OpenRecipes
‹ All models
Qwen

Training Qwen3.5-397B-A17B on H200

Ranked by tokens/s per GPU · 8,192-token SFT rows · global batch 128 · BF16
#Tokens/s per GPU$ / 1M tokensGPUsLayoutMeasuredAuthor
#1545.3$4.03128TP2 PP4 EP32 DP16
max 8,192 tokens/GPU, no offload		2026-10-11ATTUNERecipe ›
#2506.7$4.34128TP1 PP4 EP32 DP32
max 8,192 tokens/GPU, no offload		2026-10-13ATTUNERecipe ›
#3469.2$4.68128TP2 PP2 EP64 DP32
max 8,192 tokens/GPU, no offload		2026-10-13ATTUNERecipe ›

Recipes

#1attune-tp2-pp4-ep32545.3 tokens/s/GPUTP2 PP4 EP32 DP16max 8,192 tokens/GPU, no offload
Author
ATTUNE (cite)
GPUs
128 x H200 141 GB (AWS p5en.48xlarge, 8 per node)
Layout
TP 2, PP 4, CP 1, EP 32, ETP 1, DP 16; one 8,192-token row per microbatch; full recompute (uniform, 1 layer); precision-aware distributed optimizer; MoE permute fusion; alltoall dispatcher; no offload
Tokens/s per GPU
545.32
$ / 1M tokens
$4.03
MFU
5.4% (N = 16,331,407,360 active parameters per token)
Step-time CV
1.14%
Convergence check
correlation 0.99999, final loss difference 8.4e-05, 0 pointwise failures over 20 steps
Component test
precision-aware Adam on GPU (TE FusedAdam, FP32 masters stored as BF16 parameter plus remainder) 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 against a bound of 1.0e-3 after one step, and within bound after three steps at the entry's hyperparameters; reconstructed FP32 masters bitwise identical to an arm with explicit FP32 masters; the negative controls fail
Measured
2026-10-11, 14-step run scored over steps 4–13
GPU-hours
68.9
Rows
qwen35-tulu3-fixed8192-v1 (rows_sha256 7a837641dcde9ebcac9d3ed5ed84ddd941fc9bfc173606cc701e9a3f51c307f1)

