ipu_config.json

{
  "embedding_serialization_factor": 2,                                                                                         
  "recompute_checkpoint_every_layer": true,                                                                                    
  "optimizer_state_offchip": true,
  "replicated_tensor_sharding": true,
  "enable_half_first_order_momentum": true,                                                                                    
  "enable_half_partials": true,                                                                                                

  "device_iterations": 1,                                                                                                      
  "inference_device_iterations": 5,
  "replication_factor": {"pod4": 1, "pod8": 2, "pod16": 4, "pod32": 8, "pod64": 16, "default": 1},
  "inference_replication_factor": {"pod4": 1, "pod8": 2, "pod16": 4, "pod32": 8, "pod64": 16, "default": 1},
  "gradient_accumulation_steps": 512,
  "executable_cache_dir": "./exe_cache",                                                                                       

  "ipus_per_replica": 4,                                                                                                       
  "layers_per_ipu": [0, 4, 4, 4],                                                                                              
  "matmul_proportion": [0.25, 0.20, 0.20, 0.25]
}