pith. sign in

arxiv: 2606.13501 · v1 · pith:L73V6RULnew · submitted 2026-06-11 · 💻 cs.DC · cs.LG· cs.PF

GF-DiT: Scheduling Parallelism for Diffusion Transformer Serving

Xinwei Qiang , Yifan Hu , Shixuan Sun , Jing Yang , Han Zhao , Chen Chen , Yu Feng , Jingwen Leng
show 1 more author
Minyi Guo
This is my paper

Pith reviewed 2026-06-27 05:33 UTC · model grok-4.3

classification 💻 cs.DC cs.LGcs.PF
keywords diffusion transformerparallel schedulingelastic parallelismGPU servinggroup-free collectivesdynamic adaptationDiT inference
0
0 comments X

The pith

DiT serving systems achieve better performance by dynamically scheduling GPU parallelism for each request instead of fixing it in advance.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper argues that fixed parallel configurations for Diffusion Transformer requests waste GPU resources because different requests and stages need different amounts of parallelism. It shows that making parallelism elastic allows the system to adapt to workload changes on the fly. GF-DiT provides the abstractions needed to reallocate GPUs without high overhead. A sympathetic reader would care because this could make image and video generation services much more efficient and responsive under varying loads.

Core claim

GF-DiT demonstrates that treating GPU parallelism as a schedulable resource, through an asynchronous execution model using trajectory tasks and group-free collectives, enables online adaptation that delivers up to 6.01 times higher throughput and 95 percent lower mean latency than static parallelism approaches.

What carries the argument

Group-free collectives, which allow low-overhead online formation and reconfiguration of arbitrary GPU execution groups for communication.

If this is right

  • Throughput improves by up to 6.01 times compared to fixed-pipeline execution.
  • Mean latency reduces by up to 95 percent.
  • SLO violation rates drop by up to 90 percent.
  • Communication-group setup overhead falls from 778 ms to about 60 microseconds.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • This approach could extend to serving other models with heterogeneous compute patterns, such as large language models with varying generation lengths.
  • Data centers might reduce hardware requirements by using elastic scheduling to handle peak loads without dedicated overprovisioning.
  • Future schedulers could incorporate workload prediction to decide when to trigger parallelism changes.

Load-bearing premise

The workloads for diffusion transformers vary enough in their parallelism requirements across requests and stages that dynamic reconfiguration pays off despite any added complexity.

What would settle it

Running the system on a set of requests that all require the same fixed parallelism degree throughout their execution would produce no measurable improvement over static methods.

Figures

Figures reproduced from arXiv: 2606.13501 by Chen Chen, Han Zhao, Jingwen Leng, Jing Yang, Minyi Guo, Shixuan Sun, Xinwei Qiang, Yifan Hu, Yu Feng.

