bug fixed

src/rlhf/reward.py

@@ -129,7 +129,8 @@ class RewardModel(pl.LightningModule):
         # if self.deepspeed_offload:
         #     return DeepSpeedCPUAdam(optim_groups, lr=self.args.lr_init, betas=self.args.betas, eps=self.args.adam_eps, bias_correction=True, adamw_mode=False, weight_decay=0, amsgrad=False)
         # return FusedAdam(optim_groups, lr=self.args.lr_init, betas=self.args.betas, eps=self.args.adam_eps, bias_correction=True, adam_w_mode=False, weight_decay=0, amsgrad=False)
-        return ZeroOneAdam(optim_groups, lr=self.args.lr_init, betas=self.args.betas, eps=self.args.adam_eps, bias_correction=True, weight_decay=0, amsgrad=False, cuda_aware=False)
+        # return ZeroOneAdam(optim_groups, lr=self.args.lr_init, betas=self.args.betas, eps=self.args.adam_eps, bias_correction=True, weight_decay=0, amsgrad=False, cuda_aware=False)
+        return torch.optim.Adam(optim_groups, lr=1e-5, betas=(0.9, 0.95))
     
     @property
     def deepspeed_offload(self) -> bool: