DistiLLM: Towards Streamlined Distillation for Large Language Models

• URL: http://arxiv.org/abs/2402.03898v2
• Date: Wed, 3 Jul 2024 04:57:41 GMT
• Title: DistiLLM: Towards Streamlined Distillation for Large Language Models
• Authors: Jongwoo Ko, Sungnyun Kim, Tianyi Chen, Se-Young Yun,
• Abstract summary: DistiLLM is a more effective and efficient KD framework for auto-regressive language models. DisiLLM comprises two components: (1) a novel skew Kullback-Leibler divergence loss, where we unveil and leverage its theoretical properties, and (2) an adaptive off-policy approach designed to enhance the efficiency in utilizing student-generated outputs.
• Score: 53.46759297929675
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Knowledge distillation (KD) is widely used for compressing a teacher model to a smaller student model, reducing its inference cost and memory footprint while preserving model capabilities. However, current KD methods for auto-regressive sequence models (e.g., large language models) suffer from missing a standardized objective function. Moreover, the recent use of student-generated outputs to address training-inference mismatches has significantly escalated computational costs. To tackle these issues, we introduce DistiLLM, a more effective and efficient KD framework for auto-regressive language models. DistiLLM comprises two components: (1) a novel skew Kullback-Leibler divergence loss, where we unveil and leverage its theoretical properties, and (2) an adaptive off-policy approach designed to enhance the efficiency in utilizing student-generated outputs. Extensive experiments, including instruction-following tasks, demonstrate the effectiveness of DistiLLM in building high-performing student models while achieving up to 4.3$\times$ speedup compared to recent KD methods.

Related papers

• Exploring and Enhancing the Transfer of Distribution in Knowledge Distillation for Autoregressive Language Models [62.5501109475725]
  Knowledge distillation (KD) is a technique that compresses large teacher models by training smaller student models to mimic them. This paper introduces Online Knowledge Distillation (OKD), where the teacher network integrates small online modules to concurrently train with the student model. OKD achieves or exceeds the performance of leading methods in various model architectures and sizes, reducing training time by up to fourfold.
  arXiv  Detail & Related papers  (2024-09-19T07:05:26Z)
• Robust Knowledge Distillation Based on Feature Variance Against Backdoored Teacher Model [13.367731896112861]
  Knowledge distillation (KD) is one of the widely used compression techniques for edge deployment. This paper proposes RobustKD, a robust KD that compresses the model while mitigating backdoor based on feature variance.
  arXiv  Detail & Related papers  (2024-06-01T11:25:03Z)
• Revisiting Knowledge Distillation for Autoregressive Language Models [88.80146574509195]
  We propose a simple yet effective adaptive teaching approach (ATKD) to improve the knowledge distillation (KD) The core of ATKD is to reduce rote learning and make teaching more diverse and flexible. Experiments on 8 LM tasks show that, with the help of ATKD, various baseline KD methods can achieve consistent and significant performance gains.
  arXiv  Detail & Related papers  (2024-02-19T07:01:10Z)
• Co-training and Co-distillation for Quality Improvement and Compression of Language Models [88.94539115180919]
  Knowledge Distillation (KD) compresses expensive pre-trained language models (PLMs) by transferring their knowledge to smaller models. Most smaller models fail to surpass the performance of the original larger model, resulting in sacrificing performance to improve inference speed. We propose Co-Training and Co-Distillation (CTCD), a novel framework that improves performance and inference speed together by co-training two models.
  arXiv  Detail & Related papers  (2023-11-06T03:29:00Z)
• On-Policy Distillation of Language Models: Learning from Self-Generated Mistakes [44.97759066341107]
  Generalized Knowledge Distillation (GKD) trains the student on its self-generated output sequences by leveraging feedback from the teacher. We demonstrate the efficacy of GKD for distilling auto-regressive language models on summarization, translation, and arithmetic reasoning tasks.
  arXiv  Detail & Related papers  (2023-06-23T17:56:26Z)
• DisCo: Effective Knowledge Distillation For Contrastive Learning of Sentence Embeddings [36.37939188680754]
  We propose an enhanced knowledge distillation framework termed Distill-Contrast (DisCo) DisCo transfers the capability of a large sentence embedding model to a small student model on large unlabelled data. We also propose Contrastive Knowledge Distillation (CKD) to enhance the consistencies among teacher model training, KD, and student model finetuning.
  arXiv  Detail & Related papers  (2021-12-10T16:11:23Z)
• How and When Adversarial Robustness Transfers in Knowledge Distillation? [137.11016173468457]
  This paper studies how and when the adversarial robustness can be transferred from a teacher model to a student model in Knowledge distillation (KD) We show that standard KD training fails to preserve adversarial robustness, and we propose KD with input gradient alignment (KDIGA) for remedy. Under certain assumptions, we prove that the student model using our proposed KDIGA can achieve at least the same certified robustness as the teacher model.
  arXiv  Detail & Related papers  (2021-10-22T21:30:53Z)
• MixKD: Towards Efficient Distillation of Large-scale Language Models [129.73786264834894]
  We propose MixKD, a data-agnostic distillation framework, to endow the resulting model with stronger generalization ability. We prove from a theoretical perspective that under reasonable conditions MixKD gives rise to a smaller gap between the error and the empirical error. Experiments under a limited-data setting and ablation studies further demonstrate the advantages of the proposed approach.
  arXiv  Detail & Related papers  (2020-11-01T18:47:51Z)
• Pea-KD: Parameter-efficient and Accurate Knowledge Distillation on BERT [20.732095457775138]
  Knowledge Distillation (KD) is one of the widely known methods for model compression. Pea-KD consists of two main parts: Shuffled Sharing (SPS) and Pretraining with Teacher's Predictions (PTP)
  arXiv  Detail & Related papers  (2020-09-30T17:52:15Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.