Figure 1
Figure 1. Figure 1: Elastic parallelism exposes policy tradeoffs. Static SP4 preserves the long request’s latency but makes short re￾quests wait; shrinking it to SP1 admits short requests sooner at the cost of delaying the long request. Lightweight stages and small requests cannot effectively uti￾lize many GPUs, whereas insufficient parallelism prolongs large denoising workloads. As a result, existing systems struggle to simu… view at source ↗
Figure 2
Figure 2. Figure 2: Structure of a diffusion serving request. Encoding produces conditioning embeddings, denoising iteratively ad￾vances the diffusion trajectory, and VAE decoding generates the final output. Trajectory task boundaries provide seman￾tically valid rescheduling points for runtime adaptation. complete execution state that can be safely transferred, re￾sumed, or executed under a different resource configuration. E… view at source ↗
Figure 3
Figure 3. Figure 3: Motivating measurements for elastic DiT serving. (a) Different stages exhibit distinct scaling behavior and resource preferences. (b) The performance benefit of parallelism depends on request shape. (c) Different workload conditions favor different parallelism choices, indicating that no single parallel configuration is universally optimal. degree of parallelism depends on the execution stage, request char… view at source ↗
Figure 4
Figure 4. Figure 4: GF-DiT system overview. Incoming requests are converted into trajectory task graphs, scheduled by the con￾trol plane through a programmable policy interface, and dispatched to workers under dynamic execution layouts. Group-free collectives and layout-aware artifact migration make these dynamic layout choices executable at runtime. graph and request shape are largely known at admission, al￾lowing the runtim… view at source ↗
Figure 5
Figure 5. Figure 5: Signal-state designs for dynamic overlapping groups. (a) Separate slots for every logical group avoid collisions but require combinatorial per-group state. (b) Sharing slots across groups allows consecutive overlapping collectives to collide. (c) GF-DiT assigns double-buffered phase state to each rank edge, so a shared edge flips slots consistently across groups without allocating per-group signal state. a… view at source ↗
Figure 6
Figure 6. Figure 6: End-to-end serving results on H20 and A100. Each row is a platform-model pair, and each column reports one metric across the short and foreground-burst workloads. Legacy is the native vLLM-Omni fixed-pipeline execution path with static parallelism; the other policies are implemented on top of GF-DiT. SLO attainment includes failed requests as violations, while latency statistics are computed over completed… view at source ↗
Figure 8
Figure 8. Figure 8: Runtime overhead relative to the native Legacy path on 4-GPU H20. FCFS-SP4 pins GF-DiT to the same FIFO order and static SP4 layout. 6.5 Group-Free Collective Performance [PITH_FULL_IMAGE:figures/full_fig_p011_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Latency of GFC and NCCL collectives on A100 and H20. We measure BF16 all-to-all (A2A) and all-gather (AG) across different per-rank message sizes. reduces A100 4 MB all-to-all latency from 83.0 𝜇s to 50.4 𝜇s and H20 1 MB all-gather latency from 36.5 𝜇s to 24.1 𝜇s. Thus, GFC avoids serving-path process-group construction while maintaining competitive steady-state performance. 6.6 Scaling with Arrival Rate … view at source ↗
Figure 10
Figure 10. Figure 10: Performance of EDF and SRTF-SP1 under increasing arrival rates on the 4-GPU H20 Wan2.2 short workload. Target load is normalized to the estimated serving capacity. EDF SRTF-SP1 SRTF-SPmax FCFS-SP1 FCFS-SP4 0 20 40 60 80 100 overall SLO attainment (%) Simulator Real H20 [PITH_FULL_IMAGE:figures/full_fig_p013_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Simulator versus real 4-GPU H20 overall SLO attainment for the Wan2.2 foreground-burst workload. The simulator replays the exact request trace and policy logic using measured stage costs. [3] Sohaib Ahmad, Qizheng Yang, Haoliang Wang, Ramesh K. Sitara￾man, and Hui Guan. 2025. DiffServe: Efficiently Serving Text￾to-Image Diffusion Models with Query-Aware Model Scaling. arXiv:2411.15381 [cs.DC] https://arxi… view at source ↗
read the original abstract

Diffusion Transformers (DiTs) have become the dominant architecture for image and video generation, creating growing demand for efficient DiT serving. Existing systems assign each request a fixed parallel configuration throughout its lifetime. However, DiT workloads exhibit substantial heterogeneity across requests, execution stages, and system conditions, making static parallelism inefficient and often leading to poor GPU utilization and degraded service quality. This paper argues that DiT serving should treat GPU parallelism as a first-class schedulable resource. We present GF-DiT, a policy-programmable runtime for elastic DiT serving that dynamically adapts the parallelism of running requests according to workload demands and service objectives. GF-DiT introduces an asynchronous execution abstraction that decomposes requests into independently schedulable trajectory tasks and enables online GPU reallocation. To make elastic parallelism practical, GF-DiT further proposes group-free collectives, a lightweight communication abstraction that supports low-overhead online formation and reconfiguration of arbitrary execution groups. We implement GF-DiT in vLLM-Omni and evaluate it on representative image and video diffusion workloads. Compared with fixed-pipeline execution with static parallelism, GF-DiT improves throughput by up to 6.01$\times$, reduces mean latency by up to 95%, lowers SLO violation rates by up to 90%, and reduces communication-group setup overhead from 778 ms to approximately 60 $\mu$s.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper claims that DiT serving systems suffer from inefficiency due to static parallelism in the face of workload heterogeneity across requests, stages, and conditions; GF-DiT addresses this by treating GPU parallelism as a schedulable resource via an asynchronous execution abstraction that decomposes requests into independently schedulable trajectory tasks, combined with group-free collectives for low-overhead dynamic group formation, yielding up to 6.01× throughput, 95% lower mean latency, 90% fewer SLO violations, and reconfiguration overhead reduced from 778 ms to ~60 μs versus fixed-pipeline static parallelism.

