liger-kernel-nightly 0.5.10.dev20250624183504__py3-none-any.whl → 0.6.3.dev20251121010306__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/experimental/__init__.py

@@ -0,0 +1,5 @@
+from liger_kernel.transformers.experimental.embedding import LigerEmbedding  # noqa: F401
+
+__all__ = [
+    "LigerEmbedding",
+]

liger_kernel/transformers/functional.py

@@ -1,7 +1,11 @@
+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
 from liger_kernel.ops.fused_linear_cross_entropy import LigerFusedLinearCrossEntropyFunction
 from liger_kernel.ops.fused_linear_jsd import LigerFusedLinearJSDFunction
 from liger_kernel.ops.fused_neighborhood_attention import LigerFusedNeighborhoodAttentionFunction
@@ -11,6 +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
@@ -20,6 +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(
@@ -34,8 +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,
@@ -45,10 +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(
@@ -63,8 +79,11 @@ def liger_fused_linear_cross_entropy(
     reduction: str = "mean",
     softcap: Optional[float] = None,
     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,
@@ -76,10 +95,15 @@ def liger_fused_linear_cross_entropy(
         reduction,
         softcap,
         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(
@@ -253,6 +277,14 @@ 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)
+
+
 def liger_rope(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
     return LigerRopeFunction.apply(q, k, cos, sin, position_ids, unsqueeze_dim)
 

liger_kernel/transformers/fused_add_rms_norm.py

@@ -0,0 +1,39 @@
+import torch
+import torch.nn as nn
+
+from liger_kernel.ops.fused_add_rms_norm import LigerFusedAddRMSNormFunction
+
+
+class LigerFusedAddRMSNorm(nn.Module):
+    def __init__(
+        self,
+        hidden_size,
+        eps=1e-6,
+        offset=0.0,
+        casting_mode="llama",
+        init_fn="ones",
+        in_place=False,
+    ):
+        super().__init__()
+        assert init_fn in [
+            "ones",
+            "zeros",
+        ], f"init_fn must be either 'ones' or 'zeros', got {init_fn}"
+        self.weight = nn.Parameter(torch.ones(hidden_size) if init_fn == "ones" else torch.zeros(hidden_size))
+        self.variance_epsilon, self.offset, self.casting_mode, self.in_place = (eps, offset, casting_mode, in_place)
+
+    def forward(self, hidden_states, residual):
+        return LigerFusedAddRMSNormFunction.apply(
+            hidden_states,
+            residual,
+            self.weight,
+            self.variance_epsilon,
+            self.offset,
+            self.casting_mode,
+            self.in_place,
+        )
+
+    def extra_repr(self):
+        return (
+            f"{tuple(self.weight.shape)}, eps={self.variance_epsilon}, offset={self.offset}, in_place={self.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):
@@ -15,6 +16,9 @@ class LigerFusedLinearCrossEntropyLoss(torch.nn.Module):
         reduction: str = "mean",
         softcap: Optional[float] = None,
         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), (
@@ -23,7 +27,8 @@ class LigerFusedLinearCrossEntropyLoss(torch.nn.Module):
         assert reduction in {
             "mean",
             "sum",
-        }, f"reduction must be 'mean' or 'sum'. Got: {reduction}"
+            "none",
+        }, f"reduction must be 'mean' or 'sum' or 'none'. Got: {reduction}"
         assert softcap is None or softcap > 0, f"softcap must greater than 0.0 or None. Got: {softcap}"
         self.ce_weight = ce_weight
         self.ignore_index = ignore_index
@@ -32,9 +37,12 @@ class LigerFusedLinearCrossEntropyLoss(torch.nn.Module):
         self.reduction = reduction
         self.softcap = softcap
         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,
@@ -46,7 +54,11 @@ class LigerFusedLinearCrossEntropyLoss(torch.nn.Module):
             self.reduction,
             self.softcap,
             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/llama4_rope.py

