OpenRecipes
‹ All models
GLM

Training GLM-5.2-744B-A40B on H100

Ranked by tokens/s per GPU · 8,192-token SFT rows · global batch 128 · BF16
#Tokens/s per GPU$ / 1M tokensGPUsLayoutMeasuredAuthor
#168.3$28.00256TP4 PP8 EP32 DP8
max 8,192 tokens/GPU, CPU optimizer offload		2026-10-11ATTUNERecipe ›
#268.2$28.03256TP4 PP8 EP32 DP8
max 8,192 tokens/GPU, recompute 2 layers, CPU optimizer offload		2026-10-13ATTUNERecipe ›
#366.2$28.85256TP4 PP8 CP2 EP32 DP4
max 8,192 tokens/GPU, CPU optimizer offload		2026-10-13ATTUNERecipe ›

Notes for GLM-5.2-744B-A40B

Patch slime's mask_utils.py with data/patch_glm5_mask.py before building the rows.

The steps every recipe shares: Running a recipe.

Recipes

#1attune-tp4-pp8-cp168.3 tokens/s/GPUTP4 PP8 EP32 DP8max 8,192 tokens/GPU, CPU optimizer offload
Author
ATTUNE (cite)
GPUs
256 x H100 80 GB (AWS p5.48xlarge, 8 per node)
Layout
TP 4, PP 8 (first stage 14 layers, last 16), CP 1, EP 32, ETP 1, DP 8, sequence parallel; one 8,192-token row per microbatch; full recompute (uniform, 1 layer); CPU optimizer offload with overlapped copies and the precision-aware optimizer; alltoall dispatcher
Tokens/s per GPU
68.26
$ / 1M tokens
$28.00
MFU
1.7% (N = 40,297,758,720 active parameters per token)
Step-time CV
1.73%
Convergence check
correlation 0.99954, final loss difference 0.0015, 0 pointwise failures over 20 steps
Measured
2026-10-11, 14-step run scored over steps 4–13
GPU-hours
122.2
Rows
glm52-tulu3-fixed8192-v1 (rows_sha256 bb302c134974b96f3905e9c87bfc4c0bb2be674a50e6dad09121aed2a0eb3455)

Stack

framework
slime 0.3.0 with local patches; commit not recorded
megatron lm
commit 1dcf0daf (megatron-core 0.16.0rc0) with local patches
transformer engine
2.10.0
pytorch
2.11.0 (CUDA 12.9 build)
cuda
12.9
nccl
2.28.9 (PyTorch's NCCL wheel; printed by the measured run)
deepep
1.2.1
python
3.12.3
network
8 GPUs per node, EFA, 3.2 Tb/s RDMA per node

Environment

  • CUDA_DEVICE_MAX_CONNECTIONS=1
  • INDEXER_ROPE_NEOX_STYLE=0
  • NCCL_NVLS_ENABLE=$([ "$NVLINK_COUNT" -gt 0 ] && echo 1 || echo 0)
  • PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
CommandHow to run
export CUDA_DEVICE_MAX_CONNECTIONS=1
export INDEXER_ROPE_NEOX_STYLE=0
export NCCL_NVLS_ENABLE=$([ "$NVLINK_COUNT" -gt 0 ] && echo 1 || echo 0)
export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True  # set in the pod environment

