribution.
| Model (total / active) | TP | PP | EP | Notes |
|---|---|---|---|---|
| OLMoE 7B / 1B | 1 | 1 | 8 | EP only, fits single node |
| Moonlight 16B / 3B | 2 | 1 | 8 | small TP for shared layers |
| DeepSeek-V2 236B / 21B | 1 | 4 | 32 | no TP at all |
| GLM-4.5 Air 106B / 12B | 1 | 4 | 8 | no TP at all |
| Qwen3 30B-A3B | 4 | 2 | 4 | |
| GLM-4.5 355B / 32B | 2 | 8 | 16 | |
| Qwen3 235B-A22B | 4 | 16 | 8 | CP=2 for pretrain |
| DeepSeek-V3 671B / 37B | 2 | 16 | 64 | TP=2, not 8 |
| Kimi-K2 1T | 2 | 16 | 32 |
Key patterns:
TP is sized by active params, not total params. A 671B MoE with 37B active needs far less TP than a 70B dense model.
EP scales with expert count. Common: EP = num_experts or num_experts / experts_per_gpu.
PP handles depth. Large MoE models use PP=8-16 across nodes.
ETP (expert tensor parallelism) is rarely used. Llama 4 is an exception (ETP=4).
These are starting points, not hard rules. Always profile the first iteration to verify memory and communication.
Decision by Hardware Topology
Single node with NVLink:
cfg.model.tensor_model_parallel_size = 8Multiple nodes with InfiniBand:
cfg.model.tensor_model_parallel_size = 8
cfg.model.pipeline_model_parallel_size = NLimited network (Ethernet):
cfg.model.tensor_model_parallel_size = 4
cfg.model.pipeline_model_parallel_size = MThe stable rule is: keep TP within a single NVLink domain. Use PP or DP for cross-node scaling. TP across nodes is almost always a performance loss.
Decision by Sequence Length
| Sequence length | Recommendation |
|---|---|
| < 2K | standard TP + PP + DP |
| 2K-8K | add SP (sequence_parallel=True) |
| 8K-32K | add CP=2 |
| 32K+ | add CP=4-8, consider a2a+p2p for large CP |
Combined Parallelism Enablement
3D parallelism (TP + PP + DP):
cfg.model.tensor_model_parallel_size = 4
cfg.model.pipeline_model_parallel_size = 4
cfg.model.sequence_parallel = True4D parallelism (TP + PP + CP + DP):
cfg.model.tensor_model_parallel_size = 8
cfg.model.pipeline_model_parallel_size = 8
cfg.model.context_parallel_size = 2
cfg.model.sequence_parallel = TrueMoE with EP + PP (e.g. DeepSeek-V2 236B on 128 GPUs):
cfg.model.tensor_model_parallel_size = 1
cfg.model.pipeline_model_parallel_size = 4
cfg.model.expert_model_parallel_size = 32
cfg.model.sequence_parallel = FalseMoE with small TP + PP + EP (e.g. DeepSeek-V3 671B on 256 GPUs):
cfg.model.tensor_model_parallel_size = 2
cfg.model.pipeline_model_parallel_size = 16
cfg.model.expert_model_parallel_size = 64
cfg.model.sequence_parallel = TrueDP size is always implicit:
data_parallel_size = world_size / (TP * PP * CP) # dense path
expert_data_parallel_size = world_size / (PP * EP * ETP) # MoE pathMinimum GPU Count
The minimum GPUs needed to run a config (i.e. with DP=1, EDP=1)
is not the product of all parallelism dimensions. The dense path uses
a TP*CP-mesh and the MoE path uses an EP*ETP-mesh, and within each PP
stage these two meshes share the same set of GPUs — they overlap, they
don't multiply. Only PP stages multiply (they're disjoint slices of the
model). So:
min_gpus = PP * max(TP * CP, EP * ETP)Common simplification (WRONG): PP * TP * CP * EP * ETP. This
over-allocates GPUs and shows up in many READMEs and slurm sizing tables.
Don't propagate it.
The decoupling of attention and MoE parallelism (different mesh shapes for the dense and expert paths sharing the same PP-stage GPUs) is detailed in Pangu Ultra MoE (arXiv:2504.14960).
Examples
| Config | Wrong (PP·TP·CP·EP·ETP) | Correct (PP·max(TP·CP, EP·ETP)) |
|---|---|---|
| PP=1, TP=2, CP=1, EP=8, ETP=1 | 16 | 8 (1 node) |
| PP=1, TP=4, CP=1, EP=8, ETP=1 | 32 | 8 (max(4, 8)) |
| PP=1, TP=2, CP=2, EP=8, ETP=1 | 32 | 8 (max(4, 8)) |
| PP=1, TP=2, CP=4, EP=8, ETP=1 | 64 | 8 (max(8, 8)) |
| PP=2, TP=2, CP=1, EP=8, ETP=1 | 32 | 16 (2 · max(2, 8)) |
| PP=1, TP=2, CP=1, EP=4, ETP=2 | 16 | 8 (max(2, 8)) |
Scaling above the minimum
Adding GPUs scales DP and/or EDP (the world_size must satisfy
both equations simultaneously). At min_gpus the larger-mesh side has
DP (or EDP) = 1 and the smaller side absorbs the slack.