Significance. If the empirical claims hold after verification of experimental details and dependency handling, the work would offer a practical advance in elastic serving for diffusion models, improving GPU utilization for heterogeneous image/video generation workloads and providing a policy-programmable runtime that could influence future systems designs in this domain.

major comments (2)
  1. [Abstract] Abstract: The reported performance numbers (6.01× throughput, 95% latency reduction, 90% SLO improvement, 60 μs overhead) are presented without any description of experimental setup, workloads (image/video diffusion), hardware, baselines, number of trials, or error bars; this absence is load-bearing because the central claim rests entirely on these empirical comparisons rather than derivations.
  2. [Abstract] Abstract (asynchronous execution abstraction): Diffusion denoising forms a strict Markov chain in which the input to step t+1 is the output of step t. The claim that requests can be decomposed into independently schedulable trajectory tasks to enable online GPU reallocation (tensor-parallel degree, pipeline stages) therefore requires an explicit account of how per-request data dependencies are preserved without stalls or extra synchronization at step boundaries; absent this, the 60 μs reconfiguration and 6.01× throughput figures cannot be evaluated.
minor comments (2)
  1. [Abstract] Abstract: The sentence introducing 'group-free collectives' would benefit from a one-sentence parenthetical gloss on the communication primitive before the performance claims.
  2. [Abstract] Abstract: 'vLLM-Omni' is mentioned as the implementation vehicle but receives no further characterization; a brief clause on its relation to the base vLLM system would aid readers.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed and constructive review. We address each major comment below and are prepared to revise the manuscript accordingly.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The reported performance numbers (6.01× throughput, 95% latency reduction, 90% SLO improvement, 60 μs overhead) are presented without any description of experimental setup, workloads (image/video diffusion), hardware, baselines, number of trials, or error bars; this absence is load-bearing because the central claim rests entirely on these empirical comparisons rather than derivations.

    Authors: We agree that the abstract would be strengthened by a concise reference to the experimental context. In the revised version we will append a single sentence to the abstract summarizing the workloads (representative image and video diffusion models), hardware platform, and static-parallelism baselines. Full methodology, including trial counts and error bars, already appears in the Evaluation section; we will ensure the abstract points readers there explicitly. revision: yes

  2. Referee: [Abstract] Abstract (asynchronous execution abstraction): Diffusion denoising forms a strict Markov chain in which the input to step t+1 is the output of step t. The claim that requests can be decomposed into independently schedulable trajectory tasks to enable online GPU reallocation (tensor-parallel degree, pipeline stages) therefore requires an explicit account of how per-request data dependencies are preserved without stalls or extra synchronization at step boundaries; absent this, the 60 μs reconfiguration and 6.01× throughput figures cannot be evaluated.

    Authors: The manuscript (Section 3) already supplies the required account: the asynchronous execution abstraction decomposes each request into a chain of trajectory tasks whose data dependencies are tracked explicitly by the scheduler; a task for denoising step t+1 is released only after its predecessor completes, with reallocation and group formation occurring at these natural boundaries. Group-free collectives eliminate the 778 ms setup cost by allowing dynamic group formation without global barriers, keeping per-boundary overhead at ~60 μs. Because the dependency mechanism is described in the body, the abstract numbers are supported by the full text; we will nevertheless add a short clarifying clause to the abstract if the referee prefers the abstract to be self-contained. revision: partial

Circularity Check

0 steps flagged

No circularity: empirical systems paper with no derivations or fitted predictions

full rationale

The paper introduces a runtime system (GF-DiT) with asynchronous execution abstraction and group-free collectives, evaluated via throughput/latency measurements on image/video workloads. No equations, parameter fits, uniqueness theorems, or self-citation chains appear in the provided text. All performance claims (6.01× throughput, 95% latency reduction, etc.) are presented as direct empirical outcomes from implementation in vLLM-Omni, not as predictions derived from prior fitted values or self-referential definitions. The central claims rest on workload heterogeneity observations and measured reconfiguration overheads, which are externally falsifiable via benchmarks and do not reduce to the paper's own inputs by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 2 invented entities

The central claim depends on the existence of workload heterogeneity and the practicality of the new abstractions; no free parameters are visible in the abstract, but the new communication abstraction is an invented entity without independent evidence provided.

