liger-kernel-nightly 0.5.10.dev20250610174206__py3-none-any.whl → 0.5.10.dev20250611064616__py3-none-any.whl

liger_kernel/transformers/model/qwen2_5_vl.py

@@ -5,12 +5,13 @@ from typing import Union
 
 import torch
 
-from torch.nn import CrossEntropyLoss
 from transformers.models.qwen2_5_vl.modeling_qwen2_5_vl import Qwen2_5_VLCausalLMOutputWithPast
+from transformers.utils import can_return_tuple
 
 from liger_kernel.transformers.model.loss_utils import LigerForCausalLMLoss
 
 
+@can_return_tuple
 def lce_forward(
     self,
     input_ids: torch.LongTensor = None,
@@ -34,14 +35,22 @@ def lce_forward(
     **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:
+    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]`.
+    pixel_values_videos (`torch.FloatTensor` of shape `(seq_length, num_channels * temporal_size * image_size * image_size)):
+        The tensors corresponding to the input videos. Pixel values can be obtained using
+        [`AutoImageProcessor`]. See [`Qwen2_5_VLImageProcessor.__call__`] for details. [`Qwen2_5_VLProcessor`] uses
+        [`Qwen2_5_VLImageProcessor`] for processing videos.
+    image_grid_thw (`torch.LongTensor` of shape `(num_images, 3)`, *optional*):
+        The temporal, height and width of feature shape of each image in LLM.
+    video_grid_thw (`torch.LongTensor` of shape `(num_videos, 3)`, *optional*):
+        The temporal, height and width of feature shape of each video in LLM.
+    rope_deltas (`torch.LongTensor` of shape `(batch_size, )`, *optional*):
+        The rope index difference between sequence length and multimodal rope.
+    second_per_grid_ts (`torch.Tensor` of shape `(num_videos)`, *optional*):
+        The time interval (in seconds) for each grid along the temporal dimension in the 3D position IDs.
 
     Example:
 
@@ -73,78 +82,20 @@ def lce_forward(
     >>> 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,
+        input_ids=input_ids,
+        pixel_values=pixel_values,
+        pixel_values_videos=pixel_values_videos,
+        image_grid_thw=image_grid_thw,
+        video_grid_thw=video_grid_thw,
+        second_per_grid_ts=second_per_grid_ts,
         position_ids=position_ids,
         attention_mask=attention_mask,
         past_key_values=past_key_values,
@@ -180,19 +131,10 @@ def lce_forward(
         )
     else:
         logits = self.lm_head(hidden_states)
+
+        loss = None
         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)
+            loss = self.loss_function(logits=logits, labels=labels, vocab_size=self.config.vocab_size)
 
     if not return_dict:
         output = (logits,) + outputs[1:]
@@ -204,5 +146,5 @@ def lce_forward(
         past_key_values=outputs.past_key_values,
         hidden_states=outputs.hidden_states,
         attentions=outputs.attentions,
-        rope_deltas=rope_deltas,
+        rope_deltas=outputs.rope_deltas,
     )

liger_kernel/transformers/model/qwen2_vl.py

@@ -5,14 +5,13 @@ from typing import Union
 
 import torch
 
-from packaging import version
-from torch.nn import CrossEntropyLoss
-from transformers import __version__ as transformers_version
 from transformers.models.qwen2_vl.modeling_qwen2_vl import Qwen2VLCausalLMOutputWithPast
+from transformers.utils import can_return_tuple
 
 from liger_kernel.transformers.model.loss_utils import LigerForCausalLMLoss
 
 
+@can_return_tuple
 def lce_forward(
     self,
     input_ids: torch.LongTensor = None,
@@ -35,15 +34,20 @@ def lce_forward(
     **kwargs,
 ) -> Union[Tuple, Qwen2VLCausalLMOutputWithPast]:
     r"""
-    Copy paste Qwen2VL'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:
+    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]`.
+    pixel_values_videos (`torch.FloatTensor` of shape `(seq_length, num_channels * temporal_size * image_size * image_size)):
+        The tensors corresponding to the input videos. Pixel values can be obtained using
+        [`AutoImageProcessor`]. See [`Qwen2VLImageProcessor.__call__`] for details. [`Qwen2VLProcessor`] uses
+        [`Qwen2VLImageProcessor`] for processing videos.
+    image_grid_thw (`torch.LongTensor` of shape `(num_images, 3)`, *optional*):
+        The temporal, height and width of feature shape of each image in LLM.
+    video_grid_thw (`torch.LongTensor` of shape `(num_videos, 3)`, *optional*):
+        The temporal, height and width of feature shape of each video in LLM.
+    rope_deltas (`torch.LongTensor` of shape `(batch_size, )`, *optional*):
+        The rope index difference between sequence length and multimodal rope.
 
