Lei Gao
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Lei Gao has a total of 23 co-authors. Here are the top co-authors by number of shared publications:
- Murali Annavaram
University of Southern California
6 shared publications - Salman Avestimehr
Department of Electrical and Computer Engineering, University of Southern California, Los Angeles, USA
4 shared publications Hossein Entezari Zarch
University of Southern California (USC
3 shared publications- Yue Niu
University of Southern California
3 shared publications - Chaoyi Jiang
University of Southern California
2 shared publications - Tingting Tang
University of Southern California
2 shared publications
Latest Publications
- Open Access
Striking the Right Balance between Compute and Copy: Improving LLM Inferencing Under Speculative Decoding
With the skyrocketing costs of GPUs and their virtual instances in the cloud, there is a significant desire to use CPUs for large language model (LLM) inference. KV cache update, often implemented as allocation, copying, and in-place strided update for each generated token, incurs significant overhead. As the sequence length increases, the allocation and copy overheads dominate the performance. Alternate approaches may allocate large KV tensors upfront to enable in-place updates, but these matrices (with zero-padded rows) cause redundant computations. In this work, we propose a new KV cache allocation mechanism called Balancing Memory and Compute (BMC). BMC allocates, once every r iterations, KV tensors with r redundant rows, allowing in-place update without copy overhead for those iterations, but at the expense of a small amount of redundant computation. Second, we make an interesting observation that the extra rows allocated in the KV tensors and the resulting redundant computation can be repurposed for Speculative Decoding (SD) that improves token generation efficiency. Last, BMC represents a spectrum of design points with different values of r. To identify the best-performing design point(s), we derive a simple analytical model for BMC. The proposed BMC method achieves an average throughput acceleration of up to 3.2x over baseline HuggingFace (without SD). Importantly when we apply BMC with SD, it results in an additional speedup of up to 1.39x, over and above the speedup offered by SD. Further, BMC achieves a throughput acceleration of up to 1.36x and 2.29x over state-of-the-art inference servers vLLM and DeepSpeed, respectively. Although the BMC technique is evaluated extensively across different classes of CPUs (desktop and server class), we also evaluate the scheme with GPUs and demonstrate that it works well for GPUs.
Open AccessStriking the Right Balance between Compute and Copy: Improving LLM Inferencing Under Speculative Decoding
With the skyrocketing costs of GPUs and their virtual instances in the cloud, there is a significant desire to use CPUs for large language model (LLM) inference. KV cache update, often implemented as allocation, copying, and in-place strided update for each generated token, incurs significant overhead. As the sequence length increases, the allocation and copy overheads dominate the performance. Alternate approaches may allocate large KV tensors upfront to enable in-place updates, but these matrices (with zero-padded rows) cause redundant computations. In this work, we propose a new KV cache allocation mechanism called Balancing Memory and Compute (BMC). BMC allocates, once every r iterations, KV tensors with r redundant rows, allowing in-place update without copy overhead for those iterations, but at the expense of a small amount of redundant computation. Second, we make an interesting observation that the extra rows allocated in the KV tensors and the resulting redundant computation can be repurposed for Speculative Decoding (SD) that improves token generation efficiency. Last, BMC represents a spectrum of design points with different values of r. To identify the best-performing design point(s), we derive a simple analytical model for BMC. The proposed BMC method achieves an average throughput acceleration of up to 3.2x over baseline HuggingFace (without SD). Importantly when we apply BMC with SD, it results in an additional speedup of up to 1.39x, over and above the speedup offered by SD. Further, BMC achieves a throughput acceleration of up to 1.36x and 2.29x over state-of-the-art inference servers vLLM and DeepSpeed, respectively. Although the BMC technique is evaluated extensively across different classes of CPUs (desktop and server class), we also evaluate the scheme with GPUs and demonstrate that it works well for GPUs.
