Training MiniMax-M2 on H200

Author

ATTUNE (cite)

GPUs

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

Layout

TP 1, PP 8 (first and last stage 7 layers, the others 8), VPP 4, CP 1, EP 8, ETP 1, DP 8; one 8,192-token row per microbatch; full recompute (uniform, 1 layer); precision-aware distributed optimizer; no offload; alltoall dispatcher

Tokens/s per GPU

2,034.42

$ / 1M tokens

$1.08

MFU

12.8% (N = 10,415,112,192 active parameters per token)

Step-time CV

1.57%

Convergence check

correlation 0.9997869, final loss difference 0.0018, 0 pointwise failures over 20 steps

Component test

precision-aware Adam on the GPU and CPU optimizer offload (fraction 1.0, overlapped copies), each against an FP64 reference: H200, 2026-10-11. Shape: one Qwen3.5-397B-A17B MoE layer: hidden 4096, 512 experts, top-10, expert FFN 1024, EP 8, ETP 1, 4,096 tokens per rank. worst tensor (update) relative L2 error 5.12e-7 on the GPU and 7.98e-7 offloaded against a bound of 1.0e-3 after one step, and 4.99e-5 and 7.72e-5 against 1.1e-3 after three steps at the recipe's hyperparameters (bound 2 x the default GPU optimizer's error + 1e-3); reconstructed FP32 masters bitwise identical to an arm with explicit FP32 masters, the TE remainder self-test passes and the negative controls fail run on this recipe's training image (the same image configuration); the optimizer step is elementwise, so the result holds for every model on the image

Measured

2026-10-21

GPU-hours

9.07

Rows

minimax-m2-tulu3-fixed8192-v1 (rows_sha256 f731270d13a2b34202cc0266c58a599318ff5957b15c3f9b9e9c1f71f9caf400)

export CUDA_DEVICE_MAX_CONNECTIONS=1
export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
export TORCHINDUCTOR_COMPILE_THREADS=1
export NCCL_NVLS_ENABLE=$([ "$NVLINK_COUNT" -gt 0 ] && echo 1 || echo 0)

python3 -m <slime entry module> --async \
  --actor-num-nodes 8 \
  --actor-num-gpus-per-node 8 \
  ${MODEL_ARGS[@]} \
  --hf-checkpoint "$HF_CKPT" \
  --ref-load "$TORCH_DIST" \
  --rollout-function-path pretokenized_sft_rollout.generate_rollout \
  --prompt-data "$SFT_DATA" \
  --input-key messages \
  --metadata-key metadata \
  --num-rollout 20 \
  --rollout-batch-size 128 \
  --global-batch-size 128 \
  --loss-type sft_loss \
  --calculate-per-token-loss \
  --disable-compute-advantages-and-returns \
  --debug-train-only \
  --seq-length 8192 \
  --tensor-model-parallel-size 1 \
  --pipeline-model-parallel-size 8 \
  --context-parallel-size 1 \
  --expert-model-parallel-size 8 \
  --expert-tensor-parallel-size 1 \
  --recompute-granularity full \
  --recompute-method uniform \
  --recompute-num-layers 1 \
  --use-dynamic-batch-size \
  --max-tokens-per-gpu 8192 \
  --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 \
  --use-distributed-optimizer \
  --use-precision-aware-optimizer \
  --attention-dropout 0.0 --hidden-dropout 0.0 \
  --accumulate-allreduce-grads-in-fp32 --attention-softmax-in-fp32 \
  --attention-backend flash \
  --distributed-timeout-minutes 30 \
  --tensorboard-dir "$PROFILE_DIR" \
  --pipeline-model-parallel-layout 'Et|(tt|)*30tL'

