liger-kernel-nightly 0.5.5.dev20250328142430__tar.gz → 0.5.5.dev20250331042257__tar.gz

{liger_kernel_nightly-0.5.5.dev20250328142430 → liger_kernel_nightly-0.5.5.dev20250331042257}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: liger_kernel_nightly
-Version: 0.5.5.dev20250328142430
+Version: 0.5.5.dev20250331042257
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 

{liger_kernel_nightly-0.5.5.dev20250328142430 → liger_kernel_nightly-0.5.5.dev20250331042257}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "liger_kernel_nightly"
-version = "0.5.5.dev20250328142430"
+version = "0.5.5.dev20250331042257"
 description = "Efficient Triton kernels for LLM Training"
 urls = { "Homepage" = "https://github.com/linkedin/Liger-Kernel" }
 readme = { file = "README.md", content-type = "text/markdown" }

{liger_kernel_nightly-0.5.5.dev20250328142430 → liger_kernel_nightly-0.5.5.dev20250331042257}/src/liger_kernel/transformers/__init__.py

@@ -12,6 +12,7 @@ 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

liger_kernel_nightly-0.5.5.dev20250331042257/src/liger_kernel/transformers/model/llava.py

@@ -0,0 +1,383 @@
+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.models.llava.modeling_llava import logger
+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,
+    cache_position: Optional[torch.LongTensor] = None,
+    num_logits_to_keep: int = 0,
+) -> 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"
+        )
+
+    legacy_processing = False
+    if inputs_embeds is None:
+        inputs_embeds = self.get_input_embeddings()(input_ids)
+
+        # if the number of image tokens is more than image embeddings seq length, then prob we expanded it in processing
+        # not very reliable, but we don't expect one to actually pass 500+ images for one prompt
+        # In case we're in decoding stage, legacy behavior is checked by presence of pixel values even if use_cache=True
+        legacy_processing = (
+            (input_ids == self.config.image_token_index).sum(1).max() < self.config.image_seq_length
+        ) or (input_ids.shape[-1] == 1 and pixel_values is not None)
+
+    image_features = None
+    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,
+        )
+
+    if legacy_processing and image_features is not None:
+        logger.warning_once(
+            "Expanding inputs for image tokens in LLaVa should be done in processing. "
+            "Please add `patch_size` and `vision_feature_select_strategy` to the model's processing config or set directly "
+            "with `processor.patch_size = {{patch_size}}` and processor.vision_feature_select_strategy = {{vision_feature_select_strategy}}`. "
+            "Using processors without these attributes in the config is deprecated and will throw an error in v4.50."
+        )
+        # prefill stage vs decoding stage (legacy behavior copied)
+        if input_ids.shape[1] != 1:
+            inputs_embeds, attention_mask, labels, position_ids = self._merge_input_ids_with_image_features(
+                image_features, inputs_embeds, input_ids, attention_mask, labels
+            )
+            cache_position = torch.arange(attention_mask.shape[1], device=attention_mask.device)
+        else:
+            # 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
+            cache_position = torch.arange(attention_mask.shape[1], device=attention_mask.device)[-target_length:]
+
+    # 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,
+        cache_position=cache_position,
+        num_logits_to_keep=num_logits_to_keep,
+    )
+    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,
+        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,
+    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_nightly-0.5.5.dev20250328142430 → liger_kernel_nightly-0.5.5.dev20250331042257}/src/liger_kernel/transformers/monkey_patch.py

@@ -19,6 +19,8 @@ from liger_kernel.transformers.model.gemma2 import lce_forward as gemma2_lce_for
 from liger_kernel.transformers.model.gemma2 import lce_forward_deprecated as gemma2_lce_forward_deprected
 from liger_kernel.transformers.model.llama import lce_forward as llama_lce_forward
 from liger_kernel.transformers.model.llama import lce_forward_deprecated as llama_lce_forward_deprecated
+from liger_kernel.transformers.model.llava import lce_forward as llava_lce_forward
+from liger_kernel.transformers.model.llava import lce_forward_deprecated as llava_lce_forward_deprecated
 from liger_kernel.transformers.model.mistral import lce_forward as mistral_lce_forward
 from liger_kernel.transformers.model.mixtral import lce_forward as mixtral_lce_forward
 from liger_kernel.transformers.model.mixtral import lce_forward_deprecated as mixtral_lce_forward_deprecated
@@ -57,7 +59,8 @@ def _patch_rms_norm_module(module, offset=0.0, eps=1e-6, casting_mode="llama", i
 