python3 -m <slime entry module> --async \
  --actor-num-nodes 32 \
  --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 4 \
  --sequence-parallel \
  --pipeline-model-parallel-size 8 \
  --decoder-first-pipeline-num-layers 14 \
  --decoder-last-pipeline-num-layers 16 \
  --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 \
  --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 \
  --moe-token-dispatcher-type alltoall \
  --allgather-cp
recipes/glm-5.2-744b-a40b/h100/attune-tp4-pp8-cp1.yamlDownloadGitHub
# recipes/glm-5.2-744b-a40b/h100/attune-tp4-pp8-cp1.yaml
# An OpenRecipes recipe; the fields are documented in schema/recipe.schema.yaml.
schema_version: 1
model: GLM-5.2-744B-A40B
gpu:
  type: H100
  memory_gb: 80
  count: 256
  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 with local patches; commit not recorded
  megatron_lm: commit 1dcf0daf (megatron-core 0.16.0rc0) with local patches
  transformer_engine: 2.10.0
  pytorch: 2.11.0 (CUDA 12.9 build)
  cuda: '12.9'
  nccl: 2.28.9 (PyTorch's NCCL wheel; printed by the measured run)
  deepep: 1.2.1
  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), 2026-10-11: Megatron-LM git commit
    and status, installed megatron-core, TE, PyTorch, NCCL, DeepEP and slime, Python'
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: bb302c134974b96f3905e9c87bfc4c0bb2be674a50e6dad09121aed2a0eb3455
    build:
      name: glm52-tulu3-fixed8192-v1
      tokenizer: GLM-5.2 (tokenizer.json sha256 19e773648cb4e65de8660ea6365e10acca112d42a854923df93db4a6f333a82d)
      loss_mask: glm5
      rows: 12800
      row_tokens: 8192
      padding_tokens: 0
      seed: 20260828
      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: 0.661
      parquet_sha256: 5d951089abd622511a40e0b3a1e5e36ad56859459e4d8c3aff14538cadbd0743
      parquet_bytes: 169929706
      builder: data/generate_fixed_sft.py with patch_glm5_mask.py (data/README.md)
layout:
  label: TP4 PP8 EP32 DP8
  summary: TP 4, PP 8 (first stage 14 layers, last 16), CP 1, EP 32, ETP 1, DP 8, sequence parallel; one 8,192-token row
    per microbatch; full recompute (uniform, 1 layer); CPU optimizer offload with overlapped copies and the precision-aware
    optimizer; alltoall dispatcher
  highlights: max 8,192 tokens/GPU, CPU optimizer offload
  tp: 4
  pp: 8
  vpp: 1
  cp: 1
  ep: 32
  etp: 1
  dp: 8
  sequence_parallel: true
  layer_split:
    first: 14
    last: 16
  max_tokens_per_gpu: 8192
  rows_per_microbatch: 1
  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: 68.26
  usd_per_million_tokens: 28.0
  step_time_cv_percent: 1.73
  mfu: 0.0167
  mfu_note: null
  active_params:
    count: 40297758720
    source: https://huggingface.co/zai-org/GLM-5.2
    revision: cf457fa734ab149ffef225f80893eb38c6ff5cdc
  measured: 2026-10-11
  steps_per_run: 14
  window:
  - 4
  - 13
  window_tokens: 10485760
  runs: 1
  profiler: false
  gpu_hours: 122.2
  gpu_hours_note: GPUs x (job start to end), including startup
correctness:
  result: pass
  summary: passed
  reference: Not published.
  configuration: pass
  component_tests: []
  convergence:
    steps: 20
    result: pass
    correlation: 0.99954
    final_loss_difference: 0.0015
    pointwise_failures: 0
    note: on the measured run's training image
  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'
  INDEXER_ROPE_NEOX_STYLE: '0'
  NCCL_NVLS_ENABLE: $([ "$NVLINK_COUNT" -gt 0 ] && echo 1 || echo 0)
  PYTORCH_CUDA_ALLOC_CONF: expandable_segments:True
patches: []
run: |
  export CUDA_DEVICE_MAX_CONNECTIONS=1
  export INDEXER_ROPE_NEOX_STYLE=0
  export NCCL_NVLS_ENABLE=$([ "$NVLINK_COUNT" -gt 0 ] && echo 1 || echo 0)
  export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True  # set in the pod environment

  python3 -m <slime entry module> --async \
    --actor-num-nodes 32 \
    --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 4 \
    --sequence-parallel \
    --pipeline-model-parallel-size 8 \
    --decoder-first-pipeline-num-layers 14 \
    --decoder-last-pipeline-num-layers 16 \
    --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 \
    --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 \
    --moe-token-dispatcher-type alltoall \
    --allgather-cp
#2attune-tp4-pp8-cp1-rc268.2 tokens/s/GPUTP4 PP8 EP32 DP8max 8,192 tokens/GPU, recompute 2 layers, CPU optimizer offload
Author
ATTUNE (cite)
GPUs
256 x H100 80 GB (AWS p5.48xlarge, 8 per node)
Layout
TP 4, PP 8 (first stage 14 layers, last 16), CP 1, EP 32, ETP 1, DP 8, sequence parallel; one 8,192-token row per microbatch; full recompute (uniform, 2 layers); CPU optimizer offload with overlapped copies and the precision-aware optimizer; alltoall dispatcher; --allgather-cp
Tokens/s per GPU
68.18
$ / 1M tokens
$28.03
MFU
1.7% (N = 40,297,758,720 active parameters per token)
Step-time CV
1.37%
Convergence check
correlation 0.999972, final loss difference 0.00014, 0 pointwise failures over 20 steps
Measured
2026-10-13
GPU-hours
149.62
Rows
glm52-tulu3-fixed8192-v1 (rows_sha256 bb302c134974b96f3905e9c87bfc4c0bb2be674a50e6dad09121aed2a0eb3455)

