arXiv subjects

Yushan Chen

Publications and source records attributed to Yushan Chen.

4 recordsLinked to original sources

arXiv2026

FastTPS: An Optimized Method for LLM Token Phase for AI accelerators

The popularity of large language models (LLMs) escalates an ongoing demand for effective inference. However, due to the sequential processing of tokens during the token phase in decoder-only LLMs inference, the inherent low parallelism leads to reduced throughput and suboptimal utilization of the computing units on artificial intelligence (AI) accelerators, particularly when handling long-sequence inputs that impose significant memory overhead. Recently, many reported methods have been developed as potential solutions, since they emerge with numeric deviation. This paper presents FastTPS, a high performance and low-precision loss method for accelerating the token-phase in LLM inference on general AI accelerators which includes three key components: (1) AI accelerator-enabled reloading-free KV Cache concatenation which decreases memory access overhead as well as enables full fusion of Attention, (2) high-efficiency and high-accuracy 'RoPE' attention based on the tiling optimized FLAT, and (3) highly-fused MLP with fine-grain pipeline scheduling. Our results confirm that FastTPS significantly alleviates memory bottlenecks in the token phase, delivering a 6x speed improvement (compared to none-fusion) on an AMD Ryzen AI 300 series NPU with BF16 precision while sustaining 93% peak memory bandwidth utilization during Phi3-mini-4k-instruct inference.

cs.LG↗

arXiv2024

A Generic Model for Persistent Radio Source around Fast Radio Bursts

The repeated fast radio burst FRB 121102A and FRB 190520B has been reported, along with a spatially coincident, compact, persistent radio emission. In this paper, we present a parameterized one-zone model, with a basic scenario that a relativistic magnetized wind from the pulsar sweeps up the surroundings, e.g. freely expanding supernova ejecta, giving rise to a power-law distribution of electron filled between the forward shock and the termination shock. We show that via appropriate adjustment of the model parameters, we can obtain the synchrotron radio emission properties from the one-zone model bright enough to account for observation, simply and analytically fitting the observed spectra well. Through dynamical evolution of the model, we can also obtain time-varying of relevant properties. This parameterized model does not depend on concrete physical models such as central engine, instead we can constraint physical model via comparison between parameters and observation, indicating the information about the central engine and surroundings. We also discuss the synchrotron self-Compton emission in our scenario in the end, but find no clue on the counterparts at other waveband.

astro-ph.HE↗

arXiv2024

A note on the stabilizer formalism via noncommutative graphs

In this short note we formulate a stabilizer formalism in the language of noncommutative graphs. The classes of noncommutative graphs we consider are obtained via unitary representations of compact groups, and suitably chosen operators on finite-dimensional Hilbert spaces. Furthermore, in this framework, we generalize previous results in this area for determining when such noncommutative graphs have anticliques.

cs.IT↗

arXiv2011

Synthesis of Distributed Control and Communication Schemes from Global LTL Specifications

We introduce a technique for synthesis of control and communication strategies for a team of agents from a global task specification given as a Linear Temporal Logic (LTL) formula over a set of properties that can be satisfied by the agents. We consider a purely discrete scenario, in which the dynamics of each agent is modeled as a finite transition system. The proposed computational framework consists of two main steps. First, we extend results from concurrency theory to check whether the specification is distributable among the agents. Second, we generate individual control and communication strategies by using ideas from LTL model checking. We apply the method to automatically deploy a team of miniature cars in our Robotic Urban-Like Environment.

cs.RO↗