nemo-mbridge-perf-parallelism-strategies

bởi nvidia

Hướng dẫn vận hành để lựa chọn và kết hợp các chiến lược song song trong Megatron Bridge, bao gồm quy tắc định cỡ, ánh xạ cấu trúc liên kết phần cứng và kết hợp…

npx skills add https://github.com/nvidia/skills --skill nemo-mbridge-perf-parallelism-strategies

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 sizeGPUsRecommended starting point
< 1B1-8DP only
1-10B8-16TP=2-4 + DP
10-70B16-64TP=4-8 + PP=2-4 + DP
70-175B64-256TP=8 + PP=4-8 + DP
175-500B256-1024TP=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)TPPPEPNotes
OLMoE 7B / 1B118EP only, fits single node
Moonlight 16B / 3B218small TP for shared layers
DeepSeek-V2 236B / 21B1432no TP at all
GLM-4.5 Air 106B / 12B148no TP at all
Qwen3 30B-A3B424
GLM-4.5 355B / 32B2816
Qwen3 235B-A22B4168CP=2 for pretrain
DeepSeek-V3 671B / 37B21664TP=2, not 8
Kimi-K2 1T21632

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 = 8

Multiple nodes with InfiniBand:

cfg.model.tensor_model_parallel_size = 8
cfg.model.pipeline_model_parallel_size = N

Limited network (Ethernet):

cfg.model.tensor_model_parallel_size = 4
cfg.model.pipeline_model_parallel_size = M

The 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 lengthRecommendation
< 2Kstandard TP + PP + DP
2K-8Kadd SP (sequence_parallel=True)
8K-32Kadd 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 = True

4D 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 = True

MoE 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 = False

MoE 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 = True

DP size is always implicit:

data_parallel_size = world_size / (TP * PP * CP)        # dense path
expert_data_parallel_size = world_size / (PP * EP * ETP) # MoE path

Minimum 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

ConfigWrong (PP·TP·CP·EP·ETP)Correct (PP·max(TP·CP, EP·ETP))
PP=1, TP=2, CP=1, EP=8, ETP=1168 (1 node)
PP=1, TP=4, CP=1, EP=8, ETP=1328 (max(4, 8))
PP=1, TP=2, CP=2, EP=8, ETP=1328 (max(4, 8))
PP=1, TP=2, CP=4, EP=8, ETP=1648 (max(8, 8))
PP=2, TP=2, CP=1, EP=8, ETP=13216 (2 · max(2, 8))
PP=1, TP=2, CP=1, EP=4, ETP=2168 (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): dense DP = 8/2 = 4, MoE EDP = 8/8 = 1
  • 16 GPUs: dense DP = 8, MoE EDP = 2 → 2× global batch
  • 32 GPUs: dense DP = 16, MoE EDP = 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 values
  • explicitly reject the wrong PP * TP * CP * EP * ETP full product
  • give both DP formulas: dense world_size / (TP * PP * CP) and MoE world_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 GB

With 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 GPU

Code 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

  1. TP across nodes destroys throughput. Always keep TP within a single NVLink domain.

  2. PP without interleaving has large pipeline bubbles. Use virtual_pipeline_model_parallel_size when possible.

  3. SP requires tensor_model_parallel_size > 1. Enabling SP alone without TP is a config error.

  4. CP requires seq_length % (2 * context_parallel_size) == 0.

  5. EP is only for MoE models. Setting expert_model_parallel_size on a dense model is a no-op or error.

  6. The model-size-to-parallelism table above is a starting heuristic. Always profile the first iteration to check memory and communication.

  7. CUDA_DEVICE_MAX_CONNECTIONS and related env vars interact with overlap settings. See @skills/nemo-mbridge-perf-tp-dp-comm-overlap/SKILL.md.

  8. The minimum GPU count for an MoE config is PP * max(TP*CP, EP*ETP), not the product of all dimensions. The dense TP*CP-mesh and MoE EP*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=1

Success 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

Thêm skills từ nvidia

compileiq-debug
nvidia
Sử dụng khi có điều gì đó không ổn: Search() bị treo, tất cả các đánh giá đều trả về INVALID_SCORE, điểm số không cải thiện, mọi cấu hình đều trả về cùng một số, lỗi ptxas…
create-github-pr
nvidia
Tạo pull request GitHub bằng cách sử dụng gh CLI. Sử dụng khi người dùng muốn tạo PR mới, gửi mã để xem xét, hoặc mở pull request. Từ khóa kích hoạt -…
nemoclaw-maintainer-cross-issue-sweep
nvidia
Quét các vấn đề đang mở khác để tìm những vấn đề mà một PR nhất định có thể sửa hoặc vô tình làm hỏng. Đưa ra các cơ hội sửa lỗi liền kề và rủi ro mâu thuẫn với file:dòng…
fhir-basics
nvidia
Dạy các tác nhân cách hoạt động của API FHIR R4, những tài nguyên có sẵn, cách truy vấn chúng với tham số tìm kiếm, và cách phân tích chính xác tất cả các định dạng phản hồi…
compileiq-validate-result
nvidia
Sử dụng SAU KHI tìm kiếm hoàn tất và TRƯỚC KHI yêu cầu tăng tốc hoặc gửi ACF. Tải tệp CSV dump_results, trích xuất các ứng viên top-K (đơn mục tiêu)…
changelog-audit
nvidia
Kiểm tra Warp CHANGELOG.md trước khi phát hành: khôi phục các mục bị mất, sắp xếp theo tác động người dùng, tinh chỉnh ngôn ngữ mục, xuống dòng và (chế độ nhánh phát hành) so sánh bump…
maintain-dynamic-plugins
nvidia
Duy trì các bộ nạp plugin động NeMo Relay, tệp kê khai, SDK gốc Rust, giao thức worker gRPC, SDK worker Python, tài liệu, kiểm thử và phạm vi quy trình phát hành
dgx-diagnose
nvidia
Chẩn đoán các sự cố thường gặp của DGX Station GB300 — lỗi CUDA, nhắm sai GPU, lỗi container vLLM/SGLang, vấn đề trạng thái MIG, lỗi NVLink/Fabric Manager,…