Stack

framework
slime 0.3.0 with local patches; commit not recorded
megatron lm
commit 1dcf0daf (megatron-core 0.16.0rc0) with local patches
transformer engine
2.10.0
pytorch
2.11.0 (CUDA 12.9 build)
cuda
12.9
nccl
2.28.9 (PyTorch's NCCL wheel; printed by the measured run)
deepep
1.2.1
python
3.12.3
network
8 GPUs per node, EFA, 3.2 Tb/s RDMA per node

Environment

  • CUDA_DEVICE_MAX_CONNECTIONS=1
  • INDEXER_ROPE_NEOX_STYLE=0
  • NCCL_NVLS_ENABLE=$([ "$NVLINK_COUNT" -gt 0 ] && echo 1 || echo 0)
  • PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
CommandHow to run
export CUDA_DEVICE_MAX_CONNECTIONS=1
export INDEXER_ROPE_NEOX_STYLE=0
export NCCL_NVLS_ENABLE=$([ "$NVLINK_COUNT" -gt 0 ] && echo 1 || echo 0)
export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True  # set in the pod environment

python3 -m <slime entry module> --async \
  --actor-num-nodes 32 \
  --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 4 \
  --sequence-parallel \
  --pipeline-model-parallel-size 8 \
  --decoder-first-pipeline-num-layers 14 \
  --decoder-last-pipeline-num-layers 16 \
  --context-parallel-size 1 \
  --expert-model-parallel-size 32 \
  --expert-tensor-parallel-size 1 \
  --recompute-granularity full \
  --recompute-method uniform \
  --recompute-num-layers 2 \
  --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 \
  --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 \
  --moe-token-dispatcher-type alltoall \
  --allgather-cp
recipes/glm-5.2-744b-a40b/h100/attune-tp4-pp8-cp1-rc2.yamlDownloadGitHub
# recipes/glm-5.2-744b-a40b/h100/attune-tp4-pp8-cp1-rc2.yaml
# An OpenRecipes recipe; the fields are documented in schema/recipe.schema.yaml.
schema_version: 1
model: GLM-5.2-744B-A40B
gpu:
  type: H100
  memory_gb: 80
  count: 256
  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 with local patches; commit not recorded
  megatron_lm: commit 1dcf0daf (megatron-core 0.16.0rc0) with local patches
  transformer_engine: 2.10.0
  pytorch: 2.11.0 (CUDA 12.9 build)
  cuda: '12.9'
  nccl: 2.28.9 (PyTorch's NCCL wheel; printed by the measured run)
  deepep: 1.2.1
  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), 2026-10-11: Megatron-LM git commit
    and status, installed megatron-core, TE, PyTorch, NCCL, DeepEP and slime, Python'
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: bb302c134974b96f3905e9c87bfc4c0bb2be674a50e6dad09121aed2a0eb3455
    build:
      name: glm52-tulu3-fixed8192-v1
      tokenizer: GLM-5.2 (tokenizer.json sha256 19e773648cb4e65de8660ea6365e10acca112d42a854923df93db4a6f333a82d)
      loss_mask: glm5
      rows: 12800
      row_tokens: 8192
      padding_tokens: 0
      seed: 20260828
      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: 0.661
      parquet_sha256: 5d951089abd622511a40e0b3a1e5e36ad56859459e4d8c3aff14538cadbd0743
      parquet_bytes: 169929706
      builder: data/generate_fixed_sft.py with patch_glm5_mask.py (data/README.md)
layout:
  label: TP4 PP8 EP32 DP8
  summary: TP 4, PP 8 (first stage 14 layers, last 16), CP 1, EP 32, ETP 1, DP 8, sequence parallel; one 8,192-token row
    per microbatch; full recompute (uniform, 2 layers); CPU optimizer offload with overlapped copies and the precision-aware
    optimizer; alltoall dispatcher; --allgather-cp
  highlights: max 8,192 tokens/GPU, recompute 2 layers, CPU optimizer offload
  tp: 4
  pp: 8
  vpp: 1
  cp: 1
  ep: 32
  etp: 1
  dp: 8
  sequence_parallel: true
  layer_split:
    first: 14
    last: 16
  max_tokens_per_gpu: 8192
  rows_per_microbatch: 1
  recompute:
    granularity: full
    method: uniform
    num_layers: 2
  offload:
    optimizer_fraction: 1.0
    overlap_d2h_h2d: true
  dispatcher: alltoall
result:
  tokens_per_s_per_gpu: 68.18
  usd_per_million_tokens: 28.03
  step_time_cv_percent: 1.37
  mfu: 0.0167
  mfu_note: null
  active_params:
    count: 40297758720
    source: https://huggingface.co/zai-org/GLM-5.2
    revision: cf457fa734ab149ffef225f80893eb38c6ff5cdc
  measured: 2026-10-13
  steps_per_run: 20
  window:
  - 5
  - 19
  window_tokens: 15728640
  runs: 1
  profiler: false
  gpu_hours: 149.62
  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: 0.999972
    final_loss_difference: 0.00014
    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'
  INDEXER_ROPE_NEOX_STYLE: '0'
  NCCL_NVLS_ENABLE: $([ "$NVLINK_COUNT" -gt 0 ] && echo 1 || echo 0)
  PYTORCH_CUDA_ALLOC_CONF: expandable_segments:True
patches: []
run: |
  export CUDA_DEVICE_MAX_CONNECTIONS=1
  export INDEXER_ROPE_NEOX_STYLE=0
  export NCCL_NVLS_ENABLE=$([ "$NVLINK_COUNT" -gt 0 ] && echo 1 || echo 0)
  export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True  # set in the pod environment

