liger-kernel-nightly 0.5.5.dev20250402185702__py3-none-any.whl → 0.6.4.dev20260112233432__py3-none-any.whl

liger_kernel/transformers/model/mixtral.py

@@ -7,18 +7,15 @@ import torch
 
 from torch.nn import CrossEntropyLoss
 from transformers.modeling_outputs import MoeCausalLMOutputWithPast
-from transformers.models.mixtral.modeling_mixtral import _CONFIG_FOR_DOC
-from transformers.models.mixtral.modeling_mixtral import MIXTRAL_INPUTS_DOCSTRING
 from transformers.models.mixtral.modeling_mixtral import load_balancing_loss_func
-from transformers.utils import add_start_docstrings_to_model_forward
-from transformers.utils import replace_return_docstrings
+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
+from liger_kernel.transformers.model.loss_utils import unpack_cross_entropy_result
+from liger_kernel.transformers.model.output_classes import LigerMoeCausalLMOutputWithPast
 
 
-@add_start_docstrings_to_model_forward(MIXTRAL_INPUTS_DOCSTRING)
-@replace_return_docstrings(output_type=MoeCausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
 def lce_forward_deprecated(
     self,
     input_ids: torch.LongTensor = None,
@@ -144,8 +141,7 @@ def lce_forward_deprecated(
     )
 
 
-@add_start_docstrings_to_model_forward(MIXTRAL_INPUTS_DOCSTRING)
-@replace_return_docstrings(output_type=MoeCausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
+@deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
 # Ignore copy
 def lce_forward(
     self,
@@ -161,9 +157,10 @@ def lce_forward(
     output_router_logits: Optional[bool] = None,
     return_dict: Optional[bool] = None,
     cache_position: Optional[torch.LongTensor] = None,
-    num_logits_to_keep: int = 0,
-    **loss_kwargs,
-) -> Union[Tuple, MoeCausalLMOutputWithPast]:
+    logits_to_keep: Union[int, torch.Tensor] = 0,
+    skip_logits: Optional[bool] = None,
+    **kwargs,
+) -> Union[Tuple, LigerMoeCausalLMOutputWithPast]:
     r"""
     Args:
         labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
@@ -171,10 +168,12 @@ def lce_forward(
             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]`.
 
-        num_logits_to_keep (`int`, *optional*):
-            Calculate logits for the last `num_logits_to_keep` tokens. If `0`, calculate logits for all
+        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:
 
@@ -218,32 +217,50 @@ def lce_forward(
         output_router_logits=output_router_logits,
         return_dict=return_dict,
         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 in training mode, don't materialize logits
-    if self.training and (labels is not None):
-        loss = LigerForCausalLMLoss(
-            hidden_states=hidden_states,
+    token_accuracy = 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)
+
+    # Compute loss
+    if skip_logits:
+        result = 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,
-            **loss_kwargs,
+            **kwargs,
         )
+        loss, _, token_accuracy = unpack_cross_entropy_result(result)
 
-    else:  # if in inference mode materialize logits
-        logits = self.lm_head(hidden_states[:, -num_logits_to_keep:, :])
-        if labels is not None:
+    else:
+        logits = self.lm_head(kept_hidden_states)
+
+        loss = None
+        if labels is not None or shift_labels is not None:
             loss = self.loss_function(
                 logits=logits,
                 labels=labels,
-                vocab_size=self.config.vocab_size,
-                **loss_kwargs,
+                shift_labels=shift_labels,
+                vocab_size=self.vocab_size,
+                **kwargs,
             )
-
     aux_loss = None
     if output_router_logits:
         aux_loss = load_balancing_loss_func(
@@ -256,17 +273,21 @@ def lce_forward(
             loss += self.router_aux_loss_coef * aux_loss.to(loss.device)  # make sure to reside in the same device
 
     if not return_dict:
-        output = (logits,) + outputs[1:]
+        output_tuple = (logits,) + outputs[1:]
         if output_router_logits:
-            output = (aux_loss,) + output
-        return (loss,) + output if loss is not None else output
+            output_tuple = (aux_loss,) + output_tuple
+        if token_accuracy is not None:
+            output_tuple = output_tuple + (token_accuracy,)
+        return (loss,) + output_tuple if loss is not None else output_tuple
 
