{
  "ipus_per_replica": 4,
  "embedding_serialization_factor": 1,
  "recompute_checkpoint_every_layer": true,
  "optimizer_state_offchip": true,
  "enable_half_partials": true,
  "output_mode": "all",
  "replication_factor": 1,
  "gradient_accumulation_steps": 1,
  "device_iterations": 1,
  "executable_cache_dir": "disabled",
  "replicated_tensor_sharding": true
}