liger-kernel-nightly 0.4.2.dev20241119223206__tar.gz → 0.4.2.dev20241121054604__tar.gz

{liger_kernel_nightly-0.4.2.dev20241119223206/src/liger_kernel_nightly.egg-info → liger_kernel_nightly-0.4.2.dev20241121054604}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: liger_kernel_nightly
-Version: 0.4.2.dev20241119223206
+Version: 0.4.2.dev20241121054604
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 

{liger_kernel_nightly-0.4.2.dev20241119223206 → liger_kernel_nightly-0.4.2.dev20241121054604}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "liger_kernel_nightly"
-version = "0.4.2.dev20241119223206"
+version = "0.4.2.dev20241121054604"
 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-0.4.2.dev20241121054604/src/liger_kernel/chunked_loss/__init__.py

@@ -0,0 +1,4 @@
+from liger_kernel.chunked_loss.cpo_loss import LigerFusedLinearCPOLoss  # noqa: F401
+from liger_kernel.chunked_loss.dpo_loss import LigerFusedLinearDPOLoss  # noqa: F401
+from liger_kernel.chunked_loss.orpo_loss import LigerFusedLinearORPOLoss  # noqa: F401
+from liger_kernel.chunked_loss.simpo_loss import LigerFusedLinearSimPOLoss  # noqa: F401

{liger_kernel_nightly-0.4.2.dev20241119223206 → liger_kernel_nightly-0.4.2.dev20241121054604}/src/liger_kernel/chunked_loss/cpo_loss.py

@@ -1,3 +1,4 @@
+import torch
 import torch.nn.functional as F
 
 from liger_kernel.chunked_loss.fused_linear_preference import (
@@ -46,10 +47,10 @@ class LigerFusedLinearCPOFunction(LigerFusedLinearPreferenceBase):
             target,
             bias,
             loss_fn=LigerFusedLinearCPOFunction.preference_loss_fn,
-            compute_nll_loss=compute_nll_loss,
             ignore_index=ignore_index,
             alpha=alpha,
             beta=beta,
+            compute_nll_loss=compute_nll_loss,
             compiled=compiled,
         )
 
@@ -59,3 +60,42 @@ class LigerFusedLinearCPOFunction(LigerFusedLinearPreferenceBase):
         grads = LigerFusedLinearPreferenceBase.backward(ctx, grad_output)[:4]
         # Return these gradients, followed by None for the remaining inputs
         return *grads, None, None, None, None, None
+
+
+class LigerFusedLinearCPOLoss(torch.nn.Module):
+    """
+    Fused linear layer with CPO loss.
+    """
+
+    def __init__(
+        self,
+        ignore_index: int = -100,
+        beta: float = 0.1,
+        alpha: float = 1.0,
+        compute_nll_loss: bool = True,
+        compiled: bool = True,
+    ):
+        """
+        Args:
+            ignore_index (int): Index to ignore in the loss.
+            beta (float): Weight for the odds ratio loss.
+        """
+        super().__init__()
+        self.ignore_index = ignore_index
+        self.beta = beta
+        self.alpha = alpha
+        self.compute_nll_loss = compute_nll_loss
+        self.compiled = compiled
+
+    def forward(self, lin_weight, _input, target, bias=None):
+        return LigerFusedLinearCPOFunction.apply(
+            _input,
+            lin_weight,
+            target,
+            bias,
+            self.ignore_index,
+            self.beta,
+            self.alpha,
+            self.compute_nll_loss,
+            self.compiled,
+        )

{liger_kernel_nightly-0.4.2.dev20241119223206 → liger_kernel_nightly-0.4.2.dev20241121054604}/src/liger_kernel/chunked_loss/dpo_loss.py

@@ -1,3 +1,4 @@
+import torch
 import torch.nn.functional as F
 
 from liger_kernel.chunked_loss.fused_linear_preference import (
@@ -43,9 +44,9 @@ class LigerFusedLinearDPOFunction(LigerFusedLinearPreferenceBase):
             target=target,
             bias=bias,
             loss_fn=LigerFusedLinearDPOFunction.preference_loss_fn,
-            compute_nll_loss=compute_nll_loss,
             ignore_index=ignore_index,
             beta=beta,
+            compute_nll_loss=compute_nll_loss,
             compiled=compiled,
         )
 
