423A35C7/LLaMA-Factory
1
0
Fork 0
mirror of https://github.com/hiyouga/LLaMA-Factory.git synced 2025-08-02 19:52:50 +08:00
Code Issues Packages Projects Releases Wiki Activity

[trainer] fix batch processing in PPO trainer (#7576)

Browse Source
This commit is contained in:
gechengze 2025-04-02 21:17:48 +08:00 committed by GitHub
parent 903db09822
commit 11997593be
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 5 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

8
src/llamafactory/train/ppo/trainer.py
View File

@ -241,9 +241,11 @@ class CustomPPOTrainer(PPOTrainer, Trainer):
self.tokenizer.padding_side = "right" # change padding side self.tokenizer.padding_side = "right" # change padding side
queries, responses, rewards = [], [], [] queries, responses, rewards = [], [], []
for idx in range(0, self.config.batch_size, self.config.mini_batch_size): for idx in range(0, self.config.batch_size, self.config.mini_batch_size):
mini_batch_queries, mini_batch_responses = self.get_inputs( mini_batch = {
batch[idx : idx + self.config.mini_batch_size] "input_ids": batch["input_ids"][idx : idx + self.config.mini_batch_size],
) "attention_mask": batch["attention_mask"][idx : idx + self.config.mini_batch_size]
}
mini_batch_queries, mini_batch_responses = self.get_inputs(mini_batch)
mini_batch_rewards = self.get_rewards(mini_batch_queries, mini_batch_responses) mini_batch_rewards = self.get_rewards(mini_batch_queries, mini_batch_responses)
queries.extend(mini_batch_queries) queries.extend(mini_batch_queries)
responses.extend(mini_batch_responses) responses.extend(mini_batch_responses)