liger-kernel 0.5.4__py3-none-any.whl → 0.5.6__py3-none-any.whl

liger_kernel/ops/kl_div.py

@@ -185,9 +185,9 @@ class LigerKLDivLossFunction(torch.autograd.Function):
     Class implementing the forward and backward pass for the KL Divergence Loss using Triton, as defined by the following formula:
     ```python
     if log_target:
-        loss = target * (target.log() - input)
-    else:
         loss = target.exp() * (target - input)
+    else:
+        loss = target * (target.log() - input)
     ```,
     then the loss is reduced according to the `reduction` parameter.
     as defined in the PyTorch documentation: https://pytorch.org/docs/stable/generated/torch.nn.KLDivLoss.html

liger_kernel/transformers/__init__.py

@@ -1,5 +1,6 @@
 from liger_kernel.transformers.auto_model import AutoLigerKernelForCausalLM  # noqa: F401
 from liger_kernel.transformers.cross_entropy import LigerCrossEntropyLoss  # noqa: F401
+from liger_kernel.transformers.dyt import LigerDyT  # noqa: F401
 from liger_kernel.transformers.fused_linear_cross_entropy import LigerFusedLinearCrossEntropyLoss  # noqa: F401
 from liger_kernel.transformers.fused_linear_jsd import LigerFusedLinearJSD  # noqa: F401
 from liger_kernel.transformers.geglu import LigerGEGLUMLP  # noqa: F401
@@ -11,12 +12,15 @@ from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_gemma
 from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_gemma2  # noqa: F401
 from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_granite  # noqa: F401
 from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_llama  # noqa: F401
+from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_llava  # noqa: F401
 from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_mistral  # noqa: F401
 from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_mixtral  # noqa: F401
 from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_mllama  # noqa: F401
 from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_olmo2  # noqa: F401
+from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_paligemma  # noqa: F401
 from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_phi3  # noqa: F401
 from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_qwen2  # noqa: F401
+from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_qwen2_5_vl  # noqa: F401
 from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_qwen2_vl  # noqa: F401
 from liger_kernel.transformers.rms_norm import LigerRMSNorm  # noqa: F401
 from liger_kernel.transformers.rope import liger_rotary_pos_emb  # noqa: F401

liger_kernel/transformers/dyt.py

@@ -0,0 +1,20 @@
+import torch
+import torch.nn as nn
+
+from liger_kernel.ops.dyt import LigerDyTFunction
+
+
+class LigerDyT(nn.Module):
+    def __init__(self, hidden_size, init_alpha=0.5):
+        super().__init__()
+        self.hidden_size = hidden_size
+        self.init_alpha = init_alpha
+        self.alpha = nn.Parameter(torch.ones(1) * init_alpha)
+        self.gamma = nn.Parameter(torch.ones(hidden_size))
+        self.beta = nn.Parameter(torch.zeros(hidden_size))
+
+    def forward(self, x):
+        return LigerDyTFunction.apply(x, self.alpha, self.gamma, self.beta)
+
+    def extra_repr(self):
+        return f"{self.hidden_size}, init_alpha={self.init_alpha}"

liger_kernel/transformers/functional.py

@@ -1,6 +1,7 @@
 from typing import Optional
 
 from liger_kernel.ops.cross_entropy import LigerCrossEntropyFunction
+from liger_kernel.ops.dyt import LigerDyTFunction
 from liger_kernel.ops.fused_linear_cross_entropy import LigerFusedLinearCrossEntropyFunction
 from liger_kernel.ops.fused_linear_jsd import LigerFusedLinearJSDFunction
 from liger_kernel.ops.geglu import LigerGELUMulFunction
@@ -192,3 +193,7 @@ def liger_rope(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
 
 def liger_swiglu(a, b):
     return LigerSiLUMulFunction.apply(a, b)
+
+
+def liger_dyt(x, alpha, gamma, beta):
+    return LigerDyTFunction.apply(x, alpha, gamma, beta)

liger_kernel/transformers/model/gemma.py

@@ -14,6 +14,7 @@ from transformers.utils import add_start_docstrings_to_model_forward
 from transformers.utils import replace_return_docstrings
 
 from liger_kernel.transformers.fused_linear_cross_entropy import LigerFusedLinearCrossEntropyLoss
+from liger_kernel.transformers.model.loss_utils import LigerForCausalLMLoss
 