Example — TP=2, CP=1, EP=8, ETP=1, PP=1:
8 GPUs (
min_gpus): denseDP = 8/2 = 4, MoEEDP = 8/8 = 116 GPUs: dense
DP = 8, MoEEDP = 2→ 2× global batch32 GPUs: dense
DP = 16, MoEEDP = 4→ 4× global batch
When sizing slurm scripts, compute --nodes from min_gpus (or a
multiple of it for higher throughput via DP/EDP).
When answering MoE sizing prompts, include this checklist:
compute
min_gpus = PP * max(TP * CP, EP * ETP)with the requested valuesexplicitly reject the wrong
PP * TP * CP * EP * ETPfull productgive both DP formulas: dense
world_size / (TP * PP * CP)and MoEworld_size / (PP * EP * ETP)mention TP topology, SP, CP divisibility, and long-sequence CP guidance
Memory Estimation
Without parallelism (70B model, FP16):
parameters: 140 GB
gradients: 140 GB
optimizer states: 280 GB (Adam)
activations: 48 GB (batch=1, seq=4K)
total: 608 GBWith TP=4, PP=4, DP=4 (64 GPUs):
parameters: 8.75 GB per GPU
gradients: 8.75 GB per GPU
optimizer states: 17.50 GB per GPU
activations: 3.00 GB per GPU
total: ~38 GB per GPUCode Anchors
Parallelism dimensions set in model provider:
model_config = GPTModelProvider(
tensor_model_parallel_size=2,
# ... other model parameters
)DP size calculation:
data_parallel_size = world_size / (tensor_model_parallel_size × pipeline_model_parallel_size × context_parallel_size)Bridge initialization wires parallelism into process groups:
parallel_state.initialize_model_parallel(
tensor_model_parallel_size=model_config.tensor_model_parallel_size,
pipeline_model_parallel_size=model_config.pipeline_model_parallel_size,
...
context_parallel_size=model_config.context_parallel_size,
hierarchical_context_parallel_sizes=model_config.hierarchical_context_parallel_sizes,
expert_model_parallel_size=model_config.expert_model_parallel_size,
...
)Pitfalls
TP across nodes destroys throughput. Always keep TP within a single NVLink domain.
PP without interleaving has large pipeline bubbles. Use
virtual_pipeline_model_parallel_sizewhen possible.SP requires
tensor_model_parallel_size > 1. Enabling SP alone without TP is a config error.CP requires
seq_length % (2 * context_parallel_size) == 0.EP is only for MoE models. Setting
expert_model_parallel_sizeon a dense model is a no-op or error.The model-size-to-parallelism table above is a starting heuristic. Always profile the first iteration to check memory and communication.
CUDA_DEVICE_MAX_CONNECTIONSand related env vars interact with overlap settings. See @skills/nemo-mbridge-perf-tp-dp-comm-overlap/SKILL.md.The minimum GPU count for an MoE config is
PP * max(TP*CP, EP*ETP), not the product of all dimensions. The denseTP*CP-mesh and MoEEP*ETP-mesh share the same GPUs in each PP stage. See "Minimum GPU Count" section above.
Verification
Quick sanity check that combined parallelism initializes correctly using the smallest available recipe with overridden parallelism:
CUDA_VISIBLE_DEVICES=0,1,2,3 uv run python -m torch.distributed.run --nproc_per_node=4 \
scripts/training/run_recipe.py \
--recipe llama32_1b_pretrain_config \
model.tensor_model_parallel_size=2 \
model.pipeline_model_parallel_size=2 \
model.sequence_parallel=True \
train.train_iters=3 train.global_batch_size=8 train.micro_batch_size=1 \
scheduler.lr_warmup_iters=0 \
validation.eval_iters=0 validation.eval_interval=0 \
checkpoint.save_interval=0 \
logger.log_interval=1Success criteria:
exit code 0
finite loss at iteration 3 (e.g.
lm loss: 1.003808E+01)log shows TP=2 PP=2 DP=1 layout with 4 ranks
Files
- ATTRIBUTION.mdAdded by AutoGineer879 B
- BENCHMARK.md4.3 KiB
- LICENSELicence file added by AutoGineer11.2 KiB
- SKILL.md9.4 KiB
- card.yaml2.9 KiB
- evals/evals.json1.5 KiB
- skill-card.md3.7 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 Python, 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-parallelism-strategies · 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 or NVIDIA.
Download ZIP
The ZIP holds the skill’s files, its licence files and an ATTRIBUTION.md. Download nemo-mbridge-perf-parallelism-strategies.zip