Taesu Kim

Changhun Lee, Jungyu Jin, Taesu Kim et al.

Large language models (LLMs) with hundreds of billions of parameters require powerful server-grade GPUs for inference, limiting their practical deployment. To address this challenge, we introduce the outlier-aware weight quantization (OWQ) method, which aims to minimize LLM's footprint through low-precision representation. OWQ prioritizes a small subset of structured weights sensitive to quantization, storing them in high-precision, while applying highly tuned quantization to the remaining dense weights. This sensitivity-aware mixed-precision scheme reduces the quantization error notably, and extensive experiments demonstrate that 3.1-bit models using OWQ perform comparably to 4-bit models optimized by OPTQ. Furthermore, OWQ incorporates a parameter-efficient fine-tuning for task-specific adaptation, called weak column tuning (WCT), enabling accurate task-specific LLM adaptation with minimal memory overhead in the optimized format. OWQ represents a notable advancement in the flexibility, efficiency, and practicality of LLM optimization literature. The source code is available at https://github.com/xvyaward/owq

Jiwoong Choi, Minkyu Kim, Daehyun Ahn et al.

The emergence of diffusion models has greatly broadened the scope of high-fidelity image synthesis, resulting in notable advancements in both practical implementation and academic research. With the active adoption of the model in various real-world applications, the need for on-device deployment has grown considerably. However, deploying large diffusion models such as Stable Diffusion with more than one billion parameters to mobile devices poses distinctive challenges due to the limited computational and memory resources, which may vary according to the device. In this paper, we present the challenges and solutions for deploying Stable Diffusion on mobile devices with TensorFlow Lite framework, which supports both iOS and Android devices. The resulting Mobile Stable Diffusion achieves the inference latency of smaller than 7 seconds for a 512x512 image generation on Android devices with mobile GPUs.

Yookyung Shin, Younggun Lee, Suhee Jo et al.

Expressive text-to-speech has shown improved performance in recent years. However, the style control of synthetic speech is often restricted to discrete emotion categories and requires training data recorded by the target speaker in the target style. In many practical situations, users may not have reference speech recorded in target emotion but still be interested in controlling speech style just by typing text description of desired emotional style. In this paper, we propose a text-based interface for emotional style control and cross-speaker style transfer in multi-speaker TTS. We propose the bi-modal style encoder which models the semantic relationship between text description embedding and speech style embedding with a pretrained language model. To further improve cross-speaker style transfer on disjoint, multi-style datasets, we propose the novel style loss. The experimental results show that our model can generate high-quality expressive speech even in unseen style.

Suhee Jo, Younggun Lee, Yookyung Shin et al.

In recent years, emotional text-to-speech has shown considerable progress. However, it requires a large amount of labeled data, which is not easily accessible. Even if it is possible to acquire an emotional speech dataset, there is still a limitation in controlling emotion intensity. In this work, we propose a novel method for cross-speaker emotion transfer and manipulation using vector arithmetic in latent style space. By leveraging only a few labeled samples, we generate emotional speech from reading-style speech without losing the speaker identity. Furthermore, emotion strength is readily controllable using a scalar value, providing an intuitive way for users to manipulate speech. Experimental results show the proposed method affords superior performance in terms of expressiveness, naturalness, and controllability, preserving speaker identity.

Jiwon Song, Kyungseok Oh, Taesu Kim et al.

Large language models (LLMs) have proven to be highly effective across various natural language processing tasks. However, their large number of parameters poses significant challenges for practical deployment. Pruning, a technique aimed at reducing the size and complexity of LLMs, offers a potential solution by removing redundant components from the network. Despite the promise of pruning, existing methods often struggle to achieve substantial end-to-end LLM inference speedup. In this paper, we introduce SLEB, a novel approach designed to streamline LLMs by eliminating redundant transformer blocks. We choose the transformer block as the fundamental unit for pruning, because LLMs exhibit block-level redundancy with high similarity between the outputs of neighboring blocks. This choice allows us to effectively enhance the processing speed of LLMs. Our experimental results demonstrate that SLEB outperforms previous LLM pruning methods in accelerating LLM inference while also maintaining superior perplexity and accuracy, making SLEB as a promising technique for enhancing the efficiency of LLMs. The code is available at: https://github.com/jiwonsong-dev/SLEB.

