liger-kernel-nightly 0.6.2.dev20250923161735__py3-none-any.whl → 0.6.2.dev20251008084122__py3-none-any.whl

liger_kernel/transformers/__init__.py

@@ -38,6 +38,7 @@ if TYPE_CHECKING:
     from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_glm4v  # noqa: F401
     from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_glm4v_moe  # 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_internvl  # 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_llama4  # noqa: F401
     from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_llava  # noqa: F401
@@ -98,6 +99,7 @@ def __getattr__(name: str):
         "apply_liger_kernel_to_glm4v",
         "apply_liger_kernel_to_glm4v_moe",
         "apply_liger_kernel_to_granite",
+        "apply_liger_kernel_to_internvl",
         "apply_liger_kernel_to_llama",
         "apply_liger_kernel_to_llava",
         "apply_liger_kernel_to_llama4",
@@ -163,6 +165,7 @@ if _TRANSFORMERS_AVAILABLE:
             "apply_liger_kernel_to_glm4v",
             "apply_liger_kernel_to_glm4v_moe",
             "apply_liger_kernel_to_granite",
+            "apply_liger_kernel_to_internvl",
             "apply_liger_kernel_to_llama",
             "apply_liger_kernel_to_llava",
             "apply_liger_kernel_to_llama4",

liger_kernel/transformers/model/gemma.py

@@ -228,10 +228,11 @@ def lce_forward(
         )
     else:
         logits = self.lm_head(kept_hidden_states)
-        if labels is not None:
+        if labels is not None or shift_labels is not None:
             loss = self.loss_function(
                 logits=logits,
                 labels=labels,
+                shift_labels=shift_labels,
                 vocab_size=self.config.vocab_size,
                 **kwargs,
             )

liger_kernel/transformers/model/gemma2.py

@@ -252,8 +252,14 @@ def lce_forward(
             logits = logits * self.config.final_logit_softcapping
 
         loss = None
-        if labels is not None:
-            loss = self.loss_function(logits, labels, self.vocab_size, **kwargs)
+        if labels is not None or shift_labels is not None:
+            loss = self.loss_function(
+                logits=logits,
+                labels=labels,
+                shift_labels=shift_labels,
+                vocab_size=self.vocab_size,
+                **kwargs,
+            )
 
     if not return_dict:
         output = (logits,) + outputs[1:]

liger_kernel/transformers/model/gemma3.py

@@ -119,8 +119,14 @@ def causal_forward(
             logits = logits / self.config.final_logit_softcapping
             logits = torch.tanh(logits)
             logits = logits * self.config.final_logit_softcapping
-        if labels is not None:
-            loss = self.loss_function(logits, labels, self.vocab_size, **loss_kwargs)
+        if labels is not None or shift_labels is not None:
+            loss = self.loss_function(
+                logits=logits,
+                labels=labels,
+                shift_labels=shift_labels,
+                vocab_size=self.vocab_size,
+                **loss_kwargs,
+            )
 
     if not return_dict:
         output = (logits,) + outputs[1:]
@@ -275,6 +281,25 @@ def multimodal_forward(
             # Flatten the tokens
             loss_fct = nn.CrossEntropyLoss()
 
+            flat_logits = shift_logits.view(-1, self.config.text_config.vocab_size)
+            flat_labels = shift_labels.view(-1).to(shift_logits.device)
+            loss = loss_fct(flat_logits, flat_labels)
+        elif shift_labels is not None:
+            # Upcast to float if we need to compute the loss to avoid potential precision issues
+            logits = logits.float()
+            shift_logits = logits[..., :-1, :]
+            if attention_mask is not None:
+                # we use the input attention mask to shift the logits and labels, because it is 2D.
+                # we also crop attn mask in case it is longer, which happens in PrefixTuning with peft
+                shift_attention_mask = attention_mask[:, -shift_logits.shape[1] :].to(logits.device)
+                shift_logits = shift_logits[shift_attention_mask.to(logits.device) != 0].contiguous()
+                shift_labels = shift_labels[shift_attention_mask.to(shift_labels.device) != 0].contiguous()
+            else:
+                shift_logits = shift_logits.contiguous()
+                shift_labels = shift_labels.contiguous()
+            # Flatten the tokens
+            loss_fct = nn.CrossEntropyLoss()
+
             flat_logits = shift_logits.view(-1, self.config.text_config.vocab_size)
             flat_labels = shift_labels.view(-1).to(shift_logits.device)
             loss = loss_fct(flat_logits, flat_labels)

liger_kernel/transformers/model/glm4.py

@@ -111,10 +111,11 @@ def lce_forward(
 
     else:
         logits = self.lm_head(kept_hidden_states)
-        if labels is not None:
+        if labels is not None or shift_labels is not None:
             loss = self.loss_function(
                 logits=logits,
                 labels=labels,
+                shift_labels=shift_labels,
                 vocab_size=self.config.vocab_size,
                 **kwargs,
             )