axioms (1)
  • domain assumption DiT workloads exhibit substantial heterogeneity across requests, execution stages, and system conditions
    Stated directly in the abstract as the motivation for moving away from static parallelism.
invented entities (2)
  • group-free collectives no independent evidence
    purpose: Lightweight communication abstraction supporting low-overhead online formation and reconfiguration of arbitrary execution groups
    Newly proposed to enable elastic parallelism; no independent evidence or prior citation given in abstract.
  • asynchronous execution abstraction no independent evidence
    purpose: Decomposes requests into independently schedulable trajectory tasks enabling online GPU reallocation
    Core new mechanism introduced in the paper; no external validation shown in abstract.

pith-pipeline@v0.9.1-grok · 5806 in / 1318 out tokens · 23164 ms · 2026-06-27T05:33:04.258791+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

52 extracted references · 10 canonical work pages

  1. [1]

    Advanced Micro Devices, Inc. 2026. ROCm Communication Collec- tives Library (RCCL).https://rocm.docs.amd.com/projects/rccl/en/ latest/. Documentation

    2026

  2. [2]

    Gulavani, Alexey Tumanov, and Ramachandran Ramjee

    Amey Agrawal, Nitin Kedia, Ashish Panwar, Jayashree Mohan, Nipun Kwatra, Bhargav S. Gulavani, Alexey Tumanov, and Ramachandran Ramjee. 2024. Taming throughput-latency tradeoff in LLM inference with sarathi-serve. InProceedings of the 18th USENIX Conference on Operating Systems Design and Implementation(Santa Clara, CA, USA) (OSDI’24). USENIX Association, ...

    2024

  3. [3]

    Sitara- man, and Hui Guan

    Sohaib Ahmad, Qizheng Yang, Haoliang Wang, Ramesh K. Sitara- man, and Hui Guan. 2025. DiffServe: Efficiently Serving Text- to-Image Diffusion Models with Query-Aware Model Scaling. arXiv:2411.15381 [cs.DC]https://arxiv.org/abs/2411.15381

    arXiv 2025

  4. [4]

    Jason Ansel, Edward Yang, Horace He, Natalia Gimelshein, Animesh Jain, Michael Voznesensky, Bin Bao, Peter Bell, David Berard, Evgeni Burovski, Geeta Chauhan, Anjali Chourdia, Will Constable, Alban Desmaison, Zachary DeVito, Elias Ellison, Will Feng, Jiong Gong, Michael Gschwind, Brian Hirsh, Sherlock Huang, Kshiteej Kalam- barkar, Laurent Kirsch, Michael...

    work page doi:10.1145/3620665.3640366 2024

  5. [5]

    Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. 2024. Video generation models as world simulators. (2024).https://openai.com/research/video- generation-models-as-world-simulators

    2024

  6. [6]

    Zixian Cai, Zhengyang Liu, Saeed Maleki, Madanlal Musuvathi, Todd Mytkowicz, Jacob Nelson, and Olli Saarikivi. 2021. Synthesizing opti- mal collective algorithms. InProceedings of the 26th ACM SIGPLAN Symposium on Principles and Practice of Parallel Programming (PPoPP ’21). ACM, 62–75. doi:10.1145/3437801.3441620

    work page doi:10.1145/3437801.3441620 2021

  7. [7]

    DeepSeek-AI. 2025. DeepEP: An Efficient Expert-Parallel Communi- cation Library.https://github.com/deepseek-ai/DeepEP. Software library

    2025

  8. [8]

    Jiarui Fang, Jinzhe Pan, Aoyu Li, Xibo Sun, and Jiannan Wang. 2025. PipeFusion: Patch-level Pipeline Parallelism for Diffusion Transform- ers Inference. arXiv:2405.14430 [cs.CV]https://arxiv.org/abs/2405. 14430

    Pith/arXiv arXiv 2025

  9. [9]

    Jiarui Fang, Jinzhe Pan, Xibo Sun, Aoyu Li, and Jiannan Wang. 2024. xDiT: an Inference Engine for Diffusion Transformers (DiTs) with Massive Parallelism. arXiv:2411.01738 [cs.DC]https://arxiv.org/abs/ 2411.01738

    arXiv 2024

  10. [10]

    FastVideo Team. 2025. FastVideo: A Unified Inference and Post- Training Framework for Accelerated Video Generation.https://github. com/hao-ai-lab/FastVideo. Software

    2025

  11. [11]

    Yu Gao, Haoyuan Guo, Tuyen Hoang, Weilin Huang, Lu Jiang, Fangyuan Kong, Huixia Li, Jiashi Li, Liang Li, Xiaojie Li, Xunsong Li, Yifu Li, Shanchuan Lin, Zhijie Lin, Jiawei Liu, Shu Liu, Xiaonan Nie, Zhiwu Qing, Yuxi Ren, Li Sun, Zhi Tian, Rui Wang, Sen Wang, Guo- qiang Wei, Guohong Wu, Jie Wu, Ruiqi Xia, Fei Xiao, Xuefeng Xiao, Jiangqiao Yan, Ceyuan Yang,...

