make one cycle lr div factor configurable
Browse files
src/axolotl/utils/trainer.py
CHANGED
@@ -157,7 +157,7 @@ def setup_trainer(cfg, train_dataset, eval_dataset, model, tokenizer):
|
|
157 |
cfg.learning_rate,
|
158 |
total_steps=total_num_steps,
|
159 |
epochs=cfg.num_epochs,
|
160 |
-
div_factor=
|
161 |
**lr_scheduler_kwargs,
|
162 |
)
|
163 |
elif cfg.lr_scheduler == "log_sweep":
|
@@ -182,7 +182,7 @@ def setup_trainer(cfg, train_dataset, eval_dataset, model, tokenizer):
|
|
182 |
cfg.early_stopping_patience,
|
183 |
)
|
184 |
callbacks.append(early_stop_cb)
|
185 |
-
|
186 |
if cfg.local_rank == 0 and cfg.adapter == 'lora': # only save in rank 0
|
187 |
callbacks.append(SavePeftModelCallback)
|
188 |
|
|
|
157 |
cfg.learning_rate,
|
158 |
total_steps=total_num_steps,
|
159 |
epochs=cfg.num_epochs,
|
160 |
+
div_factor=cfg.lr_div_factor if cfg.lr_div_factor else 6,
|
161 |
**lr_scheduler_kwargs,
|
162 |
)
|
163 |
elif cfg.lr_scheduler == "log_sweep":
|
|
|
182 |
cfg.early_stopping_patience,
|
183 |
)
|
184 |
callbacks.append(early_stop_cb)
|
185 |
+
|
186 |
if cfg.local_rank == 0 and cfg.adapter == 'lora': # only save in rank 0
|
187 |
callbacks.append(SavePeftModelCallback)
|
188 |
|