423A35C7/LLaMA-Factory
1
0
Fork 0
mirror of https://github.com/hiyouga/LLaMA-Factory.git synced 2025-08-06 13:42:51 +08:00
Code Issues Packages Projects Releases Wiki Activity

fix ChatGLM RLHF

Browse Source 
Former-commit-id: af6c011fcb8ea9e5cf2eb4699da33d8668df04b4
This commit is contained in:
hiyouga 2023-08-15 11:19:20 +08:00
parent 603cb30887
commit a9ab8f71d7
2 changed files with 4 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
src/llmtuner/tuner/ppo/trainer.py
View File

@ -182,6 +182,8 @@ class PPOPeftTrainer(PPOTrainer, PeftTrainer):
replace_model(unwrapped_model, target="reward") replace_model(unwrapped_model, target="reward")
batch = self.prepare_model_inputs(queries, responses) batch = self.prepare_model_inputs(queries, responses)
_, _, values = self.model(**batch, output_hidden_states=True, return_dict=True) _, _, values = self.model(**batch, output_hidden_states=True, return_dict=True)
if values.size(0) != batch["input_ids"].size(0):
values = torch.transpose(values, 0, 1)
rewards = [reward for reward in values[:, -1].float().detach().cpu()] # use fp32 type rewards = [reward for reward in values[:, -1].float().detach().cpu()] # use fp32 type
replace_model(unwrapped_model, target="default") replace_model(unwrapped_model, target="default")
return rewards return rewards

2
src/llmtuner/tuner/rm/trainer.py
View File

@ -42,6 +42,8 @@ class PairwisePeftTrainer(PeftTrainer):
""" """
batch_size = inputs["input_ids"].size(0) // 2 batch_size = inputs["input_ids"].size(0) // 2
_, _, values = model(**inputs, output_hidden_states=True, return_dict=True) _, _, values = model(**inputs, output_hidden_states=True, return_dict=True)
if values.size(0) != inputs["input_ids"].size(0):
values = torch.transpose(values, 0, 1)
r_accept, r_reject = values[:, -1].split(batch_size, dim=0) r_accept, r_reject = values[:, -1].split(batch_size, dim=0)
loss = -torch.log(torch.sigmoid(r_accept - r_reject)).mean() loss = -torch.log(torch.sigmoid(r_accept - r_reject)).mean()
return (loss, [loss, r_accept, r_reject]) if return_outputs else loss return (loss, [loss, r_accept, r_reject]) if return_outputs else loss