OpenRecipes
‹ All models
Qwen

Training Qwen3.5-122B-A10B on H100

Ranked by tokens/s per GPU · 8,192-token SFT rows · global batch 128 · BF16
#Tokens/s per GPU$ / 1M tokensGPUsLayoutMeasuredAuthor
#1999.7$1.9148TP4 PP6 EP8 DP2
max 16,384 tokens/GPU, flex dispatcher		2026-10-11ATTUNERecipe ›
#2842.8$2.2748TP4 PP6 EP8 DP2
max 8,192 tokens/GPU, flex dispatcher		2026-10-13ATTUNERecipe ›
#3822.1$2.3248TP4 PP3 EP16 DP4
max 8,192 tokens/GPU		2026-10-13ATTUNERecipe ›

Recipes

#1attune-tp4-pp6-ep8-flex-mt16k999.7 tokens/s/GPUTP4 PP6 EP8 DP2max 16,384 tokens/GPU, flex dispatcher
Author
ATTUNE (cite)
GPUs
48 x H100 80 GB (AWS p5.48xlarge, 8 per node)
Layout
TP 4, PP 6, CP 1, EP 8, ETP 1, DP 2, sequence parallel; two 8,192-token rows per microbatch (max 16,384 tokens per GPU); full recompute (uniform, 1 layer); flex dispatcher; distributed optimizer with the precision-aware optimizer; TE cross-entropy fusion; no offload
Tokens/s per GPU
999.73
$ / 1M tokens
$1.91
MFU
5.5% (N = 9,008,136,192 active parameters per token)
Step-time CV
0.73%
Convergence check
correlation 0.99999, final loss difference 1.6e-05, 0 pointwise failures over 20 steps
Component test
flex dispatcher against an FP32 reference: H100, 2026-10-11. Shape: one Qwen3.5-122B-A10B MoE layer: hidden 3072, 256 experts, top-8, expert FFN 1024, EP 8, ETP 1; TP 4 with sequence parallel and TP 1; 4,096 and 16,384 tokens per rank. routing identical; worst tensor (output) relative L2 error 4.11e-3 against a bound of 8.76e-3 (2 x the alltoall dispatcher's error + 1e-3); input and weight gradients pass; the negative controls fail
Measured
2026-10-11, 14-step run scored over steps 4–13
GPU-hours
17.7
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
  • 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 NCCL_NVLS_ENABLE=$([ "$NVLINK_COUNT" -gt 0 ] && echo 1 || echo 0)

