Training gpt-oss-120b on H100

Author

ATTUNE (cite)

GPUs

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

Layout

TP 1, PP 8 (3/5/5/5/5/5/5/3), CP 1, EP 8, ETP 1, DP 8; one 8,192-token row per microbatch (max 8,192 tokens per GPU); full recompute (uniform, 1 layer); alltoall dispatcher; distributed optimizer; TE cross-entropy fusion; no offload

Tokens/s per GPU

2,569.39

$ / 1M tokens

$0.74

MFU

8.0% (N = 5,131,100,160 active parameters per token)

Step-time CV

1.35%

Convergence check

correlation 0.999998, final loss difference 8.9e-05, 0 pointwise failures over 20 steps

Measured

2026-10-14

GPU-hours

9.19

Rows

gptoss-tulu3-fixed8192-v1 (rows_sha256 de2b9bb0f64927fd4214b45c3718fec01d9f2579da462e7615b1154d88185963)

MODEL_ARGS=(
  --spec attune_gpt_oss_yarn get_gpt_oss_yarn_spec
  --num-layers 36 --hidden-size 2880 --ffn-hidden-size 2880
  --num-attention-heads 64 --group-query-attention --num-query-groups 8 --kv-channels 64
  --padded-vocab-size 201088 --tokenizer-type HuggingFaceTokenizer
  --normalization RMSNorm --norm-epsilon 1e-5 --untie-embeddings-and-output-weights
  --position-embedding-type rope --rotary-percent 1.0 --rotary-base 150000 --no-rope-fusion
  --num-experts 128 --moe-ffn-hidden-size 2880 --moe-router-topk 4 --moe-router-score-function softmax
  --moe-grouped-gemm --quick-geglu --glu-linear-offset 1.0 --activation-func-clamp-value 7.0
  --softmax-type learnable --window-attn-skip-freq 2 --window-size 127,0
)
export CUDA_DEVICE_MAX_CONNECTIONS=8 NVTE_FWD_LAYERNORM_SM_MARGIN=16 NVTE_BWD_LAYERNORM_SM_MARGIN=16
export TORCH_NCCL_AVOID_RECORD_STREAMS=1 NCCL_NVLS_ENABLE=0 TORCH_NCCL_HIGH_PRIORITY=1
export AUTOTUNE_SEQ_LEN=8192 PARITY_PROBE=0

python3 -m <slime entry module> --async \
  --actor-num-nodes 8 --actor-num-gpus-per-node 8 \
  ${MODEL_ARGS[@]} \
  --seq-length 8192 \
  --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 8 \
  --decoder-first-pipeline-num-layers 3 --decoder-last-pipeline-num-layers 3 \
  --context-parallel-size 1 --expert-model-parallel-size 8 --expert-tensor-parallel-size 1 \
  --use-dynamic-batch-size --max-tokens-per-gpu 8192 \
  --recompute-granularity full --recompute-method uniform --recompute-num-layers 1 \
  --moe-router-load-balancing-type none --moe-aux-loss-coeff 0.0 \
  --moe-token-dispatcher-type alltoall --moe-permute-fusion --moe-router-fusion --no-bias-dropout-fusion \
  --cross-entropy-loss-fusion --cross-entropy-fusion-impl te \
  --manual-gc --manual-gc-interval 100 \
  --attention-dropout 0.0 --hidden-dropout 0.0 \
  --optimizer adam --lr 5e-6 --min-lr 5e-7 --lr-decay-style cosine --lr-warmup-iters 10 --lr-decay-iters 50 \
  --adam-beta1 0.9 --adam-beta2 0.95 --adam-eps 1e-8 --clip-grad 1.0 \
  --weight-decay 0.033 --weight-decay-incr-style constant \
  --use-distributed-optimizer --grad-reduce-in-bf16 --no-check-for-nan-in-loss-and-grad --seed 1234

