Training Nemotron-3-Nano-30B-A3B on H200

Author

ATTUNE (cite)

GPUs

8 x H200 141 GB (AWS p5en.48xlarge, 8 per node)

Layout

TP 1, PP 2 (even split), CP 1, EP 4, ETP 1, DP 4; one 8,192-token row per microbatch (THD, max 8,192 tokens/GPU); full recompute (uniform, 1 layer); distributed optimizer; alltoall dispatcher; no offload

Tokens/s per GPU

7,814.02

$ / 1M tokens

$0.28

MFU

15.3% (N = 3,227,369,472 active parameters per token)

Step-time CV

1.63%

Convergence check

correlation 0.9999997, final loss difference 4.3e-05, 0 pointwise failures over 20 steps

Measured

2026-10-15

GPU-hours

1.2

Rows

nemotron3nano-tulu3-fixed8192-v1 (rows_sha256 8fc04d3ce32c2b2cf8b1255a84cd8627df647cb38b6cbbe69a8e6bbde0ce54c7)

export CUDA_DEVICE_MAX_CONNECTIONS=1
export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
export TILELANG_CACHE_DIR="$KERNEL_CACHE/tilelang"
export TRITON_CACHE_DIR="$KERNEL_CACHE/triton"
export TORCHINDUCTOR_CACHE_DIR="$KERNEL_CACHE/inductor"
export FI_PROVIDER=efa
export FI_EFA_USE_DEVICE_RDMA=1

MODEL_ARGS_LINE="$(python3 miles/utils/external_utils/model_args_utils.py nemotron-3-nano-30b-a3b)"
read -ra MODEL_ARGS <<< "$MODEL_ARGS_LINE"
python3 train.py \
  --actor-num-nodes 1 \
  --actor-num-gpus-per-node 8 \
  --num-gpus-per-node 8 \
  "${MODEL_ARGS[@]}" \
  --hf-checkpoint "$HF_CKPT" \
  --ref-load "$HF_CKPT" \
  --megatron-to-hf-mode bridge \
  --loss-type sft_loss \
  --calculate-per-token-loss \
  --disable-compute-advantages-and-returns \
  --debug-train-only \
  --rollout-function-path pretokenized_sft_rollout.generate_rollout \
  --prompt-data "$SFT_DATA" \
  --input-key messages \
  --metadata-key metadata \
  --num-rollout 20 \
  --rollout-batch-size 128 \
  --global-batch-size 128 \
  --tensor-model-parallel-size 1 \
  --pipeline-model-parallel-size 2 \
  --context-parallel-size 1 \
  --expert-model-parallel-size 4 \
  --expert-tensor-parallel-size 1 \
  --recompute-granularity full \
  --recompute-method uniform \
  --recompute-num-layers 1 \
  --use-dynamic-batch-size \
  --max-tokens-per-gpu 8192 \
  --attention-backend auto \
  --optimizer adam --lr 1e-5 --lr-decay-style cosine --min-lr 1e-6 \
  --lr-warmup-fraction 0.1 --weight-decay 0.1 --adam-beta1 0.9 --adam-beta2 0.98 \
  --attention-dropout 0.0 --hidden-dropout 0.0 \
  --accumulate-allreduce-grads-in-fp32 --attention-softmax-in-fp32 \
  --distributed-timeout-minutes 30

