Cheng Li
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Cheng Li has a total of 1000 co-authors. Here are the top co-authors by number of shared publications:
- A. Brueggemann
University of Muenster
11 shared publications - A. Guskov
Joint Institute for Nuclear Research
11 shared publications - A. Q. Zhang
Institute of High Energy Physics
11 shared publications - A. Sarantsev
Joint Institute for Nuclear Research
11 shared publications - Achim Denig
Johannes Gutenberg University of Mainz
11 shared publications - Ai-Qiang Guo
Institute of Modern Physics
11 shared publications
Latest Publications
- Open Access
- 14 November 2026
- Published by Association for Computing Machinery (ACM)
- p. 1941-1958
DittoCache: From CPU-Orchestrated to GPU-Centric KVCache for Efficient LLM Decoding
- 14 November 2026
- Published by Association for Computing Machinery (ACM)
- p. 1941-1958
Open AccessDittoCache: From CPU-Orchestrated to GPU-Centric KVCache for Efficient LLM Decoding
- Open Access
- 14 November 2026
- Published by Association for Computing Machinery (ACM)
- p. 1-15
ATP: Workload-Adaptive Scheduling for Goodput-Optimized Tensor-Parallel LLM Serving with PD Disaggregation
- 14 November 2026
- Published by Association for Computing Machinery (ACM)
- p. 1-15
Open AccessATP: Workload-Adaptive Scheduling for Goodput-Optimized Tensor-Parallel LLM Serving with PD Disaggregation
- 1 November 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems
- Vol. 45 (11) , 6332-6345
Automated Superscalar Processor Design by Learning Data Dependencies
Automated processor design, which significantly reduces manual efforts and accelerates design cycles, has been a key focus of recent research. Although prior work has achieved progress in designing simple processors, such as single-issue processors, their performance falls short of modern superscalar processors that execute multiple instructions per cycle. Current state-of-the-art methods fail to design superscalar processors due to their inability to handle inter-instruction dependencies effectively, which limits parallelism and results in inefficient execution. This work proposes a novel framework for automated superscalar processor design by introducing a hardware-friendly model: Stateful Binary Speculation Diagram (State-BSD). We observe that processor performance can be significantly improved by reusing the processor’s internal states to learn the data dependency. To address the challenge of limited hardware resources while maintaining high prediction accuracy, the proposed State-BSD includes three key components: 1) a lightweight state selector, trained with the simulated annealing method, which identifies the most reusable instructions and stores them in a small buffer, 2) an efficient state-replacer to dynamically replace unused states in the buffer under limited buffer size, and 3) a highly accurate state speculator which predicts interinstruction data dependencies using the buffered states. This work demonstrates the first automated design of superscalar processors with hardware-level optimizations. The processor designed using the proposed automated design approach achieves a 431.28× performance improvement with only 4.25× more area.
- 1 November 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems
- Vol. 45 (11) , 6332-6345
Automated Superscalar Processor Design by Learning Data Dependencies
Automated processor design, which significantly reduces manual efforts and accelerates design cycles, has been a key focus of recent research. Although prior work has achieved progress in designing simple processors, such as single-issue processors, their performance falls short of modern superscalar processors that execute multiple instructions per cycle. Current state-of-the-art methods fail to design superscalar processors due to their inability to handle inter-instruction dependencies effectively, which limits parallelism and results in inefficient execution. This work proposes a novel framework for automated superscalar processor design by introducing a hardware-friendly model: Stateful Binary Speculation Diagram (State-BSD). We observe that processor performance can be significantly improved by reusing the processor’s internal states to learn the data dependency. To address the challenge of limited hardware resources while maintaining high prediction accuracy, the proposed State-BSD includes three key components: 1) a lightweight state selector, trained with the simulated annealing method, which identifies the most reusable instructions and stores them in a small buffer, 2) an efficient state-replacer to dynamically replace unused states in the buffer under limited buffer size, and 3) a highly accurate state speculator which predicts interinstruction data dependencies using the buffered states. This work demonstrates the first automated design of superscalar processors with hardware-level optimizations. The processor designed using the proposed automated design approach achieves a 431.28× performance improvement with only 4.25× more area.