  12. [12]

    arXiv:2506.09113 [cs.CV]https://arxiv.org/abs/2506.09113

    Seedance 1.0: Exploring the Boundaries of Video Generation Models. arXiv:2506.09113 [cs.CV]https://arxiv.org/abs/2506.09113

    Pith/arXiv arXiv

  13. [13]

    Agrim Gupta, Lijun Yu, Kihyuk Sohn, Xiuye Gu, Meera Hahn, Li Fei-Fei, Irfan Essa, Lu Jiang, and José Lezama. 2023. Photorealistic Video Generation with Diffusion Models. arXiv:2312.06662 [cs.CV] https://arxiv.org/abs/2312.06662

    arXiv 2023

  14. [14]

    Yoav HaCohen, Nisan Chiprut, Benny Brazowski, Daniel Shalem, Dudu Moshe, Eitan Richardson, Eran Levin, Guy Shiran, Nir Zabari, Ori Gordon, Poriya Panet, Sapir Weissbuch, Victor Kulikov, Yaki Bitterman, Zeev Melumian, and Ofir Bibi. 2024. LTX-Video: Re- altime Video Latent Diffusion. arXiv:2501.00103 [cs.CV]https: //arxiv.org/abs/2501.00103

    Pith/arXiv arXiv 2024

  15. [15]

    Changho Hwang, Peng Cheng, Roshan Dathathri, Abhinav Jangda, Saeed Maleki, Madan Musuvathi, Olli Saarikivi, Aashaka Shah, Ziyue Yang, Binyang Li, Caio Rocha, Qinghua Zhou, Mahdieh Ghazimirsaeed, Sreevatsa Anantharamu, and Jithin Jose. 2026. MSCCL++: Rethinking GPU Communication Abstractions for AI Inference. InProceedings of the 31st ACM International Con...

    work page doi:10.1145/3779212.3790188 2026

  16. [16]

    Diederik P Kingma and Max Welling. 2022. Auto-Encoding Variational Bayes. arXiv:1312.6114 [stat.ML]https://arxiv.org/abs/1312.6114

    Pith/arXiv arXiv 2022

  17. [17]

    Weijie Kong, Qi Tian, Zijian Zhang, Rox Min, Zuozhuo Dai, Jin Zhou, Jiangfeng Xiong, Xin Li, Bo Wu, Jianwei Zhang, Kathrina Wu, Qin Lin, Junkun Yuan, Yanxin Long, Aladdin Wang, Andong Wang, Changlin Li, Duojun Huang, Fang Yang, Hao Tan, Hongmei Wang, Jacob Song, Ji- awang Bai, Jianbing Wu, Jinbao Xue, Joey Wang, Kai Wang, Mengyang Liu, Pengyu Li, Shuai Li...

    Pith/arXiv arXiv 2025

  18. [18]

    Suyi Li, Lingyun Yang, Xiaoxiao Jiang, Hanfeng Lu, Dakai An, Zhipeng Di, Weiyi Lu, Jiawei Chen, Kan Liu, Yinghao Yu, Tao Lan, Guodong Yang, Lin Qu, Liping Zhang, and Wei Wang. 2025. KATZ: efficient workflow serving for diffusion models with many adapters. InPro- ceedings of the 2025 USENIX Conference on Usenix Annual Technical Conference(Boston, MA, USA)(...

