liger-kernel-nightly 0.6.2.dev20251011154427__py3-none-any.whl → 0.6.4.dev20251202054858__py3-none-any.whl

liger_kernel/transformers/cross_entropy.py

@@ -3,6 +3,7 @@ from typing import Optional
 import torch
 
 from liger_kernel.ops.cross_entropy import LigerCrossEntropyFunction
+from liger_kernel.transformers.functional import CrossEntropyOutput
 
 
 class LigerCrossEntropyLoss(torch.nn.Module):
@@ -15,6 +16,7 @@ class LigerCrossEntropyLoss(torch.nn.Module):
         reduction: str = "mean",
         softcap: Optional[float] = None,
         return_z_loss: bool = False,
+        return_token_accuracy: bool = False,
     ):
         super().__init__()
         assert (label_smoothing >= 0) and (label_smoothing <= 1), (
@@ -33,9 +35,10 @@ class LigerCrossEntropyLoss(torch.nn.Module):
         self.reduction = reduction
         self.softcap = softcap
         self.return_z_loss = return_z_loss
+        self.return_token_accuracy = return_token_accuracy
 
     def forward(self, _input: torch.Tensor, target: torch.Tensor):
-        loss, z_loss = LigerCrossEntropyFunction.apply(
+        loss, z_loss, token_accuracy = LigerCrossEntropyFunction.apply(
             _input,
             target,
             self.weight,
@@ -45,7 +48,9 @@ class LigerCrossEntropyLoss(torch.nn.Module):
             self.reduction,
             self.softcap,
             self.return_z_loss,
+            self.return_token_accuracy,
         )
-        if not self.return_z_loss:
+        if not self.return_z_loss and not self.return_token_accuracy:
             return loss
-        return loss, z_loss
+
+        return CrossEntropyOutput(loss=loss, z_loss=z_loss, token_accuracy=token_accuracy)

liger_kernel/transformers/functional.py

@@ -1,5 +1,8 @@
+from dataclasses import dataclass
 from typing import Optional
 
+import torch
+
 from liger_kernel.ops.cross_entropy import LigerCrossEntropyFunction
 from liger_kernel.ops.dyt import LigerDyTFunction
 from liger_kernel.ops.fused_add_rms_norm import LigerFusedAddRMSNormFunction
@@ -12,6 +15,7 @@ from liger_kernel.ops.jsd import LigerJSDFunction
 from liger_kernel.ops.kl_div import LigerKLDivLossFunction
 from liger_kernel.ops.layer_norm import LigerLayerNormFunction
 from liger_kernel.ops.multi_token_attention import LigerMultiTokenAttentionFunction
+from liger_kernel.ops.poly_norm import LigerPolyNormFunction
 from liger_kernel.ops.qwen2vl_mrope import LigerQwen2VLMRopeFunction
 from liger_kernel.ops.rms_norm import LigerRMSNormFunction
 from liger_kernel.ops.rope import LigerRopeFunction
@@ -21,6 +25,13 @@ from liger_kernel.ops.swiglu import LigerSiLUMulFunction
 from liger_kernel.ops.tvd import LigerTVDLossFunction
 
 
+@dataclass
+class CrossEntropyOutput:
+    loss: torch.Tensor
+    z_loss: Optional[torch.Tensor] = None
+    token_accuracy: Optional[torch.Tensor] = None
+
+
 # conform to the function signature in https://pytorch.org/docs/stable/generated/torch.nn.functional.cross_entropy.html
 # `weight` and `size_average` are placeholders and not implemented yet
 def liger_cross_entropy(
@@ -35,8 +46,9 @@ def liger_cross_entropy(
     lse_square_scale: float = 0.0,
     softcap: Optional[float] = None,
     return_z_loss: bool = False,
+    return_token_accuracy: bool = False,
 ):
-    loss, z_loss = LigerCrossEntropyFunction.apply(
+    loss, z_loss, token_accuracy = LigerCrossEntropyFunction.apply(
         input,
         target,
         weight,
@@ -46,10 +58,13 @@ def liger_cross_entropy(
         reduction,
         softcap,
         return_z_loss,
+        return_token_accuracy,
     )
-    if not return_z_loss:
+
+    if not return_z_loss and not return_token_accuracy:
         return loss
-    return loss, z_loss
+
+    return CrossEntropyOutput(loss=loss, z_loss=z_loss, token_accuracy=token_accuracy)
 