liger_kernel/transformers/model/glm4v.py

@@ -133,10 +133,11 @@ def lce_forward(
 
     else:
         logits = self.lm_head(kept_hidden_states)
-        if labels is not None:
+        if labels is not None or shift_labels is not None:
             loss = self.loss_function(
                 logits=logits,
                 labels=labels,
+                shift_labels=shift_labels,
                 vocab_size=self.config.vocab_size,
                 **kwargs,
             )

liger_kernel/transformers/model/glm4v_moe.py

@@ -134,10 +134,11 @@ def lce_forward(
 
     else:
         logits = self.lm_head(kept_hidden_states)
-        if labels is not None:
+        if labels is not None or shift_labels is not None:
             loss = self.loss_function(
                 logits=logits,
                 labels=labels,
+                shift_labels=shift_labels,
                 vocab_size=self.config.vocab_size,
                 **kwargs,
             )

liger_kernel/transformers/model/internvl.py

@@ -0,0 +1,150 @@
+from typing import List
+from typing import Optional
+from typing import Tuple
+from typing import Union
+
+import torch
+
+from transformers.models.internvl.modeling_internvl import InternVLCausalLMOutputWithPast
+from transformers.utils import can_return_tuple
+
+from liger_kernel.transformers.model.loss_utils import LigerForCausalLMLoss
+
+
+# Copied from https://github.com/huggingface/transformers/blob/d888bd435d0c0eaabaabad5b33d52af518c7187c/src/transformers/models/internvl/modeling_internvl.py#L862
+@can_return_tuple
+def lce_forward(
+    self,
+    input_ids: torch.LongTensor = None,
+    pixel_values: Optional[torch.FloatTensor] = None,
+    attention_mask: Optional[torch.Tensor] = None,
+    position_ids: Optional[torch.LongTensor] = None,
+    past_key_values: Optional[List[torch.FloatTensor]] = None,
+    inputs_embeds: Optional[torch.FloatTensor] = None,
+    vision_feature_layer: Optional[Union[int, List[int]]] = None,
+    vision_feature_select_strategy: Optional[str] = None,
+    labels: Optional[torch.LongTensor] = None,
+    use_cache: Optional[bool] = None,
+    output_attentions: Optional[bool] = None,
+    output_hidden_states: Optional[bool] = None,
+    return_dict: Optional[bool] = None,
+    cache_position: Optional[torch.LongTensor] = None,
+    logits_to_keep: Union[int, torch.Tensor] = 0,
+    image_sizes: Optional[torch.Tensor] = None,
+    skip_logits: Optional[bool] = None,  # Added argument for liger-kernel
+    **lm_kwargs,  # renamed from kwargs
+) -> Union[Tuple, InternVLCausalLMOutputWithPast]:
+    r"""
+    Example:
+
+    ```python
+    >>> import torch
+    >>> from transformers import AutoProcessor, AutoModelForImageTextToText
+
+    >>> torch_device = "cuda"
+    >>> processor = AutoProcessor.from_pretrained("OpenGVLab/InternVL3-1B-hf")
+    >>> model = AutoModelForImageTextToText.from_pretrained(
+    ...     "OpenGVLab/InternVL3-1B-hf", dtype=torch.bfloat16, device_map=torch_device
+    ... )
+
+    >>> messages = [
+    ...     {
+    ...         "role": "user",
+    ...         "content": [
+    ...             {
+    ...                 "type": "image",
+    ...                 "url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg",
+    ...             },
+    ...             {
+    ...                 "type": "image",
+    ...                 "url": "https://thumbs.dreamstime.com/b/golden-gate-bridge-san-francisco-purple-flowers-california-echium-candicans-36805947.jpg",
+    ...             },
+    ...             {"type": "text", "text": "These images depict two different landmarks. Can you identify them?"},
+    ...         ],
+    ...     },
+    ... ]
+
+    >>> inputs = processor.apply_chat_template(messages, add_generation_prompt=True, tokenize=True, return_dict=True, return_tensors="pt").to(torch_device)
+    >>> generate_ids = model.generate(**inputs, max_new_tokens=200)
+    >>> print(processor.decode(generate_ids[0, inputs["input_ids"].shape[1] :], skip_special_tokens=True))
+    The images depict the Statue of Liberty and the Golden Gate Bridge.
+    ```"""
+
+    output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+    output_hidden_states = (
+        output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+    )
+    return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+    vision_feature_layer = (
+        vision_feature_layer if vision_feature_layer is not None else self.config.vision_feature_layer
+    )
+    vision_feature_select_strategy = (
+        vision_feature_select_strategy
+        if vision_feature_select_strategy is not None
+        else self.config.vision_feature_select_strategy
+    )
+
+    outputs = self.model(
+        input_ids=input_ids,
+        pixel_values=pixel_values,
+        attention_mask=attention_mask,
+        position_ids=position_ids,
+        past_key_values=past_key_values,
+        inputs_embeds=inputs_embeds,
+        vision_feature_layer=vision_feature_layer,
+        vision_feature_select_strategy=vision_feature_select_strategy,
+        use_cache=use_cache,
+        output_attentions=output_attentions,
+        output_hidden_states=output_hidden_states,
+        return_dict=return_dict,
+        cache_position=cache_position,
+        image_sizes=image_sizes,
+        **lm_kwargs,
+    )
+
+    # Copied from llava.py
+    hidden_states = outputs[0]
+    # Only compute necessary logits, and do not upcast them to float if we are not computing the loss
+    slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep
+    kept_hidden_states = hidden_states[:, slice_indices, :]
+
+    shift_labels = lm_kwargs.pop("shift_labels", None)
+    logits = None
+    loss = None
+
+    if skip_logits and labels is None and shift_labels is None:
+        raise ValueError("skip_logits is True, but labels and shift_labels are None")
+
+    if skip_logits is None:
+        # By default, if in training mode, don't materialize logits
+        skip_logits = self.training and (labels is not None or shift_labels is not None)
+
+    if skip_logits:
+        loss = LigerForCausalLMLoss(
+            hidden_states=kept_hidden_states,
+            lm_head_weight=self.lm_head.weight,
+            labels=labels,
+            shift_labels=shift_labels,
+            hidden_size=self.config.text_config.hidden_size,
+            **lm_kwargs,
+        )
+
+    else:
+        logits = self.lm_head(kept_hidden_states)
+        if labels is not None:
+            loss = self.loss_function(
+                logits=logits, labels=labels, vocab_size=self.config.text_config.vocab_size, **lm_kwargs
+            )
+
+    if not return_dict:
+        output = (logits,) + outputs[1:]
+        return (loss,) + output if loss is not None else output
+
+    return InternVLCausalLMOutputWithPast(
+        loss=loss,
+        logits=logits,
+        past_key_values=outputs.past_key_values,
+        hidden_states=outputs.hidden_states,
+        attentions=outputs.attentions,
+        image_hidden_states=outputs.image_hidden_states,
+    )

