ipu_config.json

{
  "device_iterations": 1,
  "embedding_serialization_factor": 2,
  "enable_half_partials": true,
  "executable_cache_dir": "./exe_cache",
  "gradient_accumulation_steps": 512,
  "inference_device_iterations": 4,
  "inference_replication_factor": 16,
  "ipus_per_replica": 4,
  "layers_per_ipu": [
    3,
    7,
    7,
    7
  ],
  "matmul_proportion": [
    0.1,
    0.15,
    0.15,
    0.15
  ],
  "optimizer_state_offchip": true,
  "optimum_version": "1.0.0",
  "output_mode": "final",
  "recompute_checkpoint_every_layer": true,
  "replicated_tensor_sharding": true,
  "replication_factor": 16,
  "seed": 42
}