# recipes/gpt-oss-120b/h100/attune-tp1-pp8-ep8.yaml
# An OpenRecipes recipe; the fields are documented in schema/recipe.schema.yaml.
schema_version: 1
model: gpt-oss-120b
gpu:
  type: H100
  memory_gb: 80
  count: 64
  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 the training image's local changes
  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 (installed; the alltoall dispatcher does not use it)
  python: 3.12.3
  network: 8 GPUs per node, EFA, 3.2 Tb/s RDMA per node
  source: 'in-pod version dump of the measured run''s training image (same digest as the gpt-oss-20b H100 recipe),
    2026-10-13: installed TE, PyTorch, DeepEP and megatron-core, Megatron-LM git commit, Python; NCCL from the measured
    run''s log'
workload:
  row_tokens: 8192
  rows: SFT conversations
  global_batch: 128
  precision: bf16
  dataset:
    name: Tulu 3 SFT mixture
    source: allenai/tulu-3-sft-mixture
    revision: b14afda60f1bbebe55d5d2fa1e4df5042f97f8be
    license: ODC-By 1.0
    rows_sha256: de2b9bb0f64927fd4214b45c3718fec01d9f2579da462e7615b1154d88185963
    build:
      name: gptoss-tulu3-fixed8192-v1
      tokenizer: gpt-oss (the gpt-oss-20b tokenizer, revision 6cee5e81ee83917806bbde320786a8fb61efebee, byte-identical
        to gpt-oss-120b's; tokenizer.json sha256 0614fe83cadab421296e664e1f48f4261fa8fef6e03e63bb75c20f38e37d07d3)
      loss_mask: gpt_oss
      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: 0.621
      parquet_sha256: 1fc671ba1a3262edeed53f82aca81c05fc694ed40dd8674a9e4c2d96055fed90
      parquet_bytes: 70055050
      builder: data/generate_fixed_sft.py with --loss-mask-type gpt_oss (patch_gptoss_mask.py; data/README.md)
layout:
  label: TP1 PP8 EP8 DP8
  summary: TP 1, PP 8 (3/5/5/5/5/5/5/3), CP 1, EP 8, ETP 1, DP 8; one 8,192-token row per microbatch (max 8,192 tokens per
    GPU); full recompute (uniform, 1 layer); alltoall dispatcher; distributed optimizer; TE cross-entropy fusion; no
    offload
  highlights: PP 8 with uneven split 3/5/5/5/5/5/5/3, full recompute
  tp: 1
  pp: 8
  vpp: 1
  cp: 1
  ep: 8
  etp: 1
  dp: 8
  sequence_parallel: false
  layer_split:
    first: 3
    last: 3
  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: 2569.39
  usd_per_million_tokens: 0.74
  step_time_cv_percent: 1.35
  mfu: 0.08
  mfu_note: null
  active_params:
    count: 5131100160
    source: https://huggingface.co/openai/gpt-oss-120b
    revision: b5c939de8f754692c1647ca79fbf85e8c1e70f8a
  measured: 2026-10-14
  steps_per_run: 20
  window:
  - 5
  - 19
  window_tokens: 15728640
  runs: 1
  profiler: false
  gpu_hours: 9.19
  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; the recipe changes only the pipeline-parallel size and split, placement arguments, from its reference
    recipe
  reference: 'Megatron-Bridge v0.4.2''s gpt-oss-120b pretraining performance recipe for 64 x H100 (TP 1, PP 4 split
    9/9/9/9, EP 8), with four changes carried by both arms: max 8,192 tokens per GPU and full
    recompute (forced by 8,192-token rows), no data-parallel overlap (--overlap-grad-reduce and --overlap-param-gather
    removed), and learning rate 5e-6 with minimum 5e-7 in place of the pretraining 3e-4 and 3e-5.'
  configuration: pass
  component_tests: []
  convergence:
    steps: 20
    result: pass
    correlation: 0.999998
    final_loss_difference: 8.9e-05
    pointwise_failures: 0
    note: 'candidate arm: the measured run; reference arm: a 20-step run of the reference recipe on the same image, rows
      and GPU type; both with the parity probe loaded and off'
  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: '8'
  NVTE_FWD_LAYERNORM_SM_MARGIN: '16'
  NVTE_BWD_LAYERNORM_SM_MARGIN: '16'
  TORCH_NCCL_AVOID_RECORD_STREAMS: '1'
  NCCL_NVLS_ENABLE: '0'
  TORCH_NCCL_HIGH_PRIORITY: '1'
  AUTOTUNE_SEQ_LEN: '8192'
  PARITY_PROBE: '0'
patches:
- patches/attune_gpt_oss_yarn.py
run: |
  MODEL_ARGS=(
    --spec attune_gpt_oss_yarn get_gpt_oss_yarn_spec
    --num-layers 36 --hidden-size 2880 --ffn-hidden-size 2880
    --num-attention-heads 64 --group-query-attention --num-query-groups 8 --kv-channels 64
    --padded-vocab-size 201088 --tokenizer-type HuggingFaceTokenizer
    --normalization RMSNorm --norm-epsilon 1e-5 --untie-embeddings-and-output-weights
    --position-embedding-type rope --rotary-percent 1.0 --rotary-base 150000 --no-rope-fusion
    --num-experts 128 --moe-ffn-hidden-size 2880 --moe-router-topk 4 --moe-router-score-function softmax
    --moe-grouped-gemm --quick-geglu --glu-linear-offset 1.0 --activation-func-clamp-value 7.0
    --softmax-type learnable --window-attn-skip-freq 2 --window-size 127,0
  )
  export CUDA_DEVICE_MAX_CONNECTIONS=8 NVTE_FWD_LAYERNORM_SM_MARGIN=16 NVTE_BWD_LAYERNORM_SM_MARGIN=16
  export TORCH_NCCL_AVOID_RECORD_STREAMS=1 NCCL_NVLS_ENABLE=0 TORCH_NCCL_HIGH_PRIORITY=1
  export AUTOTUNE_SEQ_LEN=8192 PARITY_PROBE=0