@@ -0,0 +1,93 @@
+"""
+Liger Kernel implementation of Llama4 Rotary Position Embedding (RoPE).
+Supports both text and vision RoPE variants with fused operations for optimal performance.
+"""
+
+import torch
+
+from liger_kernel.ops.llama4_rope import LigerLlama4RopeFunction
+
+
+def liger_llama4_text_rotary_pos_emb(
+    xq: torch.Tensor,
+    xk: torch.Tensor,
+    freqs_cis: torch.Tensor,
+) -> tuple[torch.Tensor, torch.Tensor]:
+    """
+    Liger-optimized implementation of Llama4 text rotary position embedding.
+
+    This implementation uses a fused Triton kernel for complex multiplication,
+    providing significant performance improvements over the original PyTorch implementation.
+
+    Args:
+        xq (torch.Tensor): Query tensor of shape (batch_size, seq_len, num_heads, head_dim)
+        xk (torch.Tensor): Key tensor of shape (batch_size, seq_len, num_heads, head_dim)
+        freqs_cis (torch.Tensor): Complex frequency tensor from Llama4TextRotaryEmbedding
+
+    Returns:
+        Tuple[torch.Tensor, torch.Tensor]: Rotated query and key tensors
+    """
+    # Use fused Triton kernel for complex RoPE
+    return LigerLlama4RopeFunction.apply(xq, xk, freqs_cis)
+
+
+def liger_llama4_vision_rotary_pos_emb(
+    query: torch.Tensor,
+    key: torch.Tensor,
+    freqs_ci: torch.Tensor,
+) -> tuple[torch.Tensor, torch.Tensor]:
+    """
+    Liger-optimized implementation of Llama4 vision rotary position embedding.
+
+    This implementation uses the same fused Triton kernel as text RoPE,
+    providing performance improvements for vision transformer attention.
+
+    Args:
+        query (torch.Tensor): Query tensor of shape (batch_size, seq_len, num_heads, head_dim)
+        key (torch.Tensor): Key tensor of shape (batch_size, seq_len, num_heads, head_dim)
+        freqs_ci (torch.Tensor): Complex frequency tensor for 2D positions
+
+    Returns:
+        Tuple[torch.Tensor, torch.Tensor]: Rotated query and key tensors
+    """
+    # Handle broadcasting for vision RoPE
+    if freqs_ci.dim() == 3:
+        try:
+            # Try the regular 3D expansion
+            freqs_ci = freqs_ci.unsqueeze(0).expand(query.shape[0], -1, -1)
+        except RuntimeError as e:
+            if "expand" in str(e) and "4" in str(e):
+                # The tensor is actually 4D internally, handle it differently
+                freqs_ci = freqs_ci.squeeze(1)  # Remove the middle dimension
+                freqs_ci = freqs_ci.unsqueeze(0).expand(query.shape[0], -1, -1)
+            else:
+                raise e
+    elif freqs_ci.dim() == 4:  # (1, seq_len, 1, head_dim//2) - already properly shaped
+        # Squeeze the middle dimension to get (1, seq_len, head_dim//2)
+        freqs_ci = freqs_ci.squeeze(2)
+    elif freqs_ci.dim() == 2:  # (seq_len, head_dim//2) - needs expansion
+        freqs_ci = freqs_ci.unsqueeze(0).expand(query.shape[0], -1, -1)
+    else:
+        raise ValueError(f"Unexpected freqs_ci shape: {freqs_ci.shape}")
+
+    # Use the same fused kernel as text RoPE
+    return LigerLlama4RopeFunction.apply(query, key, freqs_ci)
+
+
+# Note: We only patch the functions, not the classes
+# The original Llama4TextRotaryEmbedding and Llama4VisionRotaryEmbedding classes remain unchanged
+
+
+# Convenience functions for monkey patching
+def apply_liger_llama4_rope_full(modeling_module):
+    """
+    Apply Liger optimizations to Llama4 RoPE functions.
+
+    Args:
+        modeling_module: The transformers modeling module to patch
+    """
+    # Replace the text RoPE function
+    modeling_module.apply_rotary_emb = liger_llama4_text_rotary_pos_emb
+
+    # Replace the vision RoPE function
+    modeling_module.vision_apply_rotary_emb = liger_llama4_vision_rotary_pos_emb

