liger-kernel 0.6.1__py3-none-any.whl → 0.6.3__py3-none-any.whl

liger_kernel/transformers/model/phi3.py

@@ -5,131 +5,12 @@ from typing import Union
 
 import torch
 
-from torch.nn import CrossEntropyLoss
+from transformers.modeling_outputs import BaseModelOutputWithPast
 from transformers.modeling_outputs import CausalLMOutputWithPast
-from transformers.utils.deprecation import deprecate_kwarg
 
-from liger_kernel.transformers.fused_linear_cross_entropy import LigerFusedLinearCrossEntropyLoss
 from liger_kernel.transformers.model.loss_utils import LigerForCausalLMLoss
 
 
-def lce_forward_deprecated(
-    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,
-    cache_position: Optional[torch.LongTensor] = None,
-    skip_logits: Optional[bool] = None,
-) -> Union[Tuple, CausalLMOutputWithPast]:
-    r"""
-    Copy paste phi3 forward from transfomers v4.44.2 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 transformers import AutoTokenizer, Phi3ForCausalLM
-
-    >>> model = Phi3ForCausalLM.from_pretrained("microsoft/phi-3-mini-4k-instruct")
-    >>> tokenizer = AutoTokenizer.from_pretrained("microsoft/phi-3-mini-4k-instruct")
-
-    >>> prompt = "This is an example script ."
-    >>> inputs = tokenizer(prompt, return_tensors="pt")
-
-    >>> # 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]
-    'This is an example script .\n Certainly! Below is a sample script that demonstrates a simple task, such as calculating the sum'
-    ```"""
-
-    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
-
-    # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
-    outputs = self.model(
-        input_ids=input_ids,
-        attention_mask=attention_mask,
-        position_ids=position_ids,
-        past_key_values=past_key_values,
-        inputs_embeds=inputs_embeds,
-        use_cache=use_cache,
-        output_attentions=output_attentions,
-        output_hidden_states=output_hidden_states,
-        return_dict=return_dict,
-    )
-
-    hidden_states = outputs[0]
-
-    loss = None
-    logits = None
-
-    if skip_logits and labels is None:
-        raise ValueError("skip_logits is True, but labels is 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
-
-    if skip_logits:
-        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)
-    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
-
-    return CausalLMOutputWithPast(
-        loss=loss,
-        logits=logits,
-        past_key_values=outputs.past_key_values,
-        hidden_states=outputs.hidden_states,
-        attentions=outputs.attentions,
-    )
-
-
-@deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
 def lce_forward(
     self,
     input_ids: torch.LongTensor = None,
@@ -148,73 +29,41 @@ def lce_forward(
     **kwargs,
 ) -> Union[Tuple, CausalLMOutputWithPast]:
     r"""
-    Args:
-        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
-            Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
-            config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
-            (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
-
-        logits_to_keep (`int` or `torch.Tensor`, *optional*):
-            If an `int`, compute logits for the last `logits_to_keep` tokens. If `0`, calculate logits for all
-            `input_ids` (special case). Only last token logits are needed for generation, and calculating them only for that
-            token can save memory, which becomes pretty significant for long sequences or large vocabulary size.
-            If a `torch.Tensor`, must be 1D corresponding to the indices to keep in the sequence length dimension.
-            This is useful when using packed tensor format (single dimension for batch and sequence length).
-
-    Returns:
-
     Example:
 
     ```python
     >>> from transformers import AutoTokenizer, Phi3ForCausalLM
 
-    >>> model = Phi3ForCausalLM.from_pretrained("microsoft/phi-3-mini-4k-instruct")
-    >>> tokenizer = AutoTokenizer.from_pretrained("microsoft/phi-3-mini-4k-instruct")
+    >>> model = Phi3ForCausalLM.from_pretrained("meta-phi3/Phi3-2-7b-hf")
+    >>> tokenizer = AutoTokenizer.from_pretrained("meta-phi3/Phi3-2-7b-hf")
 
-    >>> prompt = "This is an example script ."
+    >>> prompt = "Hey, are you conscious? Can you talk to me?"
     >>> inputs = tokenizer(prompt, return_tensors="pt")
 
     >>> # 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]
-    'This is an example script .\n Certainly! Below is a sample script that demonstrates a simple task, such as calculating the sum'
+    "Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you."
     ```"""
 
-    from transformers.models.phi3.modeling_phi3 import logging
-
-    logger = logging.get_logger(__name__)
-
-    if (
-        use_cache
-        and self.config.rope_scaling
-        and cache_position is not None
-        and cache_position[0] == self.config.original_max_position_embeddings
-    ):
-        logger.warning(
