mcore-onboard-gb200-1node-tests

द्वारा nvidia

Onboard 1-node GitHub MR functional tests for GB200 from existing mr-scoped 2-node tests.

npx skills add https://github.com/nvidia/megatron-lm --skill mcore-onboard-gb200-1node-tests

Onboard GB200 1-Node GitHub MR Tests

Create 1-node (mr-github) variants of existing 2-node (mr-scoped) GB200 functional tests. Each GB200 node has 4 GPUs. A 2-node test uses 8 GPUs total; the 1-node variant uses 4.


Background

GB200 functional tests live in tests/test_utils/recipes/gb200/:

Recipe fileNotes
gpt.yamlGPT dense tests, nodes: 2, gpus: 4 (8 total)
moe.yamlMoE tests, nodes: 2, gpus: 4 (8 total)
moe-1node.yamlExisting 1-node MoE tests, nodes: 1, gpus: 4 (4 total)
gpt-1node.yaml1-node GPT tests (create if not present)

Model configs live at: tests/functional_tests/test_cases/{model}/{test_case}/model_config.yaml

1-node test cases use the _1node suffix: tests/functional_tests/test_cases/{model}/{test_case}_1node/model_config.yaml


Workflow

Step 1 — Find candidate tests

Scan the products: block in gpt.yaml and moe.yaml for entries with scope: [mr, ...] or scope: [mr-slim, ...]. These are the 2-node tests that need 1-node mr-github counterparts.

Ignore tests already covered in *-1node.yaml files, and ignore nightly, weekly, mr-broken scopes.

Step 2 — Read each model config

For each candidate, read its model_config.yaml and extract the key parallelism arguments:

--tensor-model-parallel-size   (TP)
--pipeline-model-parallel-size (PP)
--expert-model-parallel-size   (EP)
--expert-tensor-parallel-size  (ETP)
--context-parallel-size        (CP)
--global-batch-size
--micro-batch-size

Step 3 — Classify: trivial copy vs. needs adaptation

The world size formula is: world_size = TP × PP × DP where DP ≥ EP.

Going from 8 GPUs → 4 GPUs:

ConditionAction
TP × PP ≤ 4Trivial copy. Config unchanged; DP is halved automatically.
TP × PP = 8 (e.g. tp4 pp2)Reduce PP. Set PP = PP / 2 (e.g. pp2→1). Verify TP × PP_new ≤ 4.
EP > 4 (e.g. ep8 with tp1 pp1)Reduce EP. Set EP = 4. Experts stay at num-experts (each EP rank holds more experts).
EP > 4 and TP × PP > 4Reduce both PP and EP as above.
ETP test (ep × etp ≤ TP × DP)Check EP × ETP ≤ TP × DP_new after PP reduction. Usually satisfied when pp→1.

Do not change GBS — let gradient accumulation absorb the reduced DP.

Step 4 — Create _1node model config directories

# Trivial copy
mkdir -p tests/functional_tests/test_cases/{model}/{test_case}_1node
cp tests/functional_tests/test_cases/{model}/{test_case}/model_config.yaml \
   tests/functional_tests/test_cases/{model}/{test_case}_1node/model_config.yaml

# Then apply any parallelism changes (EP or PP) with Edit tool

Step 5 — Create or update recipe files

For GPT tests — create tests/test_utils/recipes/gb200/gpt-1node.yaml (if absent) by cloning gpt.yaml's spec block with nodes: 1. Use this template for the spec:

type: basic
format_version: 1
maintainers: [mcore]
loggers: [stdout]
spec:
  name: "{test_case}_{environment}_{platforms}"
  model: gpt          # or moe
  build: mcore-pyt-{environment}
  nodes: 1
  gpus: 4
  n_repeat: 5
  platforms: dgx_gb200
  script_setup: |    # copy verbatim from gpt.yaml / moe.yaml
    ...
  script: |-         # copy verbatim from gpt.yaml / moe.yaml
    ...

For MoE tests — append entries to the existing moe-1node.yaml.

Step 6 — Add products entries

Scope convention:

  • 1–2 most representative tests per recipe: scope: [mr-github, mr-github-slim]
  • All other tests: scope: [mr-github]
products:
  - test_case: [<test_case>_1node]
    products:
      - environment: [dev]
        scope: [mr-github, mr-github-slim]   # or [mr-github]
        platforms: [dgx_gb200]

Quick parallelism reference

Original (8 GPUs)1-node config (4 GPUs)Notes
tp1 pp1 ep1 → dp8tp1 pp1 ep1 → dp4trivial
tp2 pp1 ep1 → dp4tp2 pp1 ep1 → dp2trivial
tp1 pp2 ep1 → dp4tp1 pp2 ep1 → dp2trivial
tp4 pp1 ep1 → dp2tp4 pp1 ep1 → dp1trivial
tp1 pp4 ep1 → dp2tp1 pp4 ep1 → dp1trivial
tp1 pp1 ep8 → dp8tp1 pp1 ep4 → dp4ep 8→4
tp4 pp2 ep2 etp2 → dp1tp4 pp1 ep2 etp2 → dp1pp 2→1

Checklist

  • Identified all mr-scoped tests in gpt.yaml and moe.yaml not yet in *-1node.yaml
  • Read model config for each candidate
  • Classified trivial vs. adaptation needed
  • Created _1node/model_config.yaml for each test
  • Applied EP or PP reductions where needed
  • Created/updated recipe YAML with nodes: 1, gpus: 4
  • Assigned mr-github scope (+ mr-github-slim for 1–2 representative tests per recipe)
  • Verified no mr-github-slim overload (slim suite should stay small)

nvidia की और Skills

compileiq-debug
nvidia
Use when something is wrong: Search() hangs, all evaluations return INVALID_SCORE, scores aren't improving, every config returns the same number, ptxas errors…
official
create-github-pr
nvidia
Create GitHub pull requests using the gh CLI. Use when the user wants to create a new PR, submit code for review, or open a pull request. Trigger keywords -…
official
diagnose-perf
nvidia
First-responder performance triage for Isaac Sim and Isaac Lab. Identifies bottleneck category (GPU-bound, CPU-bound, VRAM, loading) using nvidia-smi and…
official
eagle3-review-logs
nvidia
Review EAGLE3 pipeline experiment logs from the launcher's experiments/ directory. Summarizes pass/fail status for all 4 tasks, diagnoses failures with root…
official
nemoclaw-maintainer-cross-issue-sweep
nvidia
Scans other open issues to find ones a given PR may also fix or accidentally break. Outputs adjacent-fix opportunities and contradiction risks with file:line…
official
karpathy-guidelines
nvidia
सामान्य LLM कोडिंग गलतियों को कम करने के लिए व्यवहार संबंधी दिशानिर्देश। कोड लिखते, समीक्षा करते या रिफैक्टर करते समय उपयोग करें ताकि अत्यधिक जटिलता से बचा जा सके, सर्जिकल बदलाव किए जा सकें,…
official
fhir-basics
nvidia
Teaches agents how FHIR R4 APIs work, what resources are available, how to query them with search parameters, and how to correctly parse all response formats…
official
underdeclared-agent
nvidia
A helpful assistant agent
official