# recipes/nemotron-3-nano-30b-a3b/h200/attune-tp1-pp2-ep4.yaml
# An OpenRecipes recipe; the fields are documented in schema/recipe.schema.yaml.
schema_version: 1
model: Nemotron-3-Nano-30B-A3B
gpu:
  type: H200
  memory_gb: 141
  count: 8
  per_node: 8
  instance: AWS p5en.48xlarge
author:
  name: ATTUNE
  url: https://arxiv.org/abs/2610.17528
status: verified
stack:
  framework: miles 0.1.0 (radixark/miles commit 78527b91, a slime fork), no local changes
  megatron_lm: radixark/Megatron-LM commit 37f91617 (megatron-core 0.16.0rc0), no local changes
  transformer_engine: 2.17.0
  pytorch: 2.11.0 (CUDA 13.0 build)
  cuda: '13.0'
  nccl: 2.28.9 (CUDA 13.0 build; printed by the measured run)
  deepep: 1.2.1 (installed; the alltoall dispatcher does not use it)
  python: 3.12.3
  network: 8 GPUs per node, EFA (aws-ofi-nccl 1.20), 3.2 Tb/s RDMA per node
  source: 'version dump of the measured run''s training image (same digest), 2026-10-13: installed TE, PyTorch, DeepEP,
    megatron-core and miles, Megatron-LM and miles git commits (clean trees), Python; NCCL from the measured run''s log'
workload:
  row_tokens: 8192
  rows: SFT conversations
  global_batch: 128
  precision: bf16
  dataset:
    name: Tulu 3 SFT mixture
    source: allenai/tulu-3-sft-mixture
    revision: b14afda60f1bbebe55d5d2fa1e4df5042f97f8be
    license: ODC-By 1.0
    rows_sha256: 8fc04d3ce32c2b2cf8b1255a84cd8627df647cb38b6cbbe69a8e6bbde0ce54c7
    build:
      name: nemotron3nano-tulu3-fixed8192-v1
      tokenizer: Nemotron-3-Nano-30B-A3B (revision bf77c3174f68ad409e1c2aa60daeb46e32d1c606; tokenizer.json sha256
        c6021eb6847e682f89aa52d5eb6e8c7d902a23acfc8137e25211cf84828f1592; chat_template.jinja sha256
        ab7813c3abdd9cb655905a410728b26c7884eca45ddfab8d9f931553485a7862)
      loss_mask: qwen3_5
      rows: 6400
      row_tokens: 8192
      padding_tokens: 0
      seed: 42
      packing: source conversations in a seeded global permutation, concatenated with an EOS token between them and cut
        into consecutive 8,192-token rows; rows without supervised targets are skipped
      supervised_token_ratio: 0.665
      parquet_sha256: 7c0d4c1bd9f678b0815801a876cb8c460972424707d9f6d75883c336e1f70895
      parquet_bytes: 85528499
      builder: data/generate_fixed_sft.py with --loss-mask-type qwen3_5 (data/README.md)
layout:
  label: TP1 PP2 EP4 DP4
  summary: TP 1, PP 2 (even split), CP 1, EP 4, ETP 1, DP 4; one 8,192-token row per microbatch (THD, max 8,192
    tokens/GPU); full recompute (uniform, 1 layer); distributed optimizer; alltoall dispatcher; no offload
  highlights: max 8,192 tokens/GPU, no offload
  tp: 1
  pp: 2
  vpp: 1
  cp: 1
  ep: 4
  etp: 1
  dp: 4
  sequence_parallel: false
  layer_split: null
  max_tokens_per_gpu: 8192
  rows_per_microbatch: 1
  recompute:
    granularity: full
    method: uniform
    num_layers: 1
  offload: null
  dispatcher: alltoall
result:
  tokens_per_s_per_gpu: 7814.02
  usd_per_million_tokens: 0.28
  step_time_cv_percent: 1.63
  mfu: 0.153
  mfu_note: null
  active_params:
    count: 3227369472
    source: https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16
    revision: bf77c3174f68ad409e1c2aa60daeb46e32d1c606
  measured: 2026-10-15
  steps_per_run: 20
  window:
  - 5
  - 19
  window_tokens: 15728640
  runs: 1
  profiler: false
  gpu_hours: 1.2
  gpu_hours_note: GPUs x (scheduler admission to finish), including startup; the same run is the candidate arm of the
    convergence check
correctness:
  result: pass
  summary: passed
  reference: 'miles'' layout for Nemotron-3-Nano-30B-A3B (`scripts/run_nemotron_3_nano.py`): TP 2 with sequence parallel,
    PP 2, EP 2, full recompute (uniform, 1 layer), Adam on the GPU, with max 8,192 tokens/GPU (one row) where the
    launcher''s GRPO rollouts use 1,024. Model arguments, rows and schedule are the same in both recipes.'
  configuration: pass
  component_tests: []
  convergence:
    steps: 20
    result: pass
    correlation: 0.9999997
    final_loss_difference: 4.3e-05
    pointwise_failures: 0
    note: 'candidate arm: the measured run; reference arm: a 20-step run of the reference recipe on the same image and
      rows; the parity probe off in both'
  rules: argument classes of ATTUNE's skill/scripts/correctness/rules/numerics_list.yaml (amazon-science/attune f2cb450);
    20-step check by Megatron-Bridge validate_convergence with its pipeline thresholds
env:
  CUDA_DEVICE_MAX_CONNECTIONS: '1'
  PYTORCH_CUDA_ALLOC_CONF: expandable_segments:True
  TILELANG_CACHE_DIR: $KERNEL_CACHE/tilelang
  TRITON_CACHE_DIR: $KERNEL_CACHE/triton
  TORCHINDUCTOR_CACHE_DIR: $KERNEL_CACHE/inductor
  FI_PROVIDER: efa
  FI_EFA_USE_DEVICE_RDMA: '1'
patches: []
run: |
  export CUDA_DEVICE_MAX_CONNECTIONS=1
  export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
  export TILELANG_CACHE_DIR="$KERNEL_CACHE/tilelang"
  export TRITON_CACHE_DIR="$KERNEL_CACHE/triton"
  export TORCHINDUCTOR_CACHE_DIR="$KERNEL_CACHE/inductor"
  export FI_PROVIDER=efa
  export FI_EFA_USE_DEVICE_RDMA=1