-            f"If you are not using the generate method, you may encounter nonsensical outputs after the {self.config.original_max_position_embeddings}th token, as the KV cache needs to be recomputed."
-        )
-
     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
 
-    # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
-    outputs = self.model(
+    outputs: BaseModelOutputWithPast = self.model(
         input_ids=input_ids,
         attention_mask=attention_mask,
         position_ids=position_ids,
         past_key_values=past_key_values,
         inputs_embeds=inputs_embeds,
         use_cache=use_cache,
-        output_attentions=output_attentions,
-        output_hidden_states=output_hidden_states,
-        return_dict=return_dict,
+        cache_position=cache_position,
         **kwargs,
     )
 
-    hidden_states = outputs[0]
+    hidden_states = outputs.last_hidden_state
     # 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, :]
@@ -242,10 +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/qwen3_next.py

@@ -0,0 +1,134 @@
+from typing import TYPE_CHECKING
+from typing import List
+from typing import Optional
+from typing import Union
+
+import torch
+
+from transformers.modeling_outputs import MoeCausalLMOutputWithPast
+from transformers.modeling_outputs import MoeModelOutputWithPast
+
+if TYPE_CHECKING:
+    from transformers.models.qwen3_next.modeling_qwen3_next import load_balancing_loss_func
+
+from liger_kernel.transformers.model.loss_utils import LigerForCausalLMLoss
+
+
+def lce_forward(
+    self,
+    input_ids: Optional[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,
+    output_router_logits: Optional[bool] = None,
+    cache_position: Optional[torch.LongTensor] = None,
+    logits_to_keep: Union[int, torch.Tensor] = 0,
+    skip_logits: Optional[bool] = None,
+    **kwargs,
+) -> MoeCausalLMOutputWithPast:
+    r"""
+        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
+            config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
+            (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
+
+        logits_to_keep (`int` or `torch.Tensor`, *optional*):
+            If an `int`, compute logits for the last `logits_to_keep` tokens. If `0`, calculate logits for all
+            `input_ids` (special case). Only last token logits are needed for generation, and calculating them only for that
+            token can save memory, which becomes pretty significant for long sequences or large vocabulary size.
+            If a `torch.Tensor`, must be 1D corresponding to the indices to keep in the sequence length dimension.
+            This is useful when using packed tensor format (single dimension for batch and sequence length).
+
+    Returns:
+
+    Example:
+
+    ```python
+    >>> from transformers import AutoModelForCausalLM, AutoTokenizer
+
+    >>> model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-Next-80B-A3B-Instruct")
+    >>> tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-Next-80B-A3B-Instruct")
+
+    >>> prompt = "Give me a short introduction to large language model."
+    >>> inputs = tokenizer(prompt, return_tensors="pt")
+
+    >>> # 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]
+    "Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you."
+    ```"""
+    output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+    output_router_logits = (
+        output_router_logits if output_router_logits is not None else self.config.output_router_logits
+    )
+
+    output_hidden_states = (
+        output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+    )
+
+    # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
+    outputs: MoeModelOutputWithPast = self.model(
+        input_ids=input_ids,
+        attention_mask=attention_mask,
+        position_ids=position_ids,
+        past_key_values=past_key_values,
+        inputs_embeds=inputs_embeds,
+        use_cache=use_cache,
+        output_attentions=output_attentions,
+        output_hidden_states=output_hidden_states,
+        output_router_logits=output_router_logits,
+        cache_position=cache_position,
+        **kwargs,
+    )
+
+    hidden_states = outputs.last_hidden_state
+    # 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 = kwargs.pop("shift_labels", None)
+    logits = None
+    loss = None
+
+    if skip_logits is None:
+        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.hidden_size,
+            **kwargs,
+        )
+    else:  # if in inference model materialize logits
+        logits = self.lm_head(kept_hidden_states)
+        if labels is not None or shift_labels is not None:
+            loss = self.loss_function(logits, labels, self.vocab_size, **kwargs)
+
+    aux_loss = None
+    if output_router_logits:
+        aux_loss = load_balancing_loss_func(
+            outputs.router_logits,
+            self.num_experts,
+            self.num_experts_per_tok,
+            attention_mask,
+        )
+        if labels is not None:
