1{
 2  "cuda_error": "RuntimeError('cuDNN error: CUDNN_STATUS_INTERNAL_ERROR_HOST_ALLOCATION_FAILED')",
 3  "fallback": {
 4    "device": "cpu",
 5    "steps": 32,
 6    "batch": 32,
 7    "native_autograd_saved_bytes": {
 8      "fold": 86016,
 9      "gru": 692224
10    },
11    "native_autograd_saved_tensor_count": {
12      "fold": 418,
13      "gru": 291
14    },
15    "reverse_reconstruction_relative_error": 2.248847158625722e-05,
16    "note": "Fold measurement is native autograd; a custom reversible backward would be needed for constant-memory training."
17  }
18}