What does this skill do, and when should you use it?
This skill is for engineers using Megatron-Bridge for distributed training. It provides configuration examples for tensor-parallel, data-parallel, and pipeline-parallel communication overlap. It identifies relevant Bridge code paths, launch-time environment settings, and failure conditions that can silently disable overlap. It also supplies unit-test commands for validating the configuration. Adopt it when enabling overlap or investigating a throughput regression tied to overlap settings.
Provides CommOverlapConfig examples for TP overlap, model-parallel sizes, distributed optimizer settings, and mixed-precision knobs; points to gating, pipeline-selection, data-parallel defaults, and launch-environment code in comm_overlap.py and run_plugins.py; explains prerequisites and pitfalls involving sequence parallelism, Transformer Engine, VPP, bucket size, and launch-time variables; and directs the user to run the overlap and optional recipe-plugin tests.
- A Megatron-Bridge engineer enabling TP overlap for a distributed large-model training run.
- A training maintainer checking whether distributed-optimizer gradient reduction and parameter gathering are overlapped.
- An operator using pipeline parallelism who needs to determine when Bridge selects `overlap_p2p_comm`.
- A performance investigator tracing a throughput regression to `overlap_param_gather`, `overlap_grad_reduce`, or related configuration changes.
How do you install this skill?
- Before execution, confirm that the Megatron-Bridge version, Transformer Engine, sequence parallelism, distributed optimizer, and TP/PP settings match the current environment; no automated rollback or recovery procedure is documented.
- Do not interpret bucket_size as a byte-size setting; PP overlap auto-selection depends on PP/VPP conditions, and TP overlap may silently disable itself when dependencies or sequence_parallel are missing.
- The BENCHMARK.md results are supplied as a static report and do not replace testing on the target hardware, versions, and workload; representative performance benchmarks and a PP smoke test should be added.
- Shell / CLI
- Local filesystem
Megatron-BridgeuvPythonpytestTransformer Engine for TP overlap
Install the specific skill with the repository's documented skills CLI command: npx skills add nvidia/skills --skill nemo-mbridge-perf-tp-dp-comm-overlap --yes. The CLI handles skill selection and installation destination; the source does not specify the resulting folder path.
tmp="$(mktemp -d)"
git clone --depth 1 https://github.com/NVIDIA/skills.git "$tmp"
mkdir -p ~/.claude/skills
cp -R "$tmp/skills/nemo-mbridge-perf-tp-dp-comm-overlap" ~/.claude/skills/
rm -rf "$tmp"Generated from the source repository and skill path; it copies only this skill's folder. If the author's install steps above differ, follow those first. To scope it to one project, replace ~/.claude/skills with that project's .claude/skills.
How do you use this skill?
Once installed, send your agent any of these to trigger it:
- Enable TP, DP, and PP communication overlap in Megatron-Bridge and check the sequence-parallel and VPP prerequisites.
In a compatible Agent Skills client, ask a targeted request such as: “Enable TP, DP, and PP communication overlap in Megatron-Bridge and check the sequence-parallel and VPP prerequisites.” Apply the provided Bridge configuration guidance, then run uv run python -m pytest tests/unit_tests/training/test_comm_overlap.py -q. The source defines 26 passed as the first command's success criterion. If nemo_run is available, optionally run the recipe-plugin test.
What are this skill's strengths and limitations?
- Covers TP, DP, and PP communication-overlap configuration.
- Includes copyable Python examples, source-code anchors, and failure-mode guidance.
- Defines a concrete unit-test command and expected result.
- TP overlap depends on sequence parallelism and Transformer Engine; otherwise it disables itself.
- Bridge auto-selects PP overlap only when PP is greater than 1 and VPP is greater than 1.
- The source provides no measured throughput uplift, platform matrix, or complete performance benchmark.
- Repository metadata says `NOASSERTION`, while this SKILL.md declares Apache-2.0; license scope should be checked before adoption.
How does this skill compare with similar options?
Side by side with related skills; every score comes from the same FSRS standard.
| Skill | FS score | Stars | Last updated | License |
|---|---|---|---|---|
| Megatron-Bridge Communication Overlap Guide this page ✓ NVIDIA · Official | 54 · Use with care | ★ 3.8k | today | Apache-2.0 |
| NeMo MBridge Parallelism Strategy Guide ✓ NVIDIA · Official | 58 · Recommended | ★ 3.8k | today | Apache-2.0 |
| NeMo MBridge MoE Expert-Parallel Overlap ✓ NVIDIA · Official | 57 · Use with care | ★ 3.8k | today | Apache-2.0 |
| Megatron FSDP Configuration and Performance Guide ✓ NVIDIA · Official | 54 · Use with care | ★ 3.8k | today | Apache-2.0 |
| NeMo MBridge Hierarchical Context Parallelism ✓ NVIDIA · Official | 51 · Use with care | ★ 3.8k | today | Apache-2.0 |
How did FollowSkills review this skill?
The skill provides configuration guidance, code anchors, pitfalls, and test commands; no malware, credential handling, covert exfiltration, or destructive defaults are evident. Scope is generally limited and transparent. Points are deducted because explicit user confirmation, rollback, data-flow disclosure, and failure isolation are absent, while applying training settings can have substantial resource effects.
The configuration examples, gating conditions, pitfalls, and verification commands are broadly consistent, including key conditions that disable TP overlap. Static review cannot reproduce execution; the referenced test contents are not supplied with the selected skill, and coverage of missing dependencies, abnormal inputs, and diagnostic failures is limited, so the score remains conservative under the static ceiling.
The description, when_to_use triggers, TP/DP/PP scenarios, and main non-fit conditions are reasonably clear for Megatron-Bridge performance users. Points are deducted because version, hardware, dependency, and input/output boundaries are not systematically specified, and Chinese-language guidance is absent. The core function does not entirely depend on unreachable overseas services, so no additional mainland-network deduction is applied.
The SKILL.md has clear sections for enablement, code anchors, pitfalls, and verification; the skill card states Apache 2.0 licensing, ownership, version, and known limitations, while BENCHMARK.md supplies an evaluation summary. Points are deducted for missing recommended Instructions and Examples sections, missing author/tags metadata, and incomplete installation, changelog, maintenance-owner, and update-path information.
The examples directly address the requested configuration and provide a concrete verification path; the supplied benchmark report indicates strong performance on its positive task. Static review cannot verify that report or representative outputs, and there is no benchmark-backed optimal-setting guidance or complete PP smoke test, so practical benefit and cross-environment completeness remain uncertain and the static effectiveness ceiling applies.
Line-numbered code anchors, the card.yaml evidence list, test commands, and the evaluation report provide useful audit trails and distinguish code-verified from documentation-only claims. Points are deducted because the cited source and test contents are not included, benchmark results lack independent reproduction details, and multiple corroborating evidence types cannot be independently confirmed.
Open a dimension to read why it scored that way
Evidence confidence：Low — Mostly static review, author material or a limited demo; useful for discovery, not high-risk decisions.
See the full review method →