Training GLM-4.7-Flash on H200

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 (THD, max 16,384 tokens/GPU); full recompute (uniform, 1 layer); distributed optimizer with the precision-aware optimizer; alltoall dispatcher; no offload

Tokens/s per GPU

6,178.46

$ / 1M tokens

$0.36

MFU

13.4% (N = 3,579,314,176 active parameters per token)

Step-time CV

0.95%

Convergence check

correlation 0.9779, final loss difference 0.0032, 1 pointwise failures over 20 steps

Component test

precision-aware Adam on the GPU and CPU optimizer offload (fraction 1.0, overlapped copies), each against an FP64 reference: H200, 2026-10-14. 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 (dense update) relative L2 error 5.12e-7 on the GPU and 7.98e-7 offloaded against a bound of 1.0e-3 after one step, and 4.99e-5 and 7.72e-5 against 1.1e-3 after three steps at the recipe's hyperparameters (bound 2 x the default GPU optimizer's error + 1e-3); reconstructed FP32 masters bitwise identical to an arm with explicit FP32 masters, the TE remainder self-test passes and the three negative controls fail run on this recipe's training image; the optimizer step is elementwise, so the result holds for every model on the image

Measured

2026-10-15

GPU-hours

1.34

Rows

glm47flash-tulu3-fixed8192-v1 (rows_sha256 79bfe4b7e20e66ec85c12caf4995ea9574bbf6d316772e4d69193a01efe01914)

export CUDA_DEVICE_MAX_CONNECTIONS=1
export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
export TILELANG_CACHE_DIR="$KERNEL_CACHE/tilelang"
export TRITON_CACHE_DIR="$KERNEL_CACHE/triton"
export TORCHINDUCTOR_CACHE_DIR="$KERNEL_CACHE/inductor"
export FI_PROVIDER=efa
export FI_EFA_USE_DEVICE_RDMA=1

MODEL_ARGS_LINE="$(python3 miles/utils/external_utils/model_args_utils.py glm4.7-flash)"
read -ra MODEL_ARGS <<< "$MODEL_ARGS_LINE"
python3 train.py \
  --actor-num-nodes 1 \
  --actor-num-gpus-per-node 8 \
  --num-gpus-per-node 8 \
  "${MODEL_ARGS[@]}" \
  --hf-checkpoint "$HF_CKPT" \
  --ref-load "$TORCH_DIST" \
  --loss-type sft_loss \
  --calculate-per-token-loss \
  --disable-compute-advantages-and-returns \
  --debug-train-only \
  --rollout-function-path pretokenized_sft_rollout.generate_rollout \
  --prompt-data "$SFT_DATA" \
  --input-key messages \
  --metadata-key metadata \
  --num-rollout 20 \
  --rollout-batch-size 128 \
  --global-batch-size 128 \
  --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 \
  --use-precision-aware-optimizer \
  --attention-backend flash \
  --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 \
  --attention-dropout 0.0 --hidden-dropout 0.0 \
  --accumulate-allreduce-grads-in-fp32 --attention-softmax-in-fp32 \
  --distributed-timeout-minutes 30

