| { | |
| "layers_per_ipu": [0, 12, 4, 0], | |
| "inference_layers_per_ipu": [0, 8, 8, 0], | |
| "embedding_serialization_factor": 4, | |
| "projection_serialization_factor": 4, | |
| "matmul_proportion": 0.25, | |
| "inference_matmul_proportion": 0.6, | |
| "recompute_checkpoint_every_layer": true, | |
| "optimizer_state_offchip": true, | |
| "replicated_tensor_sharding": true, | |
| "device_iterations": 10, | |
| "gradient_accumulation_steps": 128, | |
| "executable_cache_dir": "./exe_cache" | |
| } |