 
 def liger_fused_linear_cross_entropy(
@@ -66,8 +81,9 @@ def liger_fused_linear_cross_entropy(
     return_z_loss: bool = False,
     accum_dtype=None,
     use_token_scaling: bool = False,
+    return_token_accuracy: bool = False,
 ):
-    loss, z_loss = LigerFusedLinearCrossEntropyFunction.apply(
+    loss, z_loss, token_accuracy = LigerFusedLinearCrossEntropyFunction.apply(
         input,
         weight,
         target,
@@ -81,10 +97,13 @@ def liger_fused_linear_cross_entropy(
         return_z_loss,
         accum_dtype,
         use_token_scaling,
+        return_token_accuracy,
     )
-    if not return_z_loss:
+
+    if not return_z_loss and not return_token_accuracy:
         return loss
-    return loss, z_loss
+
+    return CrossEntropyOutput(loss=loss, z_loss=z_loss, token_accuracy=token_accuracy)
 
 
 def liger_fused_linear_jsd(
@@ -258,6 +277,10 @@ def liger_rms_norm(X, W, eps, offset: float = 0.0, casting_mode: str = "llama",
     return LigerRMSNormFunction.apply(X, W, eps, offset, casting_mode, in_place)
 
 
+def liger_poly_norm(X, W, B, eps=1e-6, in_place=True):
+    return LigerPolyNormFunction.apply(X, W, B, eps, in_place)
+
+
 def liger_fused_add_rms_norm(X, R, W, eps, offset: float = 0.0, casting_mode: str = "llama", in_place: bool = True):
     return LigerFusedAddRMSNormFunction.apply(X, R, W, eps, offset, casting_mode, in_place)
 

liger_kernel/transformers/fused_linear_cross_entropy.py

@@ -3,6 +3,7 @@ from typing import Optional
 import torch
 
 from liger_kernel.ops.fused_linear_cross_entropy import LigerFusedLinearCrossEntropyFunction
+from liger_kernel.transformers.functional import CrossEntropyOutput
 
 
 class LigerFusedLinearCrossEntropyLoss(torch.nn.Module):
@@ -17,6 +18,7 @@ class LigerFusedLinearCrossEntropyLoss(torch.nn.Module):
         return_z_loss: bool = False,
         accum_dtype: Optional[torch.dtype] = None,
         use_token_scaling: bool = False,
+        return_token_accuracy: bool = False,
     ):
         super().__init__()
         assert (label_smoothing >= 0) and (label_smoothing <= 1), (
@@ -37,9 +39,10 @@ class LigerFusedLinearCrossEntropyLoss(torch.nn.Module):
         self.return_z_loss = return_z_loss
         self.accum_dtype = accum_dtype
         self.use_token_scaling = use_token_scaling
+        self.return_token_accuracy = return_token_accuracy
 
     def forward(self, lin_weight, _input, target, bias=None):
-        loss, z_loss = LigerFusedLinearCrossEntropyFunction.apply(
+        loss, z_loss, token_accuracy = LigerFusedLinearCrossEntropyFunction.apply(
             _input,
             lin_weight,
             target,
@@ -53,7 +56,9 @@ class LigerFusedLinearCrossEntropyLoss(torch.nn.Module):
             self.return_z_loss,
             self.accum_dtype,
             self.use_token_scaling,
+            self.return_token_accuracy,
         )
-        if not self.return_z_loss:
+        if not self.return_z_loss and not self.return_token_accuracy:
             return loss
-        return loss, z_loss
+
+        return CrossEntropyOutput(loss=loss, z_loss=z_loss, token_accuracy=token_accuracy)

liger_kernel/transformers/grpo_loss.py

@@ -1,3 +1,6 @@
+import torch
+
+from liger_kernel.chunked_loss.fused_linear_ppo import LigerFusedLinearPPOBase
 from liger_kernel.ops.grpo_loss import GrpoLossFunction
 
