Condense, Don't Just Prune: Enhancing Efficiency and Performance in MoE Layer Pruning

Mixture-of-Experts (MOE) has garnered significant attention for their ability to scale up neural networks while utilizing the same or even fewer active parameters. However, MoE does not relieve the massive memory requirements of networks, which limits their practicality in real-world applications, especially in the era of large language models (LLMs). While recent work explores the possibility of removing entire layers of MoE to reduce memory, the performance degradation is still notable. In this paper, we propose Condense-MoE (CD-MoE} that, instead of dropping the entire MoE layer, condenses the big, sparse MoE layer into a small but dense layer with only a few experts that are activated for all tokens. Our approach is specifically designed for fine-grained MoE with shared experts, where Feed-Forward Networks are split into many small experts, with certain experts isolated to serve as shared experts that are always activated. We demonstrate the effectiveness of our method across multiple MoE models such as DeepSeekMoE and QwenMoE on various benchmarks. Specifically, for the DeepSeekMoE-16B model, our approach maintains nearly 90% of the average accuracy while reducing memory usage by 30% and enhancing inference speed by 30%. Moreover, we show that with lightweight expert fine-tuning, the pruned model can achieve further improvements on specific tasks. Our code are available at https://github.com/duterscmy/CD-MoE/tree/main.

AlphaPruning: Using Heavy-Tailed Self Regularization Theory for Improved Layer-wise Pruning of Large Language Models

Recent work on pruning large language models (LLMs) has shown that one can eliminate a large number of parameters without compromising performance, making pruning a promising strategy to reduce LLM model size. Existing LLM pruning strategies typically assign uniform pruning ratios across layers, limiting overall pruning ability; and recent work on layerwise pruning of LLMs is often based on heuristics that can easily lead to suboptimal performance. In this paper, we leverage Heavy-Tailed Self-Regularization (HT-SR) Theory, in particular the shape of empirical spectral densities (ESDs) of weight matrices, to design improved layerwise pruning ratios for LLMs. Our analysis reveals a wide variability in how well-trained, and thus relatedly how prunable, different layers of an LLM are. Based on this, we propose AlphaPruning, which uses shape metrics to allocate layerwise sparsity ratios in a more theoretically principled manner. AlphaPruning can be used in conjunction with multiple existing LLM pruning methods. Our empirical results show that AlphaPruning prunes LLaMA-7B to 80% sparsity while maintaining reasonable perplexity, marking a first in the literature on LLMs. We have open-sourced our code at https://github.com/haiquanlu/AlphaPruning.

Full-Rank No More: Low-Rank Weight Training for Modern Speech Recognition Models

This paper investigates the under-explored area of low-rank weight training for large-scale Conformer-based speech recognition models from scratch. Our study demonstrates the viability of this training paradigm for such models, yielding several notable findings. Firstly, we discover that applying a low-rank structure exclusively to the attention modules can unexpectedly enhance performance, even with a significant rank reduction of 12%. In contrast, feed-forward layers present greater challenges, as they begin to exhibit performance degradation with a moderate 50% rank reduction. Furthermore, we find that both initialization and layer-wise rank assignment play critical roles in successful low-rank training. Specifically, employing SVD initialization and linear layer-wise rank mapping significantly boosts the efficacy of low-rank weight training. Building on these insights, we introduce the Low-Rank Speech Model from Scratch (LR-SMS), an approach that achieves performance parity with full-rank training while delivering substantial reductions in parameters count (by at least 2x), and training time speedups (by 1.3x for ASR and 1.15x for AVSR).

