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

liger_kernel/transformers/model/output_classes.py

@@ -0,0 +1,147 @@
+"""
+Custom output classes for Liger-Kernel that extend transformers' ModelOutput classes
+with optional token accuracy field.
+"""
+
+from dataclasses import dataclass
+from typing import Optional
+
+import torch
+
+from transformers.modeling_outputs import CausalLMOutputWithPast
+from transformers.modeling_outputs import MoeCausalLMOutputWithPast
+
+# The following model-specific outputs are optional and depend on the installed
+# transformers version. Guard their imports so our module remains importable
+# even when those models are not available in the environment.
+try:
+    from transformers.models.gemma3.modeling_gemma3 import Gemma3CausalLMOutputWithPast as _Gemma3CausalLMOutputWithPast
+except Exception:
+    _Gemma3CausalLMOutputWithPast = None
+
+try:
+    from transformers.models.glm4v_moe.modeling_glm4v_moe import (
+        Glm4vMoeCausalLMOutputWithPast as _Glm4vMoeCausalLMOutputWithPast,
+    )
+except Exception:
+    _Glm4vMoeCausalLMOutputWithPast = None
+
+try:
+    from transformers.models.internvl.modeling_internvl import (
+        InternVLCausalLMOutputWithPast as _InternVLCausalLMOutputWithPast,
+    )
+except Exception:
+    _InternVLCausalLMOutputWithPast = None
+
+try:
+    from transformers.models.llava.modeling_llava import LlavaCausalLMOutputWithPast as _LlavaCausalLMOutputWithPast
+except Exception:
+    _LlavaCausalLMOutputWithPast = None
+
+try:
+    from transformers.models.paligemma.modeling_paligemma import (
+        PaliGemmaCausalLMOutputWithPast as _PaliGemmaCausalLMOutputWithPast,
+    )
+except Exception:
+    _PaliGemmaCausalLMOutputWithPast = None
+
+try:
+    from transformers.models.qwen2_5_vl.modeling_qwen2_5_vl import (
+        Qwen2_5_VLCausalLMOutputWithPast as _Qwen2_5_VLCausalLMOutputWithPast,
+    )
+except Exception:
+    _Qwen2_5_VLCausalLMOutputWithPast = None
+
+try:
+    from transformers.models.qwen2_vl.modeling_qwen2_vl import (
+        Qwen2VLCausalLMOutputWithPast as _Qwen2VLCausalLMOutputWithPast,
+    )
+except Exception:
+    _Qwen2VLCausalLMOutputWithPast = None
+
+try:
+    from transformers.models.qwen3_vl.modeling_qwen3_vl import (
+        Qwen3VLCausalLMOutputWithPast as _Qwen3VLCausalLMOutputWithPast,
+    )
+except Exception:
+    _Qwen3VLCausalLMOutputWithPast = None
+
+try:
+    from transformers.models.qwen3_vl_moe.modeling_qwen3_vl_moe import (
+        Qwen3VLMoeCausalLMOutputWithPast as _Qwen3VLMoeCausalLMOutputWithPast,
+    )
+except Exception:
+    _Qwen3VLMoeCausalLMOutputWithPast = None
+
+
+@dataclass
+class LigerCausalLMOutputWithPast(CausalLMOutputWithPast):
+    token_accuracy: Optional[torch.FloatTensor] = None
+
+
+@dataclass
+class LigerMoeCausalLMOutputWithPast(MoeCausalLMOutputWithPast):
+    token_accuracy: Optional[torch.FloatTensor] = None
+
+
+if _Gemma3CausalLMOutputWithPast is not None:
+
+    @dataclass
+    class LigerGemma3CausalLMOutputWithPast(_Gemma3CausalLMOutputWithPast):
+        token_accuracy: Optional[torch.FloatTensor] = None
+
+
+if _Glm4vMoeCausalLMOutputWithPast is not None:
+
+    @dataclass
+    class LigerGlm4vMoeCausalLMOutputWithPast(_Glm4vMoeCausalLMOutputWithPast):
+        token_accuracy: Optional[torch.FloatTensor] = None
+
+
+if _LlavaCausalLMOutputWithPast is not None:
+
+    @dataclass
+    class LigerLlavaCausalLMOutputWithPast(_LlavaCausalLMOutputWithPast):
+        token_accuracy: Optional[torch.FloatTensor] = None
+
+
+if _InternVLCausalLMOutputWithPast is not None:
+
+    @dataclass
+    class LigerInternVLCausalLMOutputWithPast(_InternVLCausalLMOutputWithPast):
+        token_accuracy: Optional[torch.FloatTensor] = None
+
+
+if _PaliGemmaCausalLMOutputWithPast is not None:
+
+    @dataclass
+    class LigerPaliGemmaCausalLMOutputWithPast(_PaliGemmaCausalLMOutputWithPast):
+        token_accuracy: Optional[torch.FloatTensor] = None
+
+
+if _Qwen2_5_VLCausalLMOutputWithPast is not None:
+
+    @dataclass
+    class LigerQwen2_5_VLCausalLMOutputWithPast(_Qwen2_5_VLCausalLMOutputWithPast):
+        token_accuracy: Optional[torch.FloatTensor] = None
+
+
+if _Qwen2VLCausalLMOutputWithPast is not None:
+
+    @dataclass
+    class LigerQwen2VLCausalLMOutputWithPast(_Qwen2VLCausalLMOutputWithPast):
+        token_accuracy: Optional[torch.FloatTensor] = None
+
+
+if _Qwen3VLCausalLMOutputWithPast is not None:
+
+    @dataclass
+    class LigerQwen3VLCausalLMOutputWithPast(_Qwen3VLCausalLMOutputWithPast):
+        token_accuracy: Optional[torch.FloatTensor] = None
+
+
+if _Qwen3VLMoeCausalLMOutputWithPast is not None:
+
+    @dataclass
+    class LigerQwen3VLMoeCausalLMOutputWithPast(_Qwen3VLMoeCausalLMOutputWithPast):
+        token_accuracy: Optional[torch.FloatTensor] = None