  python3 -m <slime entry module> --async \
    --actor-num-nodes 8 --actor-num-gpus-per-node 8 \
    ${MODEL_ARGS[@]} \
    --seq-length 8192 \
    --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 8 \
    --decoder-first-pipeline-num-layers 3 --decoder-last-pipeline-num-layers 3 \
    --context-parallel-size 1 --expert-model-parallel-size 8 --expert-tensor-parallel-size 1 \
    --use-dynamic-batch-size --max-tokens-per-gpu 8192 \
    --recompute-granularity full --recompute-method uniform --recompute-num-layers 1 \
    --moe-router-load-balancing-type none --moe-aux-loss-coeff 0.0 \
    --moe-token-dispatcher-type alltoall --moe-permute-fusion --moe-router-fusion --no-bias-dropout-fusion \
    --cross-entropy-loss-fusion --cross-entropy-fusion-impl te \
    --manual-gc --manual-gc-interval 100 \
    --attention-dropout 0.0 --hidden-dropout 0.0 \
    --optimizer adam --lr 5e-6 --min-lr 5e-7 --lr-decay-style cosine --lr-warmup-iters 10 --lr-decay-iters 50 \
    --adam-beta1 0.9 --adam-beta2 0.95 --adam-eps 1e-8 --clip-grad 1.0 \
    --weight-decay 0.033 --weight-decay-incr-style constant \
    --use-distributed-optimizer --grad-reduce-in-bf16 --no-check-for-nan-in-loss-and-grad --seed 1234
notes: |
  - The learning rate is 5e-6 (minimum 5e-7), from Megatron-Bridge's gpt-oss-120b SFT recipe; the tuning campaign kept it
    fixed.
  - On 80 GB GPUs this layout needs full recompute, max 8,192 tokens per GPU and the lighter end stages: in the tuning
    campaign, PP 8 without recompute, at 16,384 tokens per GPU, or with an end-heavier 6/4/4/4/4/4/4/6 split ran out of
    memory.

