Nemo Mbridge Perf Tp Dp Comm Overlap — Agent Skill by NVIDIA

From the author

Operational guide for enabling TP, DP, and PP communication overlap in Megatron-Bridge, including config knobs, code anchors, pitfalls, and verification.

— the description by NVIDIA

AutoGineer summary

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

This skill provides a comprehensive guide for enabling and tuning communication overlap (TP, DP, and PP) within the Megatron-Bridge framework. It covers essential configuration knobs, code-level gating mechanisms, common pitfalls to avoid, and verification procedures to ensure optimal performance during large-scale model training.

What it does

Provides configuration templates for enabling TP, DP, and PP communication overlap.
Identifies key code anchors that control overlap gating and P2P communication logic.
Highlights common pitfalls regarding sequence parallelism, Transformer Engine dependencies, and bucket sizing.
Details environment variable tuning for launch-time performance optimization.
Outlines verification steps using unit tests to confirm overlap configuration success.

When to use it

When configuring communication overlap for Megatron-Bridge training jobs.
When investigating throughput regressions related to communication overlap settings.
When troubleshooting why TP or PP overlap features are not activating as expected.
When tuning environment variables like CUDA_DEVICE_MAX_CONNECTIONS for performance.

Advantages

Centralizes configuration knowledge for complex parallel training setups.
Provides clear guidance on the interaction between model parallelism and overlap features.
Includes specific verification commands to validate configuration integrity.
Warns against common misconfigurations in mixed precision and environment settings.

Usage examples

“How do I enable TP communication overlap in my Megatron-Bridge config?”: The skill provides the CommOverlapConfig snippet to set tp_comm_overlap=True and notes the requirement for sequence_parallel=True and Transformer Engine availability.
“Why is my PP overlap not working?”: The guide explains that PP overlap is only auto-selected when pipeline_model_parallel_size > 1 and virtual_pipeline_model_parallel_size > 1.

Install

Download the skill as a ZIP:
https://mp-staging.autogineer.com/agent-skills/nvidia/nemo-mbridge-perf-tp-dp-comm-overlap/download

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

mkdir -p .agents/skills && unzip nemo-mbridge-perf-tp-dp-comm-overlap.zip -d .agents/skills

Documentation by NVIDIA

TP / DP / PP Communication Overlap Skill

For stable background and recommendation level, see:

@docs/training/communication-overlap.md

Enablement

Minimal Bridge override:

from megatron.bridge.training.comm_overlap import CommOverlapConfig

cfg.model.tensor_model_parallel_size = 4
cfg.model.sequence_parallel = True
cfg.model.pipeline_model_parallel_size = 4
cfg.model.virtual_pipeline_model_parallel_size = 2

cfg.comm_overlap = CommOverlapConfig(
    tp_comm_overlap=True,
)

cfg.ddp.use_distributed_optimizer = True
cfg.ddp.overlap_grad_reduce = True
cfg.ddp.overlap_param_gather = True

Optional TP preset:

from megatron.bridge.training.comm_overlap import userbuffers_bf16_h100_h12288_tp4_mbs1_seqlen2048

cfg.comm_overlap.tp_comm_overlap_cfg = userbuffers_bf16_h100_h12288_tp4_mbs1_seqlen2048

Precision knobs belong to mixed precision:

cfg.mixed_precision.grad_reduce_in_fp32 = False
cfg.mixed_precision.fp8_param_gather = False

Code Anchors

Bridge overlap gating:

if self.user_comm_overlap_cfg.tp_comm_overlap is True:
    if model_cfg.tensor_model_parallel_size < 2:
        ...
    elif not model_cfg.sequence_parallel:
        ...
    elif not HAVE_TE:
        ...

PP overlap selection:

if model_cfg.pipeline_model_parallel_size > 1:
    if vp_size > 1:
        comm_overlap_cfg.overlap_p2p_comm = True
        comm_overlap_cfg.batch_p2p_comm = False
    else:
        comm_overlap_cfg.overlap_p2p_comm = False
        comm_overlap_cfg.batch_p2p_comm = True

DP overlap defaults:

if self.data_parallel_size > 1:
    comm_overlap_cfg.bucket_size = 128 * 1024 * 1024
    comm_overlap_cfg.overlap_grad_reduce = True
    comm_overlap_cfg.overlap_param_gather = True

Launch-time env tuning:

executor.env_vars["CUDA_DEVICE_MAX_CONNECTIONS"] = str(cuda_device_max_connections)
...
executor.env_vars["NVTE_FWD_LAYERNORM_SM_MARGIN"] = str(self.layernorm_sm_margin)
executor.env_vars["NVTE_BWD_LAYERNORM_SM_MARGIN"] = str(self.layernorm_sm_margin)

Pitfalls

TP overlap silently disables itself if sequence_parallel=False or Transformer Engine is unavailable.
PP overlap is not enabled for all PP cases. Bridge only auto-selects overlap_p2p_comm=True when PP > 1 and VPP > 1.
bucket_size is a parameter-count knob, not a byte-size knob.
grad_reduce_in_fp32 and fp8_param_gather should be set through mixed precision, not as standalone DDP tuning first.
CUDA_DEVICE_MAX_CONNECTIONS and LayerNorm SM margin are launch-time plugin settings, not CommOverlapConfig fields.

Verification

Use the checked-in overlap unit coverage first:

uv run python -m pytest tests/unit_tests/training/test_comm_overlap.py -q

Optional second check if nemo_run is available:

uv run python -m pytest tests/unit_tests/recipes/test_run_plugins.py -q

Success criteria:

first command reports 26 passed
second command validates plugin-owned env wiring when not skipped

Files

ATTRIBUTION.mdAdded by AutoGineer871 B
BENCHMARK.md4.3 KiB
LICENSELicence file added by AutoGineer11.2 KiB
SKILL.md3.6 KiB
card.yaml2.1 KiB
evals/evals.json1.4 KiB
skill-card.md3.7 KiB
skill.oms.sig4.7 KiB

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

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.

Diagnosing Bugs
Matt Pocock· Community
A structured, six-phase discipline for systematically diagnosing and resolving complex software bugs and performance regressions.
Software developmentClaude CodeClaude appsOpenAI Codex+14 more
MITLow riskHas scriptsGitHub stars of the source repository317.5k
Systematic Debugging
Jesse Vincent (obra)· Community
A rigorous, four-phase framework for identifying and resolving technical issues by prioritizing root cause analysis over symptomatic fixes.
Software developmentClaude CodeClaude appsOpenAI Codex+14 more
MITLow riskHas scriptsGitHub stars of the source repository314k
Performance Optimization
Addy Osmani· Community
A rigorous, measurement-first workflow for identifying and fixing performance bottlenecks in frontend and backend applications.
Software developmentClaude CodeClaude appsOpenAI Codex+14 more
MITGitHub stars of the source repository113.9k
Debugging And Error Recovery
Addy Osmani· Community
A systematic, six-step framework for triaging, reproducing, and resolving software bugs to ensure root-cause fixes.
Software developmentClaude CodeClaude appsOpenAI Codex+14 more
MITGitHub stars of the source repository113.9k
Animation Performance
Emil Kowalski· Community
A diagnostic guide for identifying and fixing frame drops in web animations to ensure smooth, performant motion.
Software developmentClaude CodeClaude appsOpenAI Codex+14 more
MITGitHub stars of the source repository54.3k
Parallel Debugging
Seth Hobson· Community
A structured framework for debugging complex issues using parallel hypothesis testing and evidence-based arbitration.
Software developmentClaude CodeClaude appsOpenAI Codex+14 more
MITGitHub stars of the source repository42.3k