liger-kernel-nightly 0.5.5.dev20250402185702__py3-none-any.whl → 0.6.4.dev20260112233432__py3-none-any.whl

liger_kernel/transformers/model/llava.py

@@ -5,19 +5,16 @@ from typing import Union
 
 import torch
 
-from transformers.models.llava.modeling_llava import _CONFIG_FOR_DOC
-from transformers.models.llava.modeling_llava import LLAVA_INPUTS_DOCSTRING
+from torch.nn import CrossEntropyLoss
 from transformers.models.llava.modeling_llava import LlavaCausalLMOutputWithPast
-from transformers.utils import add_start_docstrings_to_model_forward
 from transformers.utils import is_torchdynamo_compiling
-from transformers.utils import replace_return_docstrings
-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 LigerLlavaCausalLMOutputWithPast
 
 
-@add_start_docstrings_to_model_forward(LLAVA_INPUTS_DOCSTRING)
-@replace_return_docstrings(output_type=LlavaCausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
 def lce_forward_deprecated(
     self,
     input_ids: torch.LongTensor = None,
@@ -33,6 +30,11 @@ def lce_forward_deprecated(
     output_attentions: Optional[bool] = None,
     output_hidden_states: Optional[bool] = None,
     return_dict: Optional[bool] = None,
+    cache_position: Optional[torch.LongTensor] = None,
+    logits_to_keep: Union[int, torch.Tensor] = 0,
+    image_sizes: torch.Tensor = None,
+    skip_logits: Optional[bool] = None,
+    **lm_kwargs,
 ) -> Union[Tuple, LlavaCausalLMOutputWithPast]:
     r"""
     Args:
@@ -41,10 +43,12 @@ def lce_forward_deprecated(
             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]`.
 
-        num_logits_to_keep (`int`, *optional*):
-            Calculate logits for the last `num_logits_to_keep` tokens. If `0`, calculate logits for all
+        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:
@@ -70,7 +74,6 @@ def lce_forward_deprecated(
     >>> processor.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
     "USER:  \nWhat's the content of the image? ASSISTANT: The image features a busy city street with a stop sign prominently displayed"
     ```"""
-
     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
@@ -94,73 +97,24 @@ def lce_forward_deprecated(
         )
 
     if inputs_embeds is None:
-        # 1. Extra the input embeddings
         inputs_embeds = self.get_input_embeddings()(input_ids)
 
-        # 2. Merge text and images
-        if pixel_values is not None and input_ids.shape[1] != 1:
-            image_outputs = self.vision_tower(pixel_values, output_hidden_states=True)
-            # this is not memory efficient at all (output_hidden_states=True) will save all the hidden stated.
-            selected_image_feature = image_outputs.hidden_states[vision_feature_layer]
-
-            if vision_feature_select_strategy == "default":
-                selected_image_feature = selected_image_feature[:, 1:]
-            elif vision_feature_select_strategy == "full":
-                selected_image_feature = selected_image_feature
-            else:
-                raise ValueError(f"Unexpected select feature strategy: {self.config.vision_feature_select_strategy}")
-
-            image_features = self.multi_modal_projector(selected_image_feature)
-            inputs_embeds = inputs_embeds.to(image_features.dtype)
-            inputs_embeds, attention_mask, labels, position_ids = self._merge_input_ids_with_image_features(
-                image_features, inputs_embeds, input_ids, attention_mask, labels
-            )
-
-        # In case input_ids.shape[1] == 1 & pixel_values==None & past_key_values != None, we are in the case of
-        # generation with cache
-        elif past_key_values is not None and pixel_values is not None and input_ids.shape[1] == 1:
-            # Retrieve the first layer to inspect the logits and mask out the hidden states
-            # that are set to 0
-            first_layer_past_key_value = past_key_values[0][0][:, :, :, 0]
-
-            # Sum all dimensions of head_dim (-2) to avoid random errors such as: https://github.com/huggingface/transformers/pull/28032#issuecomment-1863691941
-            batch_index, non_attended_tokens = torch.where(first_layer_past_key_value.float().sum(-2) == 0)
-
-            # Get the target length
-            target_length = input_ids.shape[1]
-            past_length = first_layer_past_key_value.shape[-1]
-
-            extended_attention_mask = torch.ones(
-                (attention_mask.shape[0], past_length),
-                dtype=attention_mask.dtype,
-                device=attention_mask.device,
-            )
-
-            # Filter out only the tokens that can be un-attended, this can happen
-            # if one uses Llava + Fused modules where the cache on the
-            # first iteration is already big enough, or if one passes custom cache
-            valid_indices = non_attended_tokens < extended_attention_mask.size(-1)
-            new_batch_index = batch_index[valid_indices]
-            new_non_attended_tokens = non_attended_tokens[valid_indices]
-
-            # Zero-out the places where we don't need to attend
-            extended_attention_mask[new_batch_index, new_non_attended_tokens] = 0
-
-            attention_mask = torch.cat((extended_attention_mask, attention_mask[:, -target_length:]), dim=1)
-            position_ids = torch.sum(attention_mask, dim=1).unsqueeze(-1) - 1
-
-    # TODO: @raushan retain only the new behavior after v4.47
-    elif image_features is not None:
-        n_image_tokens = (input_ids == self.config.image_token_index).sum().item()
-        n_image_features = image_features.shape[0] * image_features.shape[1]
+    if pixel_values is not None:
+        image_features = self.get_image_features(
+            pixel_values=pixel_values,
+            vision_feature_layer=vision_feature_layer,
+            vision_feature_select_strategy=vision_feature_select_strategy,
+            image_sizes=image_sizes,
+        )
 
