ipu_config.json

{
  "decompose_grad_sum": false,
  "device_iterations": 1,
  "embedding_serialization_factor": 2,
  "enable_half_first_order_momentum": true,
  "enable_half_partials": true,
  "executable_cache_dir": "./exe_cache",
  "execute_encoder_on_cpu_for_generation": false,
  "gradient_accumulation_steps": 16,
  "inference_device_iterations": 5,
  "inference_replication_factor": 4,
  "ipus_per_replica": 4,
  "layers_per_ipu": [
    0,
    4,
    4,
    4
  ],
  "matmul_proportion": [
    0.25,
    0.25,
    0.25,
    0.25
  ],
  "optimizer_state_offchip": true,
  "optimum_version": "1.3.0",
  "output_mode": "final",
  "profile_dir": "",
  "recompute_checkpoint_every_layer": true,
  "replicated_tensor_sharding": true,
  "replication_factor": 4,
  "seed": 42,
  "sharded_execution_for_inference": false,
  "transformers_version": "4.20.0",
  "use_popdist": false
}