liger-kernel-nightly 0.5.2.dev20241219211841__py3-none-any.whl → 0.5.2.dev20241220220835__py3-none-any.whl

liger_kernel/chunked_loss/cpo_loss.py

@@ -9,7 +9,9 @@ from liger_kernel.chunked_loss.fused_linear_preference import (
 class LigerFusedLinearCPOFunction(LigerFusedLinearPreferenceBase):
 
     @staticmethod
-    def preference_loss_fn(chosen_logps, rejected_logps, full_target, beta=0.1):
+    def preference_loss_fn(
+        chosen_logps, rejected_logps, full_target, beta=0.1, label_smoothing=0.0
+    ):
         """
         Paper: https://arxiv.org/pdf/2401.08417
 
@@ -30,9 +32,14 @@ class LigerFusedLinearCPOFunction(LigerFusedLinearPreferenceBase):
             rejected_logps (torch.Tensor): Avg log probabilities of rejected tokens. Shape: (batch_size,).
             full_target (torch.Tensor): Non chunked full target tensor
             beta (float): Weight for the CPO loss
+            label_smoothing (float): Label smoothing factor, will reduce to Equation above when label_smoothing -> 0.
         """
         logits = beta * (chosen_logps - rejected_logps)
-        loss = F.logsigmoid(logits).sum() / (full_target.shape[0] // 2)
+        loss = (
+            F.logsigmoid(logits) * (1 - label_smoothing)
+            + F.logsigmoid(-logits) * label_smoothing
+        ).sum() / (full_target.shape[0] // 2)
+
         return loss
 
     @staticmethod
@@ -45,6 +52,7 @@ class LigerFusedLinearCPOFunction(LigerFusedLinearPreferenceBase):
         ignore_index=-100,
         beta=0.1,
         alpha=1.0,
+        label_smoothing=0.0,
         compute_nll_loss=True,
         compiled=True,
     ):
@@ -58,6 +66,7 @@ class LigerFusedLinearCPOFunction(LigerFusedLinearPreferenceBase):
             ignore_index=ignore_index,
             alpha=alpha,
             beta=beta,
+            label_smoothing=label_smoothing,
             compute_nll_loss=compute_nll_loss,
             compiled=compiled,
         )
@@ -65,7 +74,7 @@ class LigerFusedLinearCPOFunction(LigerFusedLinearPreferenceBase):
     @staticmethod
     def backward(ctx, *grad_output):
         grads = LigerFusedLinearPreferenceBase.backward(ctx, grad_output)[:4]
-        return *grads, None, None, None, None, None
+        return *grads, None, None, None, None, None, None
 
 
 class LigerFusedLinearCPOLoss(torch.nn.Module):
@@ -78,6 +87,7 @@ class LigerFusedLinearCPOLoss(torch.nn.Module):
         ignore_index: int = -100,
         beta: float = 0.1,
         alpha: float = 1.0,
+        label_smoothing: float = 0.0,
         compute_nll_loss: bool = True,
         compiled: bool = True,
     ):
@@ -90,6 +100,7 @@ class LigerFusedLinearCPOLoss(torch.nn.Module):
         self.ignore_index = ignore_index
         self.beta = beta
         self.alpha = alpha
+        self.label_smoothing = label_smoothing
         self.compute_nll_loss = compute_nll_loss
         self.compiled = compiled
 
@@ -102,6 +113,7 @@ class LigerFusedLinearCPOLoss(torch.nn.Module):
             self.ignore_index,
             self.beta,
             self.alpha,
+            self.label_smoothing,
             self.compute_nll_loss,
             self.compiled,
         )

liger_kernel/chunked_loss/dpo_loss.py

@@ -64,7 +64,7 @@ class LigerFusedLinearDPOFunction(LigerFusedLinearPreferenceBase):
         ref_bias=None,
         ignore_index=-100,
         beta=0.1,
-        compute_nll_loss=True,
+        compute_nll_loss=False,
         compiled=True,
         use_ref_model=True,
     ):
@@ -100,7 +100,7 @@ class LigerFusedLinearDPOLoss(torch.nn.Module):
         self,
         ignore_index: int = -100,
         beta: float = 0.1,
-        compute_nll_loss: bool = True,
+        compute_nll_loss: bool = False,
         compiled: bool = True,
         use_ref_model: bool = False,
     ):

liger_kernel/chunked_loss/simpo_loss.py

@@ -10,7 +10,12 @@ class LigerFusedLinearSimPOFunction(LigerFusedLinearPreferenceBase):
 