-        if n_image_tokens != n_image_features:
+        special_image_mask = (input_ids == self.config.image_token_index).unsqueeze(-1)
+        special_image_mask = special_image_mask.expand_as(inputs_embeds).to(inputs_embeds.device)
+        if not is_torchdynamo_compiling() and inputs_embeds[special_image_mask].numel() != image_features.numel():
+            n_image_tokens = (input_ids == self.config.image_token_index).sum()
+            n_image_features = image_features.shape[0] * image_features.shape[1]
             raise ValueError(
                 f"Image features and image tokens do not match: tokens: {n_image_tokens}, features {n_image_features}"
             )
-        special_image_mask = (
-            (input_ids == self.config.image_token_index).unsqueeze(-1).expand_as(inputs_embeds).to(inputs_embeds.device)
-        )
         image_features = image_features.to(inputs_embeds.device, inputs_embeds.dtype)
         inputs_embeds = inputs_embeds.masked_scatter(special_image_mask, image_features)
 
@@ -173,13 +127,19 @@ def lce_forward_deprecated(
         output_attentions=output_attentions,
         output_hidden_states=output_hidden_states,
         return_dict=return_dict,
+        cache_position=cache_position,
+        logits_to_keep=logits_to_keep,
+        **lm_kwargs,
     )
     hidden_states = outputs[0]
 
     loss = None
     logits = None
 
-    if self.training and (labels is not None):
+    # Overwrite skip_logits, since llava never materializes logits
+    skip_logits = labels is not None
+
+    if skip_logits:
         # Shift so that tokens < n predict n
         if attention_mask is not None:
             # we use the input attention mask to shift the logits and labels, because it is 2D.
@@ -194,7 +154,33 @@ def lce_forward_deprecated(
             shift_labels = labels[..., 1:].contiguous()
 
         lce = LigerFusedLinearCrossEntropyLoss()
-        loss = lce(self.language_model.lm_head.weight, shift_hidden_states, shift_labels)
+        loss = lce(
+            self.language_model.lm_head.weight,
+            shift_hidden_states.view(-1, shift_hidden_states.size(-1)),
+            shift_labels.view(-1).to(shift_hidden_states.device),
+        )
+    else:
+        logits = self.language_model.lm_head(hidden_states)
+        if labels is not None:
+            # Upcast to float if we need to compute the loss to avoid potential precision issues
+            logits = logits.float()
+            shift_logits = logits[..., :-1, :]
+            shift_labels = labels[..., 1:]
+            if attention_mask is not None:
+                # we use the input attention mask to shift the logits and labels, because it is 2D.
+                # we also crop attn mask in case it is longer, which happens in PrefixTuning with peft
+                shift_attention_mask = attention_mask[:, -shift_logits.shape[1] :].to(logits.device)
+                shift_logits = shift_logits[shift_attention_mask.to(logits.device) != 0].contiguous()
+                shift_labels = shift_labels[shift_attention_mask.to(shift_labels.device) != 0].contiguous()
+            else:
+                shift_logits = shift_logits.contiguous()
+                shift_labels = shift_labels.contiguous()
+            # Flatten the tokens
+            loss_fct = CrossEntropyLoss()
+
+            flat_logits = shift_logits.view(-1, self.config.text_config.vocab_size)
+            flat_labels = shift_labels.view(-1).to(shift_logits.device)
+            loss = loss_fct(flat_logits, flat_labels)
 
     if not return_dict:
         # NOTE: This part has not been tested.
@@ -207,12 +193,10 @@ def lce_forward_deprecated(
         past_key_values=outputs.past_key_values,
         hidden_states=outputs.hidden_states,
         attentions=outputs.attentions,
+        image_hidden_states=image_features if pixel_values is not None else None,
     )
 