     Example:
 
@@ -75,80 +79,19 @@ def lce_forward(
     >>> 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.get_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}"
-                )
-            image_mask = (
-                (input_ids == self.config.image_token_id)
-                .unsqueeze(-1)
-                .expand_as(inputs_embeds)
-                .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.get_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}"
-                )
-            video_mask = (
-                (input_ids == self.config.video_token_id)
-                .unsqueeze(-1)
-                .expand_as(inputs_embeds)
-                .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 version.parse(transformers_version) > version.parse("4.46.3"):
-        # NOTE: this bug fix for qwen2-vl is not applied until transformers 4.47.0
-        # https://github.com/huggingface/transformers/issues/33401
-        # While correct, this breaks equivalence with past versions of Qwen2-VL from
-        # transformers and leads to failed tests or users noticing differences in results.
-        # TODO: remove above conditional when liger drops support for transformers<4.47.0
-        # 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, 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 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,
+        input_ids=input_ids,
+        pixel_values=pixel_values,
+        pixel_values_videos=pixel_values_videos,
+        image_grid_thw=image_grid_thw,
+        video_grid_thw=video_grid_thw,
         position_ids=position_ids,
         attention_mask=attention_mask,
         past_key_values=past_key_values,
@@ -184,23 +127,10 @@ def lce_forward(
         )
     else:
         logits = self.lm_head(hidden_states)
+
+        loss = None
         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
+            loss = self.loss_function(logits=logits, labels=labels, vocab_size=self.config.vocab_size)
 
     return Qwen2VLCausalLMOutputWithPast(
         loss=loss,
@@ -208,5 +138,5 @@ def lce_forward(
         past_key_values=outputs.past_key_values,
         hidden_states=outputs.hidden_states,
         attentions=outputs.attentions,
-        rope_deltas=rope_deltas,
+        rope_deltas=outputs.rope_deltas,
     )

liger_kernel/transformers/monkey_patch.py

@@ -1225,7 +1225,7 @@ def apply_liger_kernel_to_qwen2_vl(
 ) -> None:
     """
     Apply Liger kernels to replace original implementation in HuggingFace Qwen2-VL models.
-    NOTE: Qwen2-VL is not available in transformers<4.45.0
+    NOTE: Qwen2-VL is not supported in transformers<4.52.4
 
     Args:
         cross_entropy (bool): Whether to apply Liger's cross entropy loss. Default is False.
@@ -1239,12 +1239,19 @@ def apply_liger_kernel_to_qwen2_vl(
         model (PreTrainedModel): The model instance to apply Liger kernels to, if the model has already been
         loaded. Default is None.
     """
+    if transformer_version < version.parse("4.52.4"):
+        logger.warning("Qwen2-VL support is only compatible with transformers >= 4.52.4")
+        return
+
     assert not (cross_entropy and fused_linear_cross_entropy), (
         "cross_entropy and fused_linear_cross_entropy cannot both be True."
     )
 
     from transformers.models.qwen2_vl import modeling_qwen2_vl
+    from transformers.models.qwen2_vl.modeling_qwen2_vl import Qwen2VisionTransformerPretrainedModel
+    from transformers.models.qwen2_vl.modeling_qwen2_vl import Qwen2VLForConditionalGeneration
     from transformers.models.qwen2_vl.modeling_qwen2_vl import Qwen2VLModel
+    from transformers.models.qwen2_vl.modeling_qwen2_vl import Qwen2VLTextModel
 
     from liger_kernel.transformers.model.qwen2_vl import lce_forward as qwen2_vl_lce_forward
 
@@ -1266,24 +1273,38 @@ def apply_liger_kernel_to_qwen2_vl(
         # 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: Qwen2VLModel = getattr(model, model.base_model_prefix, model)
+        if isinstance(model, (Qwen2VLForConditionalGeneration, Qwen2VLModel)):
+            # Note: language_model and visual properties can be accessed throught conditional class for BC.
+            # Not sure if it is subject to changes in the future.
+            # Reference: https://github.com/huggingface/transformers/blob/v4.52.4/src/transformers/models/qwen2_vl/modeling_qwen2_vl.py#L1698
+            text_model: Qwen2VLTextModel = model.language_model
+            vision_model: Qwen2VisionTransformerPretrainedModel = model.visual
+        elif isinstance(model, Qwen2VLTextModel):
+            text_model: Qwen2VLTextModel = model
+            vision_model = None
+        else:
+            # Note: Currently there's no support for patching vision model only. Feel free to raise an issue if needed.
+            raise TypeError(
+                f"Unsupported Qwen2VL model type. `model` must be `Qwen2VLForConditionalGeneration`, `Qwen2VLModel` or `Qwen2VLTextModel`. Got: {type(model)}"
+            )
 