 
@@ -13,12 +16,20 @@ def triton_grpo_loss(
     eps_low=0.2,
     eps_high=0.4,
     inplace=True,
+    loss_type="dapo",
+    max_completion_length=None,
+    importance_sampling_level="token",
+    reduce=False,
 ):
     assert logits is not None and completion_ids is not None and advantages is not None, (
         "must provide logits、completion_ids and advantages"
     )
+    if importance_sampling_level != "token":
+        raise ValueError(
+            f"Triton GRPO loss only supports token-level importance sampling. Got {importance_sampling_level}."
+        )
 
-    return GrpoLossFunction.apply(
+    per_token_loss, per_token_kl, is_clipped = GrpoLossFunction.apply(
         logits,
         old_logp,
         ref_logp,
@@ -31,6 +42,50 @@ def triton_grpo_loss(
         eps_high,
         inplace,
     )
+    if not reduce:
+        return per_token_loss, per_token_kl, is_clipped
+
+    loss = _reduce_grpo_loss(
+        per_token_loss,
+        completion_mask,
+        loss_type=loss_type,
+        max_completion_length=max_completion_length,
+    )
+
+    metrics = []
+    if beta != 0.0 and per_token_kl is not None:
+        metrics.append(_masked_mean(per_token_kl, completion_mask))
+    metrics.append(_masked_mean(is_clipped.float(), completion_mask))
+    return loss, metrics
+
+
+def _reduce_grpo_loss(per_token_loss, completion_mask, loss_type, max_completion_length):
+    mask = completion_mask
+    if mask is None:
+        mask = torch.ones_like(per_token_loss, dtype=per_token_loss.dtype, device=per_token_loss.device)
+    mask = mask.to(per_token_loss.dtype)
+
+    if loss_type == "grpo":
+        per_seq = (per_token_loss * mask).sum(-1) / mask.sum(-1).clamp(min=1.0)
+        return per_seq.mean()
+    if loss_type == "bnpo":
+        return (per_token_loss * mask).sum() / mask.sum().clamp(min=1.0)
+    if loss_type == "dr_grpo":
+        if max_completion_length is None:
+            raise ValueError("max_completion_length must be provided when using loss_type='dr_grpo'")
+        batch = per_token_loss.shape[0]
+        return (per_token_loss * mask).sum() / (batch * max_completion_length)
+    if loss_type == "dapo":
+        normalizer = LigerFusedLinearPPOBase._compute_dapo_normalizer(mask)
+        return (per_token_loss * mask).sum() / normalizer
+    raise ValueError(f"Unsupported loss_type '{loss_type}' for Triton GRPO loss.")
+
+
+def _masked_mean(values, mask):
+    if mask is None:
+        mask = torch.ones_like(values, dtype=values.dtype, device=values.device)
+    mask = mask.to(values.dtype)
+    return (values * mask).sum() / mask.sum().clamp(min=1.0)
 
 
 # This is a demo how to use grpo_loss in GRPOTrainer. The Trl version must be 0.16

liger_kernel/transformers/model/falcon_h1.py

@@ -4,12 +4,12 @@ from typing import Union
 
 import torch
 
-from transformers.modeling_outputs import CausalLMOutputWithPast
-
 if TYPE_CHECKING:
     from transformers.models.falcon_h1.modeling_falcon_h1 import FalconHybridMambaAttentionDynamicCache
 
 from liger_kernel.transformers.model.loss_utils import LigerForCausalLMLoss
+from liger_kernel.transformers.model.loss_utils import unpack_cross_entropy_result
+from liger_kernel.transformers.model.output_classes import LigerCausalLMOutputWithPast
 