 
-@add_start_docstrings_to_model_forward(LLAVA_INPUTS_DOCSTRING)
-@deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
-@replace_return_docstrings(output_type=LlavaCausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
 def lce_forward(
     self,
     input_ids: torch.LongTensor = None,
@@ -231,8 +215,9 @@ def lce_forward(
     cache_position: Optional[torch.LongTensor] = None,
     logits_to_keep: Union[int, torch.Tensor] = 0,
     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*):
@@ -285,85 +270,75 @@ def lce_forward(
         else self.config.vision_feature_select_strategy
     )
 
-    if (input_ids is None) ^ (inputs_embeds is not None):
-        raise ValueError("You must specify exactly one of input_ids or inputs_embeds")
-
-    if pixel_values is not None and inputs_embeds is not None:
-        raise ValueError(
-            "You cannot specify both pixel_values and inputs_embeds at the same time, and must specify either one"
-        )
-
-    if inputs_embeds is None:
-        inputs_embeds = self.get_input_embeddings()(input_ids)
-
-    if pixel_values is not None:
-        image_features = self.get_image_features(
-            pixel_values=pixel_values,
-            vision_feature_layer=vision_feature_layer,
-            vision_feature_select_strategy=vision_feature_select_strategy,
-            image_sizes=image_sizes,
-        )
-
-        special_image_mask = (input_ids == self.config.image_token_index).unsqueeze(-1)
-        special_image_mask = special_image_mask.expand_as(inputs_embeds).to(inputs_embeds.device)
-        if not is_torchdynamo_compiling() and inputs_embeds[special_image_mask].numel() != image_features.numel():
-            n_image_tokens = (input_ids == self.config.image_token_index).sum()
-            n_image_features = image_features.shape[0] * image_features.shape[1]
-            raise ValueError(
-                f"Image features and image tokens do not match: tokens: {n_image_tokens}, features {n_image_features}"
-            )
-        image_features = image_features.to(inputs_embeds.device, inputs_embeds.dtype)
-        inputs_embeds = inputs_embeds.masked_scatter(special_image_mask, image_features)
-
-    outputs = self.language_model.model(
+    outputs = self.model(
+        input_ids=input_ids,
+        pixel_values=pixel_values,
         attention_mask=attention_mask,
         position_ids=position_ids,
         past_key_values=past_key_values,
         inputs_embeds=inputs_embeds,
+        vision_feature_layer=vision_feature_layer,
+        vision_feature_select_strategy=vision_feature_select_strategy,
         use_cache=use_cache,
         output_attentions=output_attentions,
         output_hidden_states=output_hidden_states,
-        return_dict=return_dict,
+        return_dict=True,
         cache_position=cache_position,
-        logits_to_keep=logits_to_keep,
+        image_sizes=image_sizes,
         **lm_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, :]
 
-    loss = None
+    shift_labels = lm_kwargs.pop("shift_labels", None)
     logits = None
-
-    if self.training and (labels is not None):
-        # Shift so that tokens < n predict n
-        if attention_mask is not None:
-            # we use the input attention mask to shift the logits and labels, because it is 2D.
-            # we also crop attn mask in case it is longer, which happens in PrefixTuning with peft
-            shift_attention_mask = attention_mask[:, -(hidden_states.shape[1] - 1) :].to(hidden_states.device)
-            shift_hidden_states = hidden_states[..., :-1, :][
-                shift_attention_mask.to(hidden_states.device) != 0
-            ].contiguous()
-            shift_labels = labels[..., 1:][shift_attention_mask.to(labels.device) != 0].contiguous()
-        else:
-            shift_hidden_states = hidden_states[..., :-1, :].contiguous()
-            shift_labels = labels[..., 1:].contiguous()
-
-        lce = LigerFusedLinearCrossEntropyLoss()
-        loss = lce(
-            self.language_model.lm_head.weight,
-            shift_hidden_states.view(-1, shift_hidden_states.size(-1)),
-            shift_labels.view(-1).to(shift_hidden_states.device),
+    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)
+
+    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.text_config.hidden_size,
+            **lm_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.text_config.vocab_size,
+                **lm_kwargs,
+            )
 
