mirror of
https://github.com/k2-fsa/icefall.git
synced 2025-08-08 09:32:20 +00:00
* add decode seamlessm4t * add requirements * add decoding with avg model * add token files * add custom tokenizer * support deepspeed to finetune large model * support large-v3 * add model saving * using monkey patch to replace models * add manifest dir option
39 lines
922 B
JSON
39 lines
922 B
JSON
{
|
|
"fp16": {
|
|
"enabled": true,
|
|
"loss_scale": 0,
|
|
"loss_scale_window": 100,
|
|
"initial_scale_power": 16,
|
|
"hysteresis": 2,
|
|
"min_loss_scale": 0.01
|
|
},
|
|
"zero_optimization": {
|
|
"stage": 1,
|
|
"allgather_partitions": true,
|
|
"allgather_bucket_size": 2e8,
|
|
"overlap_comm": true,
|
|
"reduce_scatter": true,
|
|
"reduce_bucket_size": 2e8,
|
|
"contiguous_gradients": true
|
|
},
|
|
"optimizer": {
|
|
"type": "Adam",
|
|
"params": {
|
|
"lr": 1e-5
|
|
}
|
|
},
|
|
"scheduler": {
|
|
"type": "WarmupLR",
|
|
"params": {
|
|
"warmup_min_lr": 0,
|
|
"warmup_max_lr": 1e-5,
|
|
"warmup_num_steps": 100
|
|
}
|
|
},
|
|
"gradient_accumulation_steps": 1,
|
|
"gradient_clipping": 5,
|
|
"steps_per_print": 50,
|
|
"train_micro_batch_size_per_gpu": 1,
|
|
"wall_clock_breakdown": false
|
|
}
|