training_args.json

{
  "profile": "gpt1_modern_muon",
  "batch_size": 16,
  "grad_accum_steps": 1,
  "max_steps": 20000,
  "warmup_steps": 2000,
  "max_lr": null,
  "min_lr": null,
  "weight_decay": 0.01,
  "optimizer": "muon",
  "save_every": 2000,
  "log_every": 10,
  "eval_every": 500,
  "eval_batches": 50,
  "checkpoint_dir": "checkpoints",
  "resume_from": null,
  "num_workers": 2,
  "compile": true,
  "activation_checkpointing": false,
  "chunked_loss": false,
  "no_trackio": false
}