     if not return_dict:
-        # NOTE: This part has not been tested.
-        output = outputs[1:]
-        return (loss,) + output if loss is not None else output
+        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 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=image_features if pixel_values is not None else None,
+        image_hidden_states=outputs.image_hidden_states,
+        token_accuracy=token_accuracy,
     )

liger_kernel/transformers/model/loss_utils.py

@@ -1,28 +1,61 @@
+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,
-    lm_head_weight,
-    target,
-    num_items_in_batch: int = None,
+    hidden_states: torch.Tensor,
+    lm_head_weight: torch.Tensor,
+    target: torch.Tensor,
+    num_items_in_batch: Optional[int] = None,
     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,
         reduction=reduction,
         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
 
 
@@ -31,27 +64,32 @@ def LigerForCausalLMLoss(
     lm_head_weight,
     labels,
     hidden_size: int,
-    num_items_in_batch: int = None,
+    num_items_in_batch: Optional[int] = None,
     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
-    labels = labels.to(hidden_states.device)
-    # Shift so that token < n predict n
-    labels = nn.functional.pad(labels, (0, 1), value=ignore_index)
-    shift_labels = labels[..., 1:].contiguous()
+    if shift_labels is None:
+        # Shift so that token < n predict n
+        labels = nn.functional.pad(labels, (0, 1), value=ignore_index)
+        shift_labels = labels[..., 1:].contiguous()
 
     # Flatten the tokens
     hidden_states = hidden_states.view(-1, hidden_size)
     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

liger_kernel/transformers/model/mistral.py

@@ -5,19 +5,15 @@ from typing import Union
 
 import torch
 
-from torch.nn import CrossEntropyLoss
 from transformers.cache_utils import Cache
-from transformers.modeling_outputs import CausalLMOutputWithPast
-from transformers.models.mistral.modeling_mistral import _CONFIG_FOR_DOC
-from transformers.models.mistral.modeling_mistral import MISTRAL_INPUTS_DOCSTRING
-from transformers.utils import add_start_docstrings_to_model_forward
-from transformers.utils import replace_return_docstrings
+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
 
 
-@add_start_docstrings_to_model_forward(MISTRAL_INPUTS_DOCSTRING)
-@replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
+@deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
 def lce_forward(
     self,
     input_ids: torch.LongTensor = None,
@@ -31,8 +27,10 @@ def lce_forward(
     output_hidden_states: Optional[bool] = None,
     return_dict: Optional[bool] = None,
     cache_position: Optional[torch.LongTensor] = None,
-    **loss_kwargs,
-) -> Union[Tuple, CausalLMOutputWithPast]:
+    logits_to_keep: Union[int, torch.Tensor] = 0,
+    skip_logits: Optional[bool] = None,
+    **kwargs,
+) -> Union[Tuple, LigerCausalLMOutputWithPast]:
     r"""
     Copy paste Mistral's forward but replace torch cross entropy with liger fused linear cross entropy
 
@@ -43,6 +41,12 @@ def lce_forward(
             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:
@@ -80,49 +84,62 @@ def lce_forward(
         output_hidden_states=output_hidden_states,
         return_dict=return_dict,
         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)
     loss = None
     logits = None
+    token_accuracy = None
 
-    if self.training and (labels is not None):
-        loss = LigerForCausalLMLoss(
-            hidden_states=hidden_states,
+    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:
+        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,
-            **loss_kwargs,
+            **kwargs,
         )
+        loss, _, token_accuracy = unpack_cross_entropy_result(result)
 
     else:
-        logits = self.lm_head(hidden_states)
-        if labels is not None:
-            # Upcast to float if we need to compute the loss to avoid potential precision issues
-            logits = logits.float()
-            # Shift so that tokens < n predict n
-            shift_logits = logits[..., :-1, :].contiguous()
-            shift_labels = labels[..., 1:].contiguous()
-            # Flatten the tokens
-            shift_logits = shift_logits.view(-1, self.config.vocab_size)
-            shift_labels = shift_labels.view(-1)
-            # Ensure tensors are on the same device
-            shift_labels = shift_labels.to(shift_logits.device)
-            loss_fct = CrossEntropyLoss()
-            loss = loss_fct(shift_logits, shift_labels)
+        logits = self.lm_head(kept_hidden_states)
+
+        loss = 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
+        output_tuple = (logits,) + outputs[1:]
+        output = (loss,) + output_tuple if loss is not None else output_tuple
+        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,
     )
-
-
-# Note: Grad Acc is not fixed in mistral at transformer 4.46.1