@@ -55,3 +56,39 @@ class LigerFusedLinearDPOFunction(LigerFusedLinearPreferenceBase):
         grads = LigerFusedLinearPreferenceBase.backward(ctx, grad_output)[:4]
         # Return these gradients, followed by None for the remaining inputs
         return *grads, None, None, None, None
+
+
+class LigerFusedLinearDPOLoss(torch.nn.Module):
+    """
+    Fused linear layer with DPO loss.
+    """
+
+    def __init__(
+        self,
+        ignore_index: int = -100,
+        beta: float = 0.1,
+        compute_nll_loss: bool = True,
+        compiled: bool = True,
+    ):
+        """
+        Args:
+            ignore_index (int): Index to ignore in the loss.
+            beta (float): Weight for the odds ratio loss.
+        """
+        super().__init__()
+        self.ignore_index = ignore_index
+        self.beta = beta
+        self.compute_nll_loss = compute_nll_loss
+        self.compiled = compiled
+
+    def forward(self, lin_weight, _input, target, bias=None):
+        return LigerFusedLinearDPOFunction.apply(
+            _input,
+            lin_weight,
+            target,
+            bias,
+            self.ignore_index,
+            self.beta,
+            self.compute_nll_loss,
+            self.compiled,
+        )

liger_kernel_nightly-0.4.2.dev20241121054604/src/liger_kernel/chunked_loss/functional.py

@@ -0,0 +1,9 @@
+from liger_kernel.chunked_loss.cpo_loss import LigerFusedLinearCPOFunction
+from liger_kernel.chunked_loss.dpo_loss import LigerFusedLinearDPOFunction
+from liger_kernel.chunked_loss.orpo_loss import LigerFusedLinearORPOFunction
+from liger_kernel.chunked_loss.simpo_loss import LigerFusedLinearSimPOFunction
+
+liger_fused_linear_orpo = LigerFusedLinearORPOFunction.apply
+liger_fused_linear_dpo = LigerFusedLinearDPOFunction.apply
+liger_fused_linear_cpo = LigerFusedLinearCPOFunction.apply
+liger_fused_linear_simpo = LigerFusedLinearSimPOFunction.apply

{liger_kernel_nightly-0.4.2.dev20241119223206 → liger_kernel_nightly-0.4.2.dev20241121054604}/src/liger_kernel/chunked_loss/orpo_loss.py

@@ -34,7 +34,7 @@ class LigerFusedLinearORPOFunction(LigerFusedLinearPreferenceBase):
         ignore_index=-100,
         beta=0.1,
         compute_nll_loss=True,
-        compiled=False,
+        compiled=True,
     ):
         """
         Fused linear layer with ORPO (Odds-Ratio Preference Optimization) loss.
@@ -49,9 +49,9 @@ class LigerFusedLinearORPOFunction(LigerFusedLinearPreferenceBase):
             target=target,
             bias=bias,
             loss_fn=LigerFusedLinearORPOFunction.preference_loss_fn,
-            compute_nll_loss=compute_nll_loss,
             ignore_index=ignore_index,
             beta=beta,
+            compute_nll_loss=compute_nll_loss,
             compiled=compiled,
         )
 
@@ -61,3 +61,39 @@ class LigerFusedLinearORPOFunction(LigerFusedLinearPreferenceBase):
         grads = LigerFusedLinearPreferenceBase.backward(ctx, grad_output)[:4]
         # Return these gradients, followed by None for the remaining inputs
         return *grads, None, None, None, None
+
+
+class LigerFusedLinearORPOLoss(torch.nn.Module):
+    """
+    Fused linear layer with ORPO (Odds-Ratio Preference Optimization) loss.
+    """
+
+    def __init__(
+        self,
+        ignore_index: int = -100,
+        beta: float = 0.1,
+        compute_nll_loss: bool = True,
+        compiled: bool = True,
+    ):
+        """
+        Args:
+            ignore_index (int): Index to ignore in the loss.
+            beta (float): Weight for the odds ratio loss.
+        """
+        super().__init__()
+        self.ignore_index = ignore_index
+        self.beta = beta
+        self.compute_nll_loss = compute_nll_loss
+        self.compiled = compiled
+
+    def forward(self, lin_weight, _input, target, bias=None):
+        return LigerFusedLinearORPOFunction.apply(
+            _input,
+            lin_weight,
+            target,
+            bias,
+            self.ignore_index,
+            self.beta,
+            self.compute_nll_loss,
+            self.compiled,
+        )

