{
  "embedding_serialization_factor": 2,
  "recompute_checkpoint_every_layer": true,
  "optimizer_state_offchip": true,
  "replicated_tensor_sharding": true,
  "enable_half_first_order_momentum": true,
  "enable_half_partials": true,

  "replication_factor": 4,
  "gradient_accumulation_steps": 16,
  "device_iterations": 1,
  "inference_device_iterations": 4,
  "inference_replication_factor": 4,
  "executable_cache_dir": "./exe_cache",

  "ipus_per_replica": 4,
  "layers_per_ipu": [3, 7, 7, 7],
  "matmul_proportion": [0.15, 0.18, 0.20, 0.25]
}