{
  "_comment": "Small-context smoke config for end-to-end verification. The unfused attention kernel materializes [B,H,T,T] probabilities, so keep T modest until the fused flash kernel lands.",
  "model": {
    "name": "gpt-smoke",
    "n_layers": 4,
    "d_model": 256,
    "n_heads": 4,
    "n_kv_heads": 2,
    "d_ff": 704,
    "vocab_size": 4096,
    "context_length": 128,
    "tied_embeddings": true,
    "use_rope": true,
    "rope_theta": 10000.0,
    "norm": "rmsnorm",
    "norm_eps": 1e-6,
    "activation": "swiglu",
    "dropout": 0.0
  },
  "train": {
    "lr": 1e-3,
    "min_lr_ratio": 0.1,
    "warmup_steps": 100,
    "max_steps": 2000,
    "beta1": 0.9,
    "beta2": 0.95,
    "eps": 1e-8,
    "weight_decay": 0.1,
    "grad_clip": 1.0,
    "batch_size": 32,
    "grad_accum_steps": 1,
    "precision": "f32",
    "checkpoint_every": 500,
    "eval_every": 250,
    "eval_batches": 10,
    "seed": 1337
  }
}