Training Qwen3.8-27B on H100

Author

ATTUNE (cite)

GPUs

8 x H100 80 GB (AWS p5.48xlarge, 8 per node)

Layout

TP 1, PP 4 (16/17/17/14 layers), CP 1, DP 2; one 8,192-token row per microbatch; full recompute (uniform, 1 layer); CPU optimizer offload with overlapped copies and the precision-aware optimizer; dense model, no experts

Tokens/s per GPU

1,480.40

$ / 1M tokens

$1.29

MFU

23.0% (N = 25,623,920,640 active parameters per token)

Step-time CV

8.93%

Convergence check

correlation 0.9999966, final loss difference 8.3e-05, 0 pointwise failures over 20 steps

Measured

2026-10-18

GPU-hours

5.23

Rows

qwen38-tulu3-fixed8192-v1 (rows_sha256 ee1486b28ce21df721ad61c2d3765c64a29eed0eb5680ede005bd6a321dc1a1e)

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_5 \
  --calculate-per-token-loss \
  --disable-compute-advantages-and-returns \
  --debug-train-only \
  --tensor-model-parallel-size 1 \
  --pipeline-model-parallel-size 4 \
  --decoder-first-pipeline-num-layers 16 \
  --decoder-last-pipeline-num-layers 14 \
  --context-parallel-size 1 \
  --expert-model-parallel-size 1 \
  --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 \
  --optimizer-cpu-offload \
  --overlap-cpu-optimizer-d2h-h2d \
  --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/qwen3.8-27b/h100/attune-tp1-pp4-split16.yaml
# An OpenRecipes recipe; the fields are documented in schema/recipe.schema.yaml.
schema_version: 1
model: Qwen3.8-27B
gpu:
  type: H100
  memory_gb: 80
  count: 8
  per_node: 8
  instance: AWS p5.48xlarge
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: ee1486b28ce21df721ad61c2d3765c64a29eed0eb5680ede005bd6a321dc1a1e
    build:
      name: qwen38-tulu3-fixed8192-v1
      tokenizer: Qwen3.8 (the Qwen3.8-27B tokenizer; tokenizer.json sha256 0997f410c57a1f4e53b09e4be8f4a172d90edd9564368fb0847030937229b9f3)
      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: 0.642
      parquet_sha256: a7b7b4db98e95152604a77bcf09b51502d2266ba04de98409cef6b97afd039c8
      parquet_bytes: 76722714
      builder: data/generate_fixed_sft.py (data/README.md)
layout:
  label: TP1 PP4 DP2
  summary: TP 1, PP 4 (16/17/17/14 layers), CP 1, DP 2; one 8,192-token row per microbatch; full recompute (uniform, 1
    layer); CPU optimizer offload with overlapped copies and the precision-aware optimizer; dense model, no experts
  highlights: 16/17/17/14 layer split, CPU optimizer offload
  tp: 1
  pp: 4
  vpp: 1
  cp: 1
  ep: 1
  etp: 1
  dp: 2
  sequence_parallel: false
  layer_split:
    first: 16
    last: 14
  max_tokens_per_gpu: 8192
  rows_per_microbatch: 1
  recompute:
    granularity: full
    method: uniform
    num_layers: 1
  offload:
    optimizer_fraction: 1.0
    overlap_d2h_h2d: true
  dispatcher: null
result:
  tokens_per_s_per_gpu: 1480.4
  usd_per_million_tokens: 1.29
  step_time_cv_percent: 8.93
  mfu: 0.2301
  mfu_note: null
  active_params:
    count: 25623920640
    source: https://huggingface.co/Qwen/Qwen3.8-27B
    revision: 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0
  measured: 2026-10-18
  steps_per_run: 20
  window:
  - 5
  - 19
  window_tokens: 15728640
  runs: 1
  profiler: false
  gpu_hours: 5.23
  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: 'The layout of slime''s `scripts/run-qwen3.5-27B.sh` on one 8-GPU node: TP 4 with sequence parallelism, PP
    2 (first stage 34 layers, last 30), CP 1 in place of its CP 4, full recompute and CPU optimizer offload, with this
    recipe''s optimizer settings.'
  configuration: pass
  component_tests: []
  convergence:
    steps: 20
    result: pass
    correlation: 0.9999966
    final_loss_difference: 8.3e-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
  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_5 \
    --calculate-per-token-loss \
    --disable-compute-advantages-and-returns \
    --debug-train-only \
    --tensor-model-parallel-size 1 \
    --pipeline-model-parallel-size 4 \
    --decoder-first-pipeline-num-layers 16 \
    --decoder-last-pipeline-num-layers 14 \
    --context-parallel-size 1 \
    --expert-model-parallel-size 1 \
    --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 \
    --optimizer-cpu-offload \
    --overlap-cpu-optimizer-d2h-h2d \
    --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 Qwen3.5-27B script, `scripts/run-qwen3.5-27B.sh` with `scripts/models/qwen3.5-27B.sh` at
    slime 0.3.0. slime ships it as a 4-node GRPO script and publishes no tokens/s figure for it. Ported to this
    workload on one 8-GPU node with its layout (TP 4 with sequence parallel, PP 2 with 34/30 layers, CP 1 in place of
    its CP 4, full recompute, Adam on the CPU), it measured 861.5 tokens/s/GPU on 8 H100; ATTUNE tuned it to 1,480.4
    (+72%).
  - Step times of the measured run range from 78.7 s to 105.6 s (median 87.0 s), which sets the 8.9% step-time CV.
  - The optimizer stays on the CPU: the tuning run that kept 30% of it on the GPU (`--optimizer-offload-fraction 0.7`)
    ran out of memory on the last stage, which holds the FP32 logits of the 248,320-token vocabulary.