 
 def lce_forward(
@@ -26,8 +26,9 @@ def lce_forward(
     cache_position: Optional[torch.LongTensor] = None,
     logits_to_keep: Union[int, torch.Tensor] = 0,
     skip_logits: Optional[bool] = None,
+    return_dict: Optional[bool] = None,
     **kwargs,
-) -> Union[tuple, CausalLMOutputWithPast]:
+) -> Union[tuple, LigerCausalLMOutputWithPast]:
     r"""
     labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
         Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
@@ -54,6 +55,7 @@ def lce_forward(
     output_hidden_states = (
         output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
     )
+    return_dict = return_dict if return_dict is not None else self.config.use_return_dict
 
     # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
     outputs = self.model(
@@ -77,6 +79,8 @@ def lce_forward(
     shift_labels = kwargs.pop("shift_labels", None)
     logits = None
     loss = None
+    token_accuracy = None
+
     # if in training mode, don't materialize logits
     if skip_logits and labels is None:
         raise ValueError("skip_logits is True, but labels and shift_labels are None")
@@ -85,8 +89,9 @@ def lce_forward(
         # By default, if in training mode, don't materialize logits
         skip_logits = self.training and labels is not None
 
+    # Compute loss
     if skip_logits:
-        loss = LigerForCausalLMLoss(
+        result = LigerForCausalLMLoss(
             hidden_states=kept_hidden_states,
             lm_head_weight=self.lm_head.weight,
             labels=labels,
@@ -94,15 +99,24 @@ def lce_forward(
             hidden_size=self.config.hidden_size,
             **kwargs,
         )
+        loss, _, token_accuracy = unpack_cross_entropy_result(result)
     else:
         logits = self.lm_head(kept_hidden_states)
         if labels is not None or shift_labels is not None:
             loss = self.loss_function(logits=logits, labels=labels, vocab_size=self.config.vocab_size, **kwargs)
 
-    return CausalLMOutputWithPast(
+    if not return_dict:
+        output = (logits,) + outputs[1:]
+        output = ((loss,) + output) if loss is not None else output
+        output = output + (token_accuracy,) if token_accuracy is not None else output
+        return output
+
+    # Return custom output class with token_accuracy field
+    return LigerCausalLMOutputWithPast(
         loss=loss,
         logits=logits,
         past_key_values=outputs.past_key_values,
         hidden_states=outputs.hidden_states,
         attentions=outputs.attentions,
+        token_accuracy=token_accuracy,
     )

liger_kernel/transformers/model/gemma.py

@@ -12,6 +12,8 @@ from transformers.utils.deprecation import deprecate_kwarg
 
 from liger_kernel.transformers.fused_linear_cross_entropy import LigerFusedLinearCrossEntropyLoss
 from liger_kernel.transformers.model.loss_utils import LigerForCausalLMLoss
+from liger_kernel.transformers.model.loss_utils import unpack_cross_entropy_result
+from liger_kernel.transformers.model.output_classes import LigerCausalLMOutputWithPast
 
 
 def lce_forward_deprecated(
@@ -147,7 +149,7 @@ def lce_forward(
     logits_to_keep: Union[int, torch.Tensor] = 0,
     skip_logits: Optional[bool] = None,
     **kwargs,
-) -> Union[Tuple, CausalLMOutputWithPast]:
+) -> Union[Tuple, LigerCausalLMOutputWithPast]:
     r"""
     Args:
         labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
@@ -209,6 +211,7 @@ def lce_forward(
     shift_labels = kwargs.pop("shift_labels", None)
     logits = None
     loss = None
+    token_accuracy = None
 
     if skip_logits and labels is None and shift_labels is None:
         raise ValueError("skip_logits is True, but labels and shift_labels are None")
@@ -217,8 +220,9 @@ def lce_forward(
         # By default, if in training mode, don't materialize logits
         skip_logits = self.training and (labels is not None or shift_labels is not None)
 