liger_kernel/transformers/model/paligemma.py

@@ -7,22 +7,19 @@ import torch
 
 from torch.nn import CrossEntropyLoss
 from transformers.cache_utils import Cache
-from transformers.models.paligemma.modeling_paligemma import _CONFIG_FOR_DOC
-from transformers.models.paligemma.modeling_paligemma import PALIGEMMA_INPUTS_DOCSTRING
 from transformers.models.paligemma.modeling_paligemma import PaliGemmaCausalLMOutputWithPast
-from transformers.utils import add_start_docstrings_to_model_forward
 from transformers.utils import is_torchdynamo_compiling
 from transformers.utils import logging
-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 LigerPaliGemmaCausalLMOutputWithPast
 
 logger = logging.get_logger(__name__)
 
 
-@add_start_docstrings_to_model_forward(PALIGEMMA_INPUTS_DOCSTRING)
-@replace_return_docstrings(output_type=PaliGemmaCausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
 def lce_forward_deprecated(
     self,
     input_ids: torch.LongTensor = None,
@@ -206,8 +203,6 @@ def lce_forward_deprecated(
 
 
 @deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
-@add_start_docstrings_to_model_forward(PALIGEMMA_INPUTS_DOCSTRING)
-@replace_return_docstrings(output_type=PaliGemmaCausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
 def lce_forward(
     self,
     input_ids: torch.LongTensor = None,
@@ -224,8 +219,9 @@ def lce_forward(
     output_hidden_states: Optional[bool] = None,
     return_dict: Optional[bool] = None,
     logits_to_keep: Union[int, torch.Tensor] = 0,
+    skip_logits: Optional[bool] = None,
     **lm_kwargs,
-) -> Union[Tuple, PaliGemmaCausalLMOutputWithPast]:
+) -> Union[Tuple, LigerPaliGemmaCausalLMOutputWithPast]:
     r"""
     Args:
         labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
@@ -334,12 +330,20 @@ def lce_forward(
         **lm_kwargs,
     )
 
+    shift_labels = lm_kwargs.pop("shift_labels", None)
     hidden_states = outputs[0]
 
     loss = None
     logits = None
+    token_accuracy = None
 
-    if self.training and (labels is not None):
+    if skip_logits and labels is None:
+        raise ValueError("skip_logits is True, but labels is None")
+
+    if skip_logits is None:
+        skip_logits = self.training and (labels is not None)
+
+    if skip_logits:
         shift_hidden_states = hidden_states[..., :-1, :]
         shift_labels = labels[..., 1:]
 