 
 @add_start_docstrings_to_model_forward(GEMMA_INPUTS_DOCSTRING)
@@ -200,22 +201,13 @@ def lce_forward(
     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)
-
-        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"]
-
+        loss = LigerForCausalLMLoss(
+            hidden_states=hidden_states,
+            lm_head_weight=self.lm_head.weight,
+            labels=labels,
+            hidden_size=self.config.hidden_size,
+            **loss_kwargs,
+        )
     else:  # if in inference mode materialize logits
         logits = self.lm_head(hidden_states[:, -num_logits_to_keep:, :])
         if labels is not None:

liger_kernel/transformers/model/gemma2.py

@@ -15,6 +15,7 @@ from transformers.utils import add_start_docstrings_to_model_forward
 from transformers.utils import replace_return_docstrings
 
 from liger_kernel.transformers.fused_linear_cross_entropy import LigerFusedLinearCrossEntropyLoss
+from liger_kernel.transformers.model.loss_utils import LigerForCausalLMLoss
 
 logger = logging.getLogger(__name__)
 
@@ -212,25 +213,15 @@ def lce_forward(
     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)
-
-        reduction = "sum" if "num_items_in_batch" in loss_kwargs else "mean"
-        lce = LigerFusedLinearCrossEntropyLoss(
+        loss = LigerForCausalLMLoss(
+            hidden_states=hidden_states,
+            lm_head_weight=self.lm_head.weight,
+            labels=labels,
+            hidden_size=self.config.hidden_size,
             softcap=self.config.final_logit_softcapping,
-            reduction=reduction,
+            **loss_kwargs,
         )
 
-        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 self.config.final_logit_softcapping is not None:

liger_kernel/transformers/model/llama.py

@@ -15,6 +15,7 @@ from transformers.utils import add_start_docstrings_to_model_forward
 from transformers.utils import replace_return_docstrings
 
 from liger_kernel.transformers.fused_linear_cross_entropy import LigerFusedLinearCrossEntropyLoss
+from liger_kernel.transformers.model.loss_utils import LigerForCausalLMLoss
 
 if TYPE_CHECKING:
     from transformers.cache_utils import Cache
@@ -212,21 +213,13 @@ def lce_forward(
     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)
-
-        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"]
+        loss = LigerForCausalLMLoss(
+            hidden_states=hidden_states,
+            lm_head_weight=self.lm_head.weight,
+            labels=labels,
+            hidden_size=self.config.hidden_size,
+            **loss_kwargs,
+        )
 
     else:  # if in inference mode materialize logits
         logits = self.lm_head(hidden_states[:, -num_logits_to_keep:, :])