+    # Compute loss
     if skip_logits:
-        loss = LigerForCausalLMLoss(
+        result = LigerForCausalLMLoss(
             hidden_states=kept_hidden_states,
             lm_head_weight=self.lm_head.weight,
             labels=labels,
@@ -226,6 +230,7 @@ def lce_forward(
             hidden_size=self.config.hidden_size,
             **kwargs,
         )
+        loss, _, token_accuracy = unpack_cross_entropy_result(result)
     else:
         logits = self.lm_head(kept_hidden_states)
         if labels is not None or shift_labels is not None:
@@ -238,13 +243,19 @@ def lce_forward(
             )
 
     if not return_dict:
-        output = (logits,) + outputs[1:]
-        return (loss,) + output if loss is not None else output
-
-    return CausalLMOutputWithPast(
+        output_tuple = (logits,) + outputs[1:]
+        if loss is not None:
+            output_tuple = (loss,) + output_tuple
+        if token_accuracy is not None:
+            output_tuple = output_tuple + (token_accuracy,)
+        return output_tuple
+
+    # Return custom output class with token_accuracy field
+    return LigerCausalLMOutputWithPast(
         loss=loss,
         logits=logits,
         past_key_values=outputs.past_key_values,
         hidden_states=outputs.hidden_states,
         attentions=outputs.attentions,
+        token_accuracy=token_accuracy,
     )

liger_kernel/transformers/model/gemma2.py

@@ -13,6 +13,8 @@ from transformers.utils.deprecation import deprecate_kwarg
 
 from liger_kernel.transformers.fused_linear_cross_entropy import LigerFusedLinearCrossEntropyLoss
 from liger_kernel.transformers.model.loss_utils import LigerForCausalLMLoss
+from liger_kernel.transformers.model.loss_utils import unpack_cross_entropy_result
+from liger_kernel.transformers.model.output_classes import LigerCausalLMOutputWithPast
 
 logger = logging.getLogger(__name__)
 
@@ -158,7 +160,7 @@ def lce_forward(
     logits_to_keep: Union[int, torch.Tensor] = 0,
     skip_logits: Optional[bool] = None,
     **kwargs,
-) -> Union[Tuple, CausalLMOutputWithPast]:
+) -> Union[Tuple, LigerCausalLMOutputWithPast]:
     r"""
     Args:
         labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
@@ -225,6 +227,7 @@ def lce_forward(
     shift_labels = kwargs.pop("shift_labels", None)
     logits = None
     loss = None
+    token_accuracy = None
 
     if skip_logits and labels is None and shift_labels is None:
         raise ValueError("skip_logits is True, but labels and shift_labels are None")
@@ -233,8 +236,9 @@ def lce_forward(
         # By default, if in training mode, don't materialize logits
         skip_logits = self.training and (labels is not None or shift_labels is not None)
 
+    # Compute loss
     if skip_logits:
-        loss = LigerForCausalLMLoss(
+        result = LigerForCausalLMLoss(
             hidden_states=kept_hidden_states,
             lm_head_weight=self.lm_head.weight,
             labels=labels,
@@ -243,6 +247,7 @@ def lce_forward(
             final_logit_softcapping=self.config.final_logit_softcapping,
             **kwargs,
         )
+        loss, _, token_accuracy = unpack_cross_entropy_result(result)
 
     else:
         logits = self.lm_head(kept_hidden_states)
@@ -262,13 +267,17 @@ def lce_forward(
             )
 
     if not return_dict:
-        output = (logits,) + outputs[1:]
-        return (loss,) + output if loss is not None else output
+        output_tuple = (logits,) + outputs[1:]
+        output_tuple = (loss,) + output_tuple if loss is not None else output_tuple
+        output_tuple = output_tuple + (token_accuracy,) if token_accuracy is not None else output_tuple
+        return output_tuple
 
