liger-kernel 0.6.2__py3-none-any.whl → 0.6.3__py3-none-any.whl

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

@@ -70,7 +70,7 @@ def lce_forward(
     >>> processor = AutoProcessor.from_pretrained(MODEL_PATH, use_fast=True)
     >>> model = Glm4vForConditionalGeneration.from_pretrained(
         pretrained_model_name_or_path=MODEL_PATH,
-        torch_dtype=torch.bfloat16,
+        dtype=torch.bfloat16,
         device_map="auto",
     )
     >>> inputs = processor.apply_chat_template(
@@ -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

@@ -0,0 +1,153 @@
+from typing import Optional
+from typing import Tuple
+from typing import Union
+
+import torch
+
+from transformers.models.glm4v_moe.modeling_glm4v_moe import Glm4vMoeCausalLMOutputWithPast
+from transformers.utils.deprecation import deprecate_kwarg
+
+from liger_kernel.transformers.model.loss_utils import LigerForCausalLMLoss
+
+
+@deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
+def lce_forward(
+    self,
+    input_ids: torch.LongTensor = None,
+    attention_mask: Optional[torch.Tensor] = None,
+    position_ids: Optional[torch.LongTensor] = None,
+    past_key_values: Optional[list[torch.FloatTensor]] = None,
+    inputs_embeds: Optional[torch.FloatTensor] = None,
+    labels: Optional[torch.LongTensor] = None,
+    pixel_values: Optional[torch.Tensor] = None,
+    pixel_values_videos: Optional[torch.FloatTensor] = None,
+    image_grid_thw: Optional[torch.LongTensor] = None,
+    video_grid_thw: Optional[torch.LongTensor] = None,
+    rope_deltas: Optional[torch.LongTensor] = None,
+    cache_position: Optional[torch.LongTensor] = None,
+    logits_to_keep: Union[int, torch.Tensor] = 0,
+    skip_logits: Optional[bool] = None,
+    **kwargs,
+) -> Union[Tuple, Glm4vMoeCausalLMOutputWithPast]:
+    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]`.
+        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.
+
+
+        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).
+
+    Example:
+
+    ```python
+    >>> from transformers import AutoProcessor, Glm4vMoeForConditionalGeneration
+    >>> import torch
+
+    >>> MODEL_PATH = "zai-org/GLM-4.5V"
+    >>> messages = [
+        {
+            "role": "user",
+            "content": [
+                {
+                    "type": "image",
+                    "url": "https://upload.wikimedia.org/wikipedia/commons/f/fa/Grayscale_8bits_palette_sample_image.png"
+                },
+                {
+                    "type": "text",
+                    "text": "describe this image"
+                }
+            ],
+        }
+    ]
+    >>> processor = AutoProcessor.from_pretrained(MODEL_PATH)
+    >>> model = Glm4vMoeForConditionalGeneration.from_pretrained(
+        pretrained_model_name_or_path=MODEL_PATH,
+        dtype="auto",
+        device_map="auto",
+    )
+    >>> inputs = processor.apply_chat_template(
+        messages,
+        tokenize=True,
+        add_generation_prompt=True,
+        return_dict=True,
+        return_tensors="pt"
+    ).to(model.device)
+    >>> inputs.pop("token_type_ids", None)
+    >>> generated_ids = model.generate(**inputs, max_new_tokens=8192)
+    >>> output_text = processor.decode(generated_ids[0][inputs["input_ids"].shape[1]:], skip_special_tokens=False)
+    ```
+    """
+
+    # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
+    outputs = self.model(
+        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,
+        inputs_embeds=inputs_embeds,
+        cache_position=cache_position,
+        **kwargs,
+    )
+
+    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 = 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.hidden_size,
+            **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,
+                shift_labels=shift_labels,
+                vocab_size=self.config.vocab_size,
+                **kwargs,
+            )
+
+    return Glm4vMoeCausalLMOutputWithPast(
+        loss=loss,
+        logits=logits,
+        past_key_values=outputs.past_key_values,
+        hidden_states=outputs.hidden_states,
+        attentions=outputs.attentions,
+        rope_deltas=outputs.rope_deltas,
+    )

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/loss_utils.py

@@ -25,6 +25,7 @@ def fixed_fused_linear_cross_entropy(
         ignore_index=ignore_index,
         softcap=final_logit_softcapping,
         accum_dtype=accum_dtype,
+        **kwargs,
     )
     if reduction == "sum":
         loss = loss / num_items_in_batch

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: