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

liger_kernel/transformers/model/qwen2_5_vl.py

@@ -0,0 +1,204 @@
+from typing import List
+from typing import Optional
+from typing import Tuple
+from typing import Union
+
+import torch
+
+from torch.nn import CrossEntropyLoss
+from transformers.models.qwen2_5_vl.modeling_qwen2_5_vl import _CONFIG_FOR_DOC
+from transformers.models.qwen2_5_vl.modeling_qwen2_5_vl import QWEN2_5_VL_INPUTS_DOCSTRING
+from transformers.models.qwen2_5_vl.modeling_qwen2_5_vl import Qwen2_5_VLCausalLMOutputWithPast
+from transformers.utils import add_start_docstrings_to_model_forward
+from transformers.utils import replace_return_docstrings
+
+from liger_kernel.transformers.model.loss_utils import LigerForCausalLMLoss
+
+
+@add_start_docstrings_to_model_forward(QWEN2_5_VL_INPUTS_DOCSTRING)
+@replace_return_docstrings(output_type=Qwen2_5_VLCausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
+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,
+    use_cache: Optional[bool] = None,
+    output_attentions: Optional[bool] = None,
+    output_hidden_states: Optional[bool] = None,
+    return_dict: Optional[bool] = 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,
+    second_per_grid_ts: Optional[torch.Tensor] = None,
+    **loss_kwargs,
+) -> Union[Tuple, Qwen2_5_VLCausalLMOutputWithPast]:
+    r"""
+    Copy paste Qwen2_5_VL's forward but replace torch cross entropy with liger fused linear cross entropy
+    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]`.
+
+    Returns:
+
+    Example:
+
+    ```python
+    >>> from PIL import Image
+    >>> import requests
+    >>> from transformers import AutoProcessor, Qwen2_5_VLForConditionalGeneration
+
+    >>> model = Qwen2_5_VLForConditionalGeneration.from_pretrained("Qwen/Qwen2.5-VL-7B-Instruct")
+    >>> processor = AutoProcessor.from_pretrained("Qwen/Qwen2.5-VL-7B-Instruct")
+
+    >>> messages = [
+        {
+            "role": "user",
+            "content": [
+                {"type": "image"},
+                {"type": "text", "text": "What is shown in this image?"},
+            ],
+        },
+    ]
+    >>> url = "https://www.ilankelman.org/stopsigns/australia.jpg"
+    >>> image = Image.open(requests.get(url, stream=True).raw)
+
+    >>> text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
+    >>> inputs = processor(text=[text], images=[image], vision_infos=[vision_infos])
+
+    >>> # Generate
+    >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
+    >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
+    "The image shows a street scene with a red stop sign in the foreground. In the background, there is a large red gate with Chinese characters ..."
+    ```"""
+    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
+
+    if inputs_embeds is None:
+        inputs_embeds = self.model.embed_tokens(input_ids)
+        if pixel_values is not None:
+            pixel_values = pixel_values.type(self.visual.dtype)
+            image_embeds = self.visual(pixel_values, grid_thw=image_grid_thw)
+            n_image_tokens = (input_ids == self.config.image_token_id).sum().item()
+            n_image_features = image_embeds.shape[0]
+            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}"
+                )
+
+            mask = input_ids == self.config.image_token_id
+            mask_unsqueezed = mask.unsqueeze(-1)
+            mask_expanded = mask_unsqueezed.expand_as(inputs_embeds)
+            image_mask = mask_expanded.to(inputs_embeds.device)
+
+            image_embeds = image_embeds.to(inputs_embeds.device, inputs_embeds.dtype)
+            inputs_embeds = inputs_embeds.masked_scatter(image_mask, image_embeds)
+
+        if pixel_values_videos is not None:
+            pixel_values_videos = pixel_values_videos.type(self.visual.dtype)
+            video_embeds = self.visual(pixel_values_videos, grid_thw=video_grid_thw)
+            n_video_tokens = (input_ids == self.config.video_token_id).sum().item()
+            n_video_features = video_embeds.shape[0]
+            if n_video_tokens != n_video_features:
+                raise ValueError(
+                    f"Video features and video tokens do not match: tokens: {n_video_tokens}, features {n_video_features}"
+                )
+
+            mask = input_ids == self.config.video_token_id
+            mask_unsqueezed = mask.unsqueeze(-1)
+            mask_expanded = mask_unsqueezed.expand_as(inputs_embeds)
+            video_mask = mask_expanded.to(inputs_embeds.device)
+
+            video_embeds = video_embeds.to(inputs_embeds.device, inputs_embeds.dtype)
+            inputs_embeds = inputs_embeds.masked_scatter(video_mask, video_embeds)
+
+        if attention_mask is not None:
+            attention_mask = attention_mask.to(inputs_embeds.device)
+
+    # if we get 4D attention mask we cannot calculate rope deltas anymore. TODO @raushan fixme
+    if position_ids is None and (attention_mask is None or attention_mask.ndim == 2):
+        # calculate RoPE index once per generation in the pre-fill stage only
+        if (cache_position is not None and cache_position[0] == 0) or self.rope_deltas is None:
+            position_ids, rope_deltas = self.get_rope_index(
+                input_ids,
+                image_grid_thw,
+                video_grid_thw,
+                second_per_grid_ts,
+                attention_mask,
+            )
+            self.rope_deltas = rope_deltas
+        # then use the prev pre-calculated rope-deltas to get the correct position ids
+        else:
+            batch_size, seq_length, _ = inputs_embeds.shape
+            delta = (cache_position[0] + self.rope_deltas).to(inputs_embeds.device) if cache_position is not None else 0
+            position_ids = torch.arange(seq_length, device=inputs_embeds.device)
+            position_ids = position_ids.view(1, -1).expand(batch_size, -1)
+            if cache_position is not None:  # otherwise `deltas` is an int `0`
+                delta = delta.repeat_interleave(batch_size // delta.shape[0], dim=0)
+            position_ids = position_ids.add(delta)
+            position_ids = position_ids.unsqueeze(0).expand(3, -1, -1)
+
+    outputs = self.model(
+        input_ids=None,
+        position_ids=position_ids,
+        attention_mask=attention_mask,
+        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,
+    )
+
+    hidden_states = outputs[0]
+
+    loss = None
+    logits = None
+
+    if self.training and (labels is not None):
+        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)
+        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
+            loss_fct = CrossEntropyLoss()
+            shift_logits = shift_logits.view(-1, self.config.vocab_size)
+            shift_labels = shift_labels.view(-1)
+            # Enable model parallelism
+            shift_labels = shift_labels.to(shift_logits.device)
+            loss = loss_fct(shift_logits, shift_labels)
+
+    if not return_dict:
+        output = (logits,) + outputs[1:]
+        return (loss,) + output if loss is not None else output
+
+    return Qwen2_5_VLCausalLMOutputWithPast(
+        loss=loss,
+        logits=logits,
+        past_key_values=outputs.past_key_values,
+        hidden_states=outputs.hidden_states,
+        attentions=outputs.attentions,
+        rope_deltas=rope_deltas,
+    )

