liger-kernel-nightly 0.5.2.dev20241220220835__tar.gz → 0.5.2.dev20241220231758__tar.gz

{liger_kernel_nightly-0.5.2.dev20241220220835 → liger_kernel_nightly-0.5.2.dev20241220231758}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: liger_kernel_nightly
-Version: 0.5.2.dev20241220220835
+Version: 0.5.2.dev20241220231758
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 

{liger_kernel_nightly-0.5.2.dev20241220220835 → liger_kernel_nightly-0.5.2.dev20241220231758}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "liger_kernel_nightly"
-version = "0.5.2.dev20241220220835"
+version = "0.5.2.dev20241220231758"
 description = "Efficient Triton kernels for LLM Training"
 urls = { "Homepage" = "https://github.com/linkedin/Liger-Kernel" }
 readme = { file = "README.md", content-type = "text/markdown" }

{liger_kernel_nightly-0.5.2.dev20241220220835 → liger_kernel_nightly-0.5.2.dev20241220231758}/src/liger_kernel/chunked_loss/cpo_loss.py

@@ -36,8 +36,8 @@ class LigerFusedLinearCPOFunction(LigerFusedLinearPreferenceBase):
         """
         logits = beta * (chosen_logps - rejected_logps)
         loss = (
-            F.logsigmoid(logits) * (1 - label_smoothing)
-            + F.logsigmoid(-logits) * label_smoothing
+            - F.logsigmoid(logits) * (1 - label_smoothing)
+            - F.logsigmoid(-logits) * label_smoothing
         ).sum() / (full_target.shape[0] // 2)
 
         return loss

{liger_kernel_nightly-0.5.2.dev20241220220835 → liger_kernel_nightly-0.5.2.dev20241220231758}/src/liger_kernel/chunked_loss/fused_linear_preference.py

@@ -408,7 +408,7 @@ class LigerFusedLinearPreferenceBase(torch.autograd.Function):
         else:
             preference_loss, aux_outputs = preference_loss_outputs, []
 
-        loss = alpha * chosen_nll_loss - preference_loss
+        loss = alpha * chosen_nll_loss + preference_loss
         return_vars = (
             chosen_logps,
             rejected_logps,

{liger_kernel_nightly-0.5.2.dev20241220220835 → liger_kernel_nightly-0.5.2.dev20241220231758}/src/liger_kernel/chunked_loss/orpo_loss.py

@@ -36,7 +36,7 @@ class LigerFusedLinearORPOFunction(LigerFusedLinearPreferenceBase):
             - 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

{liger_kernel_nightly-0.5.2.dev20241220220835 → liger_kernel_nightly-0.5.2.dev20241220231758}/src/liger_kernel/chunked_loss/simpo_loss.py

@@ -42,8 +42,8 @@ class LigerFusedLinearSimPOFunction(LigerFusedLinearPreferenceBase):
         """
         logits = beta * (chosen_logps - rejected_logps) - gamma
         loss = (
-            F.logsigmoid(logits) * (1 - label_smoothing)
-            + F.logsigmoid(-logits) * label_smoothing
+            - F.logsigmoid(logits) * (1 - label_smoothing)
+            - F.logsigmoid(-logits) * label_smoothing
         ).sum() / (full_target.shape[0] // 2)
 
         return loss

{liger_kernel_nightly-0.5.2.dev20241220220835 → liger_kernel_nightly-0.5.2.dev20241220231758}/src/liger_kernel_nightly.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: liger_kernel_nightly
-Version: 0.5.2.dev20241220220835
+Version: 0.5.2.dev20241220231758
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 

{liger_kernel_nightly-0.5.2.dev20241220220835 → liger_kernel_nightly-0.5.2.dev20241220231758}/test/chunked_loss/test_cpo_loss.py

@@ -60,14 +60,14 @@ class HFCPOLoss(HFAlignmentLoss):
         if self.loss_type == "sigmoid":
             # This reduces to Equation 3 from the CPO paper when label_smoothing -> 0.
             losses = (
-                F.logsigmoid(self.beta * logits) * (1 - self.label_smoothing)
-                + F.logsigmoid(-self.beta * logits) * self.label_smoothing
+                -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 = (
-                F.logsigmoid(self.beta * logits) * (1 - self.label_smoothing)
-                + F.logsigmoid(-self.beta * logits) * self.label_smoothing
+                -F.logsigmoid(self.beta * logits) * (1 - self.label_smoothing)
+                - F.logsigmoid(-self.beta * logits) * self.label_smoothing
             )
         else:
             raise ValueError(

{liger_kernel_nightly-0.5.2.dev20241220220835 → liger_kernel_nightly-0.5.2.dev20241220231758}/test/chunked_loss/test_orpo_loss.py

@@ -57,7 +57,7 @@ class HFORPOLoss(HFAlignmentLoss):
             - 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

{liger_kernel_nightly-0.5.2.dev20241220220835 → liger_kernel_nightly-0.5.2.dev20241220231758}/test/utils.py

@@ -515,7 +515,7 @@ class HFAlignmentLoss:
         else:
             losses, aggregated_aux_outputs = alignment_loss_outputs, []
         # full loss
-        loss = policy_nll_loss * self.alpha - losses.mean()
+        loss = policy_nll_loss * self.alpha + losses.mean()
         return_vars = (
             policy_chosen_logps,
             policy_rejected_logps,