OpenRecipes
‹ All models
Qwen

Training Qwen3.6-35B-A3B on H200

Ranked by tokens/s per GPU · 8,192-token SFT rows · global batch 128 · BF16
#Tokens/s per GPU$ / 1M tokensGPUsLayoutMeasuredAuthor
#14,663.1$0.478TP1 PP1 EP8 DP8
max 8,192 tokens/GPU, no offload		2026-10-13ATTUNERecipe ›
#23,466.8$0.638TP1 PP1 EP8 DP8
max 16,384 tokens/GPU		2026-10-13ATTUNERecipe ›
#33,076.8$0.718TP1 PP2 EP4 DP4
max 16,384 tokens/GPU		2026-10-13ATTUNERecipe ›

Notes for Qwen3.6-35B-A3B

MODEL_ARGS are slime's scripts/models/qwen3.5-35B-A3B.sh: Qwen3.6-35B-A3B has the Qwen3.5-35B-A3B architecture and a byte-identical tokenizer, so it trains on the Qwen3.5 rows.

The steps every recipe shares: Running a recipe.

Recipes

#1attune-tp1-pp1-ep84,663.1 tokens/s/GPUTP1 PP1 EP8 DP8max 8,192 tokens/GPU, no offload
Author
ATTUNE (cite)
GPUs
8 x H200 141 GB (AWS p5en.48xlarge, 8 per node)
Layout
TP 1, PP 1, CP 1, EP 8, ETP 1, DP 8; one 8,192-token row per microbatch; full recompute (uniform, 1 layer); alltoall dispatcher; precision-aware distributed optimizer; no offload
Tokens/s per GPU
4,663.11
$ / 1M tokens
$0.47
MFU
8.3% (N = 2,946,252,800 active parameters per token)
Step-time CV
5.55%
Convergence check
correlation 0.99999, final loss difference 0.000741, 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 run on this training image; the optimizer step is elementwise, so the result holds for every model on the image
Measured
2026-10-13
GPU-hours
1.9
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 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 1 \
  --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"
recipes/qwen3.6-35b-a3b/h200/attune-tp1-pp1-ep8.yamlDownloadGitHub
# recipes/qwen3.6-35b-a3b/h200/attune-tp1-pp1-ep8.yaml
# An OpenRecipes recipe; the fields are documented in schema/recipe.schema.yaml.
schema_version: 1
model: Qwen3.6-35B-A3B
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: 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 PP1 EP8 DP8
  summary: TP 1, PP 1, CP 1, EP 8, ETP 1, DP 8; one 8,192-token row per microbatch; full recompute (uniform, 1 layer);
    alltoall dispatcher; precision-aware distributed optimizer; no offload
  highlights: max 8,192 tokens/GPU, no offload
  tp: 1
  pp: 1
  vpp: 1
  cp: 1
  ep: 8
  etp: 1
  dp: 8
  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: 4663.11
  usd_per_million_tokens: 0.47
  step_time_cv_percent: 5.55
  mfu: 0.0833
  mfu_note: null
  active_params:
    count: 2946252800
    source: https://huggingface.co/Qwen/Qwen3.6-35B-A3B
    revision: 995ad96eacd98c81ed38be0c5b274b04031597b0
  measured: 2026-10-13
  steps_per_run: 20
  window:
  - 5
  - 19
  window_tokens: 15728640
  runs: 1
  profiler: false
  gpu_hours: 1.9
  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: 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
    note: run on this training image; the optimizer step is elementwise, so the result holds for every model on the image
  convergence:
    steps: 20
    result: pass
    correlation: 0.99999
    final_loss_difference: 0.000741
    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 1 \
    --pipeline-model-parallel-size 1 \
    --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"
#2attune-tp1-ep8-mt16k3,466.8 tokens/s/GPUTP1 PP1 EP8 DP8max 16,384 tokens/GPU
Author
ATTUNE (cite)
GPUs
8 x H200 141 GB (AWS p5en.48xlarge, 8 per node)
Layout
TP 1, PP 1, CP 1, EP 8, ETP 1, DP 8; two 8,192-token rows per microbatch; full recompute (uniform, 1 layer); alltoall dispatcher; CPU optimizer offload with D2H/H2D overlap; precision-aware distributed optimizer
Tokens/s per GPU
3,466.81
$ / 1M tokens
$0.63
MFU
6.2% (N = 2,946,252,800 active parameters per token)
Step-time CV
1.78%
Convergence check
correlation 0.999995, final loss difference 0.000147, 0 pointwise failures over 20 steps
Measured
2026-10-13
GPU-hours
2.5
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 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 1 \
  --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 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 \
  --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/qwen3.6-35b-a3b/h200/attune-tp1-ep8-mt16k.yamlDownloadGitHub
# recipes/qwen3.6-35b-a3b/h200/attune-tp1-ep8-mt16k.yaml
# An OpenRecipes recipe; the fields are documented in schema/recipe.schema.yaml.
schema_version: 1
model: Qwen3.6-35B-A3B
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: 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 PP1 EP8 DP8
  summary: TP 1, PP 1, CP 1, EP 8, ETP 1, DP 8; two 8,192-token rows per microbatch; full recompute (uniform, 1 layer);