Gyunpyo Lee, Taesu Kim, Hyeon-Jeong Suk

An automated design data archiving could reduce the time wasted by designers from working creatively and effectively. Though many datasets on classifying, detecting, and instance segmenting on car exterior exist, these large datasets are not relevant for design practices as the primary purpose lies in autonomous driving or vehicle verification. Therefore, we release GP22, composed of car styling features defined by automotive designers. The dataset contains 1480 car side profile images from 37 brands and ten car segments. It also contains annotations of design features that follow the taxonomy of the car exterior design features defined in the eye of the automotive designer. We trained the baseline model using YOLO v5 as the design feature detection model with the dataset. The presented model resulted in an mAP score of 0.995 and a recall of 0.984. Furthermore, exploration of the model performance on sketches and rendering images of the car side profile implies the scalability of the dataset for design purposes.

Yeonjoon Jung, Daehyun Ahn, Hyungjun Kim et al.

Low-Rank Adaptation (LoRA) is a popular method for parameter-efficient fine-tuning (PEFT) of generative models, valued for its simplicity and effectiveness. Despite recent enhancements, LoRA still suffers from a fundamental limitation: overfitting when the bottleneck is widened. It performs best at ranks 32-64, yet its accuracy stagnates or declines at higher ranks, still falling short of full fine-tuning (FFT) performance. We identify the root cause as LoRA's structural bottleneck, which introduces gradient entanglement to the unrelated input channels and distorts gradient propagation. To address this, we introduce a novel structure, Granular Low-Rank Adaptation (GraLoRA) that partitions weight matrices into sub-blocks, each with its own low-rank adapter. With negligible computational or storage cost, GraLoRA overcomes LoRA's limitations, effectively increases the representational capacity, and more closely approximates FFT behavior. Experiments on code generation and commonsense reasoning benchmarks show that GraLoRA consistently outperforms LoRA and other baselines, achieving up to +8.5% absolute gain in Pass@1 on HumanEval+. These improvements hold across model sizes and rank settings, making GraLoRA a scalable and robust solution for PEFT. Code, data, and scripts are available at https://github.com/SqueezeBits/GraLoRA.git

Taesu Kim, Jongho Lee, Daehyun Ahn et al.

We introduce QUICK, a group of novel optimized CUDA kernels for the efficient inference of quantized Large Language Models (LLMs). QUICK addresses the shared memory bank-conflict problem of state-of-the-art mixed precision matrix multiplication kernels. Our method interleaves the quantized weight matrices of LLMs offline to skip the shared memory write-back after the dequantization. We demonstrate up to 1.91x speedup over existing kernels of AutoAWQ on larger batches and up to 1.94x throughput gain on representative LLM models on various NVIDIA GPU devices.

Yunjae Lee, Juntaek Lim, Jehyeon Bang et al.

This paper presents a comprehensive evaluation of Intel Gaudi NPUs as an alternative to NVIDIA GPUs, which is currently the de facto standard in AI system design. First, we create a suite of microbenchmarks to compare Intel Gaudi-2 with NVIDIA A100, showing that Gaudi-2 achieves competitive performance not only in primitive AI compute, memory, and communication operations but also in executing several important AI workloads end-to-end. We then assess Gaudi NPU's programmability by discussing several software-level optimization strategies to employ for implementing critical FBGEMM operators and vLLM, evaluating their efficiency against GPU-optimized counterparts. Results indicate that Gaudi-2 achieves energy efficiency comparable to A100, though there are notable areas for improvement in terms of software maturity. Overall, we conclude that, with effective integration into high-level AI frameworks, Gaudi NPUs could challenge NVIDIA GPU's dominance in the AI server market, though further improvements are necessary to fully compete with NVIDIA's robust software ecosystem.

Dongwon Jo, Taesu Kim, Yulhwa Kim et al.

Binarization, which converts weight parameters to binary values, has emerged as an effective strategy to reduce the size of large language models (LLMs). However, typical binarization techniques significantly diminish linguistic effectiveness of LLMs. To address this issue, we introduce a novel binarization technique called Mixture of Scales (BinaryMoS). Unlike conventional methods, BinaryMoS employs multiple scaling experts for binary weights, dynamically merging these experts for each token to adaptively generate scaling factors. This token-adaptive approach boosts the representational power of binarized LLMs by enabling contextual adjustments to the values of binary weights. Moreover, because this adaptive process only involves the scaling factors rather than the entire weight matrix, BinaryMoS maintains compression efficiency similar to traditional static binarization methods. Our experimental results reveal that BinaryMoS surpasses conventional binarization techniques in various natural language processing tasks and even outperforms 2-bit quantization methods, all while maintaining similar model size to static binarization techniques.

