Merge pull request #1553 from hannlp/hans

Change the default argument settings for PPO training

Former-commit-id: 1b64678fa4979485f67c3bb1420dfdff6fcbc6e7

src/llmtuner/hparams/finetuning_args.py

@@ -74,6 +74,10 @@ class RLHFArguments:
         default=None,
         metadata={"help": "Log with either 'wandb' or 'tensorboard' in PPO training."}
     )
+    ppo_epochs: Optional[int] = field(
+        default=4,
+        metadata={"help": "Number of optimisation epochs per batch of samples"},
+    )
     ppo_score_norm: Optional[bool] = field(
         default=False,
         metadata={"help": "Use score normalization in PPO training."}

src/llmtuner/train/ppo/workflow.py

@@ -45,7 +45,7 @@ def run_ppo(
         mini_batch_size=training_args.per_device_train_batch_size,
         batch_size=training_args.per_device_train_batch_size * training_args.gradient_accumulation_steps,
         gradient_accumulation_steps=training_args.gradient_accumulation_steps,
-        ppo_epochs=1,
+        ppo_epochs=finetuning_args.ppo_epochs,
         max_grad_norm=training_args.max_grad_norm,
         seed=training_args.seed,
         optimize_device_cache=True,