Author

ATTUNE (cite)

GPUs

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

Layout

TP 1, PP 4 (9/10/10/7), CP 1, EP 8, ETP 1, DP 16; one 8,192-token row per microbatch (max 8,192 tokens per GPU); full recompute (uniform, 1 layer); alltoall dispatcher; distributed optimizer; TE cross-entropy fusion; no offload

Tokens/s per GPU

2,326.65

$ / 1M tokens

$0.82

MFU

7.2% (N = 5,131,100,160 active parameters per token)

Step-time CV

0.91%

Convergence check

correlation 0.999997, final loss difference 1.4e-05, 0 pointwise failures over 20 steps

Measured

2026-10-14

GPU-hours

8.23

Rows

gptoss-tulu3-fixed8192-v1 (rows_sha256 de2b9bb0f64927fd4214b45c3718fec01d9f2579da462e7615b1154d88185963)

MODEL_ARGS=(
  --spec attune_gpt_oss_yarn get_gpt_oss_yarn_spec
  --num-layers 36 --hidden-size 2880 --ffn-hidden-size 2880
  --num-attention-heads 64 --group-query-attention --num-query-groups 8 --kv-channels 64
  --padded-vocab-size 201088 --tokenizer-type HuggingFaceTokenizer
  --normalization RMSNorm --norm-epsilon 1e-5 --untie-embeddings-and-output-weights
  --position-embedding-type rope --rotary-percent 1.0 --rotary-base 150000 --no-rope-fusion
  --num-experts 128 --moe-ffn-hidden-size 2880 --moe-router-topk 4 --moe-router-score-function softmax
  --moe-grouped-gemm --quick-geglu --glu-linear-offset 1.0 --activation-func-clamp-value 7.0
  --softmax-type learnable --window-attn-skip-freq 2 --window-size 127,0
)
export CUDA_DEVICE_MAX_CONNECTIONS=8 NVTE_FWD_LAYERNORM_SM_MARGIN=16 NVTE_BWD_LAYERNORM_SM_MARGIN=16
export TORCH_NCCL_AVOID_RECORD_STREAMS=1 NCCL_NVLS_ENABLE=0 TORCH_NCCL_HIGH_PRIORITY=1
export AUTOTUNE_SEQ_LEN=8192 PARITY_PROBE=0

python3 -m <slime entry module> --async \
  --actor-num-nodes 8 --actor-num-gpus-per-node 8 \
  ${MODEL_ARGS[@]} \
  --seq-length 8192 \
  --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 \
  --decoder-first-pipeline-num-layers 9 --decoder-last-pipeline-num-layers 7 \
  --context-parallel-size 1 --expert-model-parallel-size 8 --expert-tensor-parallel-size 1 \
  --use-dynamic-batch-size --max-tokens-per-gpu 8192 \
  --recompute-granularity full --recompute-method uniform --recompute-num-layers 1 \
  --moe-router-load-balancing-type none --moe-aux-loss-coeff 0.0 \
  --moe-token-dispatcher-type alltoall --moe-permute-fusion --moe-router-fusion --no-bias-dropout-fusion \
  --cross-entropy-loss-fusion --cross-entropy-fusion-impl te \
  --manual-gc --manual-gc-interval 100 \
  --attention-dropout 0.0 --hidden-dropout 0.0 \
  --optimizer adam --lr 5e-6 --min-lr 5e-7 --lr-decay-style cosine --lr-warmup-iters 10 --lr-decay-iters 50 \
  --adam-beta1 0.9 --adam-beta2 0.95 --adam-eps 1e-8 --clip-grad 1.0 \
  --weight-decay 0.033 --weight-decay-incr-style constant \
  --use-distributed-optimizer --grad-reduce-in-bf16 --no-check-for-nan-in-loss-and-grad --seed 1234

# recipes/gpt-oss-120b/h100/attune-tp1-pp4-ep8.yaml
# An OpenRecipes recipe; the fields are documented in schema/recipe.schema.yaml.
schema_version: 1
model: gpt-oss-120b
gpu:
  type: H100
  memory_gb: 80
  count: 64
  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 the training image's local changes
  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 (installed; the alltoall dispatcher does not use it)
  python: 3.12.3
  network: 8 GPUs per node, EFA, 3.2 Tb/s RDMA per node
  source: 'in-pod version dump of the measured run''s training image (same digest as the gpt-oss-20b H100 recipe),
    2026-10-13: installed TE, PyTorch, DeepEP and megatron-core, Megatron-LM git commit, Python; NCCL from the measured
    run''s log'
workload:
  row_tokens: 8192
  rows: SFT conversations
  global_batch: 128
  precision: bf16
  dataset:
    name: Tulu 3 SFT mixture
    source: allenai/tulu-3-sft-mixture
    revision: b14afda60f1bbebe55d5d2fa1e4df5042f97f8be
    license: ODC-By 1.0
    rows_sha256: de2b9bb0f64927fd4214b45c3718fec01d9f2579da462e7615b1154d88185963
    build:
      name: gptoss-tulu3-fixed8192-v1
      tokenizer: gpt-oss (the gpt-oss-20b tokenizer, revision 6cee5e81ee83917806bbde320786a8fb61efebee, byte-identical
        to gpt-oss-120b's; tokenizer.json sha256 0614fe83cadab421296e664e1f48f4261fa8fef6e03e63bb75c20f38e37d07d3)
      loss_mask: gpt_oss
      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: 0.621
      parquet_sha256: 1fc671ba1a3262edeed53f82aca81c05fc694ed40dd8674a9e4c2d96055fed90
      parquet_bytes: 70055050
      builder: data/generate_fixed_sft.py with --loss-mask-type gpt_oss (patch_gptoss_mask.py; data/README.md)
layout:
  label: TP1 PP4 EP8 DP16
  summary: TP 1, PP 4 (9/10/10/7), CP 1, EP 8, ETP 1, DP 16; one 8,192-token row per microbatch (max 8,192 tokens per
    GPU); full recompute (uniform, 1 layer); alltoall dispatcher; distributed optimizer; TE cross-entropy fusion; no
    offload
  highlights: uneven pipeline split 9/10/10/7, full recompute
  tp: 1
  pp: 4
  vpp: 1
  cp: 1
  ep: 8
  etp: 1
  dp: 16
  sequence_parallel: false
  layer_split:
    first: 9
    last: 7
  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: 2326.65
  usd_per_million_tokens: 0.82
  step_time_cv_percent: 0.91
  mfu: 0.0724
  mfu_note: null
  active_params:
    count: 5131100160
    source: https://huggingface.co/openai/gpt-oss-120b
    revision: b5c939de8f754692c1647ca79fbf85e8c1e70f8a
  measured: 2026-10-14
  steps_per_run: 20
  window:
  - 5
  - 19
  window_tokens: 15728640
  runs: 1
  profiler: false
  gpu_hours: 8.23
  gpu_hours_note: GPUs x (scheduler admission to finish), including startup; the same run is the candidate arm of the
    convergence check
correctness:
  result: pass
  summary: passed; the recipe changes only the pipeline split, a placement argument, from its reference recipe
  reference: 'Megatron-Bridge v0.4.2''s gpt-oss-120b pretraining performance recipe for 64 x H100 (TP 1, PP 4 split
    9/9/9/9, EP 8), with four changes carried by both arms: max 8,192 tokens per GPU and full
    recompute (forced by 8,192-token rows), no data-parallel overlap (--overlap-grad-reduce and --overlap-param-gather
    removed), and learning rate 5e-6 with minimum 5e-7 in place of the pretraining 3e-4 and 3e-5.'
  configuration: pass
  component_tests: []
  convergence:
    steps: 20
    result: pass
    correlation: 0.999997
    final_loss_difference: 1.4e-05
    pointwise_failures: 0
    note: 'candidate arm: the measured run; reference arm: a 20-step run of the reference recipe on the same image, rows
      and GPU type; both with the parity probe loaded and off'
  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: '8'
  NVTE_FWD_LAYERNORM_SM_MARGIN: '16'
  NVTE_BWD_LAYERNORM_SM_MARGIN: '16'
  TORCH_NCCL_AVOID_RECORD_STREAMS: '1'
  NCCL_NVLS_ENABLE: '0'
  TORCH_NCCL_HIGH_PRIORITY: '1'
  AUTOTUNE_SEQ_LEN: '8192'
  PARITY_PROBE: '0'
patches:
- patches/attune_gpt_oss_yarn.py
run: |
  MODEL_ARGS=(
    --spec attune_gpt_oss_yarn get_gpt_oss_yarn_spec
    --num-layers 36 --hidden-size 2880 --ffn-hidden-size 2880
    --num-attention-heads 64 --group-query-attention --num-query-groups 8 --kv-channels 64
    --padded-vocab-size 201088 --tokenizer-type HuggingFaceTokenizer
    --normalization RMSNorm --norm-epsilon 1e-5 --untie-embeddings-and-output-weights
    --position-embedding-type rope --rotary-percent 1.0 --rotary-base 150000 --no-rope-fusion
    --num-experts 128 --moe-ffn-hidden-size 2880 --moe-router-topk 4 --moe-router-score-function softmax
    --moe-grouped-gemm --quick-geglu --glu-linear-offset 1.0 --activation-func-clamp-value 7.0
    --softmax-type learnable --window-attn-skip-freq 2 --window-size 127,0
  )
  export CUDA_DEVICE_MAX_CONNECTIONS=8 NVTE_FWD_LAYERNORM_SM_MARGIN=16 NVTE_BWD_LAYERNORM_SM_MARGIN=16
  export TORCH_NCCL_AVOID_RECORD_STREAMS=1 NCCL_NVLS_ENABLE=0 TORCH_NCCL_HIGH_PRIORITY=1
  export AUTOTUNE_SEQ_LEN=8192 PARITY_PROBE=0

