liger-kernel-nightly 0.6.2.dev20251011154427__py3-none-any.whl → 0.6.4.dev20260107111351__py3-none-any.whl

liger_kernel/transformers/model/gpt_oss.py

@@ -0,0 +1,211 @@
+from typing import List
+from typing import Optional
+from typing import Union
+
+import torch
+
+from transformers.modeling_outputs import MoeModelOutputWithPast
+from transformers.models.mixtral.modeling_mixtral import load_balancing_loss_func
+
+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 LigerMoeCausalLMOutputWithPast
+
+
+def lce_forward(
+    self,
+    input_ids: Optional[torch.LongTensor] = None,
+    attention_mask: Optional[torch.Tensor] = None,
+    position_ids: Optional[torch.LongTensor] = None,
+    past_key_values: Optional[List[torch.FloatTensor]] = 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,
+    output_router_logits: Optional[bool] = None,
+    cache_position: Optional[torch.LongTensor] = None,
+    logits_to_keep: Union[int, torch.Tensor] = 0,
+    skip_logits: Optional[bool] = None,
+    **kwargs,
+) -> LigerMoeCausalLMOutputWithPast:
+    r"""
+        Forward pass for causal language modeling with Mixture of Experts (MoE) architecture using Liger Kernel optimizations.
+
+    Args:
+        input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Indices of input sequence tokens in the vocabulary. Indices can be obtained using tokenizers.
+        attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+        position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Indices of positions of each input sequence tokens in the position embeddings.
+        past_key_values (`List[torch.FloatTensor]` or `Cache`, *optional*):
+            Pre-computed hidden-states (key and values in the self-attention blocks) that can be used to speed up
+            sequential decoding. See `past_key_values` input for more details.
+        inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+            Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation.
+            This is useful if you want more control over how to convert `input_ids` indices into associated vectors
+            than the model's internal embedding lookup matrix.
+        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]`.
+        use_cache (`bool`, *optional*):
+            If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding
+            (see `past_key_values`).
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        output_router_logits (`bool`, *optional*):
+            Whether or not to return the router logits of all MoE layers. See `router_logits` under returned tensors
+            for more detail.
+        cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*):
+            Indices depicting the position of the input sequence tokens in the sequence.
+        logits_to_keep (`int` or `torch.Tensor`, *optional*, defaults to 0):
+            If an `int`, compute logits for the last `logits_to_keep` tokens. If `0`, calculate logits for all
+            `input_ids` (special case). Only last token logits are needed for generation, and calculating them only for that
+            token can save memory, which becomes pretty significant for long sequences or large vocabulary size.
+            If a `torch.Tensor`, must be 1D corresponding to the indices to keep in the sequence length dimension.
+            This is useful when using packed tensor format (single dimension for batch and sequence length).
+        skip_logits (`bool`, *optional*):
+            Whether to skip logit computation and directly compute loss. If `None`, defaults to `True` during training
+            when labels are provided (to save memory), and `False` during inference.
+
+    Returns:
+        `LigerMoeCausalLMOutputWithPast`: An output object containing:
+            - loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` is provided):
+                Language modeling loss (for next-token prediction), including the auxiliary load balancing loss.
+            - aux_loss (`torch.FloatTensor`, *optional*, returned when `labels` is provided):
+                Auxiliary load balancing loss for the sparse MoE modules.
+            - logits (`torch.FloatTensor` of shape `(batch_size, sequence_length, config.vocab_size)`, *optional*):
+                Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
+                Note: logits are `None` during training when `skip_logits=True` to save memory.
+            - past_key_values (`Cache`, *optional*, returned when `use_cache=True` is passed):
+                Cached key and value projection states for faster sequential decoding.
+            - hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True`):
+                Tuple of `torch.FloatTensor` (one for the output of the embeddings + one for each layer) of shape
+                `(batch_size, sequence_length, hidden_size)`. Hidden-states of the model at the output of each layer.
+            - attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True`):
+                Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
+                sequence_length)`. Attentions weights after the attention softmax.
+            - router_logits (`tuple(torch.FloatTensor)`, *optional*, returned when `output_router_logits=True`):
+                Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, sequence_length, num_experts)`.
+                Router logits of the MoE layers, useful to compute the auxiliary loss and z_loss.
+            - token_accuracy (`torch.FloatTensor`, *optional*, returned when `labels` is provided):
+                Token-level prediction accuracy.