-    return MoeCausalLMOutputWithPast(
+    # Return custom output class with token_accuracy field
+    return LigerMoeCausalLMOutputWithPast(
         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,
+        router_logits=outputs.router_logits if return_dict else outputs[-1],
+        token_accuracy=token_accuracy,
     )

liger_kernel/transformers/model/mllama.py

@@ -8,16 +8,14 @@ import torch
 from torch.nn import CrossEntropyLoss
 from transformers.cache_utils import Cache
 from transformers.modeling_outputs import CausalLMOutputWithPast
-from transformers.models.mllama.modeling_mllama import MLLAMA_INPUTS_DOCSTRING
-from transformers.utils import add_start_docstrings_to_model_forward
-from transformers.utils import replace_return_docstrings
+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
+from liger_kernel.transformers.model.loss_utils import unpack_cross_entropy_result
+from liger_kernel.transformers.model.output_classes import LigerCausalLMOutputWithPast
 
 
-@add_start_docstrings_to_model_forward(MLLAMA_INPUTS_DOCSTRING)
-@replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class="MllamaTextConfig")
 def lce_forward_deprecated(
     self,
     input_ids: torch.LongTensor = None,
@@ -133,8 +131,7 @@ def lce_forward_deprecated(
     )
 
 
-@add_start_docstrings_to_model_forward(MLLAMA_INPUTS_DOCSTRING)
-@replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class="MllamaTextConfig")
+@deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
 def lce_forward(
     self,
     input_ids: torch.LongTensor = None,
@@ -151,9 +148,10 @@ def lce_forward(
     output_hidden_states: Optional[bool] = None,
     return_dict: Optional[bool] = None,
     cache_position: Optional[torch.LongTensor] = None,
-    num_logits_to_keep: int = 0,
-    **loss_kwargs,
-) -> Union[Tuple, CausalLMOutputWithPast]:
+    logits_to_keep: Union[int, torch.Tensor] = 0,
+    skip_logits: Optional[bool] = None,
+    **kwargs,
+) -> Union[Tuple, LigerCausalLMOutputWithPast]:
     r"""
     Args:
         labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
@@ -161,10 +159,12 @@ def lce_forward(
             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]`.
 
-        num_logits_to_keep (`int`, *optional*):
-            Calculate logits for the last `num_logits_to_keep` tokens. If `0`, calculate logits for all
+        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:
 
@@ -192,7 +192,9 @@ def lce_forward(
         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
-
+    # Filter out accum_dtype from kwargs for model call as MllamaTextModel doesn't accept it in transformers 4.49.0
+    # but preserve it for loss function calls
+    model_kwargs = {k: v for k, v in kwargs.items() if k != "accum_dtype"}
     # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
     outputs = self.model(
         input_ids=input_ids,
@@ -208,40 +210,60 @@ def lce_forward(
         output_hidden_states=output_hidden_states,
         return_dict=return_dict,
         cache_position=cache_position,
+        **model_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 in training mode, don't materialize logits
-    if self.training and (labels is not None):
-        loss = LigerForCausalLMLoss(
-            hidden_states=hidden_states,
+    token_accuracy = 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:
+        result = 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,
-            **loss_kwargs,
+            **kwargs,
         )
+        loss, _, token_accuracy = unpack_cross_entropy_result(result)
 
-    else:  # if in inference mode materialize logits
-        logits = self.lm_head(hidden_states[:, -num_logits_to_keep:, :])
-        if labels is not None:
+    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,
-                **loss_kwargs,
+                **kwargs,
             )
 
     if not return_dict:
         output = (logits,) + outputs[1:]
-        return (loss,) + output if loss is not None else output
+        output = (loss,) + output if loss is not None else output
+        output = output + (token_accuracy,) if token_accuracy is not None else output
+        return output
 
-    return CausalLMOutputWithPast(
+    # Return custom output class with token_accuracy field
+    return LigerCausalLMOutputWithPast(
         loss=loss,
         logits=logits,
         past_key_values=outputs.past_key_values,
         hidden_states=outputs.hidden_states,
         attentions=outputs.attentions,
+        token_accuracy=token_accuracy,
     )

liger_kernel/transformers/model/olmo2.py

@@ -5,17 +5,14 @@ from typing import Union
 
 import torch
 
-from transformers.modeling_outputs import CausalLMOutputWithPast
-from transformers.models.olmo2.modeling_olmo2 import _CONFIG_FOR_DOC
-from transformers.models.olmo2.modeling_olmo2 import OLMO2_INPUTS_DOCSTRING
-from transformers.utils import add_start_docstrings_to_model_forward
-from transformers.utils import replace_return_docstrings
+from transformers.utils.deprecation import deprecate_kwarg
 
 from liger_kernel.transformers.model.loss_utils import LigerForCausalLMLoss
+from liger_kernel.transformers.model.loss_utils import unpack_cross_entropy_result
+from liger_kernel.transformers.model.output_classes import LigerCausalLMOutputWithPast
 
