Update ipu_config.json
Browse files- ipu_config.json +18 -18
ipu_config.json
CHANGED
|
@@ -1,19 +1,19 @@
|
|
| 1 |
{
|
| 2 |
-
|
| 3 |
-
|
| 4 |
-
|
| 5 |
-
|
| 6 |
-
|
| 7 |
-
|
| 8 |
-
|
| 9 |
-
|
| 10 |
-
|
| 11 |
-
|
| 12 |
-
|
| 13 |
-
|
| 14 |
-
|
| 15 |
-
|
| 16 |
-
|
| 17 |
-
|
| 18 |
-
|
| 19 |
-
}
|
|
|
|
| 1 |
{
|
| 2 |
+
"embedding_serialization_factor": 2,
|
| 3 |
+
"recompute_checkpoint_every_layer": true,
|
| 4 |
+
"optimizer_state_offchip": true,
|
| 5 |
+
"replicated_tensor_sharding": true,
|
| 6 |
+
"enable_half_first_order_momentum": true,
|
| 7 |
+
"enable_half_partials": true,
|
| 8 |
+
|
| 9 |
+
"device_iterations": 1,
|
| 10 |
+
"inference_device_iterations": 5,
|
| 11 |
+
"replication_factor": {"pod4": 1, "pod8": 2, "pod16": 4, "pod32": 8, "pod64": 16, "default": 1},
|
| 12 |
+
"inference_replication_factor": {"pod4": 1, "pod8": 2, "pod16": 4, "pod32": 8, "pod64": 16, "default": 1},
|
| 13 |
+
"gradient_accumulation_steps": 512,
|
| 14 |
+
"executable_cache_dir": "./exe_cache",
|
| 15 |
+
|
| 16 |
+
"ipus_per_replica": 4,
|
| 17 |
+
"layers_per_ipu": [0, 4, 4, 4],
|
| 18 |
+
"matmul_proportion": [0.25, 0.25, 0.25, 0.25]
|
| 19 |
+
}
|