- Open Access
DuetServe: Harmonizing Prefill and Decode for LLM Serving via Adaptive GPU Multiplexing
Modern LLM serving systems must sustain high throughput while meeting strict latency SLOs across two distinct inference phases: compute-intensive prefill and memory-bound decode phases. Existing approaches either (1) aggregate both phases on shared GPUs, leading to interference between prefill and decode phases, which degrades time-between-tokens (TBT); or (2) disaggregate the two phases across GPUs, improving latency but wasting resources through duplicated models and KV cache transfers. We present DuetServe, a unified LLM serving framework that achieves disaggregation-level isolation within a single GPU. DuetServe operates in aggregated mode by default and dynamically activates SM-level GPU spatial multiplexing when TBT degradation is predicted. Its key idea is to decouple prefill and decode execution only when needed through fine-grained, adaptive SM partitioning that provides phase isolation only when contention threatens latency service level objectives (SLOs). DuetServe integrates (1) an attention-aware roofline model to forecast iteration latency, (2) a partitioning optimizer that selects the optimal SM split to maximize throughput under TBT constraints, and (3) an interruption-free execution engine that eliminates CPU-GPU synchronization overhead. Evaluations show that DuetServe improves total throughput by up to 1.3x while maintaining low generation latency compared to state-of-the-art frameworks.
Open AccessDuetServe: Harmonizing Prefill and Decode for LLM Serving via Adaptive GPU Multiplexing
Modern LLM serving systems must sustain high throughput while meeting strict latency SLOs across two distinct inference phases: compute-intensive prefill and memory-bound decode phases. Existing approaches either (1) aggregate both phases on shared GPUs, leading to interference between prefill and decode phases, which degrades time-between-tokens (TBT); or (2) disaggregate the two phases across GPUs, improving latency but wasting resources through duplicated models and KV cache transfers. We present DuetServe, a unified LLM serving framework that achieves disaggregation-level isolation within a single GPU. DuetServe operates in aggregated mode by default and dynamically activates SM-level GPU spatial multiplexing when TBT degradation is predicted. Its key idea is to decouple prefill and decode execution only when needed through fine-grained, adaptive SM partitioning that provides phase isolation only when contention threatens latency service level objectives (SLOs). DuetServe integrates (1) an attention-aware roofline model to forecast iteration latency, (2) a partitioning optimizer that selects the optimal SM split to maximize throughput under TBT constraints, and (3) an interruption-free execution engine that eliminates CPU-GPU synchronization overhead. Evaluations show that DuetServe improves total throughput by up to 1.3x while maintaining low generation latency compared to state-of-the-art frameworks.
- Open Access
DistilLock: Safeguarding LLMs from Unauthorized Knowledge Distillation on the Edge
Large Language Models (LLMs) have demonstrated strong performance across diverse tasks, but fine-tuning them typically relies on cloud-based, centralized infrastructures. This requires data owners to upload potentially sensitive data to external servers, raising serious privacy concerns. An alternative approach is to fine-tune LLMs directly on edge devices using local data; however, this introduces a new challenge: the model owner must transfer proprietary models to the edge, which risks intellectual property (IP) leakage. To address this dilemma, we propose DistilLock, a TEE-assisted fine-tuning framework that enables privacy-preserving knowledge distillation on the edge. In DistilLock, a proprietary foundation model is executed within a trusted execution environment (TEE) enclave on the data owner's device, acting as a secure black-box teacher. This setup preserves both data privacy and model IP by preventing direct access to model internals. Furthermore, DistilLock employs a model obfuscation mechanism to offload obfuscated weights to untrusted accelerators for efficient knowledge distillation without compromising security. We demonstrate that DistilLock prevents unauthorized knowledge distillation processes and model-stealing attacks while maintaining high computational efficiency, but offering a secure and practical solution for edge-based LLM personalization.
Open AccessDistilLock: Safeguarding LLMs from Unauthorized Knowledge Distillation on the Edge
Large Language Models (LLMs) have demonstrated strong performance across diverse tasks, but fine-tuning them typically relies on cloud-based, centralized infrastructures. This requires data owners to upload potentially sensitive data to external servers, raising serious privacy concerns. An alternative approach is to fine-tune LLMs directly on edge devices using local data; however, this introduces a new challenge: the model owner must transfer proprietary models to the edge, which risks intellectual property (IP) leakage. To address this dilemma, we propose DistilLock, a TEE-assisted fine-tuning framework that enables privacy-preserving knowledge distillation on the edge. In DistilLock, a proprietary foundation model is executed within a trusted execution environment (TEE) enclave on the data owner's device, acting as a secure black-box teacher. This setup preserves both data privacy and model IP by preventing direct access to model internals. Furthermore, DistilLock employs a model obfuscation mechanism to offload obfuscated weights to untrusted accelerators for efficient knowledge distillation without compromising security. We demonstrate that DistilLock prevents unauthorized knowledge distillation processes and model-stealing attacks while maintaining high computational efficiency, but offering a secure and practical solution for edge-based LLM personalization.