liger_kernel/transformers/model/llava.py

@@ -0,0 +1,369 @@
+from typing import List
+from typing import Optional
+from typing import Tuple
+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 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
+
+
+@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,
+    pixel_values: 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[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,
+) -> Union[Tuple, LlavaCausalLMOutputWithPast]:
+    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]`.
+
+        num_logits_to_keep (`int`, *optional*):
+            Calculate logits for the last `num_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.
+
+
+    Returns:
+
+    Example:
+
+    ```python
+    >>> from PIL import Image
+    >>> import requests
+    >>> from transformers import AutoProcessor, LlavaForConditionalGeneration
+
+    >>> model = LlavaForConditionalGeneration.from_pretrained("llava-hf/llava-1.5-7b-hf")
+    >>> processor = AutoProcessor.from_pretrained("llava-hf/llava-1.5-7b-hf")
+
+    >>> prompt = "USER: <image>\nWhat's the content of the image? ASSISTANT:"
+    >>> url = "https://www.ilankelman.org/stopsigns/australia.jpg"
+    >>> image = Image.open(requests.get(url, stream=True).raw)
+
+    >>> inputs = processor(images=image, text=prompt, return_tensors="pt")
+
+    >>> # Generate
+    >>> generate_ids = model.generate(**inputs, max_new_tokens=15)
+    >>> 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
+    )
+    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
+    )
+
+    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:
+        # 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 n_image_tokens != n_image_features:
+            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)
+
+    outputs = self.language_model.model(
+        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,
+        return_dict=return_dict,
+    )
+    hidden_states = outputs[0]
+
+    loss = 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, shift_labels)
+
+    if not return_dict:
+        # NOTE: This part has not been tested.
+        output = outputs[1:]
+        return (loss,) + output if loss is not None else output
+
+    return LlavaCausalLMOutputWithPast(
+        loss=loss,
+        logits=logits,
+        past_key_values=outputs.past_key_values,
+        hidden_states=outputs.hidden_states,
+        attentions=outputs.attentions,
+    )
+
+
+@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,
+    pixel_values: 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[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: torch.Tensor = None,
+    **lm_kwargs,
+) -> Union[Tuple, LlavaCausalLMOutputWithPast]:
+    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]`.
+
+        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 PIL import Image
+    >>> import requests
+    >>> from transformers import AutoProcessor, LlavaForConditionalGeneration
+
+    >>> model = LlavaForConditionalGeneration.from_pretrained("llava-hf/llava-1.5-7b-hf")
+    >>> processor = AutoProcessor.from_pretrained("llava-hf/llava-1.5-7b-hf")
+
+    >>> prompt = "USER: <image>\nWhat's the content of the image? ASSISTANT:"
+    >>> url = "https://www.ilankelman.org/stopsigns/australia.jpg"
+    >>> image = Image.open(requests.get(url, stream=True).raw)
+
+    >>> inputs = processor(images=image, text=prompt, return_tensors="pt")
+
+    >>> # Generate
+    >>> generate_ids = model.generate(**inputs, max_new_tokens=15)
+    >>> 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
+    )
+    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
+    )
+
+    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(
+        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,
+        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):
+        # 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),
+        )
+
+    if not return_dict:
+        # NOTE: This part has not been tested.
+        output = outputs[1:]
+        return (loss,) + output if loss is not None else output
+
+    return LlavaCausalLMOutputWithPast(
+        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,
+    )

liger_kernel/transformers/model/loss_utils.py

@@ -0,0 +1,57 @@
+import torch.nn as nn
+
+import liger_kernel.transformers.functional as F
+
+
+def fixed_fused_linear_cross_entropy(
+    hidden_states,
+    lm_head_weight,
+    target,
+    num_items_in_batch: int = None,
+    ignore_index: int = -100,
+    **kwargs,
+):
+    reduction = "sum" if num_items_in_batch is not None else "mean"
+    loss = F.liger_fused_linear_cross_entropy(
+        hidden_states,
+        lm_head_weight,
+        target,
+        reduction=reduction,
+        ignore_index=ignore_index,
+        **kwargs,
+    )
+    if reduction == "sum":
+        loss = loss / num_items_in_batch
+
+    return loss
+
+
+def LigerForCausalLMLoss(
+    hidden_states,
+    lm_head_weight,
+    labels,
+    hidden_size: int,
+    num_items_in_batch: int = None,
+    ignore_index: int = -100,
+    **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()
+
+    # 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(
+        hidden_states,
+        lm_head_weight,
+        shift_labels,
+        num_items_in_batch,
+        ignore_index,
+        **kwargs,
+    )
+    return loss

liger_kernel/transformers/model/mistral.py

@@ -13,7 +13,7 @@ from transformers.models.mistral.modeling_mistral import MISTRAL_INPUTS_DOCSTRIN
 from transformers.utils import add_start_docstrings_to_model_forward
 from transformers.utils import replace_return_docstrings
 
-from liger_kernel.transformers.fused_linear_cross_entropy import LigerFusedLinearCrossEntropyLoss
+from liger_kernel.transformers.model.loss_utils import LigerForCausalLMLoss
 
 
 @add_start_docstrings_to_model_forward(MISTRAL_INPUTS_DOCSTRING)
@@ -31,6 +31,7 @@ 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]:
     r"""
     Copy paste Mistral's forward but replace torch cross entropy with liger fused linear cross entropy
@@ -87,15 +88,13 @@ def lce_forward(
     logits = None
 
     if self.training and (labels is not None):
-        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)
-
-        lce = LigerFusedLinearCrossEntropyLoss()
-        loss = lce(self.lm_head.weight, shift_hidden_states, shift_labels)
+        loss = LigerForCausalLMLoss(
+            hidden_states=hidden_states,
+            lm_head_weight=self.lm_head.weight,
+            labels=labels,
+            hidden_size=self.config.hidden_size,
+            **loss_kwargs,
+        )
 
     else:
         logits = self.lm_head(hidden_states)