-    return CausalLMOutputWithPast(
+    # Return custom output class with token_accuracy field
+    return LigerCausalLMOutputWithPast(
         loss=loss,
         logits=logits,
         past_key_values=outputs.past_key_values,
         hidden_states=outputs.hidden_states,
         attentions=outputs.attentions,
+        token_accuracy=token_accuracy,
     )

liger_kernel/transformers/model/gemma3.py

@@ -7,12 +7,13 @@ import torch.nn as nn
 
 from transformers.cache_utils import Cache
 from transformers.cache_utils import HybridCache
-from transformers.modeling_outputs import CausalLMOutputWithPast
-from transformers.models.gemma3.modeling_gemma3 import Gemma3CausalLMOutputWithPast
 from transformers.utils import logging
 
 from liger_kernel.transformers.fused_linear_cross_entropy import LigerFusedLinearCrossEntropyLoss
 from liger_kernel.transformers.model.loss_utils import LigerForCausalLMLoss
+from liger_kernel.transformers.model.loss_utils import unpack_cross_entropy_result
+from liger_kernel.transformers.model.output_classes import LigerCausalLMOutputWithPast
+from liger_kernel.transformers.model.output_classes import LigerGemma3CausalLMOutputWithPast
 
 logger = logging.get_logger(__name__)
 
@@ -33,7 +34,7 @@ def causal_forward(
     logits_to_keep: Union[int, torch.Tensor] = 0,
     skip_logits: Optional[bool] = None,
     **loss_kwargs,
-) -> Union[Tuple, CausalLMOutputWithPast]:
+) -> Union[Tuple, LigerCausalLMOutputWithPast]:
     r"""
         labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
             Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
@@ -98,12 +99,14 @@ def causal_forward(
     shift_labels = loss_kwargs.pop("shift_labels", None)
     loss = None
     logits = None
+    token_accuracy = None
 
     if skip_logits is None:
         skip_logits = self.training and (labels is not None or shift_labels is not None)
 
+    # Compute loss
     if skip_logits:
-        loss = LigerForCausalLMLoss(
+        result = LigerForCausalLMLoss(
             hidden_states=kept_hidden_states,
             lm_head_weight=self.lm_head.weight,
             labels=labels,
@@ -112,7 +115,7 @@ def causal_forward(
             final_logit_softcapping=self.config.final_logit_softcapping,
             **loss_kwargs,
         )
-
+        loss, _, token_accuracy = unpack_cross_entropy_result(result)
     else:
         logits = self.lm_head(kept_hidden_states)
         if self.config.final_logit_softcapping is not None:
@@ -129,15 +132,19 @@ def causal_forward(
             )
 
     if not return_dict:
-        output = (logits,) + outputs[1:]
-        return (loss,) + output if loss is not None else output
+        output_tuple = (logits,) + outputs[1:]
+        output_tuple = (loss,) + output_tuple if loss is not None else output_tuple
+        output_tuple = output_tuple + (token_accuracy,) if token_accuracy is not None else output_tuple
+        return output_tuple
 
-    return CausalLMOutputWithPast(
+    # Return custom output class with token_accuracy field
+    return LigerCausalLMOutputWithPast(
         loss=loss,
         logits=logits,
         past_key_values=outputs.past_key_values,
         hidden_states=outputs.hidden_states,
         attentions=outputs.attentions,
+        token_accuracy=token_accuracy,
     )
 
 
@@ -159,7 +166,7 @@ def multimodal_forward(
     logits_to_keep: Union[int, torch.Tensor] = 0,
     skip_logits: Optional[bool] = None,
     **lm_kwargs,
-) -> Union[tuple, Gemma3CausalLMOutputWithPast]:
+) -> Union[tuple, LigerGemma3CausalLMOutputWithPast]:
     r"""
     labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
         Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
@@ -235,6 +242,7 @@ def multimodal_forward(
 
     loss = None
     logits = None
+    token_accuracy = None
     if skip_logits and labels is None:
         raise ValueError("skip_logits is True, but labels is None")
 