# recipes/glm-4.7-flash/h200/attune-tp1-pp1-ep8.yaml
# An OpenRecipes recipe; the fields are documented in schema/recipe.schema.yaml.
schema_version: 1
model: GLM-4.7-Flash
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: miles 0.1.0 (radixark/miles commit 78527b91, a slime fork), no local changes
  megatron_lm: radixark/Megatron-LM commit 37f91617 (megatron-core 0.16.0rc0), no local changes
  transformer_engine: 2.17.0
  pytorch: 2.11.0 (CUDA 13.0 build)
  cuda: '13.0'
  nccl: 2.28.9 (CUDA 13.0 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 (aws-ofi-nccl 1.20), 3.2 Tb/s RDMA per node
  source: 'version dump of the measured run''s training image (same digest), 2026-10-13: installed TE, PyTorch, DeepEP,
    megatron-core and miles, Megatron-LM and miles git commits (clean trees), 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: 79bfe4b7e20e66ec85c12caf4995ea9574bbf6d316772e4d69193a01efe01914
    build:
      name: glm47flash-tulu3-fixed8192-v1
      tokenizer: GLM-4.7-Flash (revision 7dd20894a642a0aa287e9827cb1a1f7f91386b67; tokenizer.json sha256
        19e773648cb4e65de8660ea6365e10acca112d42a854923df93db4a6f333a82d; chat_template.jinja sha256
        d63ad536c3c81880043e22ec7fd08db42b4d8fb7c89c7138bc562bfa25281375)
      loss_mask: glm5
      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: 0.674
      parquet_sha256: 33e50f503dee9999e27c16a9ae1781c822ad1acc012b4b52c2efccfd9c8bbcb4
      parquet_bytes: 83603534
      builder: data/generate_fixed_sft.py with patch_glm5_mask.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 (THD, max 16,384 tokens/GPU); full
    recompute (uniform, 1 layer); distributed optimizer with the precision-aware optimizer; alltoall dispatcher; no
    offload
  highlights: max 16,384 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: 16384
  rows_per_microbatch: 2
  recompute:
    granularity: full
    method: uniform
    num_layers: 1
  offload: null
  dispatcher: alltoall
result:
  tokens_per_s_per_gpu: 6178.46
  usd_per_million_tokens: 0.36
  step_time_cv_percent: 0.95
  mfu: 0.1342
  mfu_note: null
  active_params:
    count: 3579314176
    source: https://huggingface.co/zai-org/GLM-4.7-Flash
    revision: 7dd20894a642a0aa287e9827cb1a1f7f91386b67
  measured: 2026-10-15
  steps_per_run: 20
  window:
  - 5
  - 19
  window_tokens: 15728640
  runs: 1
  profiler: false
  gpu_hours: 1.34
  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: 'miles'' layout for GLM-4.7-Flash (`scripts/run_glm47_flash.py`): TP 4 with sequence parallel, EP 8, full
    recompute (uniform, 1 layer), max 32,768 tokens/GPU and Adam on the CPU (`--optimizer-cpu-offload
    --overlap-cpu-optimizer-d2h-h2d --use-precision-aware-optimizer`), here on 8 GPUs (DP 2). Model arguments, rows and
    schedule are the same in both recipes.'
  configuration: pass
  component_tests:
  - test: precision-aware Adam on the GPU and CPU optimizer offload (fraction 1.0, overlapped copies), each 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: 2026-10-14
    result: pass
    detail: worst tensor (dense update) relative L2 error 5.12e-7 on the GPU and 7.98e-7 offloaded against a bound of
      1.0e-3 after one step, and 4.99e-5 and 7.72e-5 against 1.1e-3 after three steps at the recipe's hyperparameters
      (bound 2 x the default GPU optimizer's error + 1e-3); reconstructed FP32 masters bitwise identical to an arm with
      explicit FP32 masters, the TE remainder self-test passes and the three negative controls fail
    note: run on this recipe's training image; the optimizer step is elementwise, so the result holds for every model on
      the image
  convergence:
    steps: 20
    result: pass
    correlation: 0.9779
    final_loss_difference: 0.0032
    pointwise_failures: 1
    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 (amazon-science/attune f2cb450);
    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
  TILELANG_CACHE_DIR: $KERNEL_CACHE/tilelang
  TRITON_CACHE_DIR: $KERNEL_CACHE/triton
  TORCHINDUCTOR_CACHE_DIR: $KERNEL_CACHE/inductor
  FI_PROVIDER: efa
  FI_EFA_USE_DEVICE_RDMA: '1'
patches: []
run: |
  export CUDA_DEVICE_MAX_CONNECTIONS=1
  export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
  export TILELANG_CACHE_DIR="$KERNEL_CACHE/tilelang"
  export TRITON_CACHE_DIR="$KERNEL_CACHE/triton"
  export TORCHINDUCTOR_CACHE_DIR="$KERNEL_CACHE/inductor"
  export FI_PROVIDER=efa
  export FI_EFA_USE_DEVICE_RDMA=1

  MODEL_ARGS_LINE="$(python3 miles/utils/external_utils/model_args_utils.py glm4.7-flash)"
  read -ra MODEL_ARGS <<< "$MODEL_ARGS_LINE"
  python3 train.py \
    --actor-num-nodes 1 \
    --actor-num-gpus-per-node 8 \
    --num-gpus-per-node 8 \
    "${MODEL_ARGS[@]}" \
    --hf-checkpoint "$HF_CKPT" \
    --ref-load "$TORCH_DIST" \
    --loss-type sft_loss \
    --calculate-per-token-loss \
    --disable-compute-advantages-and-returns \
    --debug-train-only \
    --rollout-function-path pretokenized_sft_rollout.generate_rollout \
    --prompt-data "$SFT_DATA" \
    --input-key messages \
    --metadata-key metadata \
    --num-rollout 20 \
    --rollout-batch-size 128 \
    --global-batch-size 128 \
    --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 \
    --use-precision-aware-optimizer \
    --attention-backend flash \
    --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 \
    --attention-dropout 0.0 --hidden-dropout 0.0 \
    --accumulate-allreduce-grads-in-fp32 --attention-softmax-in-fp32 \
    --distributed-timeout-minutes 30
notes: |
  - Starting point: miles' GLM-4.7-Flash launcher, `scripts/run_glm47_flash.py` at miles 0.1.0 (model page
    https://miles.radixark.com/docs/models/glm/glm4-7-flash). miles publishes it as a GRPO recipe with no tokens/s figure.
    Ported to this workload with its layout (TP 4 with sequence parallel, EP 8, full recompute, max 32,768 tokens/GPU,
    Adam on the CPU) on 8 H200, it measured 3,142.9 tokens/s/GPU; ATTUNE tuned it to 6,178.5 (+97%).
  - The loss starts 1.0% above the reference recipe's (1.1580 against 1.1465 at step 0; Hugging Face transformers gives
    1.1536 on the same 128 rows). In steps 0-3, where both recipes' gradient norms run at 16-276 before clipping to 1.0,
    it rises to 1.34 at step 3 against the reference's 0.97; from step 6 it stays within 1.2% of the reference and from
    step 12 within 0.5%.