Create ipu_config.json
Browse files- ipu_config.json +36 -0
ipu_config.json
ADDED
@@ -0,0 +1,36 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"decompose_grad_sum": false,
|
3 |
+
"device_iterations": 5,
|
4 |
+
"embedding_serialization_factor": 2,
|
5 |
+
"enable_half_first_order_momentum": true,
|
6 |
+
"enable_half_partials": true,
|
7 |
+
"executable_cache_dir": "/tmp/exe_cache/",
|
8 |
+
"execute_encoder_on_cpu_for_generation": false,
|
9 |
+
"gradient_accumulation_steps": 64,
|
10 |
+
"inference_device_iterations": 4,
|
11 |
+
"inference_replication_factor": 4,
|
12 |
+
"ipus_per_replica": 4,
|
13 |
+
"layers_per_ipu": [
|
14 |
+
3,
|
15 |
+
7,
|
16 |
+
7,
|
17 |
+
7
|
18 |
+
],
|
19 |
+
"matmul_proportion": [
|
20 |
+
0.15,
|
21 |
+
0.18,
|
22 |
+
0.2,
|
23 |
+
0.25
|
24 |
+
],
|
25 |
+
"optimizer_state_offchip": true,
|
26 |
+
"optimum_version": "1.4.1",
|
27 |
+
"output_mode": "final",
|
28 |
+
"profile_dir": "",
|
29 |
+
"recompute_checkpoint_every_layer": true,
|
30 |
+
"replicated_tensor_sharding": true,
|
31 |
+
"replication_factor": 4,
|
32 |
+
"seed": 42,
|
33 |
+
"sharded_execution_for_inference": false,
|
34 |
+
"transformers_version": "4.20.0",
|
35 |
+
"use_popdist": false
|
36 |
+
}
|