ipu_config.json

{
  "auto_loss_scaling": false,
  "device_iterations": 1,
  "embedding_serialization_factor": 1,
  "enable_half_partials": true,
  "executable_cache_dir": "/tmp/exe_cache/3.3.0/whisper",
  "execute_encoder_on_cpu_for_generation": false,
  "explicit_ir_inference": false,
  "gradient_accumulation_steps": 16,
  "inference_device_iterations": 1,
  "inference_embedding_serialization_factor": 1,
  "inference_ipus_per_replica": 2,
  "inference_layers_per_ipu": [
    4,
    4
  ],
  "inference_matmul_proportion": 0.2,
  "inference_parallelize_kwargs": {
    "on_device_generation_steps": 16,
    "use_cache": true,
    "use_encoder_output_buffer": true
  },
  "inference_projection_serialization_factor": 1,
  "inference_replication_factor": 1,
  "inference_serialized_embedding_splits_per_ipu": null,
  "inference_serialized_projection_splits_per_ipu": null,
  "ipus_per_replica": 4,
  "layers_per_ipu": [
    2,
    2,
    2,
    2
  ],
  "matmul_proportion": [
    0.2,
    0.2,
    0.6,
    0.6
  ],
  "optimizer_state_offchip": true,
  "optimum_version": "1.6.1",
  "output_mode": "final",
  "parallelize_kwargs": {},
  "projection_serialization_factor": 5,
  "recompute_checkpoint_every_layer": true,
  "replicated_tensor_sharding": false,
  "replication_factor": 1,
  "seed": 42,
  "serialized_embedding_splits_per_ipu": null,
  "serialized_projection_splits_per_ipu": null,
  "transformers_version": "4.29.2"
}