liger-kernel-nightly 0.4.0.dev20241107052928__py3-none-any.whl → 0.6.3.dev20251121010306__py3-none-any.whl

liger_kernel/transformers/model/glm4v_moe.py

@@ -0,0 +1,172 @@
+from typing import Optional
+from typing import Tuple
+from typing import Union
+
+import torch
+
+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 LigerGlm4vMoeCausalLMOutputWithPast
+
+
+@deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
+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[list[torch.FloatTensor]] = None,
+    inputs_embeds: Optional[torch.FloatTensor] = None,
+    labels: Optional[torch.LongTensor] = None,
+    pixel_values: Optional[torch.Tensor] = None,
+    pixel_values_videos: Optional[torch.FloatTensor] = None,
+    image_grid_thw: Optional[torch.LongTensor] = None,
+    video_grid_thw: Optional[torch.LongTensor] = None,
+    rope_deltas: Optional[torch.LongTensor] = 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, LigerGlm4vMoeCausalLMOutputWithPast]:
+    r"""
+    Args:
+        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]`.
+        image_grid_thw (`torch.LongTensor` of shape `(num_images, 3)`, *optional*):
+            The temporal, height and width of feature shape of each image in LLM.
+        video_grid_thw (`torch.LongTensor` of shape `(num_videos, 3)`, *optional*):
+            The temporal, height and width of feature shape of each video in LLM.
+        rope_deltas (`torch.LongTensor` of shape `(batch_size, )`, *optional*):
+            The rope index difference between sequence length and multimodal rope.
+
+
+        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).
+
+    Example:
+
+    ```python
+    >>> from transformers import AutoProcessor, Glm4vMoeForConditionalGeneration
+    >>> import torch
+
+    >>> MODEL_PATH = "zai-org/GLM-4.5V"
+    >>> messages = [
+        {
+            "role": "user",
+            "content": [
+                {
+                    "type": "image",
+                    "url": "https://upload.wikimedia.org/wikipedia/commons/f/fa/Grayscale_8bits_palette_sample_image.png"
+                },
+                {
+                    "type": "text",
+                    "text": "describe this image"
+                }
+            ],
+        }
+    ]
+    >>> processor = AutoProcessor.from_pretrained(MODEL_PATH)
+    >>> model = Glm4vMoeForConditionalGeneration.from_pretrained(
+        pretrained_model_name_or_path=MODEL_PATH,
+        dtype="auto",
+        device_map="auto",
+    )
+    >>> inputs = processor.apply_chat_template(
+        messages,
+        tokenize=True,
+        add_generation_prompt=True,
+        return_dict=True,
+        return_tensors="pt"
+    ).to(model.device)
+    >>> inputs.pop("token_type_ids", None)
+    >>> generated_ids = model.generate(**inputs, max_new_tokens=8192)
+    >>> output_text = processor.decode(generated_ids[0][inputs["input_ids"].shape[1]:], skip_special_tokens=False)
+    ```
+    """
+    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,
+        pixel_values=pixel_values,
+        pixel_values_videos=pixel_values_videos,
+        image_grid_thw=image_grid_thw,
+        video_grid_thw=video_grid_thw,
+        position_ids=position_ids,
+        attention_mask=attention_mask,
+        past_key_values=past_key_values,
+        inputs_embeds=inputs_embeds,
+        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
+
+    # Build output kwargs and include aux_loss only if present (depends on transformers version)
+    output_kwargs = dict(
+        loss=loss,
+        logits=logits,
+        past_key_values=outputs.past_key_values,
+        hidden_states=outputs.hidden_states,
+        attentions=outputs.attentions,
+        rope_deltas=outputs.rope_deltas,
+        token_accuracy=token_accuracy,
+    )
+    if hasattr(outputs, "aux_loss"):
+        output_kwargs["aux_loss"] = outputs.aux_loss
+
+    # Return GLM4V MoE output with accuracy
+    return LigerGlm4vMoeCausalLMOutputWithPast(**output_kwargs)

