OpenRecipes
‹ All models
Qwen

Training Qwen3.5-122B-A10B on B200

Ranked by tokens/s per GPU · 8,192-token SFT rows · global batch 128 · BF16
#Tokens/s per GPU$ / 1M tokensGPUsLayoutMeasuredAuthor
#12,475.8$1.6016TP1 PP4 EP4 DP4
max 8,192 tokens/GPU		2026-10-18ATTUNERecipe ›
#22,447.9$1.6216TP1 PP8 EP2 DP2
max 8,192 tokens/GPU		2026-10-18ATTUNERecipe ›

Recipes

#1attune-tp1-pp4-ep42,475.8 tokens/s/GPUTP1 PP4 EP4 DP4max 8,192 tokens/GPU

B200 runs of this model need patches/fla_blackwell_guard.py: with the image's Triton 3.6, FLA 0.4.1's autotuner picks gated-delta-net kernel configs that race on Blackwell and give a different, wrong result on every call, and the patch keeps those kernels on the configs FLA main uses there. Every training process on every node imports it before FLA (patches/README.md) and prints a [fla-guard] line.

Author
ATTUNE (cite)
GPUs
16 x B200 180 GB (AWS p6-b200.48xlarge, 8 per node)
Layout
TP 1, PP 4, CP 1, EP 4, ETP 1, DP 4; one 8,192-token row per microbatch; full recompute (uniform, 1 layer); alltoall dispatcher; TE cross-entropy fusion; precision-aware distributed optimizer; no offload
Tokens/s per GPU
2,475.79
$ / 1M tokens
$1.60
MFU
5.9% (N = 9,008,136,192 active parameters per token)
Step-time CV
4.47%
Convergence check
correlation 0.9999993, final loss difference 8.2e-05, 0 pointwise failures over 20 steps
Measured
2026-10-18
GPU-hours
7.03
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 1dcf0da (megatron-core 0.16.0rc0), with slime's patch to the MoE router (routing-replay registration)
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
not pinned
python
not pinned
fla
0.4.1 (flash-linear-attention)
triton
3.6.0
network
8 GPUs per node, EFA, 3.2 Tb/s RDMA per node
runtime fixes
patches/fla_blackwell_guard.py, imported in every training process before FLA

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)

Patches

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 2 \
  --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 4 \
  --context-parallel-size 1 \
  --expert-model-parallel-size 4 \
  --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 \
  --attention-dropout 0.0 --hidden-dropout 0.0 \
  --accumulate-allreduce-grads-in-fp32 --attention-softmax-in-fp32 \
  --attention-backend flash \
  --cross-entropy-loss-fusion --cross-entropy-fusion-impl te \
  --distributed-timeout-minutes 30
recipes/qwen3.5-122b-a10b/b200/attune-tp1-pp4-ep4.yamlDownloadGitHub
# recipes/qwen3.5-122b-a10b/b200/attune-tp1-pp4-ep4.yaml
# An OpenRecipes recipe; the fields are documented in schema/recipe.schema.yaml.
schema_version: 1
model: Qwen3.5-122B-A10B
gpu:
  type: B200
  memory_gb: 180
  count: 16
  per_node: 8
  instance: AWS p6-b200.48xlarge
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 1dcf0da (megatron-core 0.16.0rc0), with slime's patch to the MoE router (routing-replay registration)
  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: not pinned
  python: not pinned
  fla: 0.4.1 (flash-linear-attention)
  triton: 3.6.0
  network: 8 GPUs per node, EFA, 3.2 Tb/s RDMA per node
  runtime_fixes: patches/fla_blackwell_guard.py, imported in every training process before FLA
  source: 'in-pod check of the measured run''s training image (same digest) on 2026-10-01: Megatron-LM git commit and
    installed TE and PyTorch; NCCL from the measured run''s log; FLA and Triton from an in-pod check of the same digest
    on 2026-10-18'
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 PP4 EP4 DP4
  summary: TP 1, PP 4, CP 1, EP 4, ETP 1, DP 4; one 8,192-token row per microbatch; full recompute (uniform, 1 layer);
    alltoall dispatcher; TE cross-entropy fusion; precision-aware distributed optimizer; no offload
  highlights: max 8,192 tokens/GPU
  tp: 1
  pp: 4
  vpp: 1
  cp: 1
  ep: 4
  etp: 1
  dp: 4
  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: 2475.79
  usd_per_million_tokens: 1.60
  step_time_cv_percent: 4.47
  mfu: 0.0595
  mfu_note: null
  active_params:
    count: 9008136192
    source: https://huggingface.co/Qwen/Qwen3.5-122B-A10B
    revision: dc4d348443bc740c68e2d77492492c11606384d5
  measured: 2026-10-18
  steps_per_run: 20
  window:
  - 5
  - 19
  window_tokens: 15728640
  runs: 1
  profiler: false
  gpu_hours: 7.03
  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: 0.9999993
    final_loss_difference: 8.2e-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; both with patches/fla_blackwell_guard.py'
  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:
- patches/fla_blackwell_guard.py
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 2 \
    --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 4 \
    --context-parallel-size 1 \
    --expert-model-parallel-size 4 \
    --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 \
    --attention-dropout 0.0 --hidden-dropout 0.0 \
    --accumulate-allreduce-grads-in-fp32 --attention-softmax-in-fp32 \
    --attention-backend flash \
    --cross-entropy-loss-fusion --cross-entropy-fusion-impl te \
    --distributed-timeout-minutes 30
notes: |
  B200 runs of this model need `patches/fla_blackwell_guard.py`: with the image's Triton 3.6, FLA 0.4.1's autotuner
  picks gated-delta-net kernel configs that race on Blackwell and give a different, wrong result on every call, and
  the patch keeps those kernels on the configs FLA main uses there. Every training process on every node imports it
  before FLA (patches/README.md) and prints a `[fla-guard]` line.
#2attune-tp1-pp8-ep22,447.9 tokens/s/GPUTP1 PP8 EP2 DP2max 8,192 tokens/GPU