    alltoall dispatcher; CPU optimizer offload with D2H/H2D overlap; precision-aware distributed optimizer
  highlights: max 16,384 tokens/GPU
  tp: 1
  pp: 1
  vpp: 1
  cp: 1
  ep: 8
  etp: 1
  dp: 8
  sequence_parallel: false
  layer_split: null
  max_tokens_per_gpu: 16384
  rows_per_microbatch: 2
  recompute:
    granularity: full
    method: uniform
    num_layers: 1
  offload:
    optimizer_fraction: 1.0
    overlap_d2h_h2d: true
  dispatcher: alltoall
result:
  tokens_per_s_per_gpu: 3466.81
  usd_per_million_tokens: 0.63
  step_time_cv_percent: 1.78
  mfu: 0.062
  mfu_note: null
  active_params:
    count: 2946252800
    source: https://huggingface.co/Qwen/Qwen3.6-35B-A3B
    revision: 995ad96eacd98c81ed38be0c5b274b04031597b0
  measured: 2026-10-13
  steps_per_run: 20
  window:
  - 5
  - 19
  window_tokens: 15728640
  runs: 1
  profiler: false
  gpu_hours: 2.5
  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: Not published.
  configuration: pass
  component_tests: []
  convergence:
    steps: 20
    result: pass
    correlation: 0.999995
    final_loss_difference: 0.000147
    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 1 \
    --pipeline-model-parallel-size 1 \
    --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 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 \
    --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"
#3attune-tp1-pp2-ep4-mt16k3,076.8 tokens/s/GPUTP1 PP2 EP4 DP4max 16,384 tokens/GPU
Author
ATTUNE (cite)
GPUs
8 x H200 141 GB (AWS p5en.48xlarge, 8 per node)
Layout
TP 1, PP 2, CP 1, EP 4, ETP 1, DP 4; two 8,192-token rows per microbatch; full recompute (uniform, 1 layer); alltoall dispatcher; CPU optimizer offload with D2H/H2D overlap; precision-aware distributed optimizer
Tokens/s per GPU
3,076.75
$ / 1M tokens
$0.71
MFU
5.5% (N = 2,946,252,800 active parameters per token)
Step-time CV
3.02%
Convergence check
correlation 0.999997, final loss difference 6.8e-05, 0 pointwise failures over 20 steps
Measured
2026-10-13
GPU-hours
2.9
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 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 2 \
  --context-parallel-size 1 \
  --expert-model-parallel-size 4 \
  --expert-tensor-parallel-size 1 \
  --recompute-granularity full \
  --recompute-method uniform \
  --recompute-num-layers 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 \
  --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/qwen3.6-35b-a3b/h200/attune-tp1-pp2-ep4-mt16k.yamlDownloadGitHub
# recipes/qwen3.6-35b-a3b/h200/attune-tp1-pp2-ep4-mt16k.yaml
# An OpenRecipes recipe; the fields are documented in schema/recipe.schema.yaml.
schema_version: 1
model: Qwen3.6-35B-A3B
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: 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 PP2 EP4 DP4
  summary: TP 1, PP 2, CP 1, EP 4, ETP 1, DP 4; two 8,192-token rows per microbatch; full recompute (uniform, 1 layer);
    alltoall dispatcher; CPU optimizer offload with D2H/H2D overlap; precision-aware distributed optimizer
  highlights: max 16,384 tokens/GPU
  tp: 1
  pp: 2
  vpp: 1
  cp: 1
  ep: 4
  etp: 1
  dp: 4
  sequence_parallel: false
  layer_split: null
  max_tokens_per_gpu: 16384
  rows_per_microbatch: 2
  recompute:
    granularity: full
    method: uniform
    num_layers: 1
  offload:
    optimizer_fraction: 1.0
    overlap_d2h_h2d: true
  dispatcher: alltoall
result:
  tokens_per_s_per_gpu: 3076.75
  usd_per_million_tokens: 0.71
  step_time_cv_percent: 3.02
  mfu: 0.055
  mfu_note: null
  active_params:
    count: 2946252800
    source: https://huggingface.co/Qwen/Qwen3.6-35B-A3B
    revision: 995ad96eacd98c81ed38be0c5b274b04031597b0
  measured: 2026-10-13
  steps_per_run: 20
  window:
  - 5
  - 19
  window_tokens: 15728640
  runs: 1
  profiler: false
  gpu_hours: 2.9
  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: Not published.
  configuration: pass
  component_tests: []
  convergence:
    steps: 20
    result: pass
    correlation: 0.999997
    final_loss_difference: 6.8e-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 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 2 \
    --context-parallel-size 1 \
    --expert-model-parallel-size 4 \
    --expert-tensor-parallel-size 1 \
    --recompute-granularity full \
    --recompute-method uniform \
    --recompute-num-layers 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 \
    --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"

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}
}

Qwen3.6-35B-A3B on other GPUs

H100
–
No recipe yet
H200 · 8 GPUs · this page
4,663.1
tokens/s/GPU · $0.47 / 1M tokens
B200 · 8 GPUs
6,573.4
tokens/s/GPU · $0.60 / 1M tokens