- Open Access
- 27 September 2026
- Published by Association for Computing Machinery (ACM)
- p. 1092-1102
Accelerating Long-Tail Generation in Synchronous RLHF Training via Adaptive Tensor Parallelism
Reinforcement Learning from Human Feedback (RLHF) has become a key post-training paradigm for improving model quality. However, the synchronous three-stage RLHF pipeline is often bottlenecked by the generation stage, where response-length skew causes the effective batch size to shrink rapidly during decoding, leaving GPUs underutilized while a few long responses remain unfinished. Mainstream frameworks employ a static tensor parallelism (TP) configuration that cannot adapt to changing batch characteristics, leaving substantial performance headroom unexplored. We propose PAT, an adaptive TP method that dynamically reconfigures TP during the generation stage of each RLHF iteration. PAT introduces two key techniques. First, a predictor-guided online reconfiguration method decides both the reconfiguration point and the target TP configuration based on offline profiling, triggering reconfiguration only when the predicted latency benefit outweighs the reconfiguration overhead. Second, a lightweight online reconfiguration mechanism updates only the states and layouts affected by TP changes: it adapts unfinished decoding states through a cost-model-based choice between KV-cache migration and recomputation, performs in-place weight resharding, and reuses cached communication groups. We implement PAT on top of SGLang and integrate it with the VeRL framework. Evaluations on LLaMA3.1-8B and Qwen3-14B using DeepScaleR show that PAT reduces generation latency by up to 34.6% and end-to-end RLHF training iteration latency by up to 27.2% compared to the original VeRL setup.
- 27 September 2026
- Published by Association for Computing Machinery (ACM)
- p. 1092-1102
Open AccessAccelerating Long-Tail Generation in Synchronous RLHF Training via Adaptive Tensor Parallelism
Reinforcement Learning from Human Feedback (RLHF) has become a key post-training paradigm for improving model quality. However, the synchronous three-stage RLHF pipeline is often bottlenecked by the generation stage, where response-length skew causes the effective batch size to shrink rapidly during decoding, leaving GPUs underutilized while a few long responses remain unfinished. Mainstream frameworks employ a static tensor parallelism (TP) configuration that cannot adapt to changing batch characteristics, leaving substantial performance headroom unexplored. We propose PAT, an adaptive TP method that dynamically reconfigures TP during the generation stage of each RLHF iteration. PAT introduces two key techniques. First, a predictor-guided online reconfiguration method decides both the reconfiguration point and the target TP configuration based on offline profiling, triggering reconfiguration only when the predicted latency benefit outweighs the reconfiguration overhead. Second, a lightweight online reconfiguration mechanism updates only the states and layouts affected by TP changes: it adapts unfinished decoding states through a cost-model-based choice between KV-cache migration and recomputation, performs in-place weight resharding, and reuses cached communication groups. We implement PAT on top of SGLang and integrate it with the VeRL framework. Evaluations on LLaMA3.1-8B and Qwen3-14B using DeepScaleR show that PAT reduces generation latency by up to 34.6% and end-to-end RLHF training iteration latency by up to 27.2% compared to the original VeRL setup.