B200 runs of this model need patches/fla_blackwell_guard.py: with the image's Triton 3.6, FLA 0.4.1's autotuner picks gated-delta-net kernel configs that race on Blackwell and give a different, wrong result on every call, and the patch keeps those kernels on the configs FLA main uses there. Every training process on every node imports it before FLA (patches/README.md) and prints a [fla-guard] line.

Author
ATTUNE (cite)
GPUs
16 x B200 180 GB (AWS p6-b200.48xlarge, 8 per node)
Layout
TP 1, PP 8, CP 1, EP 2, ETP 1, DP 2; 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
2,447.95
$ / 1M tokens
$1.62
MFU
5.9% (N = 9,008,136,192 active parameters per token)
Step-time CV
3.16%
Convergence check
correlation 0.999999, final loss difference 0.00017, 0 pointwise failures over 20 steps
Measured
2026-10-18
GPU-hours
4.2
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 1dcf0da (megatron-core 0.16.0rc0), with slime's patch to the MoE router (routing-replay registration)
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
not pinned
python
not pinned
fla
0.4.1 (flash-linear-attention)
triton
3.6.0
network
8 GPUs per node, EFA, 3.2 Tb/s RDMA per node
runtime fixes
patches/fla_blackwell_guard.py, imported in every training process before FLA

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)

Patches

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 2 \
  --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 8 \
  --context-parallel-size 1 \
  --expert-model-parallel-size 2 \
  --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 \
  --attention-dropout 0.0 --hidden-dropout 0.0 \
  --accumulate-allreduce-grads-in-fp32 --attention-softmax-in-fp32 \
  --attention-backend flash \
  --cross-entropy-loss-fusion --cross-entropy-fusion-impl te \
  --distributed-timeout-minutes 30
recipes/qwen3.5-122b-a10b/b200/attune-tp1-pp8-ep2.yamlDownloadGitHub
# recipes/qwen3.5-122b-a10b/b200/attune-tp1-pp8-ep2.yaml
# An OpenRecipes recipe; the fields are documented in schema/recipe.schema.yaml.
schema_version: 1
model: Qwen3.5-122B-A10B
gpu:
  type: B200
  memory_gb: 180
  count: 16
  per_node: 8
  instance: AWS p6-b200.48xlarge
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 1dcf0da (megatron-core 0.16.0rc0), with slime's patch to the MoE router (routing-replay registration)
  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: not pinned
  python: not pinned
  fla: 0.4.1 (flash-linear-attention)
  triton: 3.6.0
  network: 8 GPUs per node, EFA, 3.2 Tb/s RDMA per node
  runtime_fixes: patches/fla_blackwell_guard.py, imported in every training process before FLA
  source: 'in-pod check of the measured run''s training image (same digest) on 2026-10-01: Megatron-LM git commit and
    installed TE and PyTorch; NCCL from the measured run''s log; FLA and Triton from an in-pod check of the same digest
    on 2026-10-18'
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 PP8 EP2 DP2
  summary: TP 1, PP 8, CP 1, EP 2, ETP 1, DP 2; 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: 1
  pp: 8
  vpp: 1
  cp: 1
  ep: 2
  etp: 1
  dp: 2
  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: 2447.95
  usd_per_million_tokens: 1.62
  step_time_cv_percent: 3.16
  mfu: 0.0588
  mfu_note: null
  active_params:
    count: 9008136192
    source: https://huggingface.co/Qwen/Qwen3.5-122B-A10B
    revision: dc4d348443bc740c68e2d77492492c11606384d5
  measured: 2026-10-18
  steps_per_run: 20
  window:
  - 5
  - 19
  window_tokens: 15728640
  runs: 1
  profiler: false
  gpu_hours: 4.20
  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: 0.999999
    final_loss_difference: 0.00017
    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; both with patches/fla_blackwell_guard.py'
  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:
- patches/fla_blackwell_guard.py
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 2 \
    --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 8 \
    --context-parallel-size 1 \
    --expert-model-parallel-size 2 \
    --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 \
    --attention-dropout 0.0 --hidden-dropout 0.0 \
    --accumulate-allreduce-grads-in-fp32 --attention-softmax-in-fp32 \
    --attention-backend flash \
    --cross-entropy-loss-fusion --cross-entropy-fusion-impl te \
    --distributed-timeout-minutes 30
notes: |
  B200 runs of this model need `patches/fla_blackwell_guard.py`: with the image's Triton 3.6, FLA 0.4.1's autotuner
  picks gated-delta-net kernel configs that race on Blackwell and give a different, wrong result on every call, and
  the patch keeps those kernels on the configs FLA main uses there. Every training process on every node imports it
  before FLA (patches/README.md) and prints a `[fla-guard]` line.

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
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 · this page
2,475.8
tokens/s/GPU · $1.60 / 1M tokens