fix(phi3): update monkey patch for Phi3ForCausalLM (#837)

## Summary
<!--- This is a required section; please describe the main purpose of
this proposed code change. --->
Liger backward support for transformers has been updated to 4.49.0, so
we can remove most reprecated functions that were made for
`SUPPORTED_TRANSFORMER_VERSION=4.46,1` (we should probably bump it for
next release too).

This PR merges most patching logics into one that should be able to work
with both 4.49.0 and the latest version (4.54.0)
<!---
## Details
This is an optional section; is there anything specific that reviewers
should be aware of?
--->

## Testing Done
<!--- This is a required section; please describe how this change was
tested. --->

<!-- 
Replace BLANK with your device type. For example, A100-80G-PCIe

Complete the following tasks before sending your PR, and replace `[ ]`
with
`[x]` to indicate you have done them. 
-->

- Hardware Type: <BLANK>
- [ ] run `make test` to ensure correctness
- [ ] run `make checkstyle` to ensure code style
- [ ] run `make test-convergence` to ensure convergence

---------

Signed-off-by: Tcc0403 <[email protected]>
Co-authored-by: Shao Tang <[email protected]>
Co-authored-by: Steven Shimizu <[email protected]>

Author: 3 people

src/liger_kernel/transformers/model/phi3.py

@@ -5,131 +5,12 @@
 
 import torch
 
-from torch.nn import CrossEntropyLoss
+from transformers.modeling_outputs import BaseModelOutputWithPast
 from transformers.modeling_outputs import CausalLMOutputWithPast
-from transformers.utils.deprecation import deprecate_kwarg
 
-from liger_kernel.transformers.fused_linear_cross_entropy import LigerFusedLinearCrossEntropyLoss
 from liger_kernel.transformers.model.loss_utils import LigerForCausalLMLoss
 
 
-def lce_forward_deprecated(
-    self,
-    input_ids: torch.LongTensor = None,
-    attention_mask: Optional[torch.Tensor] = None,
-    position_ids: Optional[torch.LongTensor] = None,
-    past_key_values: Optional[List[torch.FloatTensor]] = None,
-    inputs_embeds: Optional[torch.FloatTensor] = None,
-    labels: Optional[torch.LongTensor] = None,
-    use_cache: Optional[bool] = None,
-    output_attentions: Optional[bool] = None,
-    output_hidden_states: Optional[bool] = None,
-    return_dict: Optional[bool] = None,
-    cache_position: Optional[torch.LongTensor] = None,
-    skip_logits: Optional[bool] = None,
-) -> Union[Tuple, CausalLMOutputWithPast]:
-    r"""
-    Copy paste phi3 forward from transfomers v4.44.2 but replace torch cross entropy with liger fused linear cross entropy
-
-
-    Args:
-        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
-            Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
-            config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
-            (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
-
-    Returns:
-
-    Example:
-
-    ```python
-    >>> from transformers import AutoTokenizer, Phi3ForCausalLM
-
-    >>> model = Phi3ForCausalLM.from_pretrained("microsoft/phi-3-mini-4k-instruct")
-    >>> tokenizer = AutoTokenizer.from_pretrained("microsoft/phi-3-mini-4k-instruct")
-
-    >>> prompt = "This is an example script ."
-    >>> inputs = tokenizer(prompt, return_tensors="pt")
-
-    >>> # Generate
-    >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
-    >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
-    'This is an example script .\n Certainly! Below is a sample script that demonstrates a simple task, such as calculating the sum'
-    ```"""
-
-    output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
-    output_hidden_states = (
-        output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
-    )
-    return_dict = return_dict if return_dict is not None else self.config.use_return_dict
-
-    # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
-    outputs = self.model(
-        input_ids=input_ids,
-        attention_mask=attention_mask,
-        position_ids=position_ids,
-        past_key_values=past_key_values,
-        inputs_embeds=inputs_embeds,
-        use_cache=use_cache,
-        output_attentions=output_attentions,
-        output_hidden_states=output_hidden_states,
-        return_dict=return_dict,
-    )
-
-    hidden_states = outputs[0]
-
-    loss = None
-    logits = None
-
-    if skip_logits and labels is None:
-        raise ValueError("skip_logits is True, but labels is None")
-
-    if skip_logits is None:
-        # By default, if in training mode, don't materialize logits
-        skip_logits = self.training and labels is not None
-
-    if skip_logits:
-        shift_hidden_states = hidden_states[..., :-1, :].contiguous()
-        shift_labels = labels[..., 1:].contiguous()
-
-        # flatten tokens
-        shift_hidden_states = shift_hidden_states.view(-1, self.config.hidden_size)
-        shift_labels = shift_labels.view(-1)
-
-        lce = LigerFusedLinearCrossEntropyLoss()
-        loss = lce(self.lm_head.weight, shift_hidden_states, shift_labels)
-    else:
-        logits = self.lm_head(hidden_states)
-
-        loss = None
-        if labels is not None:
-            # Upcast to float if we need to compute the loss to avoid potential precision issues
-            logits = logits.float()
-            # Shift so that tokens < n predict n
-            shift_logits = logits[..., :-1, :].contiguous()
-            shift_labels = labels[..., 1:].contiguous()
-            # Flatten the tokens
-            loss_fct = CrossEntropyLoss()
-            shift_logits = shift_logits.view(-1, self.config.vocab_size)
-            shift_labels = shift_labels.view(-1)
-            # Enable model parallelism
-            shift_labels = shift_labels.to(shift_logits.device)
-            loss = loss_fct(shift_logits, shift_labels)
-
-    if not return_dict:
-        output = (logits,) + outputs[1:]
-        return (loss,) + output if loss is not None else output
-
-    return CausalLMOutputWithPast(
-        loss=loss,
-        logits=logits,
-        past_key_values=outputs.past_key_values,
-        hidden_states=outputs.hidden_states,
-        attentions=outputs.attentions,
-    )
-
-
-@deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
 def lce_forward(
     self,
     input_ids: torch.LongTensor = None,
@@ -148,36 +29,21 @@ def lce_forward(
     **kwargs,
 ) -> Union[Tuple, CausalLMOutputWithPast]:
     r"""
-    Args:
-        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
-            Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
-            config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
-            (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
-
-        logits_to_keep (`int` or `torch.Tensor`, *optional*):
-            If an `int`, compute logits for the last `logits_to_keep` tokens. If `0`, calculate logits for all
-            `input_ids` (special case). Only last token logits are needed for generation, and calculating them only for that
-            token can save memory, which becomes pretty significant for long sequences or large vocabulary size.
-            If a `torch.Tensor`, must be 1D corresponding to the indices to keep in the sequence length dimension.
-            This is useful when using packed tensor format (single dimension for batch and sequence length).
-
-    Returns:
-
     Example:
 
