Transformers
PyTorch
Graphcore
English
groupbert
Generated from Trainer
Inference Endpoints
graphcore-rahult commited on
Commit
0bbc713
1 Parent(s): 12e8885

Update ipu_config.json

Browse files

Set `replication_factor` and `inference_replication_factor` to dicts based on pod_type for flexibility. If no `pod_type` argument is provided, the default value of 4 will be retained.

Files changed (1) hide show
  1. ipu_config.json +2 -2
ipu_config.json CHANGED
@@ -7,7 +7,7 @@
7
  "execute_encoder_on_cpu_for_generation": false,
8
  "gradient_accumulation_steps": 2048,
9
  "inference_device_iterations": 5,
10
- "inference_replication_factor": 4,
11
  "ipus_per_replica": 4,
12
  "layers_per_ipu": [
13
  0,
@@ -21,7 +21,7 @@
21
  "output_mode": "final",
22
  "recompute_checkpoint_every_layer": true,
23
  "replicated_tensor_sharding": true,
24
- "replication_factor": 4,
25
  "seed": 42,
26
  "sharded_execution_for_inference": false,
27
  "transformers_version": "4.20.1"
 
7
  "execute_encoder_on_cpu_for_generation": false,
8
  "gradient_accumulation_steps": 2048,
9
  "inference_device_iterations": 5,
10
+ "inference_replication_factor": {"pod4": 1, "pod8": 2, "pod16": 4, "pod32": 8, "pod64": 16, "default": 4},
11
  "ipus_per_replica": 4,
12
  "layers_per_ipu": [
13
  0,
 
21
  "output_mode": "final",
22
  "recompute_checkpoint_every_layer": true,
23
  "replicated_tensor_sharding": true,
24
+ "replication_factor": {"pod4": 1, "pod8": 2, "pod16": 4, "pod32": 8, "pod64": 16, "default": 4},
25
  "seed": 42,
26
  "sharded_execution_for_inference": false,
27
  "transformers_version": "4.20.1"