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

Environment

  • 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)
CommandHow to 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 16 \
  --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 14 \
  --rollout-batch-size 128 \
  --global-batch-size 128 \
  --loss-type sft_loss \
  --calculate-per-token-loss \
  --disable-compute-advantages-and-returns \
  --debug-train-only \
  --tensor-model-parallel-size 2 \
  --pipeline-model-parallel-size 4 \
  --context-parallel-size 1 \
  --expert-model-parallel-size 32 \
  --expert-tensor-parallel-size 1 \
  --sequence-parallel \
  --recompute-granularity full \
  --recompute-method uniform \
  --recompute-num-layers 1 \
  --use-dynamic-batch-size \
  --max-tokens-per-gpu 8192 \
  --data-pad-size-multiplier 256 \
  --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
recipes/qwen3.5-397b-a17b/h200/attune-tp2-pp4-ep32.yamlDownloadGitHub
# recipes/qwen3.5-397b-a17b/h200/attune-tp2-pp4-ep32.yaml
# An OpenRecipes recipe; the fields are documented in schema/recipe.schema.yaml.
schema_version: 1
model: Qwen3.5-397B-A17B
gpu:
  type: H200
  memory_gb: 141
  count: 128
  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: 7a837641dcde9ebcac9d3ed5ed84ddd941fc9bfc173606cc701e9a3f51c307f1
    build:
      name: qwen35-tulu3-fixed8192-v1
      tokenizer: Qwen3.5 (the Qwen3.5-122B-A10B tokenizer; tokenizer.json sha256 5f9e4d4901a92b997e463c1f46055088b6cca5ca61a6522d1b9f64c4bb81cb42)
      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: 0.676
      parquet_sha256: f215bdc6aba60d6d0dff4471eaef805765cb393e40b95762190535376491ab01
      parquet_bytes: 76127393
      builder: data/generate_fixed_sft.py (data/README.md)
layout:
  label: TP2 PP4 EP32 DP16
  summary: TP 2, PP 4, CP 1, EP 32, ETP 1, DP 16; one 8,192-token row per microbatch; full recompute (uniform, 1 layer);
    precision-aware distributed optimizer; MoE permute fusion; alltoall dispatcher; no offload
  highlights: max 8,192 tokens/GPU, no offload
  tp: 2
  pp: 4
  vpp: 1
  cp: 1
  ep: 32
  etp: 1
  dp: 16
  sequence_parallel: true
  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: 545.32
  usd_per_million_tokens: 4.03
  step_time_cv_percent: 1.14
  mfu: 0.054
  mfu_note: null
  active_params:
    count: 16331407360
    source: https://huggingface.co/Qwen/Qwen3.5-397B-A17B
    revision: 8472618112abcbd45acbcdc58436aff4233c23f7
  measured: 2026-10-11
  steps_per_run: 14
  window:
  - 4
  - 13
  window_tokens: 10485760
  runs: 1
  profiler: false
  gpu_hours: 68.9
  gpu_hours_note: GPUs x (job start to end), including startup
correctness:
  result: pass
  summary: passed; the on-GPU optimizer passed its component test at this model's shape
  reference: Not published.
  configuration: pass
  component_tests:
  - test: precision-aware Adam on GPU (TE FusedAdam, FP32 masters stored as BF16 parameter plus remainder) 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 against a bound of 1.0e-3 after one step, and within bound
      after three steps at the entry's hyperparameters; reconstructed FP32 masters bitwise identical to an arm with explicit
      FP32 masters; the negative controls fail
  convergence:
    steps: 20
    result: pass
    correlation: 0.99999
    final_loss_difference: 8.4e-05
    pointwise_failures: 0
  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 16 \
    --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 14 \
    --rollout-batch-size 128 \
    --global-batch-size 128 \
    --loss-type sft_loss \
    --calculate-per-token-loss \
    --disable-compute-advantages-and-returns \
    --debug-train-only \
    --tensor-model-parallel-size 2 \
    --pipeline-model-parallel-size 4 \
    --context-parallel-size 1 \
    --expert-model-parallel-size 32 \
    --expert-tensor-parallel-size 1 \
    --sequence-parallel \
    --recompute-granularity full \
    --recompute-method uniform \
    --recompute-num-layers 1 \
    --use-dynamic-batch-size \
    --max-tokens-per-gpu 8192 \
    --data-pad-size-multiplier 256 \
    --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
#2attune-tp1-pp4-ep32506.7 tokens/s/GPUTP1 PP4 EP32 DP32max 8,192 tokens/GPU, no offload
Author
ATTUNE (cite)
GPUs
128 x H200 141 GB (AWS p5en.48xlarge, 8 per node)
Layout
TP 1, PP 4, CP 1, EP 32, ETP 1, DP 32; one 8,192-token row per microbatch; full recompute (uniform, 1 layer); precision-aware distributed optimizer; alltoall dispatcher; no offload
Tokens/s per GPU
506.73
$ / 1M tokens
$4.34
MFU
5.0% (N = 16,331,407,360 active parameters per token)
Step-time CV
0.77%
Convergence check
correlation 0.999999, final loss difference 1.9e-05, 0 pointwise failures over 20 steps
Component test
precision-aware Adam on GPU (TE FusedAdam, FP32 masters stored as BF16 parameter plus remainder) 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 against a bound of 1.0e-3 after one step, and within bound after three steps at the entry's hyperparameters; reconstructed FP32 masters bitwise identical to an arm with explicit FP32 masters; the negative controls fail
Measured
2026-10-13
GPU-hours
38.04
Rows
qwen35-tulu3-fixed8192-v1 (rows_sha256 7a837641dcde9ebcac9d3ed5ed84ddd941fc9bfc173606cc701e9a3f51c307f1)