Is C4 Dataset Optimal for Pruning? An Investigation of Calibration Data for LLM Pruning

Network pruning has emerged as a potential solution to make LLMs cheaper to deploy. However, existing LLM pruning approaches universally rely on the C4 dataset as the calibration data for calculating pruning scores, leaving its optimality unexplored. In this study, we evaluate the choice of calibration data on LLM pruning, across a wide range of datasets that are most commonly used in LLM training and evaluation, including four pertaining datasets as well as three categories of downstream tasks encompassing nine datasets. Each downstream dataset is prompted with In-Context Learning (ICL) and Chain-of-Thought (CoT), respectively. Besides the already intriguing observation that the choice of calibration data significantly impacts the performance of pruned LLMs, our results also uncover several subtle and often unexpected findings, summarized as follows: (1) C4 is not the optimal choice for LLM pruning, even among commonly used pre-training datasets; (2) arithmetic datasets, when used as calibration data, performs on par or even better than pre-training datasets; (3) pruning with downstream datasets does not necessarily help the corresponding downstream task, compared to pre-training data; (4) ICL is widely beneficial to all data categories, whereas CoT is only useful on certain tasks. Our findings shed light on the importance of carefully selecting calibration data for LLM pruning and pave the way for more efficient deployment of these powerful models in real-world applications. We release our code at: https://github.com/abx393/llm-pruning-calibration-data.

(PASS) Visual Prompt Locates Good Structure Sparsity through a Recurrent HyperNetwork

Large-scale neural networks have demonstrated remarkable performance in different domains like vision and language processing, although at the cost of massive computation resources. As illustrated by compression literature, structural model pruning is a prominent algorithm to encourage model efficiency, thanks to its acceleration-friendly sparsity patterns. One of the key questions of structural pruning is how to estimate the channel significance. In parallel, work on data-centric AI has shown that prompting-based techniques enable impressive generalization of large language models across diverse downstream tasks. In this paper, we investigate a charming possibility - \textit{leveraging visual prompts to capture the channel importance and derive high-quality structural sparsity}. To this end, we propose a novel algorithmic framework, namely \texttt{PASS}. It is a tailored hyper-network to take both visual prompts and network weight statistics as input, and output layer-wise channel sparsity in a recurrent manner. Such designs consider the intrinsic channel dependency between layers. Comprehensive experiments across multiple network architectures and six datasets demonstrate the superiority of \texttt{PASS} in locating good structural sparsity. For example, at the same FLOPs level, \texttt{PASS} subnetworks achieve $1\%\sim 3\%$ better accuracy on Food101 dataset; or with a similar performance of $80\%$ accuracy, \texttt{PASS} subnetworks obtain $0.35\times$ more speedup than the baselines.

From GaLore to WeLore: How Low-Rank Weights Non-uniformly Emerge from Low-Rank Gradients

Modern Large Language Models (LLMs) are composed of matrices with billions of elements, making their storage and processing quite demanding in terms of computational resources and memory usage. Being significantly large, such matrices can often be expressed in low-rank format with potential to relax resource requirements. Unlike prior works which focus on developing novel matrix decomposition algorithms, in this work we first study the emergence of low-rank structures across matrices within different layers of LLMs and establish a consequential relationship between the gradient dynamics and emerging low-rank expressiveness of matrices. Our findings reveal that different layers exhibit varying levels of converged low-rank structure, necessitating a non-uniform rank reduction across them to minimize performance drop due to compression. In view of that, we present Weight Low-Rank Projection (WeLore) that unifies weight compression and memory-efficient fine-tuning as ONE, in a data-agnostic and one-shot way. WeLore capitalizes the heavy-tail distribution of singular values to identify a suitable rank reduction ratio for matrices within LLMs. Going beyond only as a compression technique, WeLore categorizes weight matrices into Low-rank Components (LRCs) and Non-Low-rank Components (N-LRCs) based on their ability to express themselves as low-rank. Our gradient perspective and extensive experiments illustrate that LRCs tend to have better finetuning capabilities and can closely mimic (sometimes outperform) the training loss trajectory and performance of full-finetuning with notable memory and compute footprint reduction. For example, finetuning a 50\% compressed LLaMa-2 7B model using only a fraction of parameters in LRCs (WeLore) can outperform its full finetuning with ~3x better throughput and ~0.6x GPU requirement. Our codes are available at \url{https://github.com/VITA-Group/welore}

Q-GaLore: Quantized GaLore with INT4 Projection and Layer-Adaptive Low-Rank Gradients