  python3 -m <slime entry module> --async \
    --actor-num-nodes 32 \
    --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 4 \
    --sequence-parallel \
    --pipeline-model-parallel-size 8 \
    --decoder-first-pipeline-num-layers 14 \
    --decoder-last-pipeline-num-layers 16 \
    --context-parallel-size 1 \
    --expert-model-parallel-size 32 \
    --expert-tensor-parallel-size 1 \
    --recompute-granularity full \
    --recompute-method uniform \
    --recompute-num-layers 2 \
    --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 \
    --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 \
    --moe-token-dispatcher-type alltoall \
    --allgather-cp
#3attune-tp4-pp8-cp266.2 tokens/s/GPUTP4 PP8 CP2 EP32 DP4max 8,192 tokens/GPU, CPU optimizer offload
Author
ATTUNE (cite)
GPUs
256 x H100 80 GB (AWS p5.48xlarge, 8 per node)
Layout
TP 4, PP 8 (first stage 14 layers, last 16), CP 2, EP 32, ETP 1, DP 4, sequence parallel; max 8,192 tokens per GPU; full recompute (uniform, 1 layer); CPU optimizer offload with overlapped copies and the precision-aware optimizer; alltoall dispatcher; --allgather-cp
Tokens/s per GPU
66.24
$ / 1M tokens
$28.85
MFU
1.6% (N = 40,297,758,720 active parameters per token)
Step-time CV
2.59%
Convergence check
correlation 0.999974, final loss difference 0.00044, 0 pointwise failures over 20 steps
Measured
2026-10-13
GPU-hours
151.32
Rows
glm52-tulu3-fixed8192-v1 (rows_sha256 bb302c134974b96f3905e9c87bfc4c0bb2be674a50e6dad09121aed2a0eb3455)

Stack

framework
slime 0.3.0 with local patches; commit not recorded
megatron lm
commit 1dcf0daf (megatron-core 0.16.0rc0) with local patches
transformer engine
2.10.0
pytorch
2.11.0 (CUDA 12.9 build)
cuda
12.9
nccl
2.28.9 (PyTorch's NCCL wheel; printed by the measured run)
deepep
1.2.1
python
3.12.3
network
8 GPUs per node, EFA, 3.2 Tb/s RDMA per node

Environment

  • CUDA_DEVICE_MAX_CONNECTIONS=1
  • INDEXER_ROPE_NEOX_STYLE=0
  • NCCL_NVLS_ENABLE=$([ "$NVLINK_COUNT" -gt 0 ] && echo 1 || echo 0)
  • PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
CommandHow to run
export CUDA_DEVICE_MAX_CONNECTIONS=1
export INDEXER_ROPE_NEOX_STYLE=0
export NCCL_NVLS_ENABLE=$([ "$NVLINK_COUNT" -gt 0 ] && echo 1 || echo 0)
export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True  # set in the pod environment