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

Environment

  • 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)
CommandHow to 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 16 \
  --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 \
  --tensor-model-parallel-size 1 \
  --pipeline-model-parallel-size 4 \
  --context-parallel-size 1 \
  --expert-model-parallel-size 32 \
  --expert-tensor-parallel-size 1 \
  --recompute-granularity full \
  --recompute-method uniform \
  --recompute-num-layers 1 \
  --use-dynamic-batch-size \
  --max-tokens-per-gpu 8192 \
  --data-pad-size-multiplier 256 \
  --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
recipes/qwen3.5-397b-a17b/h200/attune-tp1-pp4-ep32.yamlDownloadGitHub
# recipes/qwen3.5-397b-a17b/h200/attune-tp1-pp4-ep32.yaml
# An OpenRecipes recipe; the fields are documented in schema/recipe.schema.yaml.
schema_version: 1
model: Qwen3.5-397B-A17B
gpu:
  type: H200
  memory_gb: 141
  count: 128
  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: 7a837641dcde9ebcac9d3ed5ed84ddd941fc9bfc173606cc701e9a3f51c307f1
    build:
      name: qwen35-tulu3-fixed8192-v1
      tokenizer: Qwen3.5 (the Qwen3.5-122B-A10B tokenizer; tokenizer.json sha256 5f9e4d4901a92b997e463c1f46055088b6cca5ca61a6522d1b9f64c4bb81cb42)
      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: 0.676
      parquet_sha256: f215bdc6aba60d6d0dff4471eaef805765cb393e40b95762190535376491ab01
      parquet_bytes: 76127393
      builder: data/generate_fixed_sft.py (data/README.md)
layout:
  label: TP1 PP4 EP32 DP32
  summary: TP 1, PP 4, CP 1, EP 32, ETP 1, DP 32; one 8,192-token row per microbatch; full recompute (uniform, 1 layer);
    precision-aware distributed optimizer; alltoall dispatcher; no offload
  highlights: max 8,192 tokens/GPU, no offload
  tp: 1
  pp: 4
  vpp: 1
  cp: 1
  ep: 32
  etp: 1
  dp: 32
  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: 506.73
  usd_per_million_tokens: 4.34
  step_time_cv_percent: 0.77
  mfu: 0.0502
  mfu_note: null
  active_params:
    count: 16331407360
    source: https://huggingface.co/Qwen/Qwen3.5-397B-A17B
    revision: 8472618112abcbd45acbcdc58436aff4233c23f7
  measured: 2026-10-13
  steps_per_run: 20
  window:
  - 5
  - 19
  window_tokens: 15728640
  runs: 1
  profiler: false
  gpu_hours: 38.04
  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 at this model's shape
  reference: Not published.
  configuration: pass
  component_tests:
  - test: precision-aware Adam on GPU (TE FusedAdam, FP32 masters stored as BF16 parameter plus remainder) 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 against a bound of 1.0e-3 after one step, and within bound
      after three steps at the entry's hyperparameters; reconstructed FP32 masters bitwise identical to an arm with explicit
      FP32 masters; the negative controls fail
  convergence:
    steps: 20
    result: pass
    correlation: 0.999999
    final_loss_difference: 1.9e-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; 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 16 \
    --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 \
    --tensor-model-parallel-size 1 \
    --pipeline-model-parallel-size 4 \
    --context-parallel-size 1 \
    --expert-model-parallel-size 32 \
    --expert-tensor-parallel-size 1 \
    --recompute-granularity full \
    --recompute-method uniform \
    --recompute-num-layers 1 \
    --use-dynamic-batch-size \
    --max-tokens-per-gpu 8192 \
    --data-pad-size-multiplier 256 \
    --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
#3attune-tp2-pp2-ep64469.2 tokens/s/GPUTP2 PP2 EP64 DP32max 8,192 tokens/GPU, no offload
Author
ATTUNE (cite)
GPUs
128 x H200 141 GB (AWS p5en.48xlarge, 8 per node)
Layout
TP 2, PP 2, CP 1, EP 64, ETP 1, DP 32, sequence parallel; one 8,192-token row per microbatch; full recompute (uniform, 1 layer); precision-aware distributed optimizer; alltoall dispatcher; no offload
Tokens/s per GPU
469.19
$ / 1M tokens
$4.68
MFU
4.7% (N = 16,331,407,360 active parameters per token)
Step-time CV
1.17%
Convergence check
correlation 0.999999, final loss difference 3.3e-05, 0 pointwise failures over 20 steps
Component test
precision-aware Adam on GPU (TE FusedAdam, FP32 masters stored as BF16 parameter plus remainder) 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 against a bound of 1.0e-3 after one step, and within bound after three steps at the entry's hyperparameters; reconstructed FP32 masters bitwise identical to an arm with explicit FP32 masters; the negative controls fail
Measured
2026-10-13
GPU-hours
31.25
Rows
qwen35-tulu3-fixed8192-v1 (rows_sha256 7a837641dcde9ebcac9d3ed5ed84ddd941fc9bfc173606cc701e9a3f51c307f1)

