{ "embedding_serialization_factor": 1, "recompute_checkpoint_every_layer": true, "optimizer_state_offchip": false, "replicated_tensor_sharding": true, "enable_half_partials": true, "device_iterations": 1, "inference_device_iterations": 5, "replication_factor": 2, "inference_replication_factor": 2, "gradient_accumulation_steps": 16, "executable_cache_dir": "./exe_cache", "ipus_per_replica": 2, "layers_per_ipu": [2, 4], "matmul_proportion": 0.25 }