liger_kernel/transformers/model/internvl.py

@@ -0,0 +1,157 @@
+from typing import List
+from typing import Optional
+from typing import Tuple
+from typing import Union
+
+import torch
+
+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
+@can_return_tuple
+def lce_forward(
+    self,
+    input_ids: torch.LongTensor = None,
+    pixel_values: Optional[torch.FloatTensor] = 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,
+    vision_feature_layer: Optional[Union[int, List[int]]] = None,
+    vision_feature_select_strategy: Optional[str] = None,
+    labels: Optional[torch.LongTensor] = None,
+    use_cache: Optional[bool] = None,
+    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: Optional[torch.Tensor] = None,
+    skip_logits: Optional[bool] = None,  # Added argument for liger-kernel
+    **lm_kwargs,  # renamed from kwargs
+) -> Union[Tuple, LigerInternVLCausalLMOutputWithPast]:
+    r"""
+    Example:
+
+    ```python
+    >>> import torch
+    >>> from transformers import AutoProcessor, AutoModelForImageTextToText
+
+    >>> torch_device = "cuda"
+    >>> processor = AutoProcessor.from_pretrained("OpenGVLab/InternVL3-1B-hf")
+    >>> model = AutoModelForImageTextToText.from_pretrained(
+    ...     "OpenGVLab/InternVL3-1B-hf", dtype=torch.bfloat16, device_map=torch_device
+    ... )
+
+    >>> messages = [
+    ...     {
+    ...         "role": "user",
+    ...         "content": [
+    ...             {
+    ...                 "type": "image",
+    ...                 "url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg",
+    ...             },
+    ...             {
+    ...                 "type": "image",
+    ...                 "url": "https://thumbs.dreamstime.com/b/golden-gate-bridge-san-francisco-purple-flowers-california-echium-candicans-36805947.jpg",
+    ...             },
+    ...             {"type": "text", "text": "These images depict two different landmarks. Can you identify them?"},
+    ...         ],
+    ...     },
+    ... ]
+
+    >>> inputs = processor.apply_chat_template(messages, add_generation_prompt=True, tokenize=True, return_dict=True, return_tensors="pt").to(torch_device)
+    >>> generate_ids = model.generate(**inputs, max_new_tokens=200)
+    >>> print(processor.decode(generate_ids[0, inputs["input_ids"].shape[1] :], skip_special_tokens=True))
+    The images depict the Statue of Liberty and the Golden Gate Bridge.
+    ```"""
+
+    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
+    vision_feature_layer = (
+        vision_feature_layer if vision_feature_layer is not None else self.config.vision_feature_layer
+    )
+    vision_feature_select_strategy = (
+        vision_feature_select_strategy
+        if vision_feature_select_strategy is not None
+        else self.config.vision_feature_select_strategy
+    )
+
+    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,
+        cache_position=cache_position,
+        image_sizes=image_sizes,
+        **lm_kwargs,
+    )
+
+    # Copied from llava.py
+    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 = 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")
+
+    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:
+            loss = self.loss_function(
+                logits=logits, labels=labels, vocab_size=self.config.text_config.vocab_size, **lm_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 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

@@ -1,30 +1,31 @@
-from typing import TYPE_CHECKING, List, Optional, Tuple, Union
+from typing import TYPE_CHECKING
+from typing import List
+from typing import Optional
+from typing import Tuple
+from typing import Union
 
 import torch
 import torch.nn.functional as F
+
+from torch.distributed.fsdp import FullyShardedDataParallel
 from torch.nn import CrossEntropyLoss
 from transformers.modeling_outputs import CausalLMOutputWithPast
-from transformers.models.llama.modeling_llama import (
-    _CONFIG_FOR_DOC,
-    LLAMA_INPUTS_DOCSTRING,
-)
-from transformers.utils import (
-    add_start_docstrings_to_model_forward,
-    replace_return_docstrings,
-)
-
-from liger_kernel.transformers.fused_linear_cross_entropy import (
-    LigerFusedLinearCrossEntropyLoss,
-)
+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:
     from transformers.cache_utils import Cache
 
+if PEFT_AVAILABLE:
+    from peft.utils.other import ModulesToSaveWrapper
+
 
-@add_start_docstrings_to_model_forward(LLAMA_INPUTS_DOCSTRING)
-@replace_return_docstrings(
-    output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC
-)
 def lce_forward_deprecated(
     self,
     input_ids: torch.LongTensor = None,
@@ -38,6 +39,7 @@ 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"""
     Copy paste llama forward but replace torch cross entropy with liger fused linear cross entropy
@@ -67,19 +69,11 @@ def lce_forward_deprecated(
     >>> 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_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
+        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(
@@ -100,7 +94,15 @@ def lce_forward_deprecated(
     loss = None
     logits = None
 
-    if self.training and (labels is not None):
+    # if in training mode, don't materialize logits
+    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()
 
@@ -113,13 +115,8 @@ def lce_forward_deprecated(
 
     else:
         if self.config.pretraining_tp > 1:
-            lm_head_slices = self.lm_head.weight.split(
-                self.vocab_size // self.config.pretraining_tp, dim=0
-            )
-            logits = [
-                F.linear(hidden_states, lm_head_slices[i])
-                for i in range(self.config.pretraining_tp)
-            ]
+            lm_head_slices = self.lm_head.weight.split(self.vocab_size // self.config.pretraining_tp, dim=0)
+            logits = [F.linear(hidden_states, lm_head_slices[i]) for i in range(self.config.pretraining_tp)]
             logits = torch.cat(logits, dim=-1)
         else:
             logits = self.lm_head(hidden_states)
@@ -150,10 +147,7 @@ def lce_forward_deprecated(
     )
 