 def _patch_layer_norm_module(module, eps=1e-6):
     module.variance_epsilon = getattr(module, "variance_epsilon", None) or getattr(module, "eps", None) or eps
-    module.hidden_size = module.normalized_shape
+    module.hidden_size = getattr(module, "hidden_size", None) or getattr(module, "normalized_shape", None)
+
     _bind_method_to_module(module, "forward", LigerLayerNorm.forward)
     _bind_method_to_module(module, "extra_repr", LigerLayerNorm.extra_repr)
     module.__class__.__name__ = LigerLayerNorm.__name__
@@ -224,6 +227,85 @@ def apply_liger_kernel_to_llama(
                 _patch_rms_norm_module(decoder_layer.post_attention_layernorm)
 
 
+def apply_liger_kernel_to_llava(
+    cross_entropy: bool = False,
+    fused_linear_cross_entropy: bool = True,
+    model: PreTrainedModel = None,
+    **kwargs,
+) -> None:
+    """
+    Apply Liger kernels to replace original implementation in HuggingFace Llava models.
+    Due to the characteristics of LlaVa, the model must be passed to apply Liger-Kernel's patch to other models connected to LLaVa.
+    However, if an LM not supported by Liger-Kernel is connected to LLaVa, unexpected side effects may occur.
+    NOTE: Llava is not available in transformers<4.36.0
+
+    Args:
+        rope (bool): Whether to apply Liger's rotary position embedding. Default is True.
+        cross_entropy (bool): Whether to apply Liger's cross entropy loss. Default is False.
+        fused_linear_cross_entropy (bool):
+            Whether to apply Liger's fused linear cross entropy loss. Default is True.
+            `cross_entropy` and `fused_linear_cross_entropy` cannot both be True.
+            If `fused_linear_cross_entropy` is True, the logits will not be materialized but more memory efficient.
+        rms_norm (bool): Whether to apply Liger's RMSNorm. Default is True.
+        swiglu (bool): Whether to apply Liger's SwiGLU MLP. Default is True.
+        model (PreTrainedModel): The model instance to apply Liger kernels to, if the model has already been
+        loaded. Default is None.
+    """
+    assert not (cross_entropy and fused_linear_cross_entropy), (
+        "cross_entropy and fused_linear_cross_entropy cannot both be True."
+    )
+
+    from transformers.models.llava import modeling_llava
+
+    if cross_entropy:
+        logger.warning(TRANSFORMER_DEPRECATION_WARNING)
+        modeling_llava.nn.CrossEntropyLoss = LigerCrossEntropyLoss
+    if fused_linear_cross_entropy:
+        if transformer_version >= version.parse("4.49.0"):
+            modeling_llava.LlavaForConditionalGeneration.forward = llava_lce_forward
+        else:  # if version < 4.49.0
+            logger.warning(
+                "Support for transformers versions < 4.49.0 will soon be discontinued due to issues with incorrect legacy processing. \n Please consider upgrading to avoid potential issues. See details: https://github.com/huggingface/transformers/pull/35526"
+            )
+            modeling_llava.LlavaForConditionalGeneration.forward = llava_lce_forward_deprecated
+
+    if model is not None:
+        text_model_name, vision_model_name = model.config.text_config.model_type, model.config.vision_config.model_type
+        text_liger_fn = MODEL_TYPE_TO_APPLY_LIGER_FN.get(text_model_name, None)
+        vision_liger_fn = MODEL_TYPE_TO_APPLY_LIGER_FN.get(vision_model_name, None)
+
+        kwargs = {"cross_entropy": False, "fused_linear_cross_entropy": False, **kwargs}
+        if text_liger_fn:
+            accept_params = inspect.signature(text_liger_fn).parameters
+            remain_params = set(kwargs) - (set(accept_params) & set(kwargs))
+            text_kwargs = {k: v for k, v in kwargs.items() if k not in remain_params}
+
+            if remain_params:
+                logger.warning(
+                    f"These parameters are not supported by {text_model_name}. Enter the remaining {list(text_kwargs.keys())} except for {list(remain_params)}\n"
+                    f"Parameters accepted by {text_model_name}: {list(accept_params.keys())}"
+                )
+            text_kwargs["model"] = model.language_model
+            text_liger_fn(**text_kwargs)
+        elif text_model_name not in MODEL_TYPE_TO_APPLY_LIGER_FN:
+            logger.warning(f"{text_model_name} is not supported by Liger kernel.")
+
+        if vision_liger_fn:
+            accept_params = inspect.signature(vision_liger_fn).parameters
+            remain_params = set(kwargs) - (set(accept_params) & set(kwargs))
+            vision_kwargs = {k: v for k, v in kwargs.items() if k not in remain_params}
+
+            if remain_params:
+                logger.warning(
+                    f"These parameters are not supported by {vision_model_name}. Enter the remaining {list(vision_kwargs.keys())} except for {list(remain_params)}\n"
+                    f"Parameters accepted by {vision_model_name}: {list(accept_params.keys())}"
+                )
+            vision_kwargs["model"] = model.vision_tower
+            vision_liger_fn(**vision_kwargs)
+        elif vision_model_name not in MODEL_TYPE_TO_APPLY_LIGER_FN:
+            logger.warning(f"{vision_model_name} is not supported by Liger kernel.")
+
+
 def apply_liger_kernel_to_mllama(
     rope: bool = True,
     cross_entropy: bool = False,
@@ -1071,6 +1153,7 @@ MODEL_TYPE_TO_APPLY_LIGER_FN = {
     "gemma": apply_liger_kernel_to_gemma,
     "gemma2": apply_liger_kernel_to_gemma2,
     "llama": apply_liger_kernel_to_llama,
+    "llava": apply_liger_kernel_to_llava,
     "granite": apply_liger_kernel_to_granite,
     "mllama": apply_liger_kernel_to_mllama,
     "mllama_text_model": apply_liger_kernel_to_mllama,