  MODEL_ARGS_LINE="$(python3 miles/utils/external_utils/model_args_utils.py nemotron-3-nano-30b-a3b)"
  read -ra MODEL_ARGS <<< "$MODEL_ARGS_LINE"
  python3 train.py \
    --actor-num-nodes 1 \
    --actor-num-gpus-per-node 8 \
    --num-gpus-per-node 8 \
    "${MODEL_ARGS[@]}" \
    --hf-checkpoint "$HF_CKPT" \
    --ref-load "$HF_CKPT" \
    --megatron-to-hf-mode bridge \
    --loss-type sft_loss \
    --calculate-per-token-loss \
    --disable-compute-advantages-and-returns \
    --debug-train-only \
    --rollout-function-path pretokenized_sft_rollout.generate_rollout \
    --prompt-data "$SFT_DATA" \
    --input-key messages \
    --metadata-key metadata \
    --num-rollout 20 \
    --rollout-batch-size 128 \
    --global-batch-size 128 \
    --tensor-model-parallel-size 1 \
    --pipeline-model-parallel-size 2 \
    --context-parallel-size 1 \
    --expert-model-parallel-size 4 \
    --expert-tensor-parallel-size 1 \
    --recompute-granularity full \
    --recompute-method uniform \
    --recompute-num-layers 1 \
    --use-dynamic-batch-size \
    --max-tokens-per-gpu 8192 \
    --attention-backend auto \
    --optimizer adam --lr 1e-5 --lr-decay-style cosine --min-lr 1e-6 \
    --lr-warmup-fraction 0.1 --weight-decay 0.1 --adam-beta1 0.9 --adam-beta2 0.98 \
    --attention-dropout 0.0 --hidden-dropout 0.0 \
    --accumulate-allreduce-grads-in-fp32 --attention-softmax-in-fp32 \
    --distributed-timeout-minutes 30
notes: |
  - Starting point: miles' Nemotron-3-Nano launcher, `scripts/run_nemotron_3_nano.py` at miles 0.1.0 (model page
    https://miles.radixark.com/docs/models/nemotron/nemotron-3-nano-moe). miles publishes it as a GRPO smoke test with
    no tokens/s figure. Ported to this workload with its layout (TP 2 with sequence parallel, PP 2, EP 2, full
    recompute) and max 8,192 tokens/GPU on 8 H200, it measured 5,229.2 tokens/s/GPU; ATTUNE tuned it to 7,814.0 (+49%).