Transformers
PyTorch
Graphcore
English
groupbert
Generated from Trainer
Inference Endpoints
graphcore-rahult commited on
Commit
f423cae
1 Parent(s): feeaeac

Update ipu_config.json

Browse files

Lowering matmul_proportion and moving optimizer state offchip to avoid OOM on test 'groupbert_swag'

Files changed (1) hide show
  1. ipu_config.json +2 -2
ipu_config.json CHANGED
@@ -15,8 +15,8 @@
15
  4,
16
  4
17
  ],
18
- "matmul_proportion": 0.22,
19
- "optimizer_state_offchip": false,
20
  "optimum_version": "1.4.1",
21
  "output_mode": "final",
22
  "recompute_checkpoint_every_layer": true,
 
15
  4,
16
  4
17
  ],
18
+ "matmul_proportion": 0.1,
19
+ "optimizer_state_offchip": true,
20
  "optimum_version": "1.4.1",
21
  "output_mode": "final",
22
  "recompute_checkpoint_every_layer": true,