+
+    Example:
+
+    ```python
+        >>> from transformers import AutoTokenizer, GptOssForCausalLM
+        >>> from liger_kernel.transformers import apply_liger_kernel_to_gpt_oss
+
+        >>> # Apply Liger Kernel patches for optimized performance
+        >>> apply_liger_kernel_to_gpt_oss()
+
+        >>> model = GptOssForCausalLM.from_pretrained("openai/gpt-oss-20b")
+        >>> tokenizer = AutoTokenizer.from_pretrained("openai/gpt-oss-20b")
+
+        >>> prompt = "Hey, are you conscious? Can you talk to me?"
+        >>> inputs = tokenizer(prompt, return_tensors="pt")
+
+        >>> # Inference: Forward pass returns logits
+        >>> outputs = model(**inputs)
+        >>> outputs.logits.shape
+        torch.Size([1, 12, 201088])
+
+        >>> # Get next token prediction
+        >>> next_token_logits = outputs.logits[:, -1, :]
+        >>> predicted_token_id = next_token_logits.argmax(dim=-1)
+
+        >>> # Training: Forward pass with labels returns loss
+        >>> labels = inputs.input_ids.clone()
+        >>> outputs = model(**inputs, labels=labels)
+        >>> outputs.loss
+        tensor(2.6454)
+    ```"""
+
+    output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+    output_router_logits = (
+        output_router_logits if output_router_logits is not None else self.config.output_router_logits
+    )
+
+    output_hidden_states = (
+        output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+    )
+
+    # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
+    outputs: MoeModelOutputWithPast = 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,
+        output_router_logits=output_router_logits,
+        cache_position=cache_position,
+        **kwargs,
+    )
+
+    hidden_states = outputs.last_hidden_state
+    # 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 skip_logits is None:
+        skip_logits = self.training and (labels is not None or shift_labels is not None)
+
+    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:  # if in inference model materialize logits
+        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,
+                shift_labels=shift_labels,
+                vocab_size=self.vocab_size,
+                **kwargs,
+            )
+
+    aux_loss = None
+    if output_router_logits:
+        aux_loss = load_balancing_loss_func(
+            outputs.router_logits,
+            self.num_experts,
+            self.num_experts_per_tok,
+            attention_mask,
+        )
+        if labels is not None:
+            loss += self.router_aux_loss_coef * aux_loss.to(loss.device)  # make sure to reside in the same device
+
+    return LigerMoeCausalLMOutputWithPast(
+        loss=loss,
+        aux_loss=aux_loss,
+        logits=logits,
+        past_key_values=outputs.past_key_values,
+        hidden_states=outputs.hidden_states,
+        attentions=outputs.attentions,
+        router_logits=outputs.router_logits,
+        token_accuracy=token_accuracy,
+    )

liger_kernel/transformers/model/hunyuan_v1.py

@@ -0,0 +1,134 @@
+from typing import List
+from typing import Optional
+from typing import Union
+
+import torch
+
+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: Optional[torch.LongTensor] = None,
+    attention_mask: Optional[torch.Tensor] = None,
+    position_ids: Optional[torch.LongTensor] = None,
+    past_key_values: Optional[List[torch.FloatTensor]] = 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,
+) -> 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]`.
+
+        logits_to_keep (`int` or `torch.Tensor`, *optional*):
+            If an `int`, compute logits for the last `logits_to_keep` tokens. If `0`, calculate logits for all
+            `input_ids` (special case). Only last token logits are needed for generation, and calculating them only for that
+            token can save memory, which becomes pretty significant for long sequences or large vocabulary size.
+            If a `torch.Tensor`, must be 1D corresponding to the indices to keep in the sequence length dimension.
+            This is useful when using packed tensor format (single dimension for batch and sequence length).