- Open Access
- 1 January 2025
- Published by Association for Computational Linguistics (ACL)
- p. 20217-20234
MobiZO: Enabling Efficient LLM Fine-Tuning at the Edge via Inference Engines
Lei Gao, Amir Ziashahabi, Yue Niu, Salman Avestimehr, Murali Annavaram. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.
- 1 January 2025
- Published by Association for Computational Linguistics (ACL)
- p. 20217-20234
Open AccessMobiZO: Enabling Efficient LLM Fine-Tuning at the Edge via Inference Engines
Lei Gao, Amir Ziashahabi, Yue Niu, Salman Avestimehr, Murali Annavaram. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.
- Open Access
- 1 January 2025
- Published by Association for Computational Linguistics (ACL)
- p. 19474-19488
KVPR: Efficient LLM Inference with I/O-Aware KV Cache Partial Recomputation
Chaoyi Jiang, Lei Gao, Hossein Entezari Zarch, Murali Annavaram. Findings of the Association for Computational Linguistics: ACL 2025. 2025.
- 1 January 2025
- Published by Association for Computational Linguistics (ACL)
- p. 19474-19488
Open AccessKVPR: Efficient LLM Inference with I/O-Aware KV Cache Partial Recomputation
Chaoyi Jiang, Lei Gao, Hossein Entezari Zarch, Murali Annavaram. Findings of the Association for Computational Linguistics: ACL 2025. 2025.
- Open Access
Ethos: Rectifying Language Models in Orthogonal Parameter Space
Language models (LMs) have greatly propelled the research on natural language processing. However, LMs also raise concerns regarding the generation of biased or toxic content and the potential disclosure of private information from the training dataset. In this work, we present a new efficient approach, Ethos, that rectifies LMs to mitigate toxicity and bias in outputs and avoid privacy leakage. Ethos is built on task arithmetic. However, unlike current task arithmetic algorithms, Ethos distinguishes general beneficial and undesired knowledge when reconstructing task vectors. Specifically, Ethos first obtains a set of principal components from the pre-trained models using singular value decomposition. Then, by projecting the task vector onto principal components, Ethos identifies the principal components that encode general or undesired knowledge. Ethos performs negating using the task vector with undesired knowledge only, thereby minimizing collateral damage on general model utility. We demonstrate the efficacy of our approach on three different tasks: debiasing, detoxification, and memorization unlearning. Evaluations show Ethos is more effective in removing undesired knowledge and maintaining the overall model performance compared to current task arithmetic methods.
Open AccessEthos: Rectifying Language Models in Orthogonal Parameter Space
Language models (LMs) have greatly propelled the research on natural language processing. However, LMs also raise concerns regarding the generation of biased or toxic content and the potential disclosure of private information from the training dataset. In this work, we present a new efficient approach, Ethos, that rectifies LMs to mitigate toxicity and bias in outputs and avoid privacy leakage. Ethos is built on task arithmetic. However, unlike current task arithmetic algorithms, Ethos distinguishes general beneficial and undesired knowledge when reconstructing task vectors. Specifically, Ethos first obtains a set of principal components from the pre-trained models using singular value decomposition. Then, by projecting the task vector onto principal components, Ethos identifies the principal components that encode general or undesired knowledge. Ethos performs negating using the task vector with undesired knowledge only, thereby minimizing collateral damage on general model utility. We demonstrate the efficacy of our approach on three different tasks: debiasing, detoxification, and memorization unlearning. Evaluations show Ethos is more effective in removing undesired knowledge and maintaining the overall model performance compared to current task arithmetic methods.
- Open Access
- 1 January 2024
- Published by Association for Computational Linguistics (ACL)
- p. 2054-2068
Ethos: Rectifying Language Models in Orthogonal Parameter Space
Lei Gao, Yue Niu, Tingting Tang, Salman Avestimehr, Murali Annavaram. Findings of the Association for Computational Linguistics: NAACL 2024. 2024.
- 1 January 2024
- Published by Association for Computational Linguistics (ACL)
- p. 2054-2068
Open AccessEthos: Rectifying Language Models in Orthogonal Parameter Space
Lei Gao, Yue Niu, Tingting Tang, Salman Avestimehr, Murali Annavaram. Findings of the Association for Computational Linguistics: NAACL 2024. 2024.