    2025

  19. [19]

    C. L. Liu and James W. Layland. 1973. Scheduling Algorithms for Multiprogramming in a Hard-Real-Time Environment.J. ACM20, 1 (Jan. 1973), 46–61. doi:10.1145/321738.321743

    work page doi:10.1145/321738.321743 1973

  20. [20]

    Haoyu Lu, Guoxing Yang, Nanyi Fei, Yuqi Huo, Zhiwu Lu, Ping Luo, and Mingyu Ding. 2023. VDT: General-purpose Video Diffu- sion Transformers via Mask Modeling. arXiv:2305.13311 [cs.CV] https://arxiv.org/abs/2305.13311

    arXiv 2023

  21. [21]

    Ma, Ang Chen, and Mosharaf Chowdhury

    Runyu Lu, Shiqi He, Wenxuan Tan, Shenggui Li, Ruofan Wu, Jeff J. Ma, Ang Chen, and Mosharaf Chowdhury. 2026. TetriS- erve: Efficient DiT Serving for Heterogeneous Image Generation. arXiv:2510.01565 [cs.LG]https://arxiv.org/abs/2510.01565

    Pith/arXiv arXiv 2026

  22. [22]

    Jiajun Luo, Yicheng Xiao, Jianru Xu, Yangxiu You, Rongwei Lu, Chen Tang, Jingyan Jiang, and Zhi Wang. 2025. Accelerat- ing Parallel Diffusion Model Serving with Residual Compression. arXiv:2507.17511 [cs.CV]https://arxiv.org/abs/2507.17511

    arXiv 2025

  23. [23]

    Michael Luo, Aaron Hao, Zhengxu Yan, Chengkun Cao, and Quang Lu- ong Nhat Nguyen. 2026. DiT-Serve: An Efficient Serving Engine for Dif- fusion Transformers.https://openreview.net/forum?id=NGNRc7rZBg

    2026

  24. [24]

    Xin Ma, Yaohui Wang, Xinyuan Chen, Gengyun Jia, Ziwei Liu, Yuan- Fang Li, Cunjian Chen, and Yu Qiao. 2025. Latte: Latent Diffusion Transformer for Video Generation. arXiv:2401.03048 [cs.CV]https: //arxiv.org/abs/2401.03048

    Pith/arXiv arXiv 2025

  25. [25]

    Ziming Mao, Yihan Zhang, Chihan Cui, Zhen Huang, Kaichao You, Zhongjie Chen, Zhiying Xu, Zhenyu Gu, Scott Shenker, Costin Raiciu, Yang Zhou, and Ion Stoica. 2026. UCCL-EP: Portable Expert-Parallel Communication. arXiv:2512.19849 [cs.DC]https://arxiv.org/abs/2512. 19849

    arXiv 2026

  26. [26]

    NVIDIA Corporation. 2026. NVIDIA Collective Communication Li- brary (NCCL).https://developer.nvidia.com/nccl. Software library

    2026

  27. [27]

    Lichen Pan, Juncheng Liu, Yongquan Fu, Jinhui Yuan, Rongkai Zhang, Pengze Li, and Zhen Xiao. 2025. Comprehensive Deadlock Prevention for GPU Collective Communication. InProceedings of the Twentieth Eu- ropean Conference on Computer Systems (EuroSys ’25). ACM, 541–557. doi:10.1145/3689031.3717466

    work page doi:10.1145/3689031.3717466 2025

  28. [28]

    William Peebles and Saining Xie. 2023. Scalable Diffusion Models with Transformers. arXiv:2212.09748 [cs.CV]https://arxiv.org/abs/ 2212.09748

    Pith/arXiv arXiv 2023

  29. [29]

    Sampson, Shikai Li, Simone Parmeggiani, Steve Fine, Tara Fowler, Vladan Petrovic, and Yum- ing Du

    Adam Polyak, Amit Zohar, Andrew Brown, Andros Tjandra, Animesh Sinha, Ann Lee, Apoorv Vyas, Bowen Shi, Chih-Yao Ma, Ching-Yao Chuang, David Yan, Dhruv Choudhary, Dingkang Wang, Geet Sethi, Guan Pang, Haoyu Ma, Ishan Misra, Ji Hou, Jialiang Wang, Kiran Ja- gadeesh, Kunpeng Li, Luxin Zhang, Mannat Singh, Mary Williamson, Matt Le, Matthew Yu, Mitesh Kumar Si...

