Merge pull request #1252 from anvie/neftune

add NEFTune optimization

README.md

@@ -22,6 +22,8 @@ https://github.com/hiyouga/LLaMA-Factory/assets/16256802/6ba60acc-e2e2-4bec-b846
 
 ## Changelog
 
+[23/10/21] We supported [NEFTune](https://arxiv.org/abs/2310.05914) optimization . Try `--neftune_noise_alpha` argument to activate NEFTune, e.g., `--neftune_noise_alpha 5`.
+
 [23/09/27] We supported **$S^2$-Attn** proposed by [LongLoRA](https://github.com/dvlab-research/LongLoRA) for the LLaMA models. Try `--shift_attn` argument to enable shift short attention.
 
 [23/09/23] We integrated MMLU, C-Eval and CMMLU benchmarks in this repo. See [this example](#evaluation) to evaluate your models.

README_zh.md

@@ -22,6 +22,8 @@ https://github.com/hiyouga/LLaMA-Factory/assets/16256802/6ba60acc-e2e2-4bec-b846
 
 ## 更新日志
 
+[23/10/21] 我们支持了 [NEFTune](https://arxiv.org/abs/2310.05914) 优化。试试`--neftune_noise_alpha` 参数来激活 NEFTune，例如，`--neftune_noise_alpha 5`。
+
 [23/09/27] 我们针对 LLaMA 模型支持了 [LongLoRA](https://github.com/dvlab-research/LongLoRA) 提出的 **$S^2$-Attn**。请使用 `--shift_attn` 参数以启用该功能。
 
 [23/09/23] 我们在项目中集成了 MMLU、C-Eval 和 CMMLU 评估集。使用方法请参阅[此示例](#模型评估)。

src/llmtuner/hparams/finetuning_args.py

@@ -75,6 +75,10 @@ class FinetuningArguments:
         default=0.1,
         metadata={"help": "The beta parameter for the DPO loss."}
     )
+    neftune_noise_alpha: Optional[float] = field(
+        default=None,
+        metadata={"help": "The alpha parameter for the NEFTune noise. By setting this the NEFTune optimization will be activated."}
+    )
 
     def __post_init__(self):
         if isinstance(self.lora_target, str): # support custom target modules/layers of LoRA

src/llmtuner/tuner/sft/trainer.py

@@ -3,8 +3,10 @@ import json
 import torch
 import numpy as np
 import torch.nn as nn
+from functools import wraps
 from typing import TYPE_CHECKING, Any, Dict, List, Optional, Tuple, Union
-from transformers import Seq2SeqTrainer
+from transformers import Seq2SeqTrainer, PreTrainedModel, Trainer
+from peft import PeftModel
 
 from llmtuner.extras.constants import IGNORE_INDEX
 from llmtuner.extras.logging import get_logger
@@ -21,6 +23,14 @@ class CustomSeq2SeqTrainer(Seq2SeqTrainer):
     Inherits PeftTrainer to compute generative metrics such as BLEU and ROUGE.
     """
 
+    def __init__(self, model: Union["PreTrainedModel", nn.Module] = None, neftune_noise_alpha: Optional[float] = 0, **kwargs):
+        super().__init__(model, **kwargs)
+        self.neftune_noise_alpha = neftune_noise_alpha
+        self._neftune_activated = False
+
+        if self.neftune_noise_alpha:
+            self._activate_neftune(model)
+
     def prediction_step(
         self,
         model: nn.Module,
@@ -99,3 +109,71 @@ class CustomSeq2SeqTrainer(Seq2SeqTrainer):
             for pred, label in zip(decoded_preds, decoded_labels):
                 res.append(json.dumps({"label": label, "predict": pred}, ensure_ascii=False))
             writer.write("\n".join(res))
+
+
+    @wraps(Trainer.train)
+    def train(self, *args, **kwargs):
+        output = super().train(*args, **kwargs)
+
+        # After training we make sure to retrieve back the original forward pass method
+        # for the embedding layer.
+        if self.neftune_noise_alpha is not None:
+            self._deactivate_neftune(self.model)
+
+        return output
+
+    def _toggle_neftune(self, model, activate=True):
+        """Toggle NEFTune optimization for a model (i.e. activate or deactivate).
+        This optimization based on this paper: https://arxiv.org/abs/2310.05914
+
+        Parameters:
+        model : PreTrainedModel or PeftModel
+            The model to toggle the noise for.
+        activate : bool, optional (default=True)
+            Whether to activate the noise or not.
+        """
+        if activate == self._neftune_activated:
+            return
+
+        self._neftune_activated = activate
+
+        embeddings = (model.get_input_embeddings() if isinstance(model, PreTrainedModel) 
+                    else model.base_model.get_input_embeddings() if isinstance(model, PeftModel) 
+                    else None)
+
+        if embeddings:
+            if activate:
+                embeddings.neftune_noise_alpha = self.neftune_noise_alpha
+                embeddings._trl_old_forward = embeddings.forward
+                neftune_method = _neftune_forward_function.__get__(embeddings, embeddings.__class__)
+                setattr(embeddings, "forward", neftune_method)
+                logger.info("NEFTune activated with alpha: ", self.neftune_noise_alpha)
+            elif hasattr(embeddings, "_trl_old_forward"):
+                embeddings.forward = embeddings._trl_old_forward
+                del embeddings._trl_old_forward
+                del embeddings.neftune_noise_alpha
+                logger.info("NEFTune deactivated")
+
+    _activate_neftune = lambda self, model: self._toggle_neftune(model, activate=True)
+    _deactivate_neftune = lambda self, model: self._toggle_neftune(model, activate=False)
+
+
+def _neftune_forward_function(self, input: torch.Tensor) -> torch.Tensor:
+    """
+    This code is adapted from the original source code that can be found here: https://github.com/neelsjain/NEFTune
+    """
+    embeddings = torch.nn.functional.embedding(
+        input,
+        self.weight,
+        self.padding_idx,
+        self.max_norm,
+        self.norm_type,
+        self.scale_grad_by_freq,
+        self.sparse)
+
+    if self.training:
+        dims = torch.tensor(embeddings.size(1) * embeddings.size(2))
+        mag_norm = self.neftune_noise_alpha / torch.sqrt(dims)
+        embeddings += torch.zeros_like(embeddings).uniform_(-mag_norm, mag_norm)
+
+    return embeddings

src/llmtuner/tuner/sft/workflow.py

@@ -53,6 +53,7 @@ def run_sft(
         data_collator=data_collator,
         callbacks=callbacks,
         compute_metrics=ComputeMetrics(tokenizer) if training_args.predict_with_generate else None,
+        neftune_noise_alpha=finetuning_args.neftune_noise_alpha,
         **split_dataset(dataset, data_args, training_args)
     )