Is it always necessary to compute tokens from shallow to deep layers in Transformers? The continued success of vanilla Transformers and their variants suggests an undoubted “yes”. In this work, however, we attempt to break the depth-ordered convention by proposing a novel architecture dubbed mixture-of-modules (MoM), which is motivated by an intuition that any layer, regardless of its position, can be used to compute a token as long as it possesses the needed processing capabilities. The construction of MoM starts from a finite set of modules defined by multi-head attention and feed-forward networks, each distinguished by its unique parameterization. Two routers then iteratively select attention modules and feed-forward modules from the set to process a token. The selection dynamically expands the computation graph in the forward pass of the token, culminating in an assembly of modules. We show that MoM provides not only a unified framework for Transformers and their numerous variants but also a flexible and learnable approach for reducing redundancy in Transformer parameterization. We pre-train various MoMs using OpenWebText. Empirical results demonstrate that MoMs, of different sizes, consistently outperform vanilla transformers. More interestingly, after removing 50% of the multi-head attention modules and 25% of the feed-forward modules, an MoM model still holds comparable performance. Additionally, by properly adjusting the number of modules and compressing the model depth, one can have an MoM that achieves comparable performance to GPT-2 (774M) while saving 16% TFLOPs and 42% memory usage during forward computation.

Speculative decoding has emerged as a promising technique to accelerate the inference of Large Language Models (LLMs) by employing a small language model to draft a hypothesis sequence, which is then validated by the LLM. The effectiveness of this approach heavily relies on the balance between performance and efficiency of the draft model. In our research, we focus on enhancing the proportion of draft tokens that are accepted to the final output by generating multiple hypotheses instead of just one. This allows the LLM more options to choose from and select the longest sequence that meets its standards. Our analysis reveals that hypotheses produced by the draft model share many common token sequences, suggesting a potential for optimizing computation. Leveraging this observation, we introduce an innovative approach utilizing a directed acyclic graph (DAG) to manage the drafted hypotheses. This structure enables us to efficiently predict and merge recurring token sequences, vastly reducing the computational demands of the draft model. We term this approach Graph-structured Speculative Decoding (GSD). We apply GSD across a range of LLMs, including a 70-billion parameter LLaMA-2 model, and observe a remarkable speedup of 1.70× to 1.94 ×, significantly surpassing standard speculative decoding.

While transformer-based pre-trained language models (PLMs) have dominated a number of NLP applications, these models are heavy to deploy and expensive to use. Therefore, effectively compressing large-scale PLMs becomes an increasingly important problem. Quantization, which represents high-precision tensors with low-bit fix-point format, is a viable solution. However, most existing quantization methods are task-specific, requiring customized training and quantization with a large number of trainable parameters on each individual task. Inspired by the observation that the over-parameterization nature of PLMs makes it possible to freeze most of the parameters during the fine-tuning stage, in this work, we propose a novel “quantize before fine-tuning” framework, PreQuant, that differs from both quantization-aware training and post-training quantization. {pasted macro ‘OUR’} is compatible with various quantization strategies, with outlier-aware parameter-efficient fine-tuning incorporated to correct the induced quantization error. We demonstrate the effectiveness of PreQuant on the GLUE benchmark using BERT, RoBERTa, and T5. We also provide an empirical investigation into the workflow of PreQuant, which sheds light on its efficacy.

In-context learning (ICL) is an emerging capability of large autoregressive language models where a few input-label demonstrations are appended to the input to enhance the model’s understanding of downstream NLP tasks, without directly adjusting the model parameters. The effectiveness of ICL can be attributed to the strong language modeling capabilities of large language models (LLMs), which enable them to learn the mapping between input and labels based on in-context demonstrations. Despite achieving promising results, the causal nature of language modeling in ICL restricts the attention to be backward only, i.e., a token only attends to its previous tokens, failing to capture the full input-label information and limiting the model’s performance. In this paper, we propose a novel ICL method called Repeated Demonstration with Sliding Causal Attention, (RdSca). Specifically, we duplicate later demonstrations and concatenate them to the front, allowing the model to ‘observe’ the later information even under the causal restriction. Besides, we introduce sliding causal attention, which customizes causal attention to avoid information leakage. Experimental results show that our method significantly improves the input-label mapping in ICL demonstrations. We also conduct an in-depth analysis of how to customize the causal attention without training, which has been an unexplored area in previous research.

The Lottery Ticket Hypothesis suggests that for any over-parameterized model, a small subnetwork exists to achieve competitive performance compared to the backbone architecture. In this paper, we study whether there is a winning lottery ticket for pre-trained language models, which allow the practitioners to fine-tune the parameters in the ticket but achieve good downstream performance. To achieve this, we regularize the fine-tuning process with L1 distance and explore the subnetwork structure (what we refer to as the “dominant winning ticket”). Empirically, we show that (a) the dominant winning ticket can achieve performance that is comparable with that of the full-parameter model, (b) the dominant winning ticket is transferable across different tasks, (c) and the dominant winning ticket has a natural structure within each parameter matrix. Strikingly, we find that a dominant winning ticket that takes up 0.05% of the parameters can already achieve satisfactory performance, indicating that the PLM is significantly reducible during fine-tuning.