{liger_kernel_nightly-0.5.5.dev20250328142430 → liger_kernel_nightly-0.5.5.dev20250331042257}/src/liger_kernel_nightly.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: liger_kernel_nightly
-Version: 0.5.5.dev20250328142430
+Version: 0.5.5.dev20250331042257
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 

{liger_kernel_nightly-0.5.5.dev20250328142430 → liger_kernel_nightly-0.5.5.dev20250331042257}/src/liger_kernel_nightly.egg-info/SOURCES.txt

@@ -162,6 +162,7 @@ src/liger_kernel/transformers/model/__init__.py
 src/liger_kernel/transformers/model/gemma.py
 src/liger_kernel/transformers/model/gemma2.py
 src/liger_kernel/transformers/model/llama.py
+src/liger_kernel/transformers/model/llava.py
 src/liger_kernel/transformers/model/loss_utils.py
 src/liger_kernel/transformers/model/mistral.py
 src/liger_kernel/transformers/model/mixtral.py
@@ -203,6 +204,9 @@ test/convergence/fp32/test_mini_models_multimodal.py
 test/convergence/fp32/test_mini_models_with_logits.py
 test/resources/tiny_shakespeare.txt
 test/resources/fake_configs/Google/Paligemma/paligemma-3b-pt-224/tokenizer_config.json
+test/resources/fake_configs/Llava/llava-1.5-7b-hf/preprocessor_config.json
+test/resources/fake_configs/Llava/llava-1.5-7b-hf/processor_config.json
+test/resources/fake_configs/Llava/llava-1.5-7b-hf/tokenizer_config.json
 test/resources/fake_configs/Qwen/Qwen2-VL-7B-Instruct/tokenizer_config.json
 test/resources/fake_configs/Qwen/Qwen2.5-VL-7B-Instruct/tokenizer_config.json
 test/resources/fake_configs/meta-llama/Llama-3.2-11B-Vision-Instruct/tokenizer_config.json