{
    "embedding_serialization_factor": 0,
    "recompute_checkpoint_every_layer": true,
    "optimizer_state_offchip": true,
    "replicated_tensor_sharding": true,
    "enable_half_first_order_momentum": true,
    "enable_half_partials": true,
  
    "device_iterations": 2,
    "inference_device_iterations": 4,
    "replication_factor": 4,
    "inference_replication_factor": 4,
    "gradient_accumulation_steps": 16,
    "executable_cache_dir": "./exe_cache",
  
    "ipus_per_replica": 4,
    "layers_per_ipu": [0, 4, 4, 4],
    "matmul_proportion": [0.25]
}