- Open Access
Federated Learning for Internet of Things: Applications, Challenges, and Opportunities
Billions of IoT devices will be deployed in the near future, taking advantage of faster Internet speed and the possibility of orders of magnitude more endpoints brought by 5G/6G. With the growth of IoT devices, vast quantities of data that may contain users' private information will be generated. The high communication and storage costs, mixed with privacy concerns, will increasingly challenge the traditional ecosystem of centralized over-the-cloud learning and processing for IoT platforms. Federated Learning (FL) has emerged as the most promising alternative approach to this problem. In FL, training data-driven machine learning models is an act of collaboration between multiple clients without requiring the data to be brought to a central point, hence alleviating communication and storage costs and providing a great degree of user-level privacy. However, there are still some challenges existing in the real FL system implementation on IoT networks. In this paper, we will discuss the opportunities and challenges of FL in IoT platforms, as well as how it can enable diverse IoT applications. In particular, we identify and discuss seven critical challenges of FL in IoT platforms and highlight some recent promising approaches towards addressing them.
Open AccessFederated Learning for Internet of Things: Applications, Challenges, and Opportunities
Billions of IoT devices will be deployed in the near future, taking advantage of faster Internet speed and the possibility of orders of magnitude more endpoints brought by 5G/6G. With the growth of IoT devices, vast quantities of data that may contain users' private information will be generated. The high communication and storage costs, mixed with privacy concerns, will increasingly challenge the traditional ecosystem of centralized over-the-cloud learning and processing for IoT platforms. Federated Learning (FL) has emerged as the most promising alternative approach to this problem. In FL, training data-driven machine learning models is an act of collaboration between multiple clients without requiring the data to be brought to a central point, hence alleviating communication and storage costs and providing a great degree of user-level privacy. However, there are still some challenges existing in the real FL system implementation on IoT networks. In this paper, we will discuss the opportunities and challenges of FL in IoT platforms, as well as how it can enable diverse IoT applications. In particular, we identify and discuss seven critical challenges of FL in IoT platforms and highlight some recent promising approaches towards addressing them.
- Open Access
Federated Learning for Internet of Things: A Federated Learning Framework for On-device Anomaly Data Detection
Federated learning can be a promising solution for enabling IoT cybersecurity (i.e., anomaly detection in the IoT environment) while preserving data privacy and mitigating the high communication/storage overhead (e.g., high-frequency data from time-series sensors) of centralized over-the-cloud approaches. In this paper, to further push forward this direction with a comprehensive study in both algorithm and system design, we build FedIoT platform that contains FedDetect algorithm for on-device anomaly data detection and a system design for realistic evaluation of federated learning on IoT devices. Furthermore, the proposed FedDetect learning framework improves the performance by utilizing a local adaptive optimizer (e.g., Adam) and a cross-round learning rate scheduler. In a network of realistic IoT devices (Raspberry PI), we evaluate FedIoT platform and FedDetect algorithm in both model and system performance. Our results demonstrate the efficacy of federated learning in detecting a wider range of attack types occurred at multiple devices. The system efficiency analysis indicates that both end-to-end training time and memory cost are affordable and promising for resource-constrained IoT devices. The source code is publicly available at https://github.com/FedML-AI/FedIoT.
Open AccessFederated Learning for Internet of Things: A Federated Learning Framework for On-device Anomaly Data Detection
Federated learning can be a promising solution for enabling IoT cybersecurity (i.e., anomaly detection in the IoT environment) while preserving data privacy and mitigating the high communication/storage overhead (e.g., high-frequency data from time-series sensors) of centralized over-the-cloud approaches. In this paper, to further push forward this direction with a comprehensive study in both algorithm and system design, we build FedIoT platform that contains FedDetect algorithm for on-device anomaly data detection and a system design for realistic evaluation of federated learning on IoT devices. Furthermore, the proposed FedDetect learning framework improves the performance by utilizing a local adaptive optimizer (e.g., Adam) and a cross-round learning rate scheduler. In a network of realistic IoT devices (Raspberry PI), we evaluate FedIoT platform and FedDetect algorithm in both model and system performance. Our results demonstrate the efficacy of federated learning in detecting a wider range of attack types occurred at multiple devices. The system efficiency analysis indicates that both end-to-end training time and memory cost are affordable and promising for resource-constrained IoT devices. The source code is publicly available at https://github.com/FedML-AI/FedIoT.