 
-@add_start_docstrings_to_model_forward(LLAMA_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,
@@ -167,9 +161,10 @@ def lce_forward(
     output_hidden_states: Optional[bool] = None,
     return_dict: Optional[bool] = None,
     cache_position: Optional[torch.LongTensor] = None,
-    num_logits_to_keep: int = 0,
-    **loss_kwargs,
-) -> Union[Tuple, CausalLMOutputWithPast]:
+    logits_to_keep: Union[int, torch.Tensor] = 0,
+    skip_logits: Optional[bool] = None,
+    **kwargs,
+) -> Union[Tuple, LigerCausalLMOutputWithPast]:
     r"""
     Args:
         labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
@@ -177,10 +172,12 @@ 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]`.
 
-        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:
 
@@ -201,19 +198,11 @@ def lce_forward(
     "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_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
+        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(
@@ -227,51 +216,111 @@ 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, :]
 
     if self.config.pretraining_tp > 1:
         raise Exception("Liger Kernel does not support pretraining_tp!!")
 
+    shift_labels = kwargs.pop("shift_labels", None)
     logits = None
     loss = None
-    # if in training mode, don't materialize logits
-    if self.training and (labels is not None):
-        # We do the same thing as ForCausalLMLoss but using Liger FLCE
-
-        shift_hidden_states = hidden_states[..., :-1, :].contiguous()
-        shift_labels = labels[..., 1:].contiguous()
-
-        # flatten tokens
-        shift_hidden_states = shift_hidden_states.view(-1, self.config.hidden_size)
-        shift_labels = shift_labels.view(-1)
+    token_accuracy = None
 
-        reduction = "sum" if "num_items_in_batch" in loss_kwargs else "mean"
-        lce = LigerFusedLinearCrossEntropyLoss(reduction=reduction)
-
-        loss = lce(self.lm_head.weight, shift_hidden_states, shift_labels)
-        if reduction == "sum":
-            loss /= loss_kwargs["num_items_in_batch"]
-
-    else:  # if in inference mode materialize logits
-        logits = self.lm_head(hidden_states[:, -num_logits_to_keep:, :])
-        if labels is not 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")
+
+    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 = lce_maybe_trainable_lm_head(
+            self,
+            hidden_states=kept_hidden_states,
+            hidden_size=self.config.hidden_size,
+            labels=labels,
+            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:
             loss = self.loss_function(
                 logits=logits,
                 labels=labels,
+                shift_labels=shift_labels,
                 vocab_size=self.config.vocab_size,
-                **loss_kwargs,
+                **kwargs,
             )
 
     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,
+    )
+
+
+def lce_maybe_trainable_lm_head(self, hidden_states, hidden_size, labels, shift_labels, **loss_kwargs):
+    lm_head = self.lm_head
+
+    # Unwrap the module if lm_head has been added as trainable module in PEFT LoRA configuration,
+    # i.e. listed in the modules_to_save field of LoraConfig, so the lm_head weights are read
+    # from the unwrapped module.
+    # See https://huggingface.co/docs/peft/package_reference/lora for reference.
+    if PEFT_AVAILABLE and isinstance(lm_head, ModulesToSaveWrapper):
+        lm_head = lm_head.modules_to_save.default
+
+    # If FSDP is used and lm_head is trainable, e.g., during full fine-tuning or with LoRA,
+    # reading the lm_head module weights and calling the kernel must be done within FSDP forward pass
+    # so the module entire parameters are summoned and kept in memory during the kernel execution.
+    if isinstance(lm_head, FullyShardedDataParallel):
+        return _FSDPForwardRedirection()(
+            lm_head,
+            _liger_for_causal_lm_loss,
+            lm_head.module,
+            hidden_states,
+            hidden_size,
+            labels,
+            shift_labels,
+            **loss_kwargs,
+        )
+
+    # FSDP is not used so we can read the lm_head weights and call the kernel directly
+    return _liger_for_causal_lm_loss(
+        lm_head=self.lm_head,
+        hidden_states=hidden_states,
+        hidden_size=hidden_size,
+        labels=labels,
+        shift_labels=shift_labels,
+        **loss_kwargs,
+    )
+
+
+def _liger_for_causal_lm_loss(lm_head, hidden_states, hidden_size, labels, shift_labels, **loss_kwargs):
+    return LigerForCausalLMLoss(
+        hidden_states=hidden_states,
+        lm_head_weight=lm_head.weight,
+        labels=labels,
+        hidden_size=hidden_size,
+        shift_labels=shift_labels,
+        **loss_kwargs,
     )