liger-kernel-nightly 0.5.2.dev20250104053615__tar.gz → 0.5.2.dev20250108072837__tar.gz

{liger_kernel_nightly-0.5.2.dev20250104053615 → liger_kernel_nightly-0.5.2.dev20250108072837}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: liger_kernel_nightly
-Version: 0.5.2.dev20250104053615
+Version: 0.5.2.dev20250108072837
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 

{liger_kernel_nightly-0.5.2.dev20250104053615 → liger_kernel_nightly-0.5.2.dev20250108072837}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "liger_kernel_nightly"
-version = "0.5.2.dev20250104053615"
+version = "0.5.2.dev20250108072837"
 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.dev20250104053615 → liger_kernel_nightly-0.5.2.dev20250108072837}/src/liger_kernel/chunked_loss/cpo_loss.py

@@ -33,7 +33,11 @@ class LigerFusedLinearCPOFunction(LigerFusedLinearPreferenceBase):
         loss = (-F.logsigmoid(logits) * (1 - label_smoothing) - F.logsigmoid(-logits) * label_smoothing).sum() / (
             full_target.shape[0] // 2
         )
-        return loss
+
+        chosen_rewards = beta * chosen_logps
+        rejected_rewards = beta * rejected_logps
+
+        return loss, chosen_rewards, rejected_rewards
 
     @staticmethod
     def forward(

{liger_kernel_nightly-0.5.2.dev20250104053615 → liger_kernel_nightly-0.5.2.dev20250108072837}/src/liger_kernel/chunked_loss/simpo_loss.py

@@ -42,7 +42,10 @@ class LigerFusedLinearSimPOFunction(LigerFusedLinearPreferenceBase):
             full_target.shape[0] // 2
         )
 
-        return loss
+        chosen_rewards = beta * chosen_logps
+        rejected_rewards = beta * rejected_logps
+
+        return loss, chosen_rewards, rejected_rewards
 
     @staticmethod
     def forward(

{liger_kernel_nightly-0.5.2.dev20250104053615 → liger_kernel_nightly-0.5.2.dev20250108072837}/src/liger_kernel_nightly.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: liger_kernel_nightly
-Version: 0.5.2.dev20250104053615
+Version: 0.5.2.dev20250108072837
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 

{liger_kernel_nightly-0.5.2.dev20250104053615 → liger_kernel_nightly-0.5.2.dev20250108072837}/test/chunked_loss/test_cpo_loss.py

@@ -73,7 +73,11 @@ class HFCPOLoss(HFAlignmentLoss):
             )
         else:
             raise ValueError(f"Unknown loss type: {self.loss_type}. Should be one of ['sigmoid']")
-        return losses
+
+        chosen_rewards = self.beta * policy_chosen_logps
+        rejected_rewards = self.beta * policy_rejected_logps
+
+        return losses, chosen_rewards, rejected_rewards
 
 
 class TorchLMHeadCPO(torch.nn.Module):