python3 -m <slime entry module> --async \
  --actor-num-nodes 6 --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 \
  --num-rollout 14 --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 6 \
  --context-parallel-size 1 --expert-model-parallel-size 8 --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 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 \
  --cross-entropy-loss-fusion --cross-entropy-fusion-impl te \
  --attention-dropout 0.0 --hidden-dropout 0.0 \
  --accumulate-allreduce-grads-in-fp32 --attention-softmax-in-fp32 --attention-backend flash \
  --moe-token-dispatcher-type flex --max-tokens-per-gpu 16384
recipes/qwen3.5-122b-a10b/h100/attune-tp4-pp6-ep8-flex-mt16k.yamlDownloadGitHub
# recipes/qwen3.5-122b-a10b/h100/attune-tp4-pp6-ep8-flex-mt16k.yaml
# An OpenRecipes recipe; the fields are documented in schema/recipe.schema.yaml.
schema_version: 1
model: Qwen3.5-122B-A10B
gpu:
  type: H100
  memory_gb: 80
  count: 48
  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'
  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: TP4 PP6 EP8 DP2
  summary: TP 4, PP 6, CP 1, EP 8, ETP 1, DP 2, sequence parallel; two 8,192-token rows per microbatch (max 16,384 tokens
    per GPU); full recompute (uniform, 1 layer); flex dispatcher; distributed optimizer with the precision-aware optimizer;
    TE cross-entropy fusion; no offload
  highlights: max 16,384 tokens/GPU, flex dispatcher
  tp: 4
  pp: 6
  vpp: 1
  cp: 1
  ep: 8
  etp: 1
  dp: 2
  sequence_parallel: true
  layer_split: null
  max_tokens_per_gpu: 16384
  rows_per_microbatch: 2
  recompute:
    granularity: full
    method: uniform
    num_layers: 1
  offload: null
  dispatcher: flex
result:
  tokens_per_s_per_gpu: 999.73
  usd_per_million_tokens: 1.91
  step_time_cv_percent: 0.73
  mfu: 0.0546
  mfu_note: null
  active_params:
    count: 9008136192
    source: https://huggingface.co/Qwen/Qwen3.5-122B-A10B
    revision: dc4d348443bc740c68e2d77492492c11606384d5
  measured: 2026-10-11
  steps_per_run: 14
  window:
  - 4
  - 13
  window_tokens: 10485760
  runs: 1
  profiler: false
  gpu_hours: 17.7
  gpu_hours_note: GPUs x (job start to end), including startup
correctness:
  result: pass
  summary: passed; the flex dispatcher passed its component test at this model's shape
  reference: Not published.
  configuration: pass
  component_tests:
  - test: flex dispatcher against an FP32 reference
    shape: 'one Qwen3.5-122B-A10B MoE layer: hidden 3072, 256 experts, top-8, expert FFN 1024, EP 8, ETP 1; TP 4 with
      sequence parallel and TP 1; 4,096 and 16,384 tokens per rank'
    gpu: H100
    date: 2026-10-11
    result: pass
    detail: routing identical; worst tensor (output) relative L2 error 4.11e-3 against a bound of 8.76e-3 (2 x the alltoall
      dispatcher's error + 1e-3); input and weight gradients pass; the negative controls fail
  convergence:
    steps: 20
    result: pass
    correlation: 0.99999
    final_loss_difference: 1.6e-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
  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 NCCL_NVLS_ENABLE=$([ "$NVLINK_COUNT" -gt 0 ] && echo 1 || echo 0)

  python3 -m <slime entry module> --async \
    --actor-num-nodes 6 --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 \
    --num-rollout 14 --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 6 \
    --context-parallel-size 1 --expert-model-parallel-size 8 --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 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 \
    --cross-entropy-loss-fusion --cross-entropy-fusion-impl te \
    --attention-dropout 0.0 --hidden-dropout 0.0 \
    --accumulate-allreduce-grads-in-fp32 --attention-softmax-in-fp32 --attention-backend flash \
    --moe-token-dispatcher-type flex --max-tokens-per-gpu 16384
#2attune-tp4-pp6-ep8-flex-mt8k842.8 tokens/s/GPUTP4 PP6 EP8 DP2max 8,192 tokens/GPU, flex dispatcher
Author
ATTUNE (cite)
GPUs
48 x H100 80 GB (AWS p5.48xlarge, 8 per node)
Layout
TP 4, PP 6, CP 1, EP 8, ETP 1, DP 2, sequence parallel; one 8,192-token row per microbatch; full recompute (uniform, 1 layer); precision-aware distributed optimizer; flex dispatcher; TE cross-entropy fusion; no offload
Tokens/s per GPU
842.76
$ / 1M tokens
$2.27
MFU
4.6% (N = 9,008,136,192 active parameters per token)
Step-time CV
8.31%
Convergence check
correlation 0.999997, final loss difference 2.6e-05, 0 pointwise failures over 20 steps
Component test
flex dispatcher against an FP32 reference: H100, 2026-10-11. Shape: one Qwen3.5-122B-A10B MoE layer: hidden 3072, 256 experts, top-8, expert FFN 1024, EP 8, ETP 1; TP 4 with sequence parallel and TP 1; 4,096 and 16,384 tokens per rank. routing identical; worst tensor (output) relative L2 error 4.11e-3 against a bound of 8.76e-3 (2 x the alltoall dispatcher's error + 1e-3); input and weight gradients pass; the negative controls fail
Measured
2026-10-13
GPU-hours
23.36
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
  • 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 NCCL_NVLS_ENABLE=$([ "$NVLINK_COUNT" -gt 0 ] && echo 1 || echo 0)