liger_kernel/transformers/model/llama.py

@@ -248,10 +248,11 @@ def lce_forward(
 
     else:
         logits = self.lm_head(kept_hidden_states)
-        if labels is not None:
+        if labels is not None or shift_labels is not None:
             loss = self.loss_function(
                 logits=logits,
                 labels=labels,
+                shift_labels=shift_labels,
                 vocab_size=self.config.vocab_size,
                 **kwargs,
             )

liger_kernel/transformers/model/llama4.py

@@ -91,10 +91,11 @@ def lce_forward(
 
     else:  # if in inference mode materialize logits
         logits = self.lm_head(kept_hidden_states)
-        if labels is not None:
+        if labels is not None or shift_labels is not None:
             loss = self.loss_function(
                 logits=logits,
                 labels=labels,
+                shift_labels=shift_labels,
                 vocab_size=self.config.vocab_size,
                 **kwargs,
             )

liger_kernel/transformers/model/llava.py

@@ -313,9 +313,13 @@ def lce_forward(
 
     else:
         logits = self.lm_head(kept_hidden_states)
-        if labels is not None:
+        if labels is not None or shift_labels is not None:
             loss = self.loss_function(
-                logits=logits, labels=labels, vocab_size=self.config.text_config.vocab_size, **lm_kwargs
+                logits=logits,
+                labels=labels,
+                shift_labels=shift_labels,
+                vocab_size=self.config.text_config.vocab_size,
+                **lm_kwargs,
             )
 
     if not return_dict:

liger_kernel/transformers/model/mistral.py

@@ -115,10 +115,11 @@ def lce_forward(
         logits = self.lm_head(kept_hidden_states)
 
         loss = None
-        if labels is not None:
+        if labels is not None or shift_labels is not None:
             loss = self.loss_function(
                 logits=logits,
                 labels=labels,
+                shift_labels=shift_labels,
                 vocab_size=self.config.vocab_size,
                 **kwargs,
             )

liger_kernel/transformers/model/mixtral.py

@@ -248,8 +248,14 @@ def lce_forward(
         logits = self.lm_head(kept_hidden_states)
 
         loss = None
-        if labels is not None:
-            loss = self.loss_function(logits, labels, self.vocab_size, **kwargs)
+        if labels is not None or shift_labels is not None:
+            loss = self.loss_function(
+                logits=logits,
+                labels=labels,
+                shift_labels=shift_labels,
+                vocab_size=self.vocab_size,
+                **kwargs,
+            )
     aux_loss = None
     if output_router_logits:
         aux_loss = load_balancing_loss_func(

