ros/LLaMA-Factory
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

fix #2777 #2895

Browse Source 
Former-commit-id: 54d5f62d29456a8d9d0c0dd3d0bbfffe48935803
This commit is contained in:
hiyouga
2024-03-20 17:59:45 +08:00
parent d8073488be
commit c7af26a9e3
12 changed files with 104 additions and 48 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
src/llmtuner/train/ppo/workflow.py
View File

@@ -64,16 +64,16 @@ def run_ppo(
)
# Create optimizer and scheduler
optimizer = create_custom_optimzer(model, dataset, training_args, finetuning_args)
if optimizer is None:
optimizer = AdamW(filter(lambda p: p.requires_grad, model.parameters()), lr=training_args.learning_rate)
if training_args.max_steps > 0:
num_training_steps = training_args.max_steps
else:
total_train_batch_size = backward_batch_size * finetuning_args.ppo_buffer_size * training_args.world_size
num_training_steps = training_args.num_train_epochs * math.ceil(len(dataset) / total_train_batch_size)
optimizer = create_custom_optimzer(model, training_args, finetuning_args, num_training_steps)
if optimizer is None:
optimizer = AdamW(filter(lambda p: p.requires_grad, model.parameters()), lr=training_args.learning_rate)
lr_scheduler = get_scheduler(
training_args.lr_scheduler_type,
optimizer=optimizer,