- Open Access
- 3 August 2026
- Published by Springer Nature in Journal of High Energy Physics
- Vol. 2026 (8) , 17
Measurement of the hyperon weak radiative decay Ξ0 → γΣ0 at BESIII
A bstract: The hyperon weak radiative decay Ξ 0 → γ Σ 0 is measured using a sample of Ξ 0 from $$ J/\psi \to {\Xi}^0{\overline{\Xi}}^0 $$ decays. The absolute branching fraction of Ξ 0 → γ Σ 0 is determined to be (3 . 66 ± 0 . 21 stat ± 0 . 10 syst ) × 10 − 3 , based on (10 . 087 ± 0 . 044) × 10 9 J/ψ events collected with the BESIII detector operating at the BEPCII collider. The decay asymmetry parameter is measured, with a complete angular analysis of the cascade decay chain, to be α γ = − 0 . 807 ± 0 . 095 stat ± 0 . 011 syst .- 3 August 2026
- Published by Springer Nature in Journal of High Energy Physics
- Vol. 2026 (8) , 17
Open AccessMeasurement of the hyperon weak radiative decay Ξ0 → γΣ0 at BESIII
A bstract: The hyperon weak radiative decay Ξ 0 → γ Σ 0 is measured using a sample of Ξ 0 from $$ J/\psi \to {\Xi}^0{\overline{\Xi}}^0 $$ decays. The absolute branching fraction of Ξ 0 → γ Σ 0 is determined to be (3 . 66 ± 0 . 21 stat ± 0 . 10 syst ) × 10 − 3 , based on (10 . 087 ± 0 . 044) × 10 9 J/ψ events collected with the BESIII detector operating at the BEPCII collider. The decay asymmetry parameter is measured, with a complete angular analysis of the cascade decay chain, to be α γ = − 0 . 807 ± 0 . 095 stat ± 0 . 011 syst . - Open Access
- 26 July 2026
- Published by Association for Computing Machinery (ACM)
- p. 1-7
IdealSim: Efficient Ideal Scenario Modeling for Processor Upper-Bound Performance Analysis
- 26 July 2026
- Published by Association for Computing Machinery (ACM)
- p. 1-7
Open AccessIdealSim: Efficient Ideal Scenario Modeling for Processor Upper-Bound Performance Analysis
- Open Access
- 26 July 2026
- Published by Association for Computing Machinery (ACM)
- p. 1-7
Crystal-KV: Efficient KV Cache Management for Chain-of-Thought LLMs via Answer-First Principle
- 26 July 2026
- Published by Association for Computing Machinery (ACM)
- p. 1-7
Open AccessCrystal-KV: Efficient KV Cache Management for Chain-of-Thought LLMs via Answer-First Principle
- Open Access
- 24 July 2026
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Storage
ZNSlice: Partitioning Large-Zone Commodity ZNS SSDs for Multi-Tenancy Sharing
NVMe SSDs are becoming increasingly larger and faster, making disk partitioning between multiple tenants a prevalent practice. However, co-located applications with diverse I/O patterns often trigger contention within SSD internals. In this work, we explore the potential of large-zone ZNS SSDs, which expose internal states and enable finer-grained controls, to support high-performance disk partitioning in cloud, data center, and local cluster/server environments. To address this, we first conduct a comprehensive inter-operation interference study, examining the interplay between small and large reads/writes on large-zone ZNS SSDs, alongside the impacts of I/O command selection, OPEN zone allocation, and request size manipulation strategies. Second, guided by our five key observations, we propose ZNSlice, which is the first system tailored to optimize the utilization and isolation of large-zone ZNS SSDs for multi-tenancy. It implements operation-level isolation by integrating a suite of optimizations including efficient OPEN zone allocation and management, differentiated read/write handling, zone write mode selection, I/O request size manipulation, and mitigation of background operations overhead. Evaluations using 9 workloads on 2 commodity large-zone SSDs demonstrate that ZNSlice delivers high performance disk partitions via the block interface, natively supporting conventional file systems and existing workloads without modifications. The results confirm that ZNSlice simultaneously achieves significant improvements in multiple metrics, up to 40.2% higher disk utilization, 98.4% better isolation, and 61.4%/71.8% reductions in average/tail latency compared to the best baseline when TPC-C is shared with other workloads.