liger_kernel/transformers/model/mllama.py

@@ -239,10 +239,11 @@ def lce_forward(
 
     else:
         logits = self.lm_head(kept_hidden_states)
-        if labels is not None:
+        if labels is not None or shift_labels is not None:
             loss = self.loss_function(
                 logits=logits,
                 labels=labels,
+                shift_labels=shift_labels,
                 vocab_size=self.config.vocab_size,
                 **kwargs,
             )

liger_kernel/transformers/model/olmo2.py

@@ -111,10 +111,11 @@ def lce_forward(
 
     else:
         logits = self.lm_head(kept_hidden_states)
-        if labels is not None:
+        if labels is not None or shift_labels is not None:
             loss = self.loss_function(
                 logits=logits,
                 labels=labels,
+                shift_labels=shift_labels,
                 vocab_size=self.config.vocab_size,
                 **kwargs,
             )

liger_kernel/transformers/model/paligemma.py

@@ -379,6 +379,25 @@ def lce_forward(
             # Flatten the tokens
             loss_fct = CrossEntropyLoss()
 
+            flat_logits = shift_logits.view(-1, self.config.text_config.vocab_size)
+            flat_labels = shift_labels.view(-1).to(shift_logits.device)
+            loss = loss_fct(flat_logits, flat_labels)
+        elif shift_labels is not None:
+            # Upcast to float if we need to compute the loss to avoid potential precision issues
+            logits = logits.float()
+            shift_logits = logits[..., :-1, :]
+            if attention_mask is not None:
+                # we use the input attention mask to shift the logits and labels, because it is 2D.
+                # we also crop attn mask in case it is longer, which happens in PrefixTuning with peft
+                shift_attention_mask = attention_mask[:, -shift_logits.shape[1] :].to(logits.device)
+                shift_logits = shift_logits[shift_attention_mask.to(logits.device) != 0].contiguous()
+                shift_labels = shift_labels[shift_attention_mask.to(shift_labels.device) != 0].contiguous()
+            else:
+                shift_logits = shift_logits.contiguous()
+                shift_labels = shift_labels.contiguous()
+            # Flatten the tokens
+            loss_fct = CrossEntropyLoss()
+
             flat_logits = shift_logits.view(-1, self.config.text_config.vocab_size)
             flat_labels = shift_labels.view(-1).to(shift_logits.device)
             loss = loss_fct(flat_logits, flat_labels)

liger_kernel/transformers/model/phi3.py

@@ -91,10 +91,11 @@ def lce_forward(
 
     else:
         logits = self.lm_head(kept_hidden_states)
-        if labels is not None:
+        if labels is not None or shift_labels is not None:
             loss = self.loss_function(
                 logits=logits,
                 labels=labels,
+                shift_labels=shift_labels,
                 vocab_size=self.config.vocab_size,
                 **kwargs,
             )

liger_kernel/transformers/model/qwen2.py

@@ -228,10 +228,11 @@ def lce_forward(
 
     else:
         logits = self.lm_head(kept_hidden_states)
-        if labels is not None:
+        if labels is not None or shift_labels is not None:
             loss = self.loss_function(
                 logits=logits,
                 labels=labels,
+                shift_labels=shift_labels,
                 vocab_size=self.config.vocab_size,
                 **kwargs,
             )

liger_kernel/transformers/model/qwen2_5_vl.py

@@ -133,8 +133,13 @@ def lce_forward(
         logits = self.lm_head(hidden_states)
 
         loss = None
-        if labels is not None:
-            loss = self.loss_function(logits=logits, labels=labels, vocab_size=self.config.vocab_size)
+        if labels is not None or shift_labels is not None:
+            loss = self.loss_function(
+                logits=logits,
+                labels=labels,
+                shift_labels=shift_labels,
+                vocab_size=self.config.vocab_size,
+            )
 
     if not return_dict:
         output = (logits,) + outputs[1:]

liger_kernel/transformers/model/qwen2_vl.py

@@ -129,8 +129,13 @@ def lce_forward(
         logits = self.lm_head(hidden_states)
 
         loss = None
-        if labels is not None:
-            loss = self.loss_function(logits=logits, labels=labels, vocab_size=self.config.vocab_size)
+        if labels is not None or shift_labels is not None:
+            loss = self.loss_function(
+                logits=logits,
+                labels=labels,
+                shift_labels=shift_labels,
+                vocab_size=self.config.vocab_size,
+            )
 
     return Qwen2VLCausalLMOutputWithPast(
         loss=loss,

liger_kernel/transformers/model/qwen3.py

@@ -103,10 +103,11 @@ def lce_forward(
 
     else:
         logits = self.lm_head(kept_hidden_states)
-        if labels is not None:
+        if labels is not None or shift_labels is not None:
             loss = self.loss_function(
                 logits=logits,
                 labels=labels,
+                shift_labels=shift_labels,
                 vocab_size=self.config.vocab_size,
                 **kwargs,
             )