liger_kernel/transformers/model/falcon_h1.py

@@ -0,0 +1,122 @@
+from typing import TYPE_CHECKING
+from typing import Optional
+from typing import Union
+
+import torch
+
+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(
+    self,
+    input_ids: torch.LongTensor = None,
+    attention_mask: Optional[torch.Tensor] = None,
+    position_ids: Optional[torch.LongTensor] = None,
+    past_key_values: Optional["FalconHybridMambaAttentionDynamicCache"] = None,
+    inputs_embeds: Optional[torch.FloatTensor] = None,
+    labels: Optional[torch.LongTensor] = None,
+    use_cache: Optional[bool] = None,
+    output_attentions: Optional[bool] = None,
+    output_hidden_states: Optional[bool] = None,
+    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, 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, ...,
+        config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
+        (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
+
+    Example:
+
+    ```python
+    >>> from transformers import AutoTokenizer, FalconH1ForCausalLM
+
+    >>> model = FalconH1ForCausalLM.from_pretrained("...")
+    >>> tokenizer = AutoTokenizer.from_pretrained("...")
+
+    >>> prompt = "Hey, are you conscious? Can you talk to me?"
+    >>> inputs = tokenizer(prompt, return_tensors="pt")
+
+    >>> # Generate
+    >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
+    >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
+    "Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you."
+    ```"""
+    output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+    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(
+        input_ids=input_ids,
+        attention_mask=attention_mask,
+        position_ids=position_ids,
+        past_key_values=past_key_values,
+        inputs_embeds=inputs_embeds,
+        use_cache=use_cache,
+        output_attentions=output_attentions,
+        output_hidden_states=output_hidden_states,
+        cache_position=cache_position,
+        **kwargs,
+    )
+
+    hidden_states = outputs[0]
+    # Only compute necessary logits, and do not upcast them to float if we are not computing the loss
+    slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep
+    kept_hidden_states = hidden_states[:, slice_indices, :]
+
+    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")
+
+    if skip_logits is None:
+        # By default, if in training mode, don't materialize logits
+        skip_logits = self.training and labels is not None
+
+    # Compute loss
+    if skip_logits:
+        result = LigerForCausalLMLoss(
+            hidden_states=kept_hidden_states,
+            lm_head_weight=self.lm_head.weight,
+            labels=labels,
+            shift_labels=shift_labels,
+            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)
+
+    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(
@@ -27,6 +29,7 @@ def lce_forward_deprecated(
     output_hidden_states: Optional[bool] = None,
     return_dict: Optional[bool] = None,
     cache_position: Optional[torch.LongTensor] = None,
+    skip_logits: Optional[bool] = None,
 ) -> Union[Tuple, CausalLMOutputWithPast]:
     r"""
 
@@ -81,7 +84,14 @@ def lce_forward_deprecated(
     loss = None
     logits = None
 
-    if self.training and (labels is not None):
+    if skip_logits and labels is None:
+        raise ValueError("skip_logits is True, but labels is None")
+
+    if skip_logits is None:
+        # By default, if in training mode, don't materialize logits
+        skip_logits = self.training and labels is not None
+
+    if skip_logits:
         shift_hidden_states = hidden_states[..., :-1, :].contiguous()
         shift_labels = labels[..., 1:].contiguous()
 
@@ -139,7 +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*):
@@ -201,6 +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")
@@ -209,8 +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,
@@ -218,24 +230,32 @@ 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:
+        if labels is not None or shift_labels is not None:
             loss = self.loss_function(
                 logits=logits,
                 labels=labels,
+                shift_labels=shift_labels,
                 vocab_size=self.config.vocab_size,
                 **kwargs,
             )
 
     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__)
 
@@ -30,6 +32,7 @@ def lce_forward_deprecated(
     output_hidden_states: Optional[bool] = None,
     return_dict: Optional[bool] = None,
     cache_position: Optional[torch.LongTensor] = None,
+    skip_logits: Optional[bool] = None,
     **kwargs,
 ) -> Union[Tuple, CausalLMOutputWithPast]:
     r"""
@@ -85,7 +88,14 @@ def lce_forward_deprecated(
     loss = None
     logits = None
 
-    if self.training and (labels is not None):
+    if skip_logits and labels is None:
+        raise ValueError("skip_logits is True, but labels is None")
+
+    if skip_logits is None:
+        # By default, if in training mode, don't materialize logits
+        skip_logits = self.training and labels is not None
+
+    if skip_logits:
         shift_hidden_states = hidden_states[..., :-1, :].contiguous()
         shift_labels = labels[..., 1:].contiguous()
 
@@ -150,7 +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*):
@@ -217,6 +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")
@@ -225,8 +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,
@@ -235,6 +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)
@@ -244,17 +257,27 @@ def lce_forward(
             logits = logits * self.config.final_logit_softcapping
 
         loss = None
-        if labels is not None:
-            loss = self.loss_function(logits, labels, self.vocab_size, **kwargs)
+        if labels is not None or shift_labels is not None:
+            loss = self.loss_function(
+                logits=logits,
+                labels=labels,
+                shift_labels=shift_labels,
+                vocab_size=self.vocab_size,
+                **kwargs,
+            )
 
     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,
     )