Large Language Models (LLMs) pruning seeks to remove unimportant weights for inference speedup with minimal performance impact. However, existing methods often suffer from performance loss without full-model sparsity-aware fine-tuning. This paper presents Wanda++, a novel pruning framework that outperforms the state-of-the-art methods by utilizing decoder-block-level \textbf{regional} gradients. Specifically, Wanda++ improves the pruning score with regional gradients for the first time and proposes an efficient regional optimization method to minimize pruning-induced output discrepancies between the dense and sparse decoder output. Notably, Wanda++ improves perplexity by up to 32\% over Wanda in the language modeling task and generalizes effectively to downstream tasks. Further experiments indicate our proposed method is orthogonal to sparsity-aware fine-tuning, where Wanda++ can be combined with LoRA fine-tuning to achieve a similar perplexity improvement as the Wanda method. The proposed method is lightweight, pruning a 7B LLaMA model in under 10 minutes on a single NVIDIA H100 GPU.

Language Models (LLMs) are often quantized to lower precision to reduce the memory cost and latency in inference. However, quantization often degrades model performance, thus fine-tuning is required for various down-stream tasks. Traditional fine-tuning methods such as stochastic gradient descent and Adam optimization require backpropagation, which are error-prone in the low-precision settings. To overcome these limitations, we propose the Quantized Zeroth-Order (QuZO) framework, specifically designed for fine-tuning LLMs through low-precision (e.g., 4- or 8-bit) forward passes. Our method can avoid the error-prone low-precision straight-through estimator, and utilizes optimized stochastic rounding to mitigate the increased bias. QuZO simplifies the training process, while achieving results comparable to first-order methods in ${\rm FP}8$ and superior accuracy in ${\rm INT}8$ and ${\rm INT}4$ training. Experiments demonstrate that low-bit training QuZO achieves performance comparable to MeZO optimization on GLUE, Multi-Choice, and Generation tasks, while reducing memory cost by $2.94 \times$ in LLaMA2-7B fine-tuning compared to quantized first-order methods.

Large language models have demonstrated exceptional capabilities across diverse tasks, but their fine-tuning demands significant memory, posing challenges for resource-constrained environments. Zeroth-order (ZO) optimization provides a memory-efficient alternative by eliminating the need for backpropagation. However, ZO optimization suffers from high gradient variance, and prior research has largely focused on single-task learning, leaving its application to multi-task learning unexplored. Multi-task learning is crucial for leveraging shared knowledge across tasks to improve generalization, yet it introduces unique challenges under ZO settings, such as amplified gradient variance and collinearity. In this paper, we present MaZO, the first framework specifically designed for multi-task LLM fine-tuning under ZO optimization. MaZO tackles these challenges at the parameter level through two key innovations: a weight importance metric to identify critical parameters and a multi-task weight update mask to selectively update these parameters, reducing the dimensionality of the parameter space and mitigating task conflicts. Experiments demonstrate that MaZO achieves state-of-the-art performance, surpassing even multi-task learning methods designed for first-order optimization.

Fine-tuning large language models (LLMs) has achieved remarkable performance across various natural language processing tasks, yet it demands more and more memory as model sizes keep growing. To address this issue, the recently proposed Memory-efficient Zeroth-order (MeZO) methods attempt to fine-tune LLMs using only forward passes, thereby avoiding the need for a backpropagation graph. However, significant performance drops and a high risk of divergence have limited their widespread adoption. In this paper, we propose the Adaptive Zeroth-order Tensor-Train Adaption (AdaZeta) framework, specifically designed to improve the performance and convergence of the ZO methods. To enhance dimension-dependent ZO estimation accuracy, we introduce a fast-forward, low-parameter tensorized adapter. To tackle the frequently observed divergence issue in large-scale ZO fine-tuning tasks, we propose an adaptive query number schedule that guarantees convergence. Detailed theoretical analysis and extensive experimental results on Roberta-Large and Llama-2-7B models substantiate the efficacy of our AdaZeta framework in terms of accuracy, memory efficiency, and convergence speed.