liger_kernel/transformers/model/qwen2_vl.py

@@ -14,7 +14,7 @@ from transformers.models.qwen2_vl.modeling_qwen2_vl import Qwen2VLCausalLMOutput
 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(QWEN2_VL_INPUTS_DOCSTRING)
@@ -37,6 +37,7 @@ def lce_forward(
     video_grid_thw: Optional[torch.LongTensor] = None,
     rope_deltas: Optional[torch.LongTensor] = None,
     cache_position: Optional[torch.LongTensor] = None,
+    **loss_kwargs,
 ) -> Union[Tuple, Qwen2VLCausalLMOutputWithPast]:
     r"""
     Copy paste Qwen2VL's forward but replace torch cross entropy with liger fused linear cross entropy
@@ -170,15 +171,13 @@ 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)
         if labels is not None:

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
@@ -52,13 +54,26 @@ def _patch_rms_norm_module(module, offset=0.0, eps=1e-6, casting_mode="llama", i
     module.in_place = in_place
     _bind_method_to_module(module, "forward", LigerRMSNorm.forward)
     _bind_method_to_module(module, "extra_repr", LigerRMSNorm.extra_repr)
+    module.__class__.__name__ = LigerRMSNorm.__name__
 
 
 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__
+
+
+def _patch_swiglu_module(module, liger_module):
+    _bind_method_to_module(module, "forward", liger_module.forward)
+    module.__class__.__name__ = liger_module.__name__
+
+
+def _patch_geglu_module(module):
+    _bind_method_to_module(module, "forward", LigerGEGLUMLP.forward)
+    module.__class__.__name__ = LigerGEGLUMLP.__name__
 
 
 def apply_liger_kernel_to_granite(
@@ -134,7 +149,7 @@ def apply_liger_kernel_to_granite(
 
         for decoder_layer in base_model.layers:
             if swiglu:
-                _bind_method_to_module(decoder_layer.mlp, "forward", LigerSwiGLUMLP.forward)
+                _patch_swiglu_module(decoder_layer.mlp, LigerSwiGLUMLP)
             if rms_norm:
                 _patch_rms_norm_module(decoder_layer.input_layernorm)
                 _patch_rms_norm_module(decoder_layer.post_attention_layernorm)
@@ -206,12 +221,91 @@ def apply_liger_kernel_to_llama(
 
