nemo-mbridge-perf-hierarchical-context-parallel

作成者: nvidia

Megatron-Bridgeで階層的コンテキスト並列性を有効にするための運用ガイド。設定項目、コードの該当箇所、注意点、検証方法を含む。

npx skills add https://github.com/nvidia/skills --skill nemo-mbridge-perf-hierarchical-context-parallel

Hierarchical Context Parallel Skill

This skill covers hierarchical context parallelism: nested context-parallel process groups used by cp_comm_type="a2a+p2p" and configured with hierarchical_context_parallel_sizes.

For what hierarchical CP is, when to use it, and the decision tree (a2a+p2p vs pure a2a vs p2p), see:

  • @docs/training/hierarchical-context-parallel.md
  • @skills/nemo-mbridge-perf-hierarchical-context-parallel/card.yaml

Enablement

Minimal Bridge override:

cfg.model.context_parallel_size = 4
cfg.model.cp_comm_type = "a2a+p2p"
cfg.model.hierarchical_context_parallel_sizes = [2, 2]
cfg.dist.use_decentralized_pg = False

Required constraints:

  • prod(hierarchical_context_parallel_sizes) == context_parallel_size
  • seq_length % (2 * context_parallel_size) == 0
  • Transformer Engine >= 1.12.0

Code Anchors

Upstream config and validation:

context_parallel_size: int = 1
"""Splits network input along sequence dimension across GPU ranks."""

hierarchical_context_parallel_sizes: Optional[list[int]] = None
"""Degrees of the hierarchical context parallelism. Users should provide a list to specify 
   the sizes for different levels. Taking the a2a+p2p cp comm type as example, it contains
   groups of two levels, so the first value of the list indicates the group size of the a2a
   communication type, and the second value indicates the group size of the p2p communication
   type.
"""
if args.hierarchical_context_parallel_sizes:
    from numpy import prod
    assert args.context_parallel_size == prod(args.hierarchical_context_parallel_sizes)
if "a2a+p2p" in args.cp_comm_type:
    assert args.hierarchical_context_parallel_sizes is not None, \
    "--hierarchical-context-parallel-sizes must be set when a2a+p2p is used in cp comm"

Bridge MPU path:

parallel_state.initialize_model_parallel(
    ...
    context_parallel_size=model_config.context_parallel_size,
    hierarchical_context_parallel_sizes=model_config.hierarchical_context_parallel_sizes,
    ...
)
...
return ProcessGroupCollection.use_mpu_process_groups()

Bridge decentralized-PG path:

pg_collection = ProcessGroupCollection(
    ...
    cp=cp_pg,
    tp_cp=tp_cp_pg,
    hcp=None,
    ep=ep_pg,
    ...
)

Implementation Map

The code anchors above show the config declarations and argument validation.

Validation (MCore)

TransformerConfig.__post_init__ enforces that a2a+p2p requires HCP sizes and the product matches CP.

Process group creation

parallel_state.initialize_model_parallel creates hierarchical CP sub-groups when HCP sizes are provided via create_hierarchical_groups. Bridge currently gets those groups through the MPU-backed ProcessGroupCollection.

TE integration

TEDotProductAttention passes the hierarchical groups to Transformer Engine when a2a+p2p is used. Requires Transformer Engine >= 1.12.0.

Pitfalls

  1. Bridge HCP is MPU-only today: If use_decentralized_pg=True, Bridge initializes flat CP groups and leaves HCP unset.
  2. No checked-in Bridge recipe currently exercises HCP directly.
  3. Single-GPU load helpers clear hierarchical_context_parallel_sizes.
  4. Silent broken training on old stacks: If you use a2a+p2p without setting hierarchical_context_parallel_sizes, MCore now asserts. Older versions would silently disable CP communication, so each rank attended only to its local chunk and produced artificially high throughput with broken gradients.
  5. Product must match: prod(hierarchical_context_parallel_sizes) must exactly equal context_parallel_size. A mismatch triggers an assertion.
  6. Verify in logs: Look for the process group initialization output. You should see HIERARCHICAL_CONTEXT_PARALLEL_GROUPS being created. If you only see CONTEXT_PARALLEL_GROUP, HCP is not active.

Verification

No dedicated Bridge end-to-end test exists yet for HCP (see @skills/nemo-mbridge-perf-hierarchical-context-parallel/card.yaml follow_up_validation). Use the existing unit tests and log inspection instead.

Run the decentralized-PG unit test to confirm the flat-CP behavior is preserved:

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

For a manual smoke check, launch a 4-GPU run with a small recipe and cp_comm_type=a2a+p2p plus hierarchical_context_parallel_sizes=[2,2]:

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.context_parallel_size=4 \
  model.cp_comm_type=a2a+p2p \
  "model.hierarchical_context_parallel_sizes=[2,2]" \
  train.train_iters=2

Success criteria:

  • Logs show HIERARCHICAL_CONTEXT_PARALLEL_GROUPS being created
  • Training completes at least one step without error
  • If you only see CONTEXT_PARALLEL_GROUP, HCP is not active

nvidiaのその他のスキル

compileiq-debug
nvidia
何かがおかしいときに使用:Search()がハングする、すべての評価がINVALID_SCOREを返す、スコアが改善しない、すべての設定が同じ数値を返す、ptxasエラー…
create-github-pr
nvidia
gh CLIを使用してGitHubのプルリクエストを作成します。ユーザーが新しいPRを作成したい、コードをレビューに提出したい、またはプルリクエストを開きたい場合に使用します。トリガーキーワード -…
nemoclaw-maintainer-cross-issue-sweep
nvidia
他のオープンなIssueをスキャンし、特定のPRが修正する可能性があるものや、誤って壊す可能性があるものを見つけます。隣接修正の機会や矛盾リスクをfile:line…と共に出力します。
fhir-basics
nvidia
エージェントにFHIR R4 APIの動作方法、利用可能なリソース、検索パラメータを使ったクエリ方法、およびすべてのレスポンス形式を正しく解析する方法を教えます…
compileiq-validate-result
nvidia
検索が完了した後、かつスピードアップの申請やACFの発送の前に使用します。dump_results CSVを読み込み、トップK候補(単一目的)を抽出します…
changelog-audit
nvidia
リリース前にWarp CHANGELOG.mdを監査:失われたエントリを復元、ユーザー影響で並べ替え、エントリの文言を洗練、行折り返し、および(リリースブランチモードで)比較をバンプ…
maintain-dynamic-plugins
nvidia
NeMo Relayの動的プラグインローダー、マニフェスト、RustネイティブSDK、gRPCワーカープロトコル、PythonワーカーSDK、ドキュメント、テスト、およびリリースワークフローのカバレッジを維持する
dgx-diagnose
nvidia
一般的なDGX Station GB300の問題(CUDAクラッシュ、誤ったGPUターゲット、vLLM/SGLangコンテナのバグ、MIG状態の問題、NVLink/Fabric Managerエラーなど)を診断します。