+
+    Returns:
+
+    Example:
+
+    ```python
+    >>> from transformers import AutoTokenizer, HunYuanDenseV1ForCausalLM
+
+    >>> model = HunYuanDenseV1ForCausalLM.from_pretrained("meta-hunyuan_v1_dense/HunYuanDenseV1-2-7b-hf")
+    >>> tokenizer = AutoTokenizer.from_pretrained("meta-hunyuan_v1_dense/HunYuanDenseV1-2-7b-hf")
+
+    >>> 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 skip_logits and labels is None and shift_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 or shift_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,
+                shift_labels=shift_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 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/internvl.py

@@ -5,10 +5,11 @@ from typing import Union
 
 import torch
 
-from transformers.models.internvl.modeling_internvl import InternVLCausalLMOutputWithPast
 from transformers.utils import can_return_tuple
 
 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 LigerInternVLCausalLMOutputWithPast
 
 
 # Copied from https://github.com/huggingface/transformers/blob/d888bd435d0c0eaabaabad5b33d52af518c7187c/src/transformers/models/internvl/modeling_internvl.py#L862
@@ -33,7 +34,7 @@ def lce_forward(
     image_sizes: Optional[torch.Tensor] = None,
     skip_logits: Optional[bool] = None,  # Added argument for liger-kernel
     **lm_kwargs,  # renamed from kwargs
-) -> Union[Tuple, InternVLCausalLMOutputWithPast]:
+) -> Union[Tuple, LigerInternVLCausalLMOutputWithPast]:
     r"""
     Example:
 
@@ -111,6 +112,7 @@ def lce_forward(
     shift_labels = lm_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")
@@ -120,7 +122,7 @@ def lce_forward(
         skip_logits = self.training and (labels is not None or shift_labels is not None)
 
     if skip_logits:
-        loss = LigerForCausalLMLoss(
+        result = LigerForCausalLMLoss(
             hidden_states=kept_hidden_states,
             lm_head_weight=self.lm_head.weight,
             labels=labels,
@@ -128,6 +130,7 @@ def lce_forward(
             hidden_size=self.config.text_config.hidden_size,
             **lm_kwargs,
         )
+        loss, _, token_accuracy = unpack_cross_entropy_result(result)
 
     else:
         logits = self.lm_head(kept_hidden_states)
@@ -138,13 +141,17 @@ def lce_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 InternVLCausalLMOutputWithPast(
+    # Return custom output class with token_accuracy field
+    return LigerInternVLCausalLMOutputWithPast(
         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/llama.py

@@ -15,6 +15,8 @@ from transformers.utils.deprecation import deprecate_kwarg
 from liger_kernel.transformers.fsdp import _FSDPForwardRedirection
 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.utils import PEFT_AVAILABLE
 
 if TYPE_CHECKING:
@@ -162,7 +164,7 @@ 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*):
@@ -228,6 +230,8 @@ 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 and shift_labels is None:
         raise ValueError("skip_logits is True, but labels and shift_labels are None")
@@ -236,8 +240,9 @@ 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 = lce_maybe_trainable_lm_head(
+        result = lce_maybe_trainable_lm_head(
             self,
             hidden_states=kept_hidden_states,
             hidden_size=self.config.hidden_size,
@@ -245,7 +250,7 @@ def lce_forward(
             shift_labels=shift_labels,
             **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:
@@ -259,14 +264,18 @@ def lce_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 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/llama4.py

@@ -6,9 +6,10 @@ from typing import Union
 import torch
 
 from transformers.cache_utils import Cache
-from transformers.modeling_outputs import CausalLMOutputWithPast
 
 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(
@@ -26,7 +27,7 @@ def lce_forward(
     cache_position: Optional[torch.LongTensor] = None,
     logits_to_keep: Union[int, torch.Tensor] = 0,
     **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, ...,
@@ -78,9 +79,11 @@ def lce_forward(
     shift_labels = kwargs.pop("shift_labels", None)
     logits = None
     loss = None
+    token_accuracy = None
 
+    # Compute loss
     if self.training and (labels is not None or shift_labels is not None):
-        loss = LigerForCausalLMLoss(
+        result = LigerForCausalLMLoss(
             hidden_states=kept_hidden_states,
             lm_head_weight=self.lm_head.weight,
             labels=labels,
@@ -88,6 +91,7 @@ def lce_forward(
             hidden_size=self.config.hidden_size,
