arXiv subjects

Kaihui Gao

Publications and source records attributed to Kaihui Gao.

2 recordsLinked to original sources

arXiv2026

Supercharging Packet-level Network Simulation of Large Model Training via Memoization and Fast-Forwarding

Packet-level discrete-event simulation (PLDES) is a prevalent tool for evaluating detailed performance of large model training. Although PLDES offers high fidelity and generality, its slow performance has plagued networking practitioners. Existing optimization techniques either simplify the network model, resulting in large errors; or execute it in parallel using multiple processors, with an upper bound on speedup. This paper explores an alternative optimization direction that reduces the computational loads of PLDES while maintaining high fidelity. Our key insight is that, in distributed LLM training, packet-level traffic behaviors often exhibit repetitive contention patterns and steady-states where flow rates stabilize, ignoring these redundant discrete events speeds up the simulation considerably and the error is negligible. We realize this idea by proposing Wormhole, a user-transparent PLDES kernel capable of automatically memoization for unsteady-states and skipping for steady-states. Wormhole adopts network partitioning, state memoization and reuse, and rate-based steady-state identification to accurately determine the periods of each flow's steady-state, while maintaining simulation consistency after fast-forwarding. Experiments demonstrate that Wormhole can achieve a 744x speedup over the original ns-3 (510x for MoE workload), with a bounded error of <1%. Applying current multithreading parallel techniques and Wormhole together allows a 1012x speedup, reducing the simulation time for one GPT-13B training under 128 GPUs from 9 hours to 5 minutes.

cs.NI↗

arXiv2025

Communication-Efficient Serving for Video Diffusion Models with Latent Parallelism

Video diffusion models (VDMs) perform attention computation over the 3D spatio-temporal domain. Compared to large language models (LLMs) processing 1D sequences, their memory consumption scales cubically, necessitating parallel serving across multiple GPUs. Traditional parallelism strategies partition the computational graph, requiring frequent high-dimensional activation transfers that create severe communication bottlenecks. To tackle this issue, we exploit the local spatio-temporal dependencies inherent in the diffusion denoising process and propose Latent Parallelism (LP), the first parallelism strategy tailored for VDM serving. \textcolor{black}{LP decomposes the global denoising problem into parallelizable sub-problems by dynamically rotating the partitioning dimensions (temporal, height, and width) within the compact latent space across diffusion timesteps, substantially reducing the communication overhead compared to prevailing parallelism strategies.} To ensure generation quality, we design a patch-aligned overlapping partition strategy that matches partition boundaries with visual patches and a position-aware latent reconstruction mechanism for smooth stitching. Experiments on three benchmarks demonstrate that LP reduces communication overhead by up to 97\% over baseline methods while maintaining comparable generation quality. As a non-intrusive plug-in paradigm, LP can be seamlessly integrated with existing parallelism strategies, enabling efficient and scalable video generation services.

cs.DC↗