Machine learning model weights and activations are represented in full-precision during training. This leads to performance degradation in runtime when deployed on neural network accelerator (NNA) chips, which leverage highly parallelized fixed-point arithmetic to improve runtime memory and latency. In this work, we replicate the NNA operators during the training phase, accounting for the degradation due to low-precision inference on the NNA in back-propagation. Our proposed method efficiently emulates NNA operations, thus foregoing the need to transfer quantization error-prone data to the Central Processing Unit (CPU), ultimately reducing the user perceived latency (UPL). We apply our approach to Recurrent Neural Network-Transducer (RNN-T), an attractive architecture for on-device streaming speech recognition tasks. We train and evaluate models on 270K hours of English data and show a 5-7% improvement in engine latency while saving up to 10% relative degradation in WER.

Attention-based contextual biasing approaches have shown significant improvements in the recognition of generic and/or personal rare-words in End-to-End Automatic Speech Recognition (E2E ASR) systems like neural transducers. These approaches employ cross-attention to bias the model towards specific contextual entities injected as bias-phrases to the model. Prior approaches typically relied on subword encoders for encoding the bias phrases. However, subword tokenizations are coarse and fail to capture granular pronunciation information which is crucial for biasing based on acoustic similarity. In this work, we propose to use lightweight character representations to encode fine-grained pronunciation features to improve contextual biasing guided by acoustic similarity between the audio and the contextual entities (termed acoustic biasing). We further integrate pretrained neural language model (NLM) based encoders to encode the utterance's semantic context along with contextual entities to perform biasing informed by the utterance's semantic context (termed semantic biasing). Experiments using a Conformer Transducer model on the Librispeech dataset show a 4.62% - 9.26% relative WER improvement on different biasing list sizes over the baseline contextual model when incorporating our proposed acoustic and semantic biasing approach. On a large-scale in-house dataset, we observe 7.91% relative WER improvement compared to our baseline model. On tail utterances, the improvements are even more pronounced with 36.80% and 23.40% relative WER improvements on Librispeech rare words and an in-house testset respectively.

We present dual-attention neural biasing, an architecture designed to boost Wake Words (WW) recognition and improve inference time latency on speech recognition tasks. This architecture enables a dynamic switch for its runtime compute paths by exploiting WW spotting to select which branch of its attention networks to execute for an input audio frame. With this approach, we effectively improve WW spotting accuracy while saving runtime compute cost as defined by floating point operations (FLOPs). Using an in-house de-identified dataset, we demonstrate that the proposed dual-attention network can reduce the compute cost by $90\%$ for WW audio frames, with only $1\%$ increase in the number of parameters. This architecture improves WW F1 score by $16\%$ relative and improves generic rare word error rate by $3\%$ relative compared to the baselines.

End-to-end (E2E) spoken language understanding (SLU) systems predict utterance semantics directly from speech using a single model. Previous work in this area has focused on targeted tasks in fixed domains, where the output semantic structure is assumed a priori and the input speech is of limited complexity. In this work we present our approach to developing an E2E model for generalized SLU in commercial voice assistants (VAs). We propose a fully differentiable, transformer-based, hierarchical system that can be pretrained at both the ASR and NLU levels. This is then fine-tuned on both transcription and semantic classification losses to handle a diverse set of intent and argument combinations. This leads to an SLU system that achieves significant improvements over baselines on a complex internal generalized VA dataset with a 43% improvement in accuracy, while still meeting the 99% accuracy benchmark on the popular Fluent Speech Commands dataset. We further evaluate our model on a hard test set, exclusively containing slot arguments unseen in training, and demonstrate a nearly 20% improvement, showing the efficacy of our approach in truly demanding VA scenarios.

In order to achieve high accuracy for machine learning (ML) applications, it is essential to employ models with a large number of parameters. Certain applications, such as Automatic Speech Recognition (ASR), however, require real-time interactions with users, hence compelling the model to have as low latency as possible. Deploying large scale ML applications thus necessitates model quantization and compression, especially when running ML models on resource constrained devices. For example, by forcing some of the model weight values into zero, it is possible to apply zero-weight compression, which reduces both the model size and model reading time from the memory. In the literature, such methods are referred to as sparse pruning. The fundamental questions are when and which weights should be forced to zero, i.e. be pruned. In this work, we propose a compressed sensing based pruning (CSP) approach to effectively address those questions. By reformulating sparse pruning as a sparsity inducing and compression-error reduction dual problem, we introduce the classic compressed sensing process into the ML model training process. Using ASR task as an example, we show that CSP consistently outperforms existing approaches in the literature.