| { | |
| "auto_loss_scaling": false, | |
| "device_iterations": 1, | |
| "embedding_serialization_factor": 2, | |
| "enable_half_partials": true, | |
| "executable_cache_dir": "./exe_cache", | |
| "gradient_accumulation_steps": 2048, | |
| "inference_device_iterations": 5, | |
| "inference_replication_factor": 1, | |
| "ipus_per_replica": 4, | |
| "layers_per_ipu": [ | |
| 1, | |
| 3, | |
| 4, | |
| 4 | |
| ], | |
| "matmul_proportion": [0.6, 0.2, 0.1, 0.1], | |
| "optimizer_state_offchip": true, | |
| "optimum_version": "1.4.1", | |
| "output_mode": "final", | |
| "recompute_checkpoint_every_layer": true, | |
| "replicated_tensor_sharding": true, | |
| "replication_factor": 1, | |
| "seed": 42 | |
| } | |