             **kwargs,
         )
+        loss, _, token_accuracy = unpack_cross_entropy_result(result)
 
     else:  # if in inference mode materialize logits
         logits = self.lm_head(kept_hidden_states)
@@ -100,10 +104,18 @@ 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,
     )

liger_kernel/transformers/model/llava.py

@@ -11,6 +11,8 @@ from transformers.utils import is_torchdynamo_compiling
 
 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 LigerLlavaCausalLMOutputWithPast
 
 
 def lce_forward_deprecated(
@@ -215,7 +217,7 @@ def lce_forward(
     image_sizes: torch.Tensor = None,
     skip_logits: Optional[bool] = None,
     **lm_kwargs,
-) -> Union[Tuple, LlavaCausalLMOutputWithPast]:
+) -> Union[Tuple, LigerLlavaCausalLMOutputWithPast]:
     r"""
     Args:
         labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
@@ -293,6 +295,7 @@ def lce_forward(
     shift_labels = lm_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")
@@ -302,7 +305,7 @@ def lce_forward(
         skip_logits = self.training and (labels is not None or shift_labels is not None)
 
     if skip_logits:
-        loss = LigerForCausalLMLoss(
+        result = LigerForCausalLMLoss(
             hidden_states=kept_hidden_states,
             lm_head_weight=self.lm_head.weight,
             labels=labels,
@@ -310,6 +313,7 @@ def lce_forward(
             hidden_size=self.config.text_config.hidden_size,
             **lm_kwargs,
         )
+        loss, _, token_accuracy = unpack_cross_entropy_result(result)
 
     else:
         logits = self.lm_head(kept_hidden_states)
@@ -324,13 +328,17 @@ def lce_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 LlavaCausalLMOutputWithPast(
+    # Return custom output class with token_accuracy field
+    return LigerLlavaCausalLMOutputWithPast(
         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/loss_utils.py

@@ -1,10 +1,28 @@
 from typing import Optional
+from typing import Tuple
 
 import torch
 import torch.nn as nn
 
 import liger_kernel.transformers.functional as F
 
+from liger_kernel.transformers.functional import CrossEntropyOutput
+
+
+def unpack_cross_entropy_result(
+    result,
+) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[torch.Tensor]]:
+    if isinstance(result, CrossEntropyOutput):
+        return result.loss, result.z_loss, result.token_accuracy
+
+    if isinstance(result, tuple):
+        loss = result[0]
+        z_loss = result[1] if len(result) > 1 else None
+        token_accuracy = result[2] if len(result) > 2 else None
+        return loss, z_loss, token_accuracy
+
+    return result, None, None
+
 
 def fixed_fused_linear_cross_entropy(
     hidden_states: torch.Tensor,
@@ -14,10 +32,11 @@ def fixed_fused_linear_cross_entropy(
     ignore_index: int = -100,
     final_logit_softcapping: Optional[float] = None,
     accum_dtype: Optional[torch.dtype] = None,
+    return_token_accuracy: bool = False,
     **kwargs,
 ):
     reduction = "sum" if num_items_in_batch is not None else "mean"
-    loss = F.liger_fused_linear_cross_entropy(
+    result = F.liger_fused_linear_cross_entropy(
         hidden_states,
         lm_head_weight,
         target,
@@ -25,11 +44,18 @@ def fixed_fused_linear_cross_entropy(
         ignore_index=ignore_index,
         softcap=final_logit_softcapping,
         accum_dtype=accum_dtype,
+        return_token_accuracy=return_token_accuracy,
         **kwargs,
     )
+
+    loss, _, token_accuracy = unpack_cross_entropy_result(result)
+
     if reduction == "sum":
         loss = loss / num_items_in_batch
 
+    if return_token_accuracy:
+        return CrossEntropyOutput(loss=loss, token_accuracy=token_accuracy)
+
     return loss
 
 
@@ -42,6 +68,7 @@ def LigerForCausalLMLoss(
     ignore_index: int = -100,
     shift_labels: Optional[torch.Tensor] = None,
     final_logit_softcapping: Optional[float] = None,
+    return_token_accuracy: bool = False,
     **kwargs,
 ):
     # Skip upcast since intermediate values for the loss are all fp32 in kernel
@@ -55,13 +82,14 @@ def LigerForCausalLMLoss(
     shift_labels = shift_labels.view(-1)
     # Enable model parallelism
     shift_labels = shift_labels.to(hidden_states.device)
-    loss = fixed_fused_linear_cross_entropy(
+    result = fixed_fused_linear_cross_entropy(
         hidden_states,
         lm_head_weight,
         shift_labels,
         num_items_in_batch,
         ignore_index,
         final_logit_softcapping,
+        return_token_accuracy=return_token_accuracy,
         **kwargs,
     )
-    return loss
+    return result