python3 -m <slime entry module> --async \
  --actor-num-nodes 6 --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 \
  --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 6 \
  --context-parallel-size 1 --expert-model-parallel-size 8 --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 \
  --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 \
  --cross-entropy-loss-fusion --cross-entropy-fusion-impl te \
  --attention-dropout 0.0 --hidden-dropout 0.0 \
  --accumulate-allreduce-grads-in-fp32 --attention-softmax-in-fp32 --attention-backend flash \
  --moe-token-dispatcher-type flex
recipes/qwen3.5-122b-a10b/h100/attune-tp4-pp6-ep8-flex-mt8k.yamlDownloadGitHub
# recipes/qwen3.5-122b-a10b/h100/attune-tp4-pp6-ep8-flex-mt8k.yaml
# An OpenRecipes recipe; the fields are documented in schema/recipe.schema.yaml.
schema_version: 1
model: Qwen3.5-122B-A10B
gpu:
  type: H100
  memory_gb: 80
  count: 48
  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'
  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: TP4 PP6 EP8 DP2
  summary: TP 4, PP 6, CP 1, EP 8, ETP 1, DP 2, sequence parallel; one 8,192-token row per microbatch; full recompute
    (uniform, 1 layer); precision-aware distributed optimizer; flex dispatcher; TE cross-entropy fusion; no offload
  highlights: max 8,192 tokens/GPU, flex dispatcher
  tp: 4
  pp: 6
  vpp: 1
  cp: 1
  ep: 8
  etp: 1
  dp: 2
  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: flex
result:
  tokens_per_s_per_gpu: 842.76
  usd_per_million_tokens: 2.27
  step_time_cv_percent: 8.31
  mfu: 0.0461
  mfu_note: null
  active_params:
    count: 9008136192
    source: https://huggingface.co/Qwen/Qwen3.5-122B-A10B
    revision: dc4d348443bc740c68e2d77492492c11606384d5
  measured: 2026-10-13
  steps_per_run: 20
  window:
  - 5
  - 19
  window_tokens: 15728640
  runs: 1
  profiler: false
  gpu_hours: 23.36
  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: flex dispatcher against an FP32 reference
    shape: 'one Qwen3.5-122B-A10B MoE layer: hidden 3072, 256 experts, top-8, expert FFN 1024, EP 8, ETP 1; TP 4 with
      sequence parallel and TP 1; 4,096 and 16,384 tokens per rank'
    gpu: H100
    date: 2026-10-11
    result: pass
    detail: routing identical; worst tensor (output) relative L2 error 4.11e-3 against a bound of 8.76e-3 (2 x the alltoall
      dispatcher's error + 1e-3); input and weight gradients pass; the negative controls fail
  convergence:
    steps: 20
    result: pass
    correlation: 0.999997
    final_loss_difference: 2.6e-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
  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 NCCL_NVLS_ENABLE=$([ "$NVLINK_COUNT" -gt 0 ] && echo 1 || echo 0)