+            loss += self.router_aux_loss_coef * aux_loss.to(loss.device)  # make sure to reside in the same device
+
+    return MoeCausalLMOutputWithPast(
+        loss=loss,
+        aux_loss=aux_loss,
+        logits=logits,
+        past_key_values=outputs.past_key_values,
+        hidden_states=outputs.hidden_states,
+        attentions=outputs.attentions,
+        router_logits=outputs.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/model/smolvlm.py

@@ -0,0 +1,158 @@
+from typing import TYPE_CHECKING
+from typing import Optional
+from typing import Union
+
+import torch
+
+from transformers.models.smolvlm.modeling_smolvlm import SmolVLMCausalLMOutputWithPast
+from transformers.processing_utils import Unpack
+from transformers.utils.generic import can_return_tuple
+
+from liger_kernel.transformers.model.loss_utils import LigerForCausalLMLoss
+
+if TYPE_CHECKING:
+    from transformers.cache_utils import Cache
+    from transformers.utils.generic import TransformersKwargs
+
+
+# Forward adapted to enable fused Linear + CE without materializing logits.
+# Mirrors the pattern used for other multimodal models (e.g., InternVL, LLaVA).
+@can_return_tuple
+def lce_forward(
+    self,
+    input_ids: Optional[torch.LongTensor] = None,
+    attention_mask: Optional[torch.Tensor] = None,
+    position_ids: Optional[torch.LongTensor] = None,
+    past_key_values: Optional["Cache"] = None,
+    inputs_embeds: Optional[torch.FloatTensor] = None,
+    pixel_values: Optional[torch.FloatTensor] = None,
+    pixel_attention_mask: Optional[torch.BoolTensor] = None,
+    image_hidden_states: 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,
+    cache_position: Optional[torch.LongTensor] = None,
+    return_dict: Optional[bool] = None,
+    logits_to_keep: Union[int, torch.Tensor] = 0,
+    skip_logits: Optional[bool] = None,  # Added argument for liger-kernel
+    **lm_kwargs: Unpack["TransformersKwargs"],  # renamed from kwargs
+) -> Union[tuple, SmolVLMCausalLMOutputWithPast]:
+    r"""
+    pixel_attention_mask (`torch.Tensor` of shape `(batch_size, image_size, image_size)`, *optional*):
+        Mask to avoid performing attention on padding pixel indices.
+    image_hidden_states (`torch.FloatTensor` of shape `(batch_size, num_channels, image_size, image_size)`):
+        The hidden states of the image encoder after modality projection.
+    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 `model.image_token_id`. Tokens with indices set to `model.image_token_id` are
+        ignored (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
+
+    Example:
+
+    ```python
+    >>> import requests
+    >>> import torch
+    >>> from PIL import Image
+    >>> from io import BytesIO
+
+    >>> from transformers import AutoProcessor, AutoModelForImageTextToText
+    >>> from transformers.image_utils import load_image
+
+    >>> # Note that passing the image urls (instead of the actual pil images) to the processor is also possible
+    >>> image1 = load_image("https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg")
+    >>> image2 = load_image("https://cdn.britannica.com/59/94459-050-DBA42467/Skyline-Chicago.jpg")
+    >>> image3 = load_image("https://cdn.britannica.com/68/170868-050-8DDE8263/Golden-Gate-Bridge-San-Francisco.jpg")
+
+    >>> processor = AutoProcessor.from_pretrained("HuggingFaceTB/SmolVLM2-2.2B-Instruct")
+    >>> model = AutoModelForImageTextToText.from_pretrained("HuggingFaceTB/SmolVLM2-2.2B-Instruct", dtype=torch.bfloat16, device_map="auto")
+
+    >>> # Create inputs
+    >>> messages = [
+    ...     {
+    ...         "role": "user",
+    ...         "content": [
+    ...             {"type": "video", "path": path/to/video},
+    ...             {"type": "text", "text": "What is happening in this video?"},
+    ...         ]
+    ...     }
+    ... ]
+
+    >>> inputs = processor.apply_chat_template([messages], add_generation_prompt=True)
+
+    >>> # Generate
+    >>> generated_ids = model.generate(**inputs, max_new_tokens=256)
+    >>> generated_texts = processor.batch_decode(generated_ids, skip_special_tokens=True)
+
+    >>> print(generated_texts)
+    ```"""
+    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
+
+    # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
+    outputs = self.model(
+        input_ids=input_ids,
+        attention_mask=attention_mask,
+        position_ids=position_ids,
+        past_key_values=past_key_values,
+        inputs_embeds=inputs_embeds,
+        pixel_values=pixel_values,
+        pixel_attention_mask=pixel_attention_mask,
+        image_hidden_states=image_hidden_states,
+        use_cache=use_cache,
+        output_attentions=output_attentions,
+        output_hidden_states=output_hidden_states,
+        cache_position=cache_position,
+        return_dict=True,
+        **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 or shift_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 SmolVLMCausalLMOutputWithPast(
+        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,
+    )