    Pith/arXiv arXiv 2025

  30. [30]

    PyTorch Contributors. 2025. PyTorch Symmetric Memory.https: //docs.pytorch.org/docs/stable/symmetric_memory.html. Accessed: 2026-06-08

    2025

  31. [31]

    Linus Schrage. 1968. Letter to the Editor—A Proof of the Optimality of the Shortest Remaining Processing Time Discipline.Oper. Res.16, 3 (June 1968), 687–690. doi:10.1287/opre.16.3.687

    work page doi:10.1287/opre.16.3.687 1968

  32. [32]

    SGLang Diffusion Team. 2025. SGLang Diffusion: Accelerating Video and Image Generation.https://www.lmsys.org/blog/2025-11- 07-sglang-diffusion/. LMSYS Blog

    2025

  33. [33]

    Aashaka Shah, Vijay Chidambaram, Meghan Cowan, Saeed Maleki, Madan Musuvathi, Todd Mytkowicz, Jacob Nelson, Olli Saarikivi, and Rachee Singh. 2023. TACCL: Guiding Collective Algorithm Synthe- sis using Communication Sketches. In20th USENIX Symposium on Networked Systems Design and Implementation (NSDI 23). USENIX As- sociation, Boston, MA, 593–612.https:/...

    2023

  34. [34]

    Min Si, Pavan Balaji, Yongzhou Chen, Ching-Hsiang Chu, Adi Gangidi, Saif Hasan, Subodh Iyengar, Dan Johnson, Bingzhe Liu, Regina Ren, Deep Shah, Ashmitha Jeevaraj Shetty, Greg Steinbrecher, Yulun Wang, Bruce Wu, Xinfeng Xie, Jingyi Yang, Mingran Yang, Kenny Yu, Min- lan Yu, Cen Zhao, Wes Bland, Denis Boyda, Suman Gumudavelli, Prashanth Kannan, Cristian Lu...

    arXiv 2026

  35. [35]

    Desen Sun, Zepeng Zhao, and Yuke Wang. 2025. PATCHEDSERVE: A Patch Management Framework for SLO-Optimized Hybrid Resolution Diffusion Serving. arXiv:2501.09253 [cs.DC]https://arxiv.org/abs/ 2501.09253

    arXiv 2025

  36. [36]

    Desen Sun, Zepeng Zhao, and Yuke Wang. 2026. MixFusion: A Patch- Level Parallel Serving System for Mixed-Resolution Diffusion Models. InProceedings of the 31st ACM SIGPLAN Annual Symposium on Prin- ciples and Practice of Parallel Programming(Sydney, NSW, Australia) (PPoPP ’26). Association for Computing Machinery, New York, NY, USA, 522–536. doi:10.1145/3...

    work page doi:10.1145/3774934.3786420 2026

  37. [37]

    Meituan LongCat Team, Xunliang Cai, Qilong Huang, Zhuoliang Kang, Hongyu Li, Shijun Liang, Liya Ma, Siyu Ren, Xiaoming Wei, Rixu Xie, and Tong Zhang. 2025. LongCat-Video Technical Report. arXiv:2510.22200 [cs.CV]https://arxiv.org/abs/2510.22200

    arXiv 2025

  38. [38]

    Philippe Tillet, H. T. Kung, and David Cox. 2019. Triton: an interme- diate language and compiler for tiled neural network computations. InProceedings of the 3rd ACM SIGPLAN International Workshop on Machine Learning and Programming Languages(Phoenix, AZ, USA) (MAPL 2019). Association for Computing Machinery, New York, NY, USA, 10–19. doi:10.1145/3315508....

    work page doi:10.1145/3315508.3329973 2019

  39. [39]

    Manjunath Gorentla Venkata, Valentine Petrov, Sergey Lebedev, De- vendar Bureddy, Ferrol Aderholdt, Joshua Ladd, Gil Bloch, Mike Dub- man, and Gilad Shainer. 2025. Unified Collective Communication: A Unified Library for CPU, GPU, and DPU Collectives.IEEE Micro45, 2 (March 2025), 26–35. doi:10.1109/MM.2025.3534638

    work page doi:10.1109/mm.2025.3534638 2025

  40. [40]

    Team Wan, Ang Wang, Baole Ai, Bin Wen, Chaojie Mao, Chen-Wei Xie, Di Chen, Feiwu Yu, Haiming Zhao, Jianxiao Yang, Jianyuan Zeng, Jiayu Wang, Jingfeng Zhang, Jingren Zhou, Jinkai Wang, Jixuan Chen, Kai Zhu, Kang Zhao, Keyu Yan, Lianghua Huang, Mengyang Feng, Ningyi Zhang, Pandeng Li, Pingyu Wu, Ruihang Chu, Ruili Feng, Shiwei Zhang, Siyang Sun, Tao Fang, T...

    Pith/arXiv arