Transformers
PyTorch
Graphcore
bert
generated_from_trainer
Inference Endpoints
bert-large-uncased / ipu_config.json
sergiopperez's picture
Update BERT large uncased checkpoint after running phase 1 (SL 128) and phase 2 (SL 512)
387585b
raw history blame
No virus
704 Bytes
{
"device_iterations": 1,
"embedding_serialization_factor": 2,
"enable_half_first_order_momentum": true,
"enable_half_partials": true,
"executable_cache_dir": "./exe_cache",
"gradient_accumulation_steps": 512,
"inference_device_iterations": 4,
"inference_replication_factor": 16,
"ipus_per_replica": 4,
"layers_per_ipu": [
3,
7,
7,
7
],
"matmul_proportion": [
0.1,
0.15,
0.15,
0.15
],
"optimizer_state_offchip": true,
"optimum_version": "1.0.0",
"output_mode": "final",
"profile_dir": "",
"recompute_checkpoint_every_layer": true,
"replicated_tensor_sharding": true,
"replication_factor": 16,
"seed": 42,
"use_popdist": false
}