-        if hasattr(model, "visual"):
-            # Patch Qwen2VisionTransformerPretrainedModel
-            for vision_block in model.visual.blocks:
+        # Patch Qwen2VisionTransformerPretrainedModel
+        if vision_model is not None:
+            for vision_block in vision_model.blocks:
                 if layer_norm:
                     _patch_layer_norm_module(vision_block.norm1)
                     _patch_layer_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)
+        # Patch Qwen2VisionTextModel
+        if text_model is not None:
             if rms_norm:
-                _patch_rms_norm_module(decoder_layer.input_layernorm)
-                _patch_rms_norm_module(decoder_layer.post_attention_layernorm)
+                _patch_rms_norm_module(text_model.norm)
+            for decoder_layer in text_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)
 
 
 def apply_liger_kernel_to_qwen2_5_vl(
@@ -1309,12 +1330,19 @@ def apply_liger_kernel_to_qwen2_5_vl(
         model (PreTrainedModel): The model instance to apply Liger kernels to, if the model has already been
         loaded. Default is None.
     """
+    if transformer_version < version.parse("4.52.4"):
+        logger.warning("Qwen2.5-VL support is only compatible with transformers >= 4.52.4")
+        return
+
     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_VisionTransformerPretrainedModel
+    from transformers.models.qwen2_5_vl.modeling_qwen2_5_vl import Qwen2_5_VLForConditionalGeneration
     from transformers.models.qwen2_5_vl.modeling_qwen2_5_vl import Qwen2_5_VLModel
+    from transformers.models.qwen2_5_vl.modeling_qwen2_5_vl import Qwen2_5_VLTextModel
 
     from liger_kernel.transformers.model.qwen2_5_vl import lce_forward as qwen2_5_vl_lce_forward
 
@@ -1333,24 +1361,37 @@ def apply_liger_kernel_to_qwen2_5_vl(
         # 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 isinstance(model, (Qwen2_5_VLForConditionalGeneration, Qwen2_5_VLModel)):
+            # Note: language_model and visual properties can be accessed throught conditional class for BC.
+            # Not sure if it is subject to changes in the future.
+            # Reference: https://github.com/huggingface/transformers/blob/v4.52.4/src/transformers/models/qwen2_5_vl/modeling_qwen2_5_vl.py#L1823
+            text_model: Qwen2_5_VLTextModel = model.language_model
+            vision_model: Qwen2_5_VisionTransformerPretrainedModel = model.visual
+        elif isinstance(model, Qwen2_5_VLTextModel):
+            text_model: Qwen2_5_VLTextModel = model
+            vision_model = None
+        else:
+            # Note: Currently there's no support for patching vision model only. Feel free to raise an issue if needed.
+            raise TypeError(
+                f"Unsupported Qwen2VL model type. `model` must be `Qwen2VLForConditionalGeneration`, `Qwen2VLModel` or `Qwen2VLTextModel`. Got: {type(model)}"
+            )
 
-        if hasattr(model, "visual"):
+        if vision_model is not None:
             # 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 text_model is not None:
             if rms_norm:
-                _patch_rms_norm_module(decoder_layer.input_layernorm)
-                _patch_rms_norm_module(decoder_layer.post_attention_layernorm)
+                _patch_rms_norm_module(text_model.norm)
+            for decoder_layer in text_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)
 
 
 def apply_liger_kernel_to_phi3(
@@ -1571,7 +1612,9 @@ MODEL_TYPE_TO_APPLY_LIGER_FN = {
     "qwen3": apply_liger_kernel_to_qwen3,
     "qwen3_moe": apply_liger_kernel_to_qwen3_moe,
     "qwen2_vl": apply_liger_kernel_to_qwen2_vl,
+    "qwen2_vl_text": apply_liger_kernel_to_qwen2_vl,
     "qwen2_5_vl": apply_liger_kernel_to_qwen2_5_vl,
+    "qwen2_5_vl_text": apply_liger_kernel_to_qwen2_5_vl,
     "phi3": apply_liger_kernel_to_phi3,
     "paligemma": apply_liger_kernel_to_paligemma,
 }

{liger_kernel_nightly-0.5.10.dev20250610174206.dist-info → liger_kernel_nightly-0.5.10.dev20250611064616.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: liger_kernel_nightly
-Version: 0.5.10.dev20250610174206
+Version: 0.5.10.dev20250611064616
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 

{liger_kernel_nightly-0.5.10.dev20250610174206.dist-info → liger_kernel_nightly-0.5.10.dev20250611064616.dist-info}/RECORD