liger_kernel/transformers/model/qwen3_moe.py

@@ -107,8 +107,14 @@ def lce_forward(
         )
     else:  # if in inference model materialize logits
         logits = self.lm_head(kept_hidden_states)
-        if labels is not None:
-            loss = self.loss_function(logits, labels, self.vocab_size, **kwargs)
+        if labels is not None or shift_labels is not None:
+            loss = self.loss_function(
+                logits=logits,
+                labels=labels,
+                shift_labels=shift_labels,
+                vocab_size=self.vocab_size,
+                **kwargs,
+            )
 
     aux_loss = None
     if output_router_logits:

liger_kernel/transformers/model/smollm3.py

@@ -121,10 +121,11 @@ def lce_forward(
 
     else:
         logits = self.lm_head(kept_hidden_states)
-        if labels is not None:
+        if labels is not None or shift_labels is not None:
             loss = self.loss_function(
                 logits=logits,
                 labels=labels,
+                shift_labels=shift_labels,
                 vocab_size=self.config.vocab_size,
                 **kwargs,
             )

liger_kernel/transformers/monkey_patch.py

@@ -4,6 +4,7 @@ import logging
 from functools import partial
 from types import MethodType
 from typing import Callable
+from typing import Optional
 
 import transformers
 
@@ -1334,7 +1335,6 @@ def apply_liger_kernel_to_qwen2(
             if rms_norm:
                 _patch_rms_norm_module(decoder_layer.input_layernorm)
                 _patch_rms_norm_module(decoder_layer.post_attention_layernorm)
-    print("Applied Liger kernels to Qwen2")
 
 
 def apply_liger_kernel_to_qwen3(
@@ -2029,6 +2029,84 @@ def apply_liger_kernel_to_glm4v_moe(
                     _patch_rms_norm_module(decoder_layer.post_attention_layernorm)
 
 
+def apply_liger_kernel_to_internvl(
+    cross_entropy: bool = False,
+    fused_linear_cross_entropy: bool = True,
+    rms_norm: bool = True,
+    model: Optional[PreTrainedModel] = None,
+    **kwargs,
+) -> None:
+    """
+    Apply Liger kernels to replace original implementation in HuggingFace InternVL models.
+    Due to the characteristics of InternVL, the model must be passed to apply Liger-Kernel's patch to other models connected to InternVL.
+    However, if an LM not supported by Liger-Kernel is connected to InternVL, unexpected side effects may occur.
+    NOTE: InternVL is not available in transformers<4.52.1
+
+    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.internvl import modeling_internvl
+
+    from liger_kernel.transformers.model.internvl import lce_forward as internvl_lce_forward
+
+    if cross_entropy:
+        logger.warning(TRANSFORMER_DEPRECATION_WARNING)
+        modeling_internvl.nn.CrossEntropyLoss = LigerCrossEntropyLoss
+    if fused_linear_cross_entropy:
+        modeling_internvl.InternVLForConditionalGeneration.forward = internvl_lce_forward
+    if rms_norm:
+        modeling_internvl.InternVLVisionRMSNorm = LigerRMSNorm
+
+    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} | {"rms_norm": rms_norm}
+        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.")
+
+
 # Model type corresponds to the keys defined in transformers/models/auto/modeling_auto.py
 MODEL_TYPE_TO_APPLY_LIGER_FN = {
     "gemma": apply_liger_kernel_to_gemma,
@@ -2038,6 +2116,7 @@ MODEL_TYPE_TO_APPLY_LIGER_FN = {
     "glm4": apply_liger_kernel_to_glm4,
     "glm4v": apply_liger_kernel_to_glm4v,
     "glm4v_moe": apply_liger_kernel_to_glm4v_moe,
+    "internvl": apply_liger_kernel_to_internvl,
     "llama": apply_liger_kernel_to_llama,
     "llama4_text": apply_liger_kernel_to_llama4,
     "llama4": apply_liger_kernel_to_llama4,

{liger_kernel_nightly-0.6.2.dev20250923161735.dist-info → liger_kernel_nightly-0.6.2.dev20251008084122.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: liger_kernel_nightly
-Version: 0.6.2.dev20250923161735
+Version: 0.6.2.dev20251008084122
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 
@@ -311,6 +311,7 @@ loss.backward()
 | Granite 3.0 & 3.1   | `liger_kernel.transformers.apply_liger_kernel_to_granite`     | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss |
 | OLMo2   | `liger_kernel.transformers.apply_liger_kernel_to_olmo2`     | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy |
 | GLM-4   | `liger_kernel.transformers.apply_liger_kernel_to_glm4`     | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy |
+| InternVL3   | `liger_kernel.transformers.apply_liger_kernel_to_internvl`     | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy |
 
 
 ## Low-level APIs

{liger_kernel_nightly-0.6.2.dev20250923161735.dist-info → liger_kernel_nightly-0.6.2.dev20251008084122.dist-info}/RECORD