{liger_kernel_nightly-0.4.2.dev20241119223206 → liger_kernel_nightly-0.4.2.dev20241121054604}/src/liger_kernel/chunked_loss/simpo_loss.py

@@ -1,3 +1,4 @@
+import torch
 import torch.nn.functional as F
 
 from liger_kernel.chunked_loss.fused_linear_preference import (
@@ -62,3 +63,45 @@ class LigerFusedLinearSimPOFunction(LigerFusedLinearPreferenceBase):
         grads = LigerFusedLinearPreferenceBase.backward(ctx, grad_output)[:4]
         # Return these gradients, followed by None for the remaining inputs
         return *grads, None, None, None, None, None, None
+
+
+class LigerFusedLinearSimPOLoss(torch.nn.Module):
+    """
+    Fused linear layer with SimPO loss.
+    """
+
+    def __init__(
+        self,
+        ignore_index: int = -100,
+        beta: float = 0.1,
+        alpha: float = 1.0,
+        compute_nll_loss: bool = True,
+        compiled: bool = True,
+        gamma: float = 0.5,
+    ):
+        """
+        Args:
+            ignore_index (int): Index to ignore in the loss.
+            beta (float): Weight for the odds ratio loss.
+        """
+        super().__init__()
+        self.ignore_index = ignore_index
+        self.beta = beta
+        self.alpha = alpha
+        self.compute_nll_loss = compute_nll_loss
+        self.compiled = compiled
+        self.gamma = gamma
+
+    def forward(self, lin_weight, _input, target, bias=None):
+        return LigerFusedLinearSimPOFunction.apply(
+            _input,
+            lin_weight,
+            target,
+            bias,
+            self.ignore_index,
+            self.beta,
+            self.alpha,
+            self.compute_nll_loss,
+            self.compiled,
+            self.gamma,
+        )

{liger_kernel_nightly-0.4.2.dev20241119223206 → liger_kernel_nightly-0.4.2.dev20241121054604/src/liger_kernel_nightly.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: liger_kernel_nightly
-Version: 0.4.2.dev20241119223206
+Version: 0.4.2.dev20241121054604
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 

{liger_kernel_nightly-0.4.2.dev20241119223206 → liger_kernel_nightly-0.4.2.dev20241121054604}/src/liger_kernel_nightly.egg-info/SOURCES.txt

@@ -6,6 +6,7 @@ src/liger_kernel/env_report.py
 src/liger_kernel/chunked_loss/__init__.py
 src/liger_kernel/chunked_loss/cpo_loss.py
 src/liger_kernel/chunked_loss/dpo_loss.py
+src/liger_kernel/chunked_loss/functional.py
 src/liger_kernel/chunked_loss/fused_linear_preference.py
 src/liger_kernel/chunked_loss/orpo_loss.py
 src/liger_kernel/chunked_loss/simpo_loss.py

{liger_kernel_nightly-0.4.2.dev20241119223206 → liger_kernel_nightly-0.4.2.dev20241121054604}/src/liger_kernel/chunked_loss/fused_linear_preference.py

@@ -27,10 +27,10 @@ class LigerFusedLinearPreferenceBase(torch.autograd.Function):
         bias=None,
         loss_fn=None,
         chunk_size=1,
-        compute_nll_loss=True,
         ignore_index=-100,
         alpha=1.0,
         beta=0.1,
+        compute_nll_loss=True,
         compiled=True,
         **loss_kwargs,
     ):