| { | |
| "strategy": "ekfac", | |
| "use_empirical_fisher": false, | |
| "immediate_gradient_removal": false, | |
| "ignore_bias": false, | |
| "distributed_sync_steps": 1000, | |
| "covariance_max_examples": 10000000, | |
| "covariance_data_partition_size": 1, | |
| "covariance_module_partition_size": 1, | |
| "activation_covariance_dtype": "torch.float32", | |
| "gradient_covariance_dtype": "torch.float32", | |
| "eigendecomposition_dtype": "torch.float64", | |
| "lambda_max_examples": 10000000, | |
| "lambda_data_partition_size": 1, | |
| "lambda_module_partition_size": 1, | |
| "lambda_iterative_aggregate": false, | |
| "cached_activation_cpu_offload": false, | |
| "lambda_dtype": "torch.float32" | |
| } |