Training Large Language Models (LLMs) is memory-intensive due to the large number of parameters and associated optimization states. GaLore, a recent method, reduces memory usage by projecting weight gradients into a low-rank subspace without compromising performance. However, GaLore relies on time-consuming Singular Value Decomposition (SVD) operations to identify the subspace, and the frequent subspace updates lead to significant training time overhead. Moreover, GaLore offers minimal improvements in accuracy and efficiency compared to LoRA in more accessible fine-tuning scenarios. To address these limitations, we introduce Q-Galore, a novel approach that substantially reduces memory usage by combining quantization and low-rank projection, surpassing the benefits of GaLore. Our method is based on two key observations: (i) the gradient subspace exhibits diverse properties, with some layers converging early in training while others are subject to frequent changes; (ii) the projection matrices are highly resilient to low-bit quantization. Leveraging these insights, Q-GaLore adaptively updates the gradient subspace based on its convergence statistics, achieving comparable performance while significantly reducing the number of SVD operations. We maintain the projection matrices in INT4 format and weights in INT8 format, incorporating stochastic rounding to capture accumulated gradient information. This approach enables a high-precision training trajectory using only low-precision weights. We demonstrate that Q-GaLore achieves highly competitive performance with exceptional memory efficiency. At pre-training, Q-GaLore facilitates training a LLaMA-7B model from scratch on a single NVIDIA RTX 4060 Ti with only 16 GB memory. At fine-tuning, it reduces memory consumption by up to 50% compared to LoRA and GaLore, while consistently outperforming QLoRA at the same memory cost.

Composable Interventions for Language Models

Test-time interventions for language models can enhance factual accuracy, mitigate harmful outputs, and improve model efficiency without costly retraining. But despite a flood of new methods, different types of interventions are largely developing independently. In practice, multiple interventions must be applied sequentially to the same model, yet we lack standardized ways to study how interventions interact. We fill this gap by introducing composable interventions, a framework to study the effects of using multiple interventions on the same language models, featuring new metrics and a unified codebase. Using our framework, we conduct extensive experiments and compose popular methods from three emerging intervention categories -- Knowledge Editing, Model Compression, and Machine Unlearning. Our results from 310 different compositions uncover meaningful interactions: compression hinders editing and unlearning, composing interventions hinges on their order of application, and popular general-purpose metrics are inadequate for assessing composability. Taken together, our findings showcase clear gaps in composability, suggesting a need for new multi-objective interventions. All of our code is public: https://github.com/hartvigsen-group/composable-interventions.

Dynamic Data Pruning for Automatic Speech Recognition

The recent success of Automatic Speech Recognition (ASR) is largely attributed to the ever-growing amount of training data. However, this trend has made model training prohibitively costly and imposed computational demands. While data pruning has been proposed to mitigate this issue by identifying a small subset of relevant data, its application in ASR has been barely explored, and existing works often entail significant overhead to achieve meaningful results. To fill this gap, this paper presents the first investigation of dynamic data pruning for ASR, finding that we can reach the full-data performance by dynamically selecting 70% of data. Furthermore, we introduce Dynamic Data Pruning for ASR (DDP-ASR), which offers several fine-grained pruning granularities specifically tailored for speech-related datasets, going beyond the conventional pruning of entire time sequences. Our intensive experiments show that DDP-ASR can save up to 1.6x training time with negligible performance loss.

MSRS: Training Multimodal Speech Recognition Models from Scratch with Sparse Mask Optimization

Pre-trained models have been a foundational approach in speech recognition, albeit with associated additional costs. In this study, we propose a regularization technique that facilitates the training of visual and audio-visual speech recognition models (VSR and AVSR) from scratch. This approach, abbreviated as \textbf{MSRS} (Multimodal Speech Recognition from Scratch), introduces a sparse regularization that rapidly learns sparse structures within the dense model at the very beginning of training, which receives healthier gradient flow than the dense equivalent. Once the sparse mask stabilizes, our method allows transitioning to a dense model or keeping a sparse model by updating non-zero values. MSRS achieves competitive results in VSR and AVSR with 21.1% and 0.9% WER on the LRS3 benchmark, while reducing training time by at least 2x. We explore other sparse approaches and show that only MSRS enables training from scratch by implicitly masking the weights affected by vanishing gradients.