{ "optimizer_state_offchip": false, "replicated_tensor_sharding": true, "enable_half_partials": true, "recompute_checkpoint_every_layer": false, "device_iterations": 1, "inference_device_iterations": 4, "replication_factor": 1, "inference_replication_factor": 1, "gradient_accumulation_steps": 32, "executable_cache_dir": "./exe_cache", "ipus_per_replica": 4, "layers_per_ipu": [6, 5, 5, 7], "matmul_proportion": [0.2, 1.0, 1.0, 1.0] }