         for decoder_layer in base_model.layers:
             if swiglu:
-                _bind_method_to_module(decoder_layer.mlp, "forward", LigerSwiGLUMLP.forward)
+                _patch_swiglu_module(decoder_layer.mlp, LigerSwiGLUMLP)
             if rms_norm:
                 _patch_rms_norm_module(decoder_layer.input_layernorm)
                 _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,
@@ -296,7 +390,7 @@ def apply_liger_kernel_to_mllama(
                 _patch_rms_norm_module(text_model.norm)
             for decoder_layer in text_model.layers:
                 if swiglu:
-                    _bind_method_to_module(decoder_layer.mlp, "forward", LigerSwiGLUMLP.forward)
+                    _patch_swiglu_module(decoder_layer.mlp, LigerSwiGLUMLP)
                 if rms_norm:
                     _patch_rms_norm_module(decoder_layer.input_layernorm)
                     _patch_rms_norm_module(decoder_layer.post_attention_layernorm)
@@ -370,7 +464,7 @@ def apply_liger_kernel_to_mistral(
 
         for decoder_layer in base_model.layers:
             if swiglu:
-                _bind_method_to_module(decoder_layer.mlp, "forward", LigerSwiGLUMLP.forward)
+                _patch_swiglu_module(decoder_layer.mlp, LigerSwiGLUMLP)
             if rms_norm:
                 _patch_rms_norm_module(decoder_layer.input_layernorm)
                 _patch_rms_norm_module(decoder_layer.post_attention_layernorm)
@@ -442,7 +536,7 @@ def apply_liger_kernel_to_mixtral(
         for decoder_layer in base_model.layers:
             if swiglu:
                 for expert in decoder_layer.block_sparse_moe.experts:
-                    _bind_method_to_module(expert, "forward", LigerBlockSparseTop2MLP.forward)
+                    _patch_swiglu_module(expert, LigerBlockSparseTop2MLP)
             if rms_norm:
                 _patch_rms_norm_module(decoder_layer.input_layernorm)
                 _patch_rms_norm_module(decoder_layer.post_attention_layernorm)
@@ -516,7 +610,7 @@ def apply_liger_kernel_to_gemma(
 
         for decoder_layer in base_model.layers:
             if geglu:
-                _bind_method_to_module(decoder_layer.mlp, "forward", LigerGEGLUMLP.forward)
+                _patch_geglu_module(decoder_layer.mlp)
             if rms_norm:
                 _patch_rms_norm_module_for_gemma(decoder_layer.input_layernorm)
                 _patch_rms_norm_module_for_gemma(decoder_layer.post_attention_layernorm)
@@ -592,7 +686,7 @@ def apply_liger_kernel_to_gemma2(
 
         for decoder_layer in base_model.layers:
             if geglu:
-                _bind_method_to_module(decoder_layer.mlp, "forward", LigerGEGLUMLP.forward)
+                _patch_geglu_module(decoder_layer.mlp)
             if rms_norm:
                 _patch_rms_norm_module_for_gemma2(decoder_layer.input_layernorm)
                 _patch_rms_norm_module_for_gemma2(decoder_layer.post_attention_layernorm)
@@ -600,6 +694,116 @@ def apply_liger_kernel_to_gemma2(
                 _patch_rms_norm_module_for_gemma2(decoder_layer.post_feedforward_layernorm)
 
