Training Qwen3-32B on H200

Author

ATTUNE (cite)

GPUs

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

Layout

TP 4, PP 2 (32 layers per stage), CP 1, DP 1, sequence parallel; two 8,192-token rows per microbatch (max 16,384 tokens per GPU); no recompute; precision-aware distributed optimizer; no offload; dense model, no experts

Tokens/s per GPU

2,200.16

$ / 1M tokens

$1.00

MFU

42.7% (N = 31,983,534,080 active parameters per token)

Step-time CV

0.45%

Convergence check

correlation 0.9999998, final loss difference 0.00012, 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

2026-10-18

GPU-hours

3.02

Rows

qwen3-tulu3-fixed8192-v1 (rows_sha256 8f76b822a284adbe24faaf3cfe725e65cb5e94f98ba54aceade45b275345f983)

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 1 \
  --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 \
  --loss-mask-type qwen3 \
  --calculate-per-token-loss \
  --disable-compute-advantages-and-returns \
  --debug-train-only \
  --tensor-model-parallel-size 4 \
  --sequence-parallel \
  --pipeline-model-parallel-size 2 \
  --context-parallel-size 1 \
  --expert-model-parallel-size 1 \
  --expert-tensor-parallel-size 1 \
  --use-dynamic-batch-size \
  --max-tokens-per-gpu 16384 \
  --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"

# recipes/qwen3-32b/h200/attune-tp4-pp2-mt16k.yaml
# An OpenRecipes recipe; the fields are documented in schema/recipe.schema.yaml.
schema_version: 1
model: Qwen3-32B
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: '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
  python: 3.12.3
  network: 8 GPUs per node, EFA, 3.2 Tb/s RDMA per node
  source: 'in-pod version checks of the measured run''s training image, 2026-10-13 and 2026-10-18: installed TE, PyTorch,
    DeepEP and megatron-core, Megatron-LM git commit, 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: 8f76b822a284adbe24faaf3cfe725e65cb5e94f98ba54aceade45b275345f983
    build:
      name: qwen3-tulu3-fixed8192-v1
      tokenizer: Qwen3-32B (revision 9216db5781bf21249d130ec9da846c4624c16137; tokenizer.json sha256 aeb13307a71acd8fe81861d94ad54ab689df773318809eed3cbe794b4492dae4;
        tokenizer_config.json sha256 d5d09f07b48c3086c508b30d1c9114bd1189145b74e982a265350c923acd8101)
      loss_mask: qwen3
      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: 0.677
      parquet_sha256: 7ea7f5dca51f6be280e2f7cd7ee786ac05f546570e5ac1b07236c920338241fe
      parquet_bytes: 88808580
      builder: data/generate_fixed_sft.py with --loss-mask-type qwen3 (data/README.md)
layout:
  label: TP4 PP2
  summary: TP 4, PP 2 (32 layers per stage), CP 1, DP 1, sequence parallel; two 8,192-token rows per microbatch (max 16,384
    tokens per GPU); no recompute; precision-aware distributed optimizer; no offload; dense model, no experts
  highlights: max 16,384 tokens/GPU, no recompute, no offload
  tp: 4
  pp: 2
  vpp: 1
  cp: 1
  ep: 1
  etp: 1
  dp: 1
  sequence_parallel: true
  layer_split: null
  max_tokens_per_gpu: 16384
  rows_per_microbatch: 2
  recompute: null
  offload: null
  dispatcher: null
result:
  tokens_per_s_per_gpu: 2200.16
  usd_per_million_tokens: 1.0
  step_time_cv_percent: 0.45
  mfu: 0.4269
  mfu_note: null
  active_params:
    count: 31983534080
    source: https://huggingface.co/Qwen/Qwen3-32B
    revision: 9216db5781bf21249d130ec9da846c4624c16137
  measured: 2026-10-18
  steps_per_run: 20
  window:
  - 5
  - 19
  window_tokens: 15728640
  runs: 1
  profiler: false
  gpu_hours: 3.02
  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-qwen3-32B.sh` (one 8-GPU node): TP 8 with sequence parallelism, PP 1,
    full recompute (uniform, 1 layer), max 20,480 tokens/GPU 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: 0.9999998
    final_loss_difference: 0.00012
    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
  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 1 \
    --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 \
    --loss-mask-type qwen3 \
    --calculate-per-token-loss \
    --disable-compute-advantages-and-returns \
    --debug-train-only \
    --tensor-model-parallel-size 4 \
    --sequence-parallel \
    --pipeline-model-parallel-size 2 \
    --context-parallel-size 1 \
    --expert-model-parallel-size 1 \
    --expert-tensor-parallel-size 1 \
    --use-dynamic-batch-size \
    --max-tokens-per-gpu 16384 \
    --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"
notes: |
  - Starting point: slime's Qwen3-32B script, `scripts/run-qwen3-32B.sh` with `scripts/models/qwen3-32B.sh` at slime
    0.3.0. slime ships it as a one-node GRPO script that shares the GPUs with its rollout engine and publishes no
    tokens/s figure for it. Ported to this workload with its layout (TP 8 with sequence parallel, PP 1, full
    recompute, max 20,480 tokens/GPU, Adam on the CPU) on 8 H200, it measured 1,327.8 tokens/s/GPU; ATTUNE tuned it to
    2,200.2 (+66%).
  - No layer is recomputed, so activation memory grows with the tokens per GPU: the tuning runs at max 24,576 tokens
    per GPU, and at max 16,384 with PP 1 and DP 2, ran out of memory. Check the memory before raising
    `--max-tokens-per-gpu`.