     @staticmethod
     def preference_loss_fn(
-        chosen_logps, rejected_logps, full_target, beta=0.1, gamma=0.5
+        chosen_logps,
+        rejected_logps,
+        full_target,
+        beta=0.1,
+        gamma=0.5,
+        label_smoothing=0.0,
     ):
         """
         Paper: https://arxiv.org/pdf/2405.14734
@@ -33,9 +38,14 @@ class LigerFusedLinearSimPOFunction(LigerFusedLinearPreferenceBase):
             full_target: Non chunked full target tensor
             beta (float): beta weight
             gamma (float): gemma margin term
+            label_smoothing (float): Label smoothing factor, will reduce to Equation above when label_smoothing -> 0.
         """
         logits = beta * (chosen_logps - rejected_logps) - gamma
-        loss = F.logsigmoid(logits).sum() / (full_target.shape[0] // 2)
+        loss = (
+            F.logsigmoid(logits) * (1 - label_smoothing)
+            + F.logsigmoid(-logits) * label_smoothing
+        ).sum() / (full_target.shape[0] // 2)
+
         return loss
 
     @staticmethod
@@ -48,6 +58,7 @@ class LigerFusedLinearSimPOFunction(LigerFusedLinearPreferenceBase):
         ignore_index=-100,
         beta=0.1,
         alpha=1.0,
+        label_smoothing=0.0,
         compute_nll_loss=False,
         compiled=True,
         gamma=0.5,
@@ -63,6 +74,7 @@ class LigerFusedLinearSimPOFunction(LigerFusedLinearPreferenceBase):
             ignore_index=ignore_index,
             alpha=alpha,
             beta=beta,
+            label_smoothing=label_smoothing,
             compiled=compiled,
             gamma=gamma,
         )
@@ -70,7 +82,7 @@ class LigerFusedLinearSimPOFunction(LigerFusedLinearPreferenceBase):
     @staticmethod
     def backward(ctx, *grad_output):
         grads = LigerFusedLinearPreferenceBase.backward(ctx, grad_output)[:4]
-        return *grads, None, None, None, None, None, None
+        return *grads, None, None, None, None, None, None, None
 
 
 class LigerFusedLinearSimPOLoss(torch.nn.Module):
@@ -83,6 +95,7 @@ class LigerFusedLinearSimPOLoss(torch.nn.Module):
         ignore_index: int = -100,
         beta: float = 0.1,
         alpha: float = 1.0,
+        label_smoothing: float = 0.0,
         compute_nll_loss: bool = True,
         compiled: bool = True,
         gamma: float = 0.5,
@@ -96,6 +109,7 @@ class LigerFusedLinearSimPOLoss(torch.nn.Module):
         self.ignore_index = ignore_index
         self.beta = beta
         self.alpha = alpha
+        self.label_smoothing = label_smoothing
         self.compute_nll_loss = compute_nll_loss
         self.compiled = compiled
         self.gamma = gamma
@@ -109,6 +123,7 @@ class LigerFusedLinearSimPOLoss(torch.nn.Module):
             self.ignore_index,
             self.beta,
             self.alpha,
+            self.label_smoothing,
             self.compute_nll_loss,
             self.compiled,
             self.gamma,

{liger_kernel_nightly-0.5.2.dev20241219211841.dist-info → liger_kernel_nightly-0.5.2.dev20241220220835.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: liger_kernel_nightly
-Version: 0.5.2.dev20241219211841
+Version: 0.5.2.dev20241220220835
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 

{liger_kernel_nightly-0.5.2.dev20241219211841.dist-info → liger_kernel_nightly-0.5.2.dev20241220220835.dist-info}/RECORD