@@ -41,7 +41,7 @@ liger_kernel/ops/tvd.py,sha256=FHJtLQI95ijqgg9UtaHpMAjSCiPxB6CduPwPMcGxelc,6405
 liger_kernel/ops/utils.py,sha256=uoFKQqo-34N2TWQNvXMFywqGiOMMXNEVBxVojzlUAa0,3836
 liger_kernel/ops/experimental/embedding.py,sha256=tolj3tItkzpSb30zWqDN2_yX4ectflaQ8HMyKyFIQc8,4172
 liger_kernel/ops/experimental/mm_int8int2.py,sha256=TrS9lpwekrik_w5qE7AhMJD1bcq-OidjtbsW80oZ6IM,13314
-liger_kernel/transformers/__init__.py,sha256=GA4GdoP0VKKGzXoLN51DacuTtNuwciS96ZAblkDYl3w,8631
+liger_kernel/transformers/__init__.py,sha256=1HPWpJGAZl4TmPzG5kgvNMR7g9Ifvj5usTP4x-RsJ38,8819
 liger_kernel/transformers/auto_model.py,sha256=0qCTRZt280Bj_LcFdzo9hlaR-BWNazawXOGgoCZjgEg,1545
 liger_kernel/transformers/cross_entropy.py,sha256=z3KTWQnFxr_IZaVjtYt0ZNEWQdDdYThN35xWkHlDGH0,1683
 liger_kernel/transformers/dyt.py,sha256=i-4GPaMrl-jab9TVI5qN0-H9qycn_mCbV82ozU4nbmU,723
@@ -58,7 +58,7 @@ liger_kernel/transformers/jsd.py,sha256=DGqRnxIZxsvxo0_tbbxX3b-sDbDjC_yKufyRIHCc
 liger_kernel/transformers/kl_div.py,sha256=WLffFbh1EExD2Eb1F7lN11fo9JJC-0751WJjZAF1Fj8,409
 liger_kernel/transformers/layer_norm.py,sha256=c9pk3PEasOKYR0rhe5e5nNrnYKVCEW4VC8S6LpCq9EQ,906
 liger_kernel/transformers/llama4_rope.py,sha256=kS6PSHEwf3dS7hD7C7p8S0geugx2EMCiP0h0F7LsUoY,3639
-liger_kernel/transformers/monkey_patch.py,sha256=DpYF1mVjmErtkFgnxi20rnyeXr7bvVSg3BR166cMF5I,98472
+liger_kernel/transformers/monkey_patch.py,sha256=kGytMCLZdn-hPiNB_FhwojgBye0wmEKK6ZqsyKyCVAo,102745
 liger_kernel/transformers/multi_token_attention.py,sha256=K3NIY9_5TPgZ4_Rahn0xnkMXxD_fmlJHK4CWGYvGQp0,1752
 liger_kernel/transformers/qwen2vl_mrope.py,sha256=5EwSqrMdsL9MYspeBMXBsNJKvH0MOmRrtJXAJlnnlOI,1047
 liger_kernel/transformers/rms_norm.py,sha256=vkekcvTeWY8vL4H6hg3t0XeY0Ew_3OFMPHuzqlxPPVw,2719
@@ -71,35 +71,36 @@ liger_kernel/transformers/tvd.py,sha256=XrRfyJIqN6HFxXk8MYyFVZM1OLz3mtSbRZvWfZ_J
 liger_kernel/transformers/experimental/__init__.py,sha256=oQqk-f32JYgWEP9DJCj6ty6bbJSGrdXsFDQFwGeX6vI,127
 liger_kernel/transformers/experimental/embedding.py,sha256=2P0QYdlFyFrG5OqTzTa1wcRgDSyjBMv5i1a7BrDPDQw,881
 liger_kernel/transformers/model/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-liger_kernel/transformers/model/gemma.py,sha256=mNX-mIwV6jI4zfbrUHp0C468pOmjzsL7mjXipGt-eS0,10007
-liger_kernel/transformers/model/gemma2.py,sha256=R_JFPyWTk7RyA7D05ZiIaNO5pX8gWcvfWf-6rdCRMxs,11296
-liger_kernel/transformers/model/gemma3.py,sha256=FKO4j3t4W_5uECRA1lhVnXC-It2GhirHm4tpCf9ApAc,12785
-liger_kernel/transformers/model/glm4.py,sha256=GlnEhdGJuDIqp2R9qC54biY3HwV1tWmfpJm6ijoAsrM,5257
-liger_kernel/transformers/model/glm4v.py,sha256=dE9rRx1bOIr4T9xSXj58dtukaR80szeuclbgX0A0ovg,5951
-liger_kernel/transformers/model/glm4v_moe.py,sha256=Nij7tIYGkcBhJUMvQhYoFd1tVT4pbkxc2k4MSvN2pak,6166
