liger-kernel-nightly 0.4.0.dev20241107052928__py3-none-any.whl → 0.6.3.dev20251121010306__py3-none-any.whl

liger_kernel/transformers/model/phi3.py

@@ -1,146 +1,17 @@
-from typing import List, Optional, Tuple, Union
+from typing import List
+from typing import Optional
+from typing import Tuple
+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,
-    PHI3_INPUTS_DOCSTRING,
-)
-from transformers.utils import (
-    add_start_docstrings_to_model_forward,
-    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(
-    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,
-) -> 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 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,
@@ -154,121 +25,96 @@ 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*):
-            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")
+    >>> 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_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
+        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, :]
 
+    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):
-        # We do the same thing as ForCausalLMLoss but using Liger FLCE
-
-        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)
-
-        reduction = "sum" if "num_items_in_batch" in loss_kwargs else "mean"
-        lce = LigerFusedLinearCrossEntropyLoss(reduction=reduction)
-
-        loss = lce(self.lm_head.weight, shift_hidden_states, shift_labels)
-        if reduction == "sum":
-            loss /= loss_kwargs["num_items_in_batch"]
-
-    else:  # if in inference mode materialize logits
-        logits = self.lm_head(hidden_states[:, -num_logits_to_keep:, :])
-        if labels is not 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,
-                **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,
     )

liger_kernel/transformers/model/qwen2.py

@@ -1,26 +1,20 @@
-from typing import List, Optional, Tuple, Union
+from typing import List
+from typing import Optional
+from typing import Tuple
+from typing import Union
 
 import torch
+
 from torch.nn import CrossEntropyLoss
 from transformers.modeling_outputs import CausalLMOutputWithPast
-from transformers.models.qwen2.modeling_qwen2 import (
-    _CONFIG_FOR_DOC,
-    QWEN2_INPUTS_DOCSTRING,
-)
-from transformers.utils import (
-    add_start_docstrings_to_model_forward,
-    replace_return_docstrings,
-)
-
-from liger_kernel.transformers.fused_linear_cross_entropy import (
-    LigerFusedLinearCrossEntropyLoss,
-)
-
-
-@add_start_docstrings_to_model_forward(QWEN2_INPUTS_DOCSTRING)
-@replace_return_docstrings(
-    output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC
-)
+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
+
+
 def lce_forward_deprecated(
     self,
     input_ids: torch.LongTensor = None,
@@ -34,6 +28,7 @@ def lce_forward_deprecated(
     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 Qwen2's forward but replace torch cross entropy with liger fused linear cross entropy
@@ -63,19 +58,11 @@ def lce_forward_deprecated(
     >>> 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_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
+        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(
@@ -96,6 +83,13 @@ def lce_forward_deprecated(
     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 self.training and (labels is not None):
         shift_hidden_states = hidden_states[..., :-1, :].contiguous()
         shift_labels = labels[..., 1:].contiguous()
@@ -136,10 +130,7 @@ def lce_forward_deprecated(
     )
 
 
-@add_start_docstrings_to_model_forward(QWEN2_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,
@@ -153,9 +144,10 @@ 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*):
@@ -163,10 +155,12 @@ 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:
 
@@ -187,19 +181,11 @@ def lce_forward(
     "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_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
+        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(
@@ -213,44 +199,61 @@ 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):
-        # We do the same thing as ForCausalLMLoss but using Liger FLCE
+    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)
 
-        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)
-
-        reduction = "sum" if "num_items_in_batch" in loss_kwargs else "mean"
-        lce = LigerFusedLinearCrossEntropyLoss(reduction=reduction)
-
-        loss = lce(self.lm_head.weight, shift_hidden_states, shift_labels)
-        if reduction == "sum":
-            loss /= loss_kwargs["num_items_in_batch"]
-
-    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_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 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,
     )