Format

Signed-off-by: Austin Liu <austin362667@gmail.com>

src/liger_kernel/chunked_loss/cpo_loss.py

@@ -32,7 +32,7 @@ def preference_loss_fn(chosen_logps, rejected_logps, full_target, beta=0.1):
             beta (float): Weight for the CPO loss
         """
         logits = beta * (chosen_logps - rejected_logps)
-        loss = - F.logsigmoid(logits).sum() / (full_target.shape[0] // 2)
+        loss = -F.logsigmoid(logits).sum() / (full_target.shape[0] // 2)
         return loss
 
     @staticmethod

src/liger_kernel/chunked_loss/dpo_loss.py

@@ -49,7 +49,7 @@ def preference_loss_fn(
         rejected_logratios = rejected_logps - ref_rejected_logps
 
         logits_diff = beta * (chosen_logratios - rejected_logratios)
-        loss = - F.logsigmoid(logits_diff).sum() / (full_target.shape[0] // 2)
+        loss = -F.logsigmoid(logits_diff).sum() / (full_target.shape[0] // 2)
         return loss
 
     @staticmethod

src/liger_kernel/chunked_loss/orpo_loss.py

@@ -36,7 +36,7 @@ def preference_loss_fn(chosen_logps, rejected_logps, full_target, beta=0.1):
             - torch.log1p(-torch.exp(rejected_logps))
         )
         ratio = F.logsigmoid(log_odds)
-        loss = - beta * ratio.sum() / (full_target.shape[0] // 2)
+        loss = -beta * ratio.sum() / (full_target.shape[0] // 2)
 
         chosen_rewards = beta * chosen_logps
         rejected_rewards = beta * rejected_logps

src/liger_kernel/chunked_loss/simpo_loss.py

@@ -35,7 +35,7 @@ def preference_loss_fn(
             gamma (float): gemma margin term
         """
         logits = beta * (chosen_logps - rejected_logps) - gamma
-        loss = - F.logsigmoid(logits).sum() / (full_target.shape[0] // 2)
+        loss = -F.logsigmoid(logits).sum() / (full_target.shape[0] // 2)
         return loss
 
     @staticmethod

test/chunked_loss/test_cpo_loss.py

@@ -59,13 +59,13 @@ def alignment_loss(
         # calculates a conservative CPO loss.
         if self.loss_type == "sigmoid":
             # This reduces to Equation 3 from the CPO paper when label_smoothing -> 0.
-            losses = - (
+            losses = -(
                 F.logsigmoid(self.beta * logits) * (1 - self.label_smoothing)
                 + F.logsigmoid(-self.beta * logits) * self.label_smoothing
             )
         elif self.loss_type == "simpo":
             logits = logits - (self.simpo_gamma / self.beta)
-            losses = - (
+            losses = -(
                 F.logsigmoid(self.beta * logits) * (1 - self.label_smoothing)
                 + F.logsigmoid(-self.beta * logits) * self.label_smoothing
             )

test/chunked_loss/test_orpo_loss.py

@@ -57,7 +57,7 @@ def alignment_loss(
             - torch.log1p(-torch.exp(policy_rejected_logps))
         )
         ratio = F.logsigmoid(log_odds)
-        losses = - self.beta * ratio
+        losses = -self.beta * ratio
 
         chosen_rewards = self.beta * policy_chosen_logps
         rejected_rewards = self.beta * policy_rejected_logps