-liger_kernel/transformers/model/llama.py,sha256=i8jJgyZsMKWQ-zKloETLugtwFpUOdaWxLDceciFXKd4,12832
-liger_kernel/transformers/model/llama4.py,sha256=IgbB8sTh3dlETQnaNNy1bZLuXy-Nt7qmeAjF27ydGpg,4210
-liger_kernel/transformers/model/llava.py,sha256=bLCioday_SOm69ogMDBhy_4UsVkH2-BSl93-EXY6-7I,15076
+liger_kernel/transformers/model/gemma.py,sha256=WryzpVmCm2H_XgLKNu3jJ6gVawjQDjapTetg4WHlbR4,10078
+liger_kernel/transformers/model/gemma2.py,sha256=eOQEfJBKezJNNrirhkPSagGxr9qj_y4lENOZgjUZKpE,11471
+liger_kernel/transformers/model/gemma3.py,sha256=-tvZw88S-STqmvdim-xrZZRJ17KLWoge_73ilIvhpIU,14157
+liger_kernel/transformers/model/glm4.py,sha256=2TBM5-4URpj6uX96G1AZ_DrjAmQtgLwXGzBvaXtfwdk,5328
+liger_kernel/transformers/model/glm4v.py,sha256=nlgEMOBjFEOu7a-cwwp9mWhTFqIs3QrOvcxW-uaPq-s,6022
+liger_kernel/transformers/model/glm4v_moe.py,sha256=q3-R_FoQPayS85AriJWWebblXB6Ix9fvxhSrI3mHiz4,6237
+liger_kernel/transformers/model/internvl.py,sha256=Uv8KGXOz9NhiKVZDeRNzAJH5kRuMZikUbswWM9u5KM0,6069
+liger_kernel/transformers/model/llama.py,sha256=L_VuaxxFJpzEmpLnaqwBbI5-Q14Qgfj-ufhLydCWgdk,12903
+liger_kernel/transformers/model/llama4.py,sha256=epEO_VD1gJCDovabSIQLxxncoh-TQTBfj-UgIlR5c7U,4281
+liger_kernel/transformers/model/llava.py,sha256=t6kMiyBkteVam-ltiod2f1mevj8l8ZHxYDvfu9C_lEk,15196
 liger_kernel/transformers/model/loss_utils.py,sha256=02RVkPI7Qs4ZP4yU_udCAvD_2hgIaHmxremRKe3N7EE,1885
-liger_kernel/transformers/model/mistral.py,sha256=syYNL8dLThX2-4uC13Lu0krEZ5zw3InviDUR3AJmc-I,5500
-liger_kernel/transformers/model/mixtral.py,sha256=VY-y73IyjcCyWyI7ahxXLw0fJrhgjYfr1xwRYtsHX0o,11396
-liger_kernel/transformers/model/mllama.py,sha256=NhJtlXiuszJHo5YSJOvSGYH47ly7Hse8r-5BKznBg9s,11522
-liger_kernel/transformers/model/olmo2.py,sha256=6L_bo-ZUgO1lYppdJneOtYxNIylQKS6BiGp13g7Uq9E,5259
-liger_kernel/transformers/model/paligemma.py,sha256=xuIx3oOwTgftU3jqLfWOxUxgCLBNJh0yNC21an9qDjo,18773
-liger_kernel/transformers/model/phi3.py,sha256=AwScxUe3LjmHHyQg4gW9bMoUI7uA6fUEMXJ3YhBiHtQ,4046
-liger_kernel/transformers/model/qwen2.py,sha256=3fpOTEOkniQmkCfN1KUa3KhseHJVzhj2Ht9FdYPUy-E,9962
-liger_kernel/transformers/model/qwen2_5_vl.py,sha256=zEVVwotCXnAm3RRc8-1Nc8uitSWrwW4B9dYY2uOZDwg,6331
-liger_kernel/transformers/model/qwen2_vl.py,sha256=5vK-vtCDpKZ2w33xYp2BS8kQYWUbKMqaiKvQcI27Mss,5884
-liger_kernel/transformers/model/qwen3.py,sha256=w2jBHuK9kK9EmOr5dnEIXNQXUgUSV_sJUkXSEwxLPHs,4885
-liger_kernel/transformers/model/qwen3_moe.py,sha256=BkpfFH3fOH0yRfA7LF-AoHTLut2GV0Y4MOlkiIYewfU,5511
-liger_kernel/transformers/model/smollm3.py,sha256=mqayvpwpMbp2yd_Ue7IPzy-dA4KHSDi_ROZW5vHCHfQ,7596
+liger_kernel/transformers/model/mistral.py,sha256=XmM4N21RIOkJ9PJ4PZ3DcRUhGUczn_lbx0plf1zeHb0,5571
+liger_kernel/transformers/model/mixtral.py,sha256=SLdLO81AZL7zror0LXLkn2PHqKzjwMMs4kALNqoaT00,11571