python3 -m <slime entry module> --async \
  --actor-num-nodes 32 \
  --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 4 \
  --sequence-parallel \
  --pipeline-model-parallel-size 8 \
  --decoder-first-pipeline-num-layers 14 \
  --decoder-last-pipeline-num-layers 16 \
  --context-parallel-size 2 \
  --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 \
  --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 \
  --moe-token-dispatcher-type alltoall \
  --allgather-cp
recipes/glm-5.2-744b-a40b/h100/attune-tp4-pp8-cp2.yamlDownloadGitHub
# recipes/glm-5.2-744b-a40b/h100/attune-tp4-pp8-cp2.yaml
# An OpenRecipes recipe; the fields are documented in schema/recipe.schema.yaml.
schema_version: 1
model: GLM-5.2-744B-A40B
gpu:
  type: H100
  memory_gb: 80
  count: 256
  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 with local patches; commit not recorded
  megatron_lm: commit 1dcf0daf (megatron-core 0.16.0rc0) with local patches
  transformer_engine: 2.10.0
  pytorch: 2.11.0 (CUDA 12.9 build)
  cuda: '12.9'
  nccl: 2.28.9 (PyTorch's NCCL wheel; printed by the measured run)
  deepep: 1.2.1
  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), 2026-10-11: Megatron-LM git commit
    and status, installed megatron-core, TE, PyTorch, NCCL, DeepEP and slime, Python'
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: bb302c134974b96f3905e9c87bfc4c0bb2be674a50e6dad09121aed2a0eb3455
    build:
      name: glm52-tulu3-fixed8192-v1
      tokenizer: GLM-5.2 (tokenizer.json sha256 19e773648cb4e65de8660ea6365e10acca112d42a854923df93db4a6f333a82d)
      loss_mask: glm5
      rows: 12800
      row_tokens: 8192
      padding_tokens: 0
      seed: 20260828
      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: 0.661
      parquet_sha256: 5d951089abd622511a40e0b3a1e5e36ad56859459e4d8c3aff14538cadbd0743
      parquet_bytes: 169929706
      builder: data/generate_fixed_sft.py with patch_glm5_mask.py (data/README.md)
layout:
  label: TP4 PP8 CP2 EP32 DP4
  summary: TP 4, PP 8 (first stage 14 layers, last 16), CP 2, EP 32, ETP 1, DP 4, sequence parallel; max 8,192 tokens
    per GPU; full recompute (uniform, 1 layer); CPU optimizer offload with overlapped copies and the precision-aware optimizer;
    alltoall dispatcher; --allgather-cp
  highlights: max 8,192 tokens/GPU, CPU optimizer offload
  tp: 4
  pp: 8
  vpp: 1
  cp: 2
  ep: 32
  etp: 1
  dp: 4
  sequence_parallel: true
  layer_split:
    first: 14
    last: 16
  max_tokens_per_gpu: 8192
  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: 66.24
  usd_per_million_tokens: 28.85
  step_time_cv_percent: 2.59
  mfu: 0.0162
  mfu_note: null
  active_params:
    count: 40297758720
    source: https://huggingface.co/zai-org/GLM-5.2
    revision: cf457fa734ab149ffef225f80893eb38c6ff5cdc
  measured: 2026-10-13
  steps_per_run: 20
  window:
  - 5
  - 19
  window_tokens: 15728640
  runs: 1
  profiler: false
  gpu_hours: 151.32
  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: 0.999974
    final_loss_difference: 0.00044
    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'
  INDEXER_ROPE_NEOX_STYLE: '0'
  NCCL_NVLS_ENABLE: $([ "$NVLINK_COUNT" -gt 0 ] && echo 1 || echo 0)
  PYTORCH_CUDA_ALLOC_CONF: expandable_segments:True
patches: []
run: |
  export CUDA_DEVICE_MAX_CONNECTIONS=1
  export INDEXER_ROPE_NEOX_STYLE=0
  export NCCL_NVLS_ENABLE=$([ "$NVLINK_COUNT" -gt 0 ] && echo 1 || echo 0)
  export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True  # set in the pod environment

  python3 -m <slime entry module> --async \
    --actor-num-nodes 32 \
    --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 4 \
    --sequence-parallel \
    --pipeline-model-parallel-size 8 \
    --decoder-first-pipeline-num-layers 14 \
    --decoder-last-pipeline-num-layers 16 \
    --context-parallel-size 2 \
    --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 \
    --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 \
    --moe-token-dispatcher-type alltoall \
    --allgather-cp

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}
}

GLM-5.2-744B-A40B on other GPUs

H100 · 256 GPUs · this page
68.3
tokens/s/GPU · $28.00 / 1M tokens
H200 · 128 GPUs
161.8
tokens/s/GPU · $13.58 / 1M tokens
B200 · 64 GPUs
146.8
tokens/s/GPU · $26.95 / 1M tokens