 
+def apply_liger_kernel_to_paligemma(
+    rope: bool = True,
+    cross_entropy: bool = False,
+    fused_linear_cross_entropy: bool = True,
+    layer_norm: bool = True,
+    rms_norm: bool = True,
+    geglu: bool = True,
+    model: PreTrainedModel = None,
+) -> None:
+    """
+    Apply Liger kernels to replace original implementation in HuggingFace PaliGemma
+
+    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.
+        layer_norm (bool): Whether to apply Liger's LayerNorm. Default is True.
+        rms_norm (bool): Whether to apply Liger's RMSNorm. Default is True.
+        geglu (bool): Whether to apply Liger's GeGLU 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."
+    )
+
+    # PaliGemma submodules are ['vision_tower', 'multi_modal_projector', 'language_model']
+
+    from transformers.models.gemma.modeling_gemma import GemmaForCausalLM
+    from transformers.models.gemma2.modeling_gemma2 import Gemma2ForCausalLM
+    from transformers.models.paligemma import modeling_paligemma
+    from transformers.models.paligemma.modeling_paligemma import PaliGemmaForConditionalGeneration
+    from transformers.models.siglip import modeling_siglip
+    from transformers.models.siglip.modeling_siglip import SiglipEncoderLayer
+    from transformers.models.siglip.modeling_siglip import SiglipVisionModel
+
+    from liger_kernel.transformers.model.paligemma import lce_forward
+    from liger_kernel.transformers.model.paligemma import lce_forward_deprecated
+
+    # The vision_tower is a SiglipVisionModel
+    if layer_norm:
+        modeling_siglip.nn.LayerNorm = LigerLayerNorm
+
+    # SiglipMLP is standard FFN so LigerGEGLUMLP is not compatible
+    # The multi_modal_projector is Linear, nothing to do
+
+    # The language_model is GemmaForCausalLM or Gemma2ForCausalLM
+    apply_liger_kernel_to_gemma(
+        rope=rope, cross_entropy=False, fused_linear_cross_entropy=False, rms_norm=rms_norm, geglu=geglu
+    )
+    apply_liger_kernel_to_gemma2(
+        rope=rope, cross_entropy=False, fused_linear_cross_entropy=False, rms_norm=rms_norm, geglu=geglu
+    )
+    # Handle loss function
+    if cross_entropy:
+        modeling_paligemma.nn.CrossEntropyLoss = LigerCrossEntropyLoss
+    if fused_linear_cross_entropy:
+        if transformer_version >= version.parse(SUPPORTED_TRANSFORMER_VERSION):
+            modeling_paligemma.PaliGemmaForConditionalGeneration.forward = lce_forward
+        else:  # if version < 4.46.1
+            logger.warning(TRANSFORMER_DEPRECATION_WARNING)
+            modeling_paligemma.PaliGemmaForConditionalGeneration.forward = lce_forward_deprecated
+
+    if model is not None:
+        # The model instance already exists, so we need to additionally patch the
+        # instance variables that reference already-instantiated modules
+
+        if not isinstance(model, PaliGemmaForConditionalGeneration):
+            raise TypeError("model have to be of type PaliGemmaForConditionalGeneration")
+
+        vision_tower: SiglipVisionModel = model.vision_tower
+
+        _patch_layer_norm_module(vision_tower.vision_model.post_layernorm)
+
+        for layer in vision_tower.vision_model.encoder.layers:
+            layer: SiglipEncoderLayer
+            if layer_norm:
+                _patch_layer_norm_module(layer.layer_norm1)
+                _patch_layer_norm_module(layer.layer_norm2)
+
+        language_model = model.language_model
+
+        if isinstance(language_model, GemmaForCausalLM):
+            apply_liger_kernel_to_gemma(
+                rope=rope,
+                cross_entropy=False,
+                fused_linear_cross_entropy=False,
+                rms_norm=rms_norm,
+                geglu=geglu,
+                model=language_model,
+            )
+
+        elif isinstance(language_model, Gemma2ForCausalLM):
+            apply_liger_kernel_to_gemma2(
+                rope=rope,
+                cross_entropy=False,
+                fused_linear_cross_entropy=False,
+                rms_norm=rms_norm,
+                geglu=geglu,
+                model=language_model,
+            )
+        else:
+            raise TypeError(
+                "The language_model of a PaliGemma model must be either GemmaForCausalLM or Gemma2ForCausalLM."
+            )
+
+
 def apply_liger_kernel_to_qwen2(
     rope: bool = True,
     cross_entropy: bool = False,
@@ -666,7 +870,7 @@ def apply_liger_kernel_to_qwen2(
 