@@ -359,8 +363,16 @@ def lce_forward(
         shift_hidden_states = shift_hidden_states.view(-1, self.config.text_config.hidden_size)
         shift_labels = shift_labels.view(-1).to(hidden_device)
 
-        lce = LigerFusedLinearCrossEntropyLoss()
-        loss = lce(self.language_model.lm_head.weight, shift_hidden_states, shift_labels)
+        # Use LigerForCausalLMLoss with accuracy support and pass already shifted labels
+        result = LigerForCausalLMLoss(
+            hidden_states=shift_hidden_states,
+            lm_head_weight=self.language_model.lm_head.weight,
+            labels=None,
+            shift_labels=shift_labels,
+            hidden_size=self.config.text_config.hidden_size,
+            **lm_kwargs,
+        )
+        loss, _, token_accuracy = unpack_cross_entropy_result(result)
     else:
         logits = self.language_model.lm_head(hidden_states)
         if labels is not None:
@@ -383,15 +395,39 @@ def lce_forward(
             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)
+
     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 PaliGemmaCausalLMOutputWithPast(
+    # Return PaliGemma output with token_accuracy field
+    return LigerPaliGemmaCausalLMOutputWithPast(
         loss=loss,
         logits=logits,
         past_key_values=outputs.past_key_values,
         hidden_states=outputs.hidden_states,
         attentions=outputs.attentions,
         image_hidden_states=image_features if pixel_values is not None else None,
+        token_accuracy=token_accuracy,
     )

liger_kernel/transformers/model/phi3.py

@@ -5,20 +5,14 @@ from typing import Union
 
 import torch
 
-from torch.nn import CrossEntropyLoss
-from transformers.modeling_outputs import CausalLMOutputWithPast
-from transformers.models.phi3.modeling_phi3 import _CONFIG_FOR_DOC
-from transformers.models.phi3.modeling_phi3 import PHI3_INPUTS_DOCSTRING
-from transformers.utils import add_start_docstrings_to_model_forward
-from transformers.utils import replace_return_docstrings
-
-from liger_kernel.transformers.fused_linear_cross_entropy import LigerFusedLinearCrossEntropyLoss
+from transformers.modeling_outputs import BaseModelOutputWithPast
+
 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(PHI3_INPUTS_DOCSTRING)
-@replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
-def lce_forward_deprecated(
+def lce_forward(
     self,
     input_ids: torch.LongTensor = None,
     attention_mask: Optional[torch.Tensor] = None,
@@ -31,34 +25,26 @@ def lce_forward_deprecated(
     output_hidden_states: Optional[bool] = None,
     return_dict: Optional[bool] = None,
     cache_position: Optional[torch.LongTensor] = None,
-) -> Union[Tuple, CausalLMOutputWithPast]:
+    logits_to_keep: Union[int, torch.Tensor] = 0,
+    skip_logits: Optional[bool] = None,
+    **kwargs,
+) -> Union[Tuple, LigerCausalLMOutputWithPast]:
     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")
+    >>> 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."
     ```"""
 
     output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
@@ -67,179 +53,68 @@ def lce_forward_deprecated(
     )
     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, :]
 
-    loss = None
+    shift_labels = kwargs.pop("shift_labels", None)
     logits = None
+    loss = None
+    token_accuracy = None
 
-    if self.training and labels is not None:
-        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,
-    )
-
-
-@add_start_docstrings_to_model_forward(PHI3_INPUTS_DOCSTRING)
-@replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
-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,
-    num_logits_to_keep: int = 0,
-    **loss_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]`.
-
-        num_logits_to_keep (`int`, *optional*):
-            Calculate logits for the last `num_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.
-
-    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'
-    ```"""
-
-    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(
-        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,
-    )
+    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")
 
-    hidden_states = outputs[0]
+    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)
 
-    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,
+    # 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,
         )
-
-    else:  # if in inference mode materialize logits
-        logits = self.lm_head(hidden_states[:, -num_logits_to_keep:, :])
-        if labels is not None:
+        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,
-                **loss_kwargs,
+                **kwargs,
             )
 
     if not return_dict:
-        output = (logits,) + outputs[1:]
-        return (loss,) + output if loss is not None else output
+        output_tuple = (logits,) + outputs[1:]
+        output = (loss,) + output_tuple if loss is not None else output_tuple
+        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,
     )