@@ -3,13 +3,13 @@ liger_kernel/env_report.py,sha256=ok9PMXtO-8uLj_feCJI4h9hz2NtolZ2AG_OJTW5qmo4,18
 liger_kernel/utils.py,sha256=HJa-xVKOohDn6pLVIx-Fv0V9h0QAL3qZGQNRICI-OpI,249
 liger_kernel/chunked_loss/README.md,sha256=K6rucm6nqHpWCmxUOhBYcE3apwQxAy0TfRUippR7Icw,2243
 liger_kernel/chunked_loss/__init__.py,sha256=R2wCcz4Y0kTAve926DH3k182XKezpXeACMHj05g9Mm8,346
-liger_kernel/chunked_loss/cpo_loss.py,sha256=Qu1Ul2A12sp6CqIT-atPbHWFb_LLtINEA9mOpIRx_0g,3097
-liger_kernel/chunked_loss/dpo_loss.py,sha256=9S67SzKkLyoBmHGx8bkmthSNHlCT2ikBy9CCdb7wGj0,4381
+liger_kernel/chunked_loss/cpo_loss.py,sha256=_0phkDQg_SG239xNVlaAvGoL33myt0Zt6avDJfg6HC0,3552
+liger_kernel/chunked_loss/dpo_loss.py,sha256=jbTno1pKEc-HxAGFY3NSycBzdWyTacyRCzH3FhrMUMo,4383
 liger_kernel/chunked_loss/functional.py,sha256=9Gr-YXIuEzEJkBUhDx3G2fuQayckLor7cC7svhmPML4,549
 liger_kernel/chunked_loss/fused_linear_distillation.py,sha256=2BH6DCPjsR2zS6zcwFPcIIZRhLF8SohjGdKsAJ_301o,10222
 liger_kernel/chunked_loss/fused_linear_preference.py,sha256=AsovMdfsOjgWVxtDhZ_rXqpahMsKTg8YueXnZcHt1XQ,16376
 liger_kernel/chunked_loss/orpo_loss.py,sha256=ZuKGjbkIYzV4UzvupNdq6vyxCp7-BztQkUt8ZnFvKos,3531
-liger_kernel/chunked_loss/simpo_loss.py,sha256=Wa4LOlDG9PbJkOOkKg8hbKvnKgg7OTBz6-qIkwPK1yw,3275
+liger_kernel/chunked_loss/simpo_loss.py,sha256=Lb7v-a1nmapiCZGxgoWtoLRBiSLI3arFJ-WstjTqIAs,3757
 liger_kernel/ops/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 liger_kernel/ops/cross_entropy.py,sha256=oG5hfrlmnlF5lOoZRhHRglObxgH4B0KadjWMJj9EWPM,15860
 liger_kernel/ops/fused_linear_cross_entropy.py,sha256=Tnw4gyAYVVdnCOqhOuLEzbUQ3goOTnoAfk3pqSIM5ac,9301
@@ -58,9 +58,9 @@ liger_kernel/transformers/trainer/__init__.py,sha256=c4OQVJmhNOloj0JYSEc0j_cQuBb
 liger_kernel/transformers/trainer/orpo_trainer.py,sha256=O2k2vdHl-O1S-U61aEmyUFu3QrEuNAipQa2oUBb3HAA,7679
 liger_kernel/triton/__init__.py,sha256=yfRe0zMb47QnqjecZWG7LnanfCTzeku7SgWRAwNVmzU,101
 liger_kernel/triton/monkey_patch.py,sha256=5BcGKTtdqeYchypBIBopGIWPx1-cFALz7sOKoEsqXJ0,1584
-liger_kernel_nightly-0.5.2.dev20241219211841.dist-info/LICENSE,sha256=OhzLDHJ0to4a8sodVLELZiCFylZ1NAAYLs-HrjPy0ag,1312
-liger_kernel_nightly-0.5.2.dev20241219211841.dist-info/METADATA,sha256=53BLTRM2B5gVYzBVvKfDi4nWijlv2xXT63LDGXTlvQc,21055
-liger_kernel_nightly-0.5.2.dev20241219211841.dist-info/NOTICE,sha256=njwnoPZLh9AN8SJQzxvCGLHi-8X__AvWRze6joNXIY8,2066
-liger_kernel_nightly-0.5.2.dev20241219211841.dist-info/WHEEL,sha256=P9jw-gEje8ByB7_hXoICnHtVCrEwMQh-630tKvQWehc,91
-liger_kernel_nightly-0.5.2.dev20241219211841.dist-info/top_level.txt,sha256=2eghu4hA3LnkM7ElW92tQ8zegWKgSbeo-k-aGe1YnvY,13
-liger_kernel_nightly-0.5.2.dev20241219211841.dist-info/RECORD,,
+liger_kernel_nightly-0.5.2.dev20241220220835.dist-info/LICENSE,sha256=OhzLDHJ0to4a8sodVLELZiCFylZ1NAAYLs-HrjPy0ag,1312
+liger_kernel_nightly-0.5.2.dev20241220220835.dist-info/METADATA,sha256=TKTsG1OaMEqrbqun7C_JsCMT3Ui6ED2G0CPLnno79QA,21055
+liger_kernel_nightly-0.5.2.dev20241220220835.dist-info/NOTICE,sha256=njwnoPZLh9AN8SJQzxvCGLHi-8X__AvWRze6joNXIY8,2066
+liger_kernel_nightly-0.5.2.dev20241220220835.dist-info/WHEEL,sha256=P9jw-gEje8ByB7_hXoICnHtVCrEwMQh-630tKvQWehc,91
+liger_kernel_nightly-0.5.2.dev20241220220835.dist-info/top_level.txt,sha256=2eghu4hA3LnkM7ElW92tQ8zegWKgSbeo-k-aGe1YnvY,13
+liger_kernel_nightly-0.5.2.dev20241220220835.dist-info/RECORD,,