ipu_config.json

{
  "auto_loss_scaling": false,
  "device_iterations": 1,
  "embedding_serialization_factor": 1,
  "enable_half_partials": true,
  "executable_cache_dir": "",
  "execute_encoder_on_cpu_for_generation": false,
  "gradient_accumulation_steps": 1,
  "inference_device_iterations": 1,
  "inference_replication_factor": 1,
  "ipus_per_replica": 16,
  "layers_per_ipu": [
    0,
    4,
    4,
    4,
    4,
    4,
    4,
    4,
    4,
    4,
    4,
    4,
    4,
    4,
    4,
    4,
    4
  ],
  "matmul_proportion": 0.2,
  "optimizer_state_offchip": true,
  "optimum_version": "1.6.1",
  "output_mode": "final",
  "recompute_checkpoint_every_layer": false,
  "replicated_tensor_sharding": false,
  "replication_factor": 1,
  "seed": null,
  "transformers_version": "4.25.1"
}