fix llama & gemma's forward patch

Author: hiyouga

src/liger_kernel/transformers/model/gemma.py

@@ -35,6 +35,8 @@ 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]:
     r"""
 
@@ -106,11 +108,25 @@ def lce_forward(
         shift_hidden_states = shift_hidden_states.view(-1, self.config.hidden_size)
         shift_labels = shift_labels.view(-1)
 
-        lce = LigerFusedLinearCrossEntropyLoss()
+        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"]
+
+    elif hasattr(self, "loss_function"):  # 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,
+                **loss_kwargs,
+            )
 
     else:
-        logits = self.lm_head(hidden_states)
+        logits = self.lm_head(hidden_states[:, -num_logits_to_keep:, :])
         if labels is not None:
             # Upcast to float if we need to compute the loss to avoid potential precision issues
             logits = logits.float()

src/liger_kernel/transformers/model/llama.py

@@ -1,8 +1,8 @@
 from typing import List, Optional, Tuple, Union
 
 import torch
-import torch.nn.functional as F
 from torch.nn import CrossEntropyLoss
+from transformers.cache_utils import Cache
 from transformers.modeling_outputs import CausalLMOutputWithPast
 from transformers.models.llama.modeling_llama import (
     _CONFIG_FOR_DOC,
@@ -22,150 +22,21 @@
 @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,
     position_ids: Optional[torch.LongTensor] = None,
-    past_key_values: Optional[List[torch.FloatTensor]] = None,
+    past_key_values: Optional[Union[Cache, 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 llama forward 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, 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."
-    ```"""
-    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,
-    )
-
-    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:
-        if self.config.pretraining_tp > 1:
-            lm_head_slices = self.lm_head.weight.split(
-                self.vocab_size // self.config.pretraining_tp, dim=0
-            )
-            logits = [
-                F.linear(hidden_states, lm_head_slices[i])
-                for i in range(self.config.pretraining_tp)
-            ]
-            logits = torch.cat(logits, dim=-1)
-        else:
-            logits = self.lm_head(hidden_states)
-        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(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,
+    num_logits_to_keep: int = 0,
+    **loss_kwargs,
 ) -> Union[Tuple, CausalLMOutputWithPast]:
     r"""
     Args:
@@ -224,7 +95,6 @@ def lce_forward(
         output_hidden_states=output_hidden_states,
         return_dict=return_dict,
         cache_position=cache_position,
-        **kwargs,
     )
 
     hidden_states = outputs[0]
@@ -245,22 +115,37 @@ def lce_forward(
         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"
+        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 /= kwargs["num_items_in_batch"]
+            loss /= loss_kwargs["num_items_in_batch"]
 
-    else:  # if in inference mode materialize logits
+    elif hasattr(self, "loss_function"):  # 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,
+                **loss_kwargs,
             )
+    else:
+        logits = self.lm_head(hidden_states[:, -num_logits_to_keep:, :])
+        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:]

src/liger_kernel/transformers/monkey_patch.py

@@ -12,9 +12,6 @@
 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
@@ -95,14 +92,13 @@ def apply_liger_kernel_to_llama(
     if cross_entropy:
         modeling_llama.CrossEntropyLoss = LigerCrossEntropyLoss
     if fused_linear_cross_entropy:
-        if transformer_version >= version.parse("4.46.0"):
-            modeling_llama.LlamaForCausalLM.forward = llama_lce_forward
-        else:  # if version < 4.46.0
+        if transformer_version < version.parse("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
+
+        modeling_llama.LlamaForCausalLM.forward = llama_lce_forward
 
     if model is not None:
         # The model instance already exists, so we need to additionally patch the