 
-@add_start_docstrings_to_model_forward(OLMO2_INPUTS_DOCSTRING)
-@replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
+@deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
 def lce_forward(
     self,
     input_ids: torch.LongTensor = None,
@@ -29,9 +26,10 @@ def lce_forward(
     output_hidden_states: Optional[bool] = None,
     return_dict: Optional[bool] = None,
     cache_position: Optional[torch.LongTensor] = None,
-    num_logits_to_keep: int = 0,
-    **loss_kwargs,
-) -> Union[Tuple, CausalLMOutputWithPast]:
+    logits_to_keep: Union[int, torch.Tensor] = 0,
+    skip_logits: Optional[bool] = None,
+    **kwargs,
+) -> Union[Tuple, LigerCausalLMOutputWithPast]:
     r"""
     Args:
         labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
@@ -39,10 +37,12 @@ def lce_forward(
             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]`.
 
-        num_logits_to_keep (`int`, *optional*):
-            Calculate logits for the last `num_logits_to_keep` tokens. If `0`, calculate logits for all
+        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:
 
@@ -81,36 +81,61 @@ def lce_forward(
         output_hidden_states=output_hidden_states,
         return_dict=return_dict,
         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 in training mode, don't materialize logits
-    if self.training and (labels is not None):
-        loss = LigerForCausalLMLoss(
-            hidden_states=hidden_states,
+    token_accuracy = 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)
+
+    # Compute loss
+    if skip_logits:
+        result = 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,
-            **loss_kwargs,
+            **kwargs,
         )
+        loss, _, token_accuracy = unpack_cross_entropy_result(result)
 
-    else:  # if in inference mode materialize logits
-        logits = self.lm_head(hidden_states[:, -num_logits_to_keep:, :])
-        if labels is not None:
+    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,
-                **loss_kwargs,
+                **kwargs,
             )
 
-    return CausalLMOutputWithPast(
+    if not return_dict:
+        output = (logits,) + outputs[1:]
+        output = ((loss,) + output) if loss is not None else output
+        output = output + (token_accuracy,) if token_accuracy is not None else output
+        return output
+
+    # Return custom output class with token_accuracy field
+    return LigerCausalLMOutputWithPast(
         loss=loss,
         logits=logits,
         past_key_values=outputs.past_key_values,
         hidden_states=outputs.hidden_states,
         attentions=outputs.attentions,
+        token_accuracy=token_accuracy,
     )

liger_kernel/transformers/model/olmo3.py

@@ -0,0 +1,142 @@
+from typing import List
+from typing import Optional
+from typing import Tuple
+from typing import Union
+
+import torch
+
+from transformers.modeling_outputs import BaseModelOutputWithPast
+from transformers.utils.deprecation import deprecate_kwarg
+
+from liger_kernel.transformers.model.loss_utils import LigerForCausalLMLoss
+from liger_kernel.transformers.model.loss_utils import unpack_cross_entropy_result
+from liger_kernel.transformers.model.output_classes import LigerCausalLMOutputWithPast
+
+
+@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,
+    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,
+    skip_logits: Optional[bool] = None,
+    **kwargs,
+) -> Union[Tuple, LigerCausalLMOutputWithPast]:
+    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, Olmo3ForCausalLM
+
+    >>> model = Olmo3ForCausalLM.from_pretrained("allenai/Olmo-3-7B-Instruct")
+    >>> tokenizer = AutoTokenizer.from_pretrained("allenai/Olmo-3-7B-Instruct")
+
+    >>> 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]
+    'Hey, are you conscious? Can you talk to me?\nI’m not sure if you’re conscious of this, but I’m'
+    ```
+    """
+    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: 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]
+    # 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
+    token_accuracy = 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)
+
+    # Compute loss
+    if skip_logits:
+        result = 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,
+        )
+        loss, _, token_accuracy = unpack_cross_entropy_result(result)
+
+    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,
+            )
+
+    if not return_dict:
+        output = (logits,) + outputs[1:]
+        output = ((loss,) + output) if loss is not None else output
+        output = output + (token_accuracy,) if token_accuracy is not None else output
+        return output
+
+    # Return custom output class with token_accuracy field
+    return LigerCausalLMOutputWithPast(
+        loss=loss,
+        logits=logits,
+        past_key_values=outputs.past_key_values,
+        hidden_states=outputs.hidden_states,
+        attentions=outputs.attentions,
+        token_accuracy=token_accuracy,
+    )