{
  "cuda_error": "RuntimeError('cuDNN error: CUDNN_STATUS_INTERNAL_ERROR_HOST_ALLOCATION_FAILED')",
  "fallback": {
    "device": "cpu",
    "steps": 32,
    "batch": 32,
    "native_autograd_saved_bytes": {
      "fold": 86016,
      "gru": 692224
    },
    "native_autograd_saved_tensor_count": {
      "fold": 418,
      "gru": 291
    },
    "reverse_reconstruction_relative_error": 2.248847158625722e-05,
    "note": "Fold measurement is native autograd; a custom reversible backward would be needed for constant-memory training."
  }
}