# recipes/minimax-m2/h200/attune-tp1-pp8-ep8-vpp4.yaml
# An OpenRecipes recipe; the fields are documented in schema/recipe.schema.yaml.
schema_version: 1
model: MiniMax-M2
gpu:
  type: H200
  memory_gb: 141
  count: 64
  per_node: 8
  instance: AWS p5en.48xlarge
author:
  name: ATTUNE
  url: https://arxiv.org/abs/2610.17528
status: verified
stack:
  framework: 'slime 0.3.0: THUDM/slime a897e1f plus 8 commits that do not change the training step'
  megatron_lm: commit 1dcf0daf (megatron-core 0.16.0rc0), with a backported upstream fix for the Qwen3.5 attention output
    gate at TP above the query-group count
  transformer_engine: 2.10.0
  pytorch: 2.11.0 (CUDA 12.9 build)
  cuda: '12.9'
  nccl: 2.28.9 (CUDA 12.9 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, 3.2 Tb/s RDMA per node
  source: 'in-pod version check of the measured run''s training image (same digest), 2026-10-22: installed TE, PyTorch,
    DeepEP and megatron-core, Megatron-LM git commit, Python, and slime''s MiniMax-M2 model and launcher scripts; 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: f731270d13a2b34202cc0266c58a599318ff5957b15c3f9b9e9c1f71f9caf400
    build:
      name: minimax-m2-tulu3-fixed8192-v1
      tokenizer: MiniMax-M2 (revision 757303d492a50514c312788b5247a4f696a4c6a3; tokenizer.json sha256 757622126525aeeb131756849d93298070ff3f0319c455ec8c5bb0f6b1cebbe8;
        chat_template.jinja sha256 86cd6e574dec52bf6d7e678c5b46ceb5223d68b3a1130bbbc09d0a2dc0909438)
      loss_mask: minimax_m2
      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: 0.64
      parquet_sha256: 2fb29ca63b79dc153defaacba1b41c783840de528a3ec353e5f6505144575036
      parquet_bytes: 75539521
      builder: data/generate_fixed_sft.py with --loss-mask-type minimax_m2 (data/README.md)
layout:
  label: TP1 PP8 VPP4 EP8 DP8
  summary: TP 1, PP 8 (first and last stage 7 layers, the others 8), VPP 4, CP 1, EP 8, ETP 1, DP 8; one 8,192-token row
    per microbatch; full recompute (uniform, 1 layer); precision-aware distributed optimizer; no offload; alltoall dispatcher
  highlights: VPP 4, no offload
  tp: 1
  pp: 8
  vpp: 4
  cp: 1
  ep: 8
  etp: 1
  dp: 8
  sequence_parallel: false
  layer_split:
    first: 7
    last: 7
  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: 2034.42
  usd_per_million_tokens: 1.08
  step_time_cv_percent: 1.57
  mfu: 0.1285
  mfu_note: null
  active_params:
    count: 10415112192
    source: https://huggingface.co/MiniMaxAI/MiniMax-M2
    revision: 757303d492a50514c312788b5247a4f696a4c6a3
  measured: 2026-10-21
  steps_per_run: 20
  window:
  - 5
  - 19
  window_tokens: 15728640
  runs: 1
  profiler: false
  gpu_hours: 9.07
  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; 1 component test passed on this image
  reference: 'The layout of slime''s `scripts/run-minimax-m2.sh` on 64 GPUs: TP 2 with sequence parallelism, PP 2, EP
    8 in place of its EP 4, ETP 1, full recompute (uniform, 1 layer), max 8,192 tokens/GPU (one row per microbatch) and
    CPU optimizer offload with overlapped copies and the precision-aware optimizer. Model arguments, rows and schedule
    are the same in both recipes.'
  configuration: pass
  component_tests:
  - test: precision-aware Adam on the GPU and CPU optimizer offload (fraction 1.0, overlapped copies), each against an
      FP64 reference
    shape: 'one Qwen3.5-397B-A17B MoE layer: hidden 4096, 512 experts, top-10, expert FFN 1024, EP 8, ETP 1, 4,096 tokens
      per rank'
    gpu: H200
    date: 2026-10-11
    result: pass
    detail: worst tensor (update) relative L2 error 5.12e-7 on the GPU and 7.98e-7 offloaded against a bound of 1.0e-3
      after one step, and 4.99e-5 and 7.72e-5 against 1.1e-3 after three steps at the recipe's hyperparameters (bound
      2 x the default GPU optimizer's error + 1e-3); reconstructed FP32 masters bitwise identical to an arm with explicit
      FP32 masters, the TE remainder self-test passes and the negative controls fail
    note: run on this recipe's training image (the same image configuration); the optimizer step is elementwise, so the
      result holds for every model on the image
  convergence:
    steps: 20
    result: pass
    correlation: 0.9997869
    final_loss_difference: 0.0018
    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 aff5b81);
    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
  TORCHINDUCTOR_COMPILE_THREADS: '1'
  NCCL_NVLS_ENABLE: $([ "$NVLINK_COUNT" -gt 0 ] && echo 1 || echo 0)
patches: []
run: |
  export CUDA_DEVICE_MAX_CONNECTIONS=1
  export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
  export TORCHINDUCTOR_COMPILE_THREADS=1
  export NCCL_NVLS_ENABLE=$([ "$NVLINK_COUNT" -gt 0 ] && echo 1 || echo 0)