- 24 July 2026
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Storage
Open AccessZNSlice: Partitioning Large-Zone Commodity ZNS SSDs for Multi-Tenancy Sharing
NVMe SSDs are becoming increasingly larger and faster, making disk partitioning between multiple tenants a prevalent practice. However, co-located applications with diverse I/O patterns often trigger contention within SSD internals. In this work, we explore the potential of large-zone ZNS SSDs, which expose internal states and enable finer-grained controls, to support high-performance disk partitioning in cloud, data center, and local cluster/server environments. To address this, we first conduct a comprehensive inter-operation interference study, examining the interplay between small and large reads/writes on large-zone ZNS SSDs, alongside the impacts of I/O command selection, OPEN zone allocation, and request size manipulation strategies. Second, guided by our five key observations, we propose ZNSlice, which is the first system tailored to optimize the utilization and isolation of large-zone ZNS SSDs for multi-tenancy. It implements operation-level isolation by integrating a suite of optimizations including efficient OPEN zone allocation and management, differentiated read/write handling, zone write mode selection, I/O request size manipulation, and mitigation of background operations overhead. Evaluations using 9 workloads on 2 commodity large-zone SSDs demonstrate that ZNSlice delivers high performance disk partitions via the block interface, natively supporting conventional file systems and existing workloads without modifications. The results confirm that ZNSlice simultaneously achieves significant improvements in multiple metrics, up to 40.2% higher disk utilization, 98.4% better isolation, and 61.4%/71.8% reductions in average/tail latency compared to the best baseline when TPC-C is shared with other workloads.
- Open Access
- 11 June 2026
- Published by American Physical Society (APS) in Physical Review Letters
- Vol. 136 (23) , 231903
Precise Measurement of the Chromoelectric Dipole Moment of the Charm Quark
The combined symmetry of charge conjugation and parity ( C P ) is tested in the hadronic transition ψ ( 3686 ) → π + π − J / ψ , using a dataset of ( 2712.4 ± 14.3 ) × 10 6 ψ ( 3686 ) events collected by the BESIII detector at the BEPCII collider. The resulting asymmetry observable is A C P = ( 0.6 ± 1.8 ± 0.1 ) × 10 − 4 , which is determined by combining the two channels J / ψ → e + e − and J / ψ → μ + μ − . Additionally, by considering the relationship between the chromoelectric dipole moment (CEDM) and the A C P observable derived from the quantum chromodynamics multipole expansion theory based on the Chen-Kuang model and the Cornell potential model, we obtain the results of charm quark’s CEDM with d c ′ = ( 2.6 ± 7.8 ± 0.4 ± 0.6 ) × 10 − 16 e cm , and d c ′ = ( 3.5 ± 10.5 ± 0.6 ± 0.5 ) × 10 − 16 e cm , respectively. The uncertainties are statistical, systematic, and theoretical, respectively. These results correspond to an upper limit of | d c ′ | < 2.1 × 10 − 15 e cm at 90% confidence level, representing an order of magnitude improvement in sensitivity compared to the previous direct bound using the same decay process.
- 11 June 2026
- Published by American Physical Society (APS) in Physical Review Letters
- Vol. 136 (23) , 231903
Open AccessPrecise Measurement of the Chromoelectric Dipole Moment of the Charm Quark
The combined symmetry of charge conjugation and parity ( C P ) is tested in the hadronic transition ψ ( 3686 ) → π + π − J / ψ , using a dataset of ( 2712.4 ± 14.3 ) × 10 6 ψ ( 3686 ) events collected by the BESIII detector at the BEPCII collider. The resulting asymmetry observable is A C P = ( 0.6 ± 1.8 ± 0.1 ) × 10 − 4 , which is determined by combining the two channels J / ψ → e + e − and J / ψ → μ + μ − . Additionally, by considering the relationship between the chromoelectric dipole moment (CEDM) and the A C P observable derived from the quantum chromodynamics multipole expansion theory based on the Chen-Kuang model and the Cornell potential model, we obtain the results of charm quark’s CEDM with d c ′ = ( 2.6 ± 7.8 ± 0.4 ± 0.6 ) × 10 − 16 e cm , and d c ′ = ( 3.5 ± 10.5 ± 0.6 ± 0.5 ) × 10 − 16 e cm , respectively. The uncertainties are statistical, systematic, and theoretical, respectively. These results correspond to an upper limit of | d c ′ | < 2.1 × 10 − 15 e cm at 90% confidence level, representing an order of magnitude improvement in sensitivity compared to the previous direct bound using the same decay process.