         for decoder_layer in base_model.layers:
             if swiglu:
-                _bind_method_to_module(decoder_layer.mlp, "forward", LigerSwiGLUMLP.forward)
+                _patch_swiglu_module(decoder_layer.mlp, LigerSwiGLUMLP)
             if rms_norm:
                 _patch_rms_norm_module(decoder_layer.input_layernorm)
                 _patch_rms_norm_module(decoder_layer.post_attention_layernorm)
@@ -739,7 +943,74 @@ def apply_liger_kernel_to_qwen2_vl(
             _patch_rms_norm_module(base_model.norm)
         for decoder_layer in base_model.layers:
             if swiglu:
-                _bind_method_to_module(decoder_layer.mlp, "forward", LigerSwiGLUMLP.forward)
+                _patch_swiglu_module(decoder_layer.mlp, LigerSwiGLUMLP)
+            if rms_norm:
+                _patch_rms_norm_module(decoder_layer.input_layernorm)
+                _patch_rms_norm_module(decoder_layer.post_attention_layernorm)
+
+
+def apply_liger_kernel_to_qwen2_5_vl(
+    rope: bool = True,
+    cross_entropy: bool = False,
+    fused_linear_cross_entropy: bool = True,
+    rms_norm: bool = True,
+    swiglu: bool = True,
+    model: PreTrainedModel = None,
+) -> None:
+    """
+    Apply Liger kernels to replace original implementation in HuggingFace Qwen2.5-VL models.
+    NOTE: Qwen2.5-VL is not available in transformers<4.48.2
+
+    Args:
+        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.qwen2_5_vl import modeling_qwen2_5_vl
+    from transformers.models.qwen2_5_vl.modeling_qwen2_5_vl import Qwen2_5_VLModel
+
+    from liger_kernel.transformers.model.qwen2_5_vl import lce_forward as qwen2_5_vl_lce_forward
+
+    if rope:
+        modeling_qwen2_5_vl.apply_multimodal_rotary_pos_emb = liger_multimodal_rotary_pos_emb
+    if rms_norm:
+        modeling_qwen2_5_vl.Qwen2RMSNorm = LigerRMSNorm
+    if cross_entropy:
+        modeling_qwen2_5_vl.CrossEntropyLoss = LigerCrossEntropyLoss
+    if fused_linear_cross_entropy:
+        modeling_qwen2_5_vl.Qwen2_5_VLForConditionalGeneration.forward = qwen2_5_vl_lce_forward
+    if swiglu:
+        modeling_qwen2_5_vl.Qwen2MLP = LigerSwiGLUMLP
+
+    if model is not None:
+        # The model instance already exists, so we need to additionally patch the
+        # instance variables that reference already-instantiated modules
+
+        # get the base model from the model instance
+        base_model: Qwen2_5_VLModel = getattr(model, model.base_model_prefix, model)
+
+        if hasattr(model, "visual"):
+            # Patch Qwen2_5_VisionTransformerPretrainedModel
+            for vision_block in model.visual.blocks:
+                if rms_norm:
+                    _patch_rms_norm_module(vision_block.norm1)
+                    _patch_rms_norm_module(vision_block.norm2)
+
+        if rms_norm:
+            _patch_rms_norm_module(base_model.norm)
+        for decoder_layer in base_model.layers:
+            if swiglu:
+                _patch_swiglu_module(decoder_layer.mlp, LigerSwiGLUMLP)
             if rms_norm:
                 _patch_rms_norm_module(decoder_layer.input_layernorm)
                 _patch_rms_norm_module(decoder_layer.post_attention_layernorm)
@@ -808,7 +1079,7 @@ def apply_liger_kernel_to_phi3(
 
         for decoder_layer in base_model.layers:
             if swiglu:
-                _bind_method_to_module(decoder_layer.mlp, "forward", LigerPhi3SwiGLUMLP.forward)
+                _patch_swiglu_module(decoder_layer.mlp, LigerPhi3SwiGLUMLP)
             if rms_norm:
                 _patch_rms_norm_module(decoder_layer.input_layernorm)
                 _patch_rms_norm_module(decoder_layer.post_attention_layernorm)
@@ -871,7 +1142,7 @@ def apply_liger_kernel_to_olmo2(
 
         for decoder_layer in base_model.layers:
             if swiglu:
-                _bind_method_to_module(decoder_layer.mlp, "forward", LigerSwiGLUMLP.forward)
+                _patch_swiglu_module(decoder_layer.mlp, LigerSwiGLUMLP)
             if rms_norm:
                 _patch_rms_norm_module(decoder_layer.post_attention_layernorm, in_place=False)
                 _patch_rms_norm_module(decoder_layer.post_feedforward_layernorm, in_place=False)
@@ -882,6 +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,
@@ -890,7 +1162,9 @@ MODEL_TYPE_TO_APPLY_LIGER_FN = {
     "olmo2": apply_liger_kernel_to_olmo2,
     "qwen2": apply_liger_kernel_to_qwen2,
     "qwen2_vl": apply_liger_kernel_to_qwen2_vl,
+    "qwen2_5_vl": apply_liger_kernel_to_qwen2_5_vl,
     "phi3": apply_liger_kernel_to_phi3,
+    "paligemma": apply_liger_kernel_to_paligemma,
 }
 
 

liger_kernel/utils.py

@@ -5,12 +5,10 @@ def infer_device():
     """
     Get current device name based on available devices
     """
-    if torch.cuda.is_available():
+    if torch.cuda.is_available():  # Works for both Nvidia and AMD
         return "cuda"
     elif torch.xpu.is_available():
         return "xpu"
-    elif torch.hip.is_available():
-        return "hip"
     else:
         return "cpu"