  python3 -m <slime entry module> --async \
    --actor-num-nodes 6 --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 \
    --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 6 \
    --context-parallel-size 1 --expert-model-parallel-size 8 --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 \
    --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 \
    --cross-entropy-loss-fusion --cross-entropy-fusion-impl te \
    --attention-dropout 0.0 --hidden-dropout 0.0 \
    --accumulate-allreduce-grads-in-fp32 --attention-softmax-in-fp32 --attention-backend flash \
    --moe-token-dispatcher-type flex
#3attune-tp4-pp3-ep16822.1 tokens/s/GPUTP4 PP3 EP16 DP4max 8,192 tokens/GPU
Author
ATTUNE (cite)
GPUs
48 x H100 80 GB (AWS p5.48xlarge, 8 per node)
Layout
TP 4, PP 3, CP 1, EP 16, ETP 1, DP 4, sequence parallel; one 8,192-token row per microbatch; full recompute (uniform, 1 layer); precision-aware distributed optimizer; alltoall dispatcher; TE cross-entropy fusion; no offload
Tokens/s per GPU
822.08
$ / 1M tokens
$2.32
MFU
4.5% (N = 9,008,136,192 active parameters per token)
Step-time CV
6.52%
Convergence check
correlation 0.999998, final loss difference 4.1e-05, 0 pointwise failures over 20 steps
Measured
2026-10-13
GPU-hours
18.72
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
  • 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 NCCL_NVLS_ENABLE=$([ "$NVLINK_COUNT" -gt 0 ] && echo 1 || echo 0)

python3 -m <slime entry module> --async \
  --actor-num-nodes 6 --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 \
  --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 3 \
  --context-parallel-size 1 --expert-model-parallel-size 16 --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 \
  --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 \
  --cross-entropy-loss-fusion --cross-entropy-fusion-impl te \
  --attention-dropout 0.0 --hidden-dropout 0.0 \
  --accumulate-allreduce-grads-in-fp32 --attention-softmax-in-fp32 --attention-backend flash
recipes/qwen3.5-122b-a10b/h100/attune-tp4-pp3-ep16.yamlDownloadGitHub
# recipes/qwen3.5-122b-a10b/h100/attune-tp4-pp3-ep16.yaml
# An OpenRecipes recipe; the fields are documented in schema/recipe.schema.yaml.
schema_version: 1
model: Qwen3.5-122B-A10B
gpu:
  type: H100
  memory_gb: 80
  count: 48
  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'
  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: TP4 PP3 EP16 DP4
  summary: TP 4, PP 3, CP 1, EP 16, ETP 1, DP 4, sequence parallel; one 8,192-token row per microbatch; full recompute
    (uniform, 1 layer); precision-aware distributed optimizer; alltoall dispatcher; TE cross-entropy fusion; no offload
  highlights: max 8,192 tokens/GPU
  tp: 4
  pp: 3
  vpp: 1
  cp: 1
  ep: 16
  etp: 1
  dp: 4
  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: 822.08
  usd_per_million_tokens: 2.32
  step_time_cv_percent: 6.52
  mfu: 0.0449
  mfu_note: null
  active_params:
    count: 9008136192
    source: https://huggingface.co/Qwen/Qwen3.5-122B-A10B
    revision: dc4d348443bc740c68e2d77492492c11606384d5
  measured: 2026-10-13
  steps_per_run: 20
  window:
  - 5
  - 19
  window_tokens: 15728640
  runs: 1
  profiler: false
  gpu_hours: 18.72
  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: 0.999998
    final_loss_difference: 4.1e-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
  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 NCCL_NVLS_ENABLE=$([ "$NVLINK_COUNT" -gt 0 ] && echo 1 || echo 0)

  python3 -m <slime entry module> --async \
    --actor-num-nodes 6 --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 \
    --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 3 \
    --context-parallel-size 1 --expert-model-parallel-size 16 --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 \
    --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 \
    --cross-entropy-loss-fusion --cross-entropy-fusion-impl te \
    --attention-dropout 0.0 --hidden-dropout 0.0 \
    --accumulate-allreduce-grads-in-fp32 --attention-softmax-in-fp32 --attention-backend flash

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}
}

Qwen3.5-122B-A10B on other GPUs

H100 · 48 GPUs · this page
999.7
tokens/s/GPU · $1.91 / 1M tokens
H200 · 32 GPUs
2,451.4
tokens/s/GPU · $0.90 / 1M tokens
B200 · 16 GPUs
2,475.8
tokens/s/GPU · $1.60 / 1M tokens