ipu_config.json

{
"embedding_serialization_factor": 0,
  "recompute_checkpoint_every_layer": true,
  "optimizer_state_offchip": false,
  "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
}