  python3 -m <slime entry module> --async \
    --actor-num-nodes 8 --actor-num-gpus-per-node 8 \
    ${MODEL_ARGS[@]} \
    --seq-length 8192 \
    --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 \
    --decoder-first-pipeline-num-layers 9 --decoder-last-pipeline-num-layers 7 \
    --context-parallel-size 1 --expert-model-parallel-size 8 --expert-tensor-parallel-size 1 \
    --use-dynamic-batch-size --max-tokens-per-gpu 8192 \
    --recompute-granularity full --recompute-method uniform --recompute-num-layers 1 \
    --moe-router-load-balancing-type none --moe-aux-loss-coeff 0.0 \
    --moe-token-dispatcher-type alltoall --moe-permute-fusion --moe-router-fusion --no-bias-dropout-fusion \
    --cross-entropy-loss-fusion --cross-entropy-fusion-impl te \
    --manual-gc --manual-gc-interval 100 \
    --attention-dropout 0.0 --hidden-dropout 0.0 \
    --optimizer adam --lr 5e-6 --min-lr 5e-7 --lr-decay-style cosine --lr-warmup-iters 10 --lr-decay-iters 50 \
    --adam-beta1 0.9 --adam-beta2 0.95 --adam-eps 1e-8 --clip-grad 1.0 \
    --weight-decay 0.033 --weight-decay-incr-style constant \
    --use-distributed-optimizer --grad-reduce-in-bf16 --no-check-for-nan-in-loss-and-grad --seed 1234
notes: |
  - The learning rate is 5e-6 (minimum 5e-7), from Megatron-Bridge's gpt-oss-120b SFT recipe; ATTUNE's tuning campaign
    ran at 3e-4.
  - Keep the uneven 9/10/10/7 split and full recompute on 80 GB GPUs: with the even 9/9/9/9 split the last stage, which
    also holds the output layer and the loss, runs short of memory.