| { | |
| "recompute_checkpoint_every_layer": true, | |
| "optimizer_state_offchip": true, | |
| "replicated_tensor_sharding": true, | |
| "enable_half_first_order_momentum": true, | |
| "enable_half_partials": true, | |
| "replication_factor": {"pod4": 1, "pod8": 2, "pod16": 4, "pod32": 8, "pod64": 16, "pod128": 32, "pod256": 64, "default": 1}, | |
| "gradient_accumulation_steps": 512, | |
| "device_iterations": 1, | |
| "executable_cache_dir": "./exe_cache", | |
| "inference_device_iterations": 4, | |
| "inference_replication_factor": {"pod4": 1, "pod8": 2, "pod16": 4, "pod32": 8, "pod64": 16, "pod128": 32, "pod256": 64, "default": 1}, | |
| "ipus_per_replica": 4, | |
| "layers_per_ipu": [0, 4, 4, 4], | |
| "matmul_proportion": [0.25, 0.25, 0.25, 0.25] | |
| } |