gpt2-wikitext-103 / ipu_config.json
Jinchen's picture
Add fine-tuned model
e76c1eb
raw
history blame
663 Bytes
{
"device_iterations": 2,
"embedding_serialization_factor": 4,
"enable_half_first_order_momentum": true,
"enable_half_partials": true,
"executable_cache_dir": "./exe_cache",
"gradient_accumulation_steps": 128,
"inference_device_iterations": 5,
"inference_replication_factor": 4,
"ipus_per_replica": 4,
"layers_per_ipu": [
0,
4,
4,
4
],
"matmul_proportion": 0.25,
"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": 4,
"seed": 42,
"use_popdist": false
}