Jaesung Tae, Hyeongju Kim, Taesu Kim

We present EdiTTS, an off-the-shelf speech editing methodology based on score-based generative modeling for text-to-speech synthesis. EdiTTS allows for targeted, granular editing of audio, both in terms of content and pitch, without the need for any additional training, task-specific optimization, or architectural modifications to the score-based model backbone. Specifically, we apply coarse yet deliberate perturbations in the Gaussian prior space to induce desired behavior from the diffusion model while applying masks and softening kernels to ensure that iterative edits are applied only to the target region. Through listening tests and speech-to-text back transcription, we show that EdiTTS outperforms existing baselines and produces robust samples that satisfy user-imposed requirements.

Suwon Shon, Younggun Lee, Taesu Kim

This paper describes a fast speaker search system to retrieve segments of the same voice identity in the large-scale data. A recent study shows that Locality Sensitive Hashing (LSH) enables quick retrieval of a relevant voice in the large-scale data in conjunction with i-vector while maintaining accuracy. In this paper, we proposed Random Speaker-variability Subspace (RSS) projection to map a data into LSH based hash tables. We hypothesized that rather than projecting on completely random subspace without considering data, projecting on randomly generated speaker variability space would give more chance to put the same speaker representation into the same hash bins, so we can use less number of hash tables. Multiple RSS can be generated by randomly selecting a subset of speakers from a large speaker cohort. From the experimental result, the proposed approach shows 100 times and 7 times faster than the linear search and LSH, respectively

Younggun Lee, Suwon Shon, Taesu Kim

Although there are more than 6,500 languages in the world, the pronunciations of many phonemes sound similar across the languages. When people learn a foreign language, their pronunciation often reflects their native language's characteristics. This motivates us to investigate how the speech synthesis network learns the pronunciation from datasets from different languages. In this study, we are interested in analyzing and taking advantage of multilingual speech synthesis network. First, we train the speech synthesis network bilingually in English and Korean and analyze how the network learns the relations of phoneme pronunciation between the languages. Our experimental result shows that the learned phoneme embedding vectors are located closer if their pronunciations are similar across the languages. Consequently, the trained networks can synthesize the English speakers' Korean speech and vice versa. Using this result, we propose a training framework to utilize information from a different language. To be specific, we pre-train a speech synthesis network using datasets from both high-resource language and low-resource language, then we fine-tune the network using the low-resource language dataset. Finally, we conducted more simulations on 10 different languages to show it is generally extendable to other languages.

Younggun Lee, Taesu Kim

We propose prosody embeddings for emotional and expressive speech synthesis networks. The proposed methods introduce temporal structures in the embedding networks, thus enabling fine-grained control of the speaking style of the synthesized speech. The temporal structures can be designed either on the speech side or the text side, leading to different control resolutions in time. The prosody embedding networks are plugged into end-to-end speech synthesis networks and trained without any other supervision except for the target speech for synthesizing. It is demonstrated that the prosody embedding networks learned to extract prosodic features. By adjusting the learned prosody features, we could change the pitch and amplitude of the synthesized speech both at the frame level and the phoneme level. We also introduce the temporal normalization of prosody embeddings, which shows better robustness against speaker perturbations during prosody transfer tasks.

Younggun Lee, Taesu Kim, Soo-Young Lee

We propose a neural text-to-speech (TTS) model that can imitate a new speaker's voice using only a small amount of speech sample. We demonstrate voice imitation using only a 6-seconds long speech sample without any other information such as transcripts. Our model also enables voice imitation instantly without additional training of the model. We implemented the voice imitating TTS model by combining a speaker embedder network with a state-of-the-art TTS model, Tacotron. The speaker embedder network takes a new speaker's speech sample and returns a speaker embedding. The speaker embedding with a target sentence are fed to Tacotron, and speech is generated with the new speaker's voice. We show that the speaker embeddings extracted by the speaker embedder network can represent the latent structure in different voices. The generated speech samples from our model have comparable voice quality to the ones from existing multi-speaker TTS models.

Hyungjun Kim, Taesu Kim, Jinseok Kim et al.

Artificial Neural Network computation relies on intensive vector-matrix multiplications. Recently, the emerging nonvolatile memory (NVM) crossbar array showed a feasibility of implementing such operations with high energy efficiency, thus there are many works on efficiently utilizing emerging NVM crossbar array as analog vector-matrix multiplier. However, its nonlinear I-V characteristics restrain critical design parameters, such as the read voltage and weight range, resulting in substantial accuracy loss. In this paper, instead of optimizing hardware parameters to a given neural network, we propose a methodology of reconstructing a neural network itself optimized to resistive memory crossbar arrays. To verify the validity of the proposed method, we simulated various neural network with MNIST and CIFAR-10 dataset using two different specific Resistive Random Access Memory (RRAM) model. Simulation results show that our proposed neural network produces significantly higher inference accuracies than conventional neural network when the synapse devices have nonlinear I-V characteristics.