Training Qwen3.8-27B 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; one 8,192-token row per microbatch; no recompute; CPU optimizer offload with overlapped copies and the precision-aware optimizer; dense model, no experts

Tokens/s per GPU

1,194.30

$ / 1M tokens

$1.84

MFU

18.6% (N = 25,623,920,640 active parameters per token)

Step-time CV

1.89%

Convergence check

correlation 0.999997, final loss difference 0.000236, 0 pointwise failures over 20 steps

Measured

2026-10-14

GPU-hours

5.8

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 4 \
  --pipeline-model-parallel-size 2 \
  --context-parallel-size 1 \
  --expert-model-parallel-size 1 \
  --expert-tensor-parallel-size 1 \
  --sequence-parallel \
  --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/h200/attune-tp4-pp2.yaml
# An OpenRecipes recipe; the fields are documented in schema/recipe.schema.yaml.
schema_version: 1
model: Qwen3.8-27B
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'
  network: 8 GPUs per node, EFA, 3.2 Tb/s RDMA per node
  source: 'in-pod environment dump of the measured run''s training image (same digest), taken by the component tests on 2026-10-11:
    installed TE, PyTorch and DeepEP, Megatron-LM git commit; 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: TP4 PP2
  summary: TP 4, PP 2 (32 layers per stage), CP 1, DP 1, sequence parallel; one 8,192-token row per microbatch; no recompute;
    CPU optimizer offload with overlapped copies and the precision-aware optimizer; dense model, no experts
  highlights: max 8,192 tokens/GPU, no recompute, CPU optimizer 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: 8192
  rows_per_microbatch: 1
  recompute: null
  offload:
    optimizer_fraction: 1.0
    overlap_d2h_h2d: true
  dispatcher: null
result:
  tokens_per_s_per_gpu: 1194.3
  usd_per_million_tokens: 1.84
  step_time_cv_percent: 1.89
  mfu: 0.1857
  mfu_note: null
  active_params:
    count: 25623920640
    source: https://huggingface.co/Qwen/Qwen3.8-27B
    revision: 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0
  measured: 2026-10-14
  steps_per_run: 20
  window:
  - 5
  - 19
  window_tokens: 15728640
  runs: 1
  profiler: false
  gpu_hours: 5.8
  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: 0.999997
    final_loss_difference: 0.000236
    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; 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 4 \
    --pipeline-model-parallel-size 2 \
    --context-parallel-size 1 \
    --expert-model-parallel-size 1 \
    --expert-tensor-parallel-size 1 \
    --sequence-parallel \
    --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: |
  - Recompute off and the optimizer on the GPU each speed this model up, but together they do not fit in 141 GB: the
    tuning run with both changes ran out of memory. This recipe keeps the optimizer on the CPU.
  - No layer is recomputed, so activation memory grows with the tokens per GPU: the tuning run at max 16,384 tokens per
    GPU ran out of memory. Check the memory before raising `--max-tokens-per-gpu`.