Fix huggingface GA issue for llama (#333)

## Summary
<!--- This is a required section; please describe the main purpose of
this proposed code change. --->

<!---
## Details
This is an optional section; is there anything specific that reviewers
should be aware of?
--->

To fix #322

This PR introduces a new `lce_forward` compatible with
`transformers>=4.46.0` (after grad acc fix) while ensuring backward
compatibilty.

To be specific, i keep the original flce untouched and write a new one
for `4.46.0`. If HF version is `<4.46.0`, it will show a warning for
deprecation, and fallback to the old flce.


```python
        if transformer_version >= version.parse("4.46.0"):
            modeling_llama.LlamaForCausalLM.forward = llama_lce_forward
        else:  # if version < 4.46.0
            logger.warning(
                "Support for transformers versions < 4.46.0 will soon be discontinued due to issues with incorrect gradient accumulation. "
                "Please consider upgrading to avoid potential issues. See details: huggingface/transformers#34191"
            )
            modeling_llama.LlamaForCausalLM.forward = llama_lce_forward_deprecated
```


For more context of grad acc fix, please see
huggingface/transformers#34191

## TODO

- [ ] broadcast the changes to all models once the effect is verified.


## 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>
- [x] run `make test` to ensure correctness
- [x] run `make checkstyle` to ensure code style
- [x] run `make test-convergence` to ensure convergence

src/liger_kernel/transformers/model/llama.py

@@ -22,7 +22,7 @@
 @replace_return_docstrings(
     output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC
 )
-def lce_forward(
+def lce_forward_deprecated(
     self,
     input_ids: torch.LongTensor = None,
     attention_mask: Optional[torch.Tensor] = None,
@@ -145,3 +145,137 @@ def lce_forward(
         hidden_states=outputs.hidden_states,
         attentions=outputs.attentions,
     )
+
+
+@add_start_docstrings_to_model_forward(LLAMA_INPUTS_DOCSTRING)
+@replace_return_docstrings(
+    output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC
+)
+def lce_forward(
+    self,
+    input_ids=None,
+    attention_mask=None,
+    position_ids=None,
+    past_key_values=None,
+    inputs_embeds=None,
+    labels=None,
+    use_cache=None,
+    output_attentions=None,
+    output_hidden_states=None,
+    return_dict=None,
+    cache_position=None,
+    num_logits_to_keep=0,
+    **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, LlamaForCausalLM
+
+    >>> model = LlamaForCausalLM.from_pretrained("meta-llama/Llama-2-7b-hf")
+    >>> tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-2-7b-hf")
+
+    >>> 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 conscious, but I can talk to you."
+    ```"""
+
+    """
+    My TODO:
+    1. Run e2e example with hf < 4.46.0, GA on/GA off
+    2. Run e2e example with hf >= 4.46.0, GA on/GA off
+    """
+
+    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,
+        cache_position=cache_position,
+        **kwargs,
+    )
+
+    hidden_states = outputs[0]
+
+    if self.config.pretraining_tp > 1:
+        raise Exception("Liger Kernel does not support pretraining_tp!!")
+
+    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 kwargs else "mean"
+        lce = LigerFusedLinearCrossEntropyLoss(reduction=reduction)
+
+        loss = lce(self.lm_head.weight, shift_hidden_states, shift_labels)
+        if reduction == "sum":
+            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:
+            loss = self.loss_function(
+                logits=logits,
+                labels=labels,
+                vocab_size=self.config.vocab_size,
+                **kwargs,
+            )
+
+    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,
+    )

src/liger_kernel/transformers/monkey_patch.py

@@ -3,13 +3,18 @@
 from functools import partial
 from typing import Callable
 
+import transformers
+from packaging import version
 from transformers import PreTrainedModel
 
 from liger_kernel.transformers.cross_entropy import LigerCrossEntropyLoss
 from liger_kernel.transformers.geglu import LigerGEGLUMLP
 from liger_kernel.transformers.layer_norm import LigerLayerNorm
 from liger_kernel.transformers.model.gemma import lce_forward as gemma_lce_forward
 from liger_kernel.transformers.model.llama import lce_forward as llama_lce_forward
+from liger_kernel.transformers.model.llama import (
+    lce_forward_deprecated as llama_lce_forward_deprecated,
+)
 from liger_kernel.transformers.model.mistral import lce_forward as mistral_lce_forward
 from liger_kernel.transformers.model.mixtral import lce_forward as mixtral_lce_forward
 from liger_kernel.transformers.model.phi3 import lce_forward as phi3_lce_forward
@@ -22,6 +27,8 @@
     LigerSwiGLUMLP,
 )
 
+transformer_version = version.parse(transformers.__version__)
+
 logger = logging.getLogger(__name__)
 
 
@@ -88,7 +95,14 @@ def apply_liger_kernel_to_llama(
     if cross_entropy:
         modeling_llama.CrossEntropyLoss = LigerCrossEntropyLoss
     if fused_linear_cross_entropy:
-        modeling_llama.LlamaForCausalLM.forward = llama_lce_forward
+        if transformer_version >= version.parse("4.46.0"):
+            modeling_llama.LlamaForCausalLM.forward = llama_lce_forward
+        else:  # if version < 4.46.0
+            logger.warning(
+                "Support for transformers versions < 4.46.0 will soon be discontinued due to issues with incorrect gradient accumulation. "
+                "Please consider upgrading to avoid potential issues. See details: https://github.com/huggingface/transformers/pull/34191"
+            )
+            modeling_llama.LlamaForCausalLM.forward = llama_lce_forward_deprecated
 
     if model is not None:
         # The model instance already exists, so we need to additionally patch the