     ```python
     >>> from transformers import AutoTokenizer, Phi3ForCausalLM
 
-    >>> model = Phi3ForCausalLM.from_pretrained("microsoft/phi-3-mini-4k-instruct")
-    >>> tokenizer = AutoTokenizer.from_pretrained("microsoft/phi-3-mini-4k-instruct")
+    >>> model = Phi3ForCausalLM.from_pretrained("meta-phi3/Phi3-2-7b-hf")
+    >>> tokenizer = AutoTokenizer.from_pretrained("meta-phi3/Phi3-2-7b-hf")
 
-    >>> prompt = "This is an example script ."
+    >>> prompt = "Hey, are you conscious? Can you talk to me?"
     >>> inputs = tokenizer(prompt, return_tensors="pt")
 
     >>> # Generate
     >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
     >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
-    'This is an example script .\n Certainly! Below is a sample script that demonstrates a simple task, such as calculating the sum'
+    "Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you."
     ```"""
 
     output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
@@ -186,21 +52,18 @@ def lce_forward(
     )
     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, :]

src/liger_kernel/transformers/monkey_patch.py

@@ -26,7 +26,6 @@
 from liger_kernel.transformers.model.mixtral import lce_forward as mixtral_lce_forward
 from liger_kernel.transformers.model.mixtral import lce_forward_deprecated as mixtral_lce_forward_deprecated
 from liger_kernel.transformers.model.phi3 import lce_forward as phi3_lce_forward
-from liger_kernel.transformers.model.phi3 import lce_forward_deprecated as phi3_lce_forward_deprecated
 from liger_kernel.transformers.model.qwen2 import lce_forward as qwen2_lce_forward
 from liger_kernel.transformers.model.qwen2 import lce_forward_deprecated as qwen2_lce_forward_deprecated
 from liger_kernel.transformers.model.smollm3 import lce_forward as smollm3_lce_forward
@@ -1677,25 +1676,14 @@ def apply_liger_kernel_to_phi3(
     if swiglu:
         modeling_phi3.Phi3MLP = LigerPhi3SwiGLUMLP
     if cross_entropy:
-        if transformer_version >= version.parse(SUPPORTED_TRANSFORMER_VERSION):
-            from transformers.loss.loss_utils import nn
+        from transformers.loss.loss_utils import nn
 
-            nn.functional.cross_entropy = liger_cross_entropy
-        else:
-            logger.warning(TRANSFORMER_DEPRECATION_WARNING)
-            modeling_phi3.CrossEntropyLoss = LigerCrossEntropyLoss
+        nn.functional.cross_entropy = liger_cross_entropy
     if fused_linear_cross_entropy:
-        if transformer_version >= version.parse(SUPPORTED_TRANSFORMER_VERSION):
-            if model is not None:
-                model.forward = MethodType(phi3_lce_forward, model)
-            else:
-                modeling_phi3.Phi3ForCausalLM.forward = phi3_lce_forward
-        else:  # if version < 4.46.1
-            logger.warning(TRANSFORMER_DEPRECATION_WARNING)
-            if model is not None:
-                model.forward = MethodType(phi3_lce_forward_deprecated, model)
-            else:
-                modeling_phi3.Phi3ForCausalLM.forward = phi3_lce_forward_deprecated
+        if model is not None:
+            model.forward = MethodType(phi3_lce_forward, model)
+        else:
+            modeling_phi3.Phi3ForCausalLM.forward = phi3_lce_forward
 
     if model is not None:
         # The model instance already exists, so we need to additionally patch the