diff --git a/‎src/liger_kernel/transformers/model/qwen2_5_vl.py‎
Lines changed: 29 additions & 87 deletions b/‎src/liger_kernel/transformers/model/qwen2_5_vl.py‎
Lines changed: 29 additions & 87 deletions
diff --git a/‎src/liger_kernel/transformers/model/qwen2_vl.py‎
Lines changed: 26 additions & 96 deletions b/‎src/liger_kernel/transformers/model/qwen2_vl.py‎
Lines changed: 26 additions & 96 deletions
@@ -5,12 +5,13 @@
 
 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,
     )
@@ -5,14 +5,13 @@
 
 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,
@@ -184,29 +127,16 @@ 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,
         logits=logits,
         past_key_values=outputs.past_key_values,
         hidden_states=outputs.hidden_states,
         attentions=outputs.attentions,
-        rope_deltas=rope_deltas,
+        rope_deltas=outputs.rope_deltas,
     )