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

Environment

  • 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)
CommandHow to 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 16 \
  --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 \
  --tensor-model-parallel-size 2 \
  --pipeline-model-parallel-size 2 \
  --context-parallel-size 1 \
  --expert-model-parallel-size 64 \
  --expert-tensor-parallel-size 1 \
  --sequence-parallel \
  --recompute-granularity full \
  --recompute-method uniform \
  --recompute-num-layers 1 \
  --use-dynamic-batch-size \
  --max-tokens-per-gpu 8192 \
  --data-pad-size-multiplier 256 \
  --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
recipes/qwen3.5-397b-a17b/h200/attune-tp2-pp2-ep64.yamlDownloadGitHub
# recipes/qwen3.5-397b-a17b/h200/attune-tp2-pp2-ep64.yaml
# An OpenRecipes recipe; the fields are documented in schema/recipe.schema.yaml.
schema_version: 1
model: Qwen3.5-397B-A17B
gpu:
  type: H200
  memory_gb: 141
  count: 128
  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: 7a837641dcde9ebcac9d3ed5ed84ddd941fc9bfc173606cc701e9a3f51c307f1
    build:
      name: qwen35-tulu3-fixed8192-v1
      tokenizer: Qwen3.5 (the Qwen3.5-122B-A10B tokenizer; tokenizer.json sha256 5f9e4d4901a92b997e463c1f46055088b6cca5ca61a6522d1b9f64c4bb81cb42)
      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: 0.676
      parquet_sha256: f215bdc6aba60d6d0dff4471eaef805765cb393e40b95762190535376491ab01
      parquet_bytes: 76127393
      builder: data/generate_fixed_sft.py (data/README.md)
layout:
  label: TP2 PP2 EP64 DP32
  summary: TP 2, PP 2, CP 1, EP 64, ETP 1, DP 32, sequence parallel; one 8,192-token row per microbatch; full recompute
    (uniform, 1 layer); precision-aware distributed optimizer; alltoall dispatcher; no offload
  highlights: max 8,192 tokens/GPU, no offload
  tp: 2
  pp: 2
  vpp: 1
  cp: 1
  ep: 64
  etp: 1
  dp: 32
  sequence_parallel: true
  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: 469.19
  usd_per_million_tokens: 4.68
  step_time_cv_percent: 1.17
  mfu: 0.0465
  mfu_note: null
  active_params:
    count: 16331407360
    source: https://huggingface.co/Qwen/Qwen3.5-397B-A17B
    revision: 8472618112abcbd45acbcdc58436aff4233c23f7
  measured: 2026-10-13
  steps_per_run: 20
  window:
  - 5
  - 19
  window_tokens: 15728640
  runs: 1
  profiler: false
  gpu_hours: 31.25
  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 at this model's shape
  reference: Not published.
  configuration: pass
  component_tests:
  - test: precision-aware Adam on GPU (TE FusedAdam, FP32 masters stored as BF16 parameter plus remainder) 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 against a bound of 1.0e-3 after one step, and within bound
      after three steps at the entry's hyperparameters; reconstructed FP32 masters bitwise identical to an arm with explicit
      FP32 masters; the negative controls fail
  convergence:
    steps: 20
    result: pass
    correlation: 0.999999
    final_loss_difference: 3.3e-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; 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 16 \
    --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 \
    --tensor-model-parallel-size 2 \
    --pipeline-model-parallel-size 2 \
    --context-parallel-size 1 \
    --expert-model-parallel-size 64 \
    --expert-tensor-parallel-size 1 \
    --sequence-parallel \
    --recompute-granularity full \
    --recompute-method uniform \
    --recompute-num-layers 1 \
    --use-dynamic-batch-size \
    --max-tokens-per-gpu 8192 \
    --data-pad-size-multiplier 256 \
    --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

Citation

If you use a recipe by ATTUNE, please cite the paper.

BibTeX
@article{fan2026attune,
  title   = {{ATTUNE: Agentic Throughput Tuning for LLM Training}},
  author  = {Leo Fan and Hongyang Wang and Wenbo Chen and Jingguo Liu and Xian Li and Xuan Wang and
             Ruochen Wen and Yiming Yan and Hao Wu and Zheshen Wang and Benoit Dumoulin},
  journal = {arXiv preprint arXiv:2610.17528},
  year    = {2026},
  url     = {https://arxiv.org/abs/2610.17528}
}

Qwen3.5-397B-A17B on other GPUs

H100 · 128 GPUs
313.3
tokens/s/GPU · $6.10 / 1M tokens
H200 · 128 GPUs · this page
545.3
tokens/s/GPU · $4.03 / 1M tokens
B200 · 64 GPUs
1,232.5
tokens/s/GPU · $3.21 / 1M tokens