@@ -261,7 +269,9 @@ def multimodal_forward(
         shift_labels = shift_labels.view(-1).to(hidden_device)
 
         lce = LigerFusedLinearCrossEntropyLoss()
-        loss = lce(self.lm_head.weight, shift_hidden_states, shift_labels)
+        result = lce(self.lm_head.weight, shift_hidden_states, shift_labels)
+        loss, _, token_accuracy = unpack_cross_entropy_result(result)
+
     else:
         logits = self.lm_head(kept_hidden_states)
         if labels is not None:
@@ -306,13 +316,16 @@ def multimodal_forward(
 
     if not return_dict:
         output = (logits,) + outputs[1:]
-        return (loss,) + output if loss is not None else output
+        output = (loss,) + output if loss is not None else output
+        output = output + (token_accuracy,) if token_accuracy is not None else output
+        return output
 
-    return Gemma3CausalLMOutputWithPast(
+    return LigerGemma3CausalLMOutputWithPast(
         loss=loss,
         logits=logits,
         past_key_values=outputs.past_key_values,
         hidden_states=outputs.hidden_states,
         attentions=outputs.attentions,
         image_hidden_states=outputs.image_hidden_states,
+        token_accuracy=token_accuracy,
     )

liger_kernel/transformers/model/glm4.py

@@ -5,10 +5,11 @@ from typing import Union
 
 import torch
 
-from transformers.modeling_outputs import CausalLMOutputWithPast
 from transformers.utils.deprecation import deprecate_kwarg
 
 from liger_kernel.transformers.model.loss_utils import LigerForCausalLMLoss
+from liger_kernel.transformers.model.loss_utils import unpack_cross_entropy_result
+from liger_kernel.transformers.model.output_classes import LigerCausalLMOutputWithPast
 
 
 @deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
@@ -28,7 +29,7 @@ def lce_forward(
     logits_to_keep: Union[int, torch.Tensor] = 0,
     skip_logits: Optional[bool] = None,
     **kwargs,
-) -> Union[Tuple, CausalLMOutputWithPast]:
+) -> Union[Tuple, LigerCausalLMOutputWithPast]:
     r"""
     Args:
         labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
@@ -91,6 +92,7 @@ def lce_forward(
     shift_labels = kwargs.pop("shift_labels", None)
     logits = None
     loss = None
+    token_accuracy = None
 
     if skip_logits and labels is None and shift_labels is None:
         raise ValueError("skip_logits is True, but labels and shift_labels are None")
@@ -99,8 +101,9 @@ def lce_forward(
         # By default, if in training mode, don't materialize logits
         skip_logits = self.training and (labels is not None or shift_labels is not None)
 
+    # Compute loss
     if skip_logits:
-        loss = LigerForCausalLMLoss(
+        result = LigerForCausalLMLoss(
             hidden_states=kept_hidden_states,
             lm_head_weight=self.lm_head.weight,
             labels=labels,
@@ -108,6 +111,7 @@ def lce_forward(
             hidden_size=self.config.hidden_size,
             **kwargs,
         )
+        loss, _, token_accuracy = unpack_cross_entropy_result(result)
 
     else:
         logits = self.lm_head(kept_hidden_states)
@@ -120,10 +124,18 @@ def lce_forward(
                 **kwargs,
             )
 
-    return CausalLMOutputWithPast(
+    if not return_dict:
+        output = (logits,) + outputs[1:]
+        output = ((loss,) + output) if loss is not None else output
+        output = output + (token_accuracy,) if token_accuracy is not None else output
+        return output
+
+    # Return custom output class with token_accuracy field
+    return LigerCausalLMOutputWithPast(
         loss=loss,
         logits=logits,
         past_key_values=outputs.past_key_values,
         hidden_states=outputs.hidden_states,
         attentions=outputs.attentions,
+        token_accuracy=token_accuracy,
     )