@@ -53,7 +53,7 @@ liger_kernel/transformers/grpo_loss.py,sha256=uAkUNKSnUGEOqa82L9w2e6AI1kcmG8K45-
 liger_kernel/transformers/jsd.py,sha256=DGqRnxIZxsvxo0_tbbxX3b-sDbDjC_yKufyRIHCcScY,2979
 liger_kernel/transformers/kl_div.py,sha256=WLffFbh1EExD2Eb1F7lN11fo9JJC-0751WJjZAF1Fj8,409
 liger_kernel/transformers/layer_norm.py,sha256=c9pk3PEasOKYR0rhe5e5nNrnYKVCEW4VC8S6LpCq9EQ,906
-liger_kernel/transformers/monkey_patch.py,sha256=zeqmbU__X965iSZ4ZO0Zq3kq6qvqfSgU7B3acxynL3Y,74605
+liger_kernel/transformers/monkey_patch.py,sha256=IWqNiimHL0895yo0TjQ3lN_Y8fKGesxC-bF5He6zB2g,77536
 liger_kernel/transformers/multi_token_attention.py,sha256=l9VDICK0dfmifUDW668hGscP8AHq2rYcM2oGUa3baRQ,1751
 liger_kernel/transformers/qwen2vl_mrope.py,sha256=5EwSqrMdsL9MYspeBMXBsNJKvH0MOmRrtJXAJlnnlOI,1047
 liger_kernel/transformers/rms_norm.py,sha256=eErIr1n-13oVrc1VJY07lqazYelw_vlu9Az__RmXPSE,2717
@@ -79,17 +79,17 @@ liger_kernel/transformers/model/olmo2.py,sha256=6L_bo-ZUgO1lYppdJneOtYxNIylQKS6B
 liger_kernel/transformers/model/paligemma.py,sha256=xuIx3oOwTgftU3jqLfWOxUxgCLBNJh0yNC21an9qDjo,18773
 liger_kernel/transformers/model/phi3.py,sha256=m-MD_OuTaYMGZhHOvl-RHOVEObrL8tL5cBv3VTNd4F0,10376
 liger_kernel/transformers/model/qwen2.py,sha256=SdN7V-MI3eX9s2DAFRvC1g-G146uG_5n1fnNdY9QwYk,9658
-liger_kernel/transformers/model/qwen2_5_vl.py,sha256=k6jt1bTCJsKsZVGhTxqIbDzmnL8-B3CpWJOjLazswbo,9203
-liger_kernel/transformers/model/qwen2_vl.py,sha256=Cgs7-nPlKFifiDO9gqSI6np4vRUVCKiqoospT_vIi_M,9614
+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/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-0.5.10.dev20250610174206.dist-info/LICENSE,sha256=OhzLDHJ0to4a8sodVLELZiCFylZ1NAAYLs-HrjPy0ag,1312
-liger_kernel_nightly-0.5.10.dev20250610174206.dist-info/METADATA,sha256=T6CCI8j-_GLD4_OTFov5VFLiGK7sITnt6Ht6zVDPhqw,24309
-liger_kernel_nightly-0.5.10.dev20250610174206.dist-info/NOTICE,sha256=njwnoPZLh9AN8SJQzxvCGLHi-8X__AvWRze6joNXIY8,2066
-liger_kernel_nightly-0.5.10.dev20250610174206.dist-info/WHEEL,sha256=iAkIy5fosb7FzIOwONchHf19Qu7_1wCWyFNR5gu9nU0,91
-liger_kernel_nightly-0.5.10.dev20250610174206.dist-info/top_level.txt,sha256=2eghu4hA3LnkM7ElW92tQ8zegWKgSbeo-k-aGe1YnvY,13
-liger_kernel_nightly-0.5.10.dev20250610174206.dist-info/RECORD,,
+liger_kernel_nightly-0.5.10.dev20250611064616.dist-info/LICENSE,sha256=OhzLDHJ0to4a8sodVLELZiCFylZ1NAAYLs-HrjPy0ag,1312
+liger_kernel_nightly-0.5.10.dev20250611064616.dist-info/METADATA,sha256=49fgwei-BXjeHZEaeZCYnr3xfKN_psmRALV3GmK6YCk,24309
+liger_kernel_nightly-0.5.10.dev20250611064616.dist-info/NOTICE,sha256=njwnoPZLh9AN8SJQzxvCGLHi-8X__AvWRze6joNXIY8,2066
+liger_kernel_nightly-0.5.10.dev20250611064616.dist-info/WHEEL,sha256=iAkIy5fosb7FzIOwONchHf19Qu7_1wCWyFNR5gu9nU0,91
+liger_kernel_nightly-0.5.10.dev20250611064616.dist-info/top_level.txt,sha256=2eghu4hA3LnkM7ElW92tQ8zegWKgSbeo-k-aGe1YnvY,13
+liger_kernel_nightly-0.5.10.dev20250611064616.dist-info/RECORD,,