+liger_kernel/transformers/model/mllama.py,sha256=5q8q2BxQR_8hNZ83XrJIbndw-l6T7ZyFLM7OCv_uPK0,11593
+liger_kernel/transformers/model/olmo2.py,sha256=9O1Cze2B6ON-i1jgjQwjpS_WsDEK0PzL003s-MkevWA,5330
+liger_kernel/transformers/model/paligemma.py,sha256=mnTnSmEDla_bbVmPFmqhNVT__Cuf-TM-KLGFUa1sU-4,19967
+liger_kernel/transformers/model/phi3.py,sha256=L4gG8htOABmaxzcmHph0bBFCACRvL9r6wuDVFXi2o7Q,4117
+liger_kernel/transformers/model/qwen2.py,sha256=lgn0X6EzAZUhOv17ZDD9choIDdaPVIAsIrrdvwzWXqs,10033
+liger_kernel/transformers/model/qwen2_5_vl.py,sha256=Ea3zvL1FJfjlaerpeXCq-1zmorrajwNsR-XsgWr4fFQ,6465
+liger_kernel/transformers/model/qwen2_vl.py,sha256=ZeasFPGs-bxm2Y_E15mo0YNx5wwtKYDV-bjVKjkLPBk,6018
+liger_kernel/transformers/model/qwen3.py,sha256=Q2aOg5erPrgVgRcqJm8sefLSDtvU1AD5B7aJnP7mRMM,4956
+liger_kernel/transformers/model/qwen3_moe.py,sha256=1CwTMCNFDYsjGoa_aHFBagtC5HuJTV-s0__5UvcjD3A,5686
+liger_kernel/transformers/model/smollm3.py,sha256=0KWVkDtXbjsBKhJnaquV6vUUYyLtfmNwYH0sxJt-qTk,7667
 liger_kernel/transformers/trainer/__init__.py,sha256=p7yQfklV8-467qSz_ZMimkbDF7HHWHwku25A-GYL0WU,193
 liger_kernel/transformers/trainer/orpo_trainer.py,sha256=tX0h63aOFe3rNqTmk6JpMf75UPo981yzEa6TghnjS0Q,5370
 liger_kernel/triton/__init__.py,sha256=qCiCamzCRv6lpV8IqpAc9YMdNKC7GKurClWceQPnlis,92
 liger_kernel/triton/monkey_patch.py,sha256=Rd0hUHAzDkFfHvnX7-PBaNK5EKnZhtfM_h-fgQH9HPY,1568
-liger_kernel_nightly-0.6.2.dev20250923161735.dist-info/LICENSE,sha256=OhzLDHJ0to4a8sodVLELZiCFylZ1NAAYLs-HrjPy0ag,1312
-liger_kernel_nightly-0.6.2.dev20250923161735.dist-info/METADATA,sha256=Zqk8XfPYutZA2NQgoD6XKfMFbCEwY8sEKDbM_2LwLRs,24629
-liger_kernel_nightly-0.6.2.dev20250923161735.dist-info/NOTICE,sha256=njwnoPZLh9AN8SJQzxvCGLHi-8X__AvWRze6joNXIY8,2066
-liger_kernel_nightly-0.6.2.dev20250923161735.dist-info/WHEEL,sha256=iAkIy5fosb7FzIOwONchHf19Qu7_1wCWyFNR5gu9nU0,91
-liger_kernel_nightly-0.6.2.dev20250923161735.dist-info/top_level.txt,sha256=2eghu4hA3LnkM7ElW92tQ8zegWKgSbeo-k-aGe1YnvY,13
-liger_kernel_nightly-0.6.2.dev20250923161735.dist-info/RECORD,,
+liger_kernel_nightly-0.6.2.dev20251008084122.dist-info/LICENSE,sha256=OhzLDHJ0to4a8sodVLELZiCFylZ1NAAYLs-HrjPy0ag,1312
+liger_kernel_nightly-0.6.2.dev20251008084122.dist-info/METADATA,sha256=Uk7nue5xFD5LoCoKeAD8zHpI2-c-taOdvk1KyBdBUXE,24777
+liger_kernel_nightly-0.6.2.dev20251008084122.dist-info/NOTICE,sha256=njwnoPZLh9AN8SJQzxvCGLHi-8X__AvWRze6joNXIY8,2066
+liger_kernel_nightly-0.6.2.dev20251008084122.dist-info/WHEEL,sha256=iAkIy5fosb7FzIOwONchHf19Qu7_1wCWyFNR5gu9nU0,91
+liger_kernel_nightly-0.6.2.dev20251008084122.dist-info/top_level.txt,sha256=2eghu4hA3LnkM7ElW92tQ8zegWKgSbeo-k-aGe1YnvY,13
+liger_kernel_nightly-0.6.2.dev20251008084122.dist-info/RECORD,,