  python3 -m <slime entry module> --async \
    --actor-num-nodes 8 \
    --actor-num-gpus-per-node 8 \
    ${MODEL_ARGS[@]} \
    --hf-checkpoint "$HF_CKPT" \
    --ref-load "$TORCH_DIST" \
    --rollout-function-path pretokenized_sft_rollout.generate_rollout \
    --prompt-data "$SFT_DATA" \
    --input-key messages \
    --metadata-key metadata \
    --num-rollout 20 \
    --rollout-batch-size 128 \
    --global-batch-size 128 \
    --loss-type sft_loss \
    --calculate-per-token-loss \
    --disable-compute-advantages-and-returns \
    --debug-train-only \
    --seq-length 8192 \
    --tensor-model-parallel-size 1 \
    --pipeline-model-parallel-size 8 \
    --context-parallel-size 1 \
    --expert-model-parallel-size 8 \
    --expert-tensor-parallel-size 1 \
    --recompute-granularity full \
    --recompute-method uniform \
    --recompute-num-layers 1 \
    --use-dynamic-batch-size \
    --max-tokens-per-gpu 8192 \
    --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 \
    --use-distributed-optimizer \
    --use-precision-aware-optimizer \
    --attention-dropout 0.0 --hidden-dropout 0.0 \
    --accumulate-allreduce-grads-in-fp32 --attention-softmax-in-fp32 \
    --attention-backend flash \
    --distributed-timeout-minutes 30 \
    --tensorboard-dir "$PROFILE_DIR" \
    --pipeline-model-parallel-layout 'Et|(tt|)*30tL'
notes: |
  - Starting point: slime's MiniMax-M2 script, `scripts/run-minimax-m2.sh` with `scripts/models/minimax-m2.sh` at
    slime 0.3.0. slime ships it as a 16-node GRPO script that shares the GPUs with its rollout engine and publishes no
    tokens/s figure for it. Ported to this workload on 64 H200 with its layout (TP 2 with sequence parallel, PP 2,
    full recompute, Adam on the CPU) and EP 8 in place of its EP 4, at which each GPU's expert weights and FP32
    gradients alone exceed its memory, it measured 557.8 tokens/s/GPU; ATTUNE tuned it to 2,034.4 (+265%).
  - `--pipeline-model-parallel-layout 'Et|(tt|)*30tL'` cuts the 62 layers into 32 virtual stages, 4 per pipeline rank:
    the first holds the embedding and 1 layer, the last 1 layer and the loss, and the 30 in between 2 layers each, so
    ranks 0 and 7 hold 7 layers and the others 8.
  - The reference recipe runs at TP 2 with sequence parallelism, where slime's MiniMax-M2 layer spec gives the
    full-width q and k norm weights twice their gradient: each TP rank applies the norm to the gathered q and k, and
    Megatron sums the gradients of these sequence-parallel weights over TP. The forward pass is unchanged, and this
    recipe, at TP 1, is not affected.