run_path: /mnt/ssd-2/lucia/paper_runs/experiments/plan_muon_eps1e17_4k_bs256
steps:
- magic:
    run_path: /mnt/ssd-2/lucia/paper_runs/experiments/plan_muon_eps1e17_4k_bs256
    overwrite: false
    resume: true
    model: gpt2
    model_kwargs: resid_pdrop=0.0,attn_pdrop=0.0,embd_pdrop=0.0
    precision: fp32
    use_tf32_matmuls: false
    seed: 42
    cleanup_ckpts: false
    distributed:
      nnode: 1
      nproc_per_node: 2
    data:
      dataset: /mnt/ssd-1/lucia/bergson-damping/runs/ekfac_vs_n/datasets/train_4k.hf
      split: train
      chunk_length: 0
    query:
      dataset: /mnt/ssd-1/lucia/bergson-damping/runs/ekfac_vs_n/datasets/query_20.hf
      split: train
      chunk_length: 0
    batch_size: 256
    grad_accum_steps: 8
    num_epochs: 2
    lr_schedule:
      lr: 0.0004
      lr_scheduler_type: polynomial
      lr_start: 1.0e-06
      lr_end: 4.0e-05
      warmup_steps: 0.25
    optimizer: muon
    adam_beta1: 0.95
    adam_beta2: 0.975
    adam_eps: 1.0e-08
    eps_root: 1.0e-17
    weight_decay: 0.01
    loss_reduction: mean
    train_mode: false
    num_subsets: 100
    subset_fraction: 0.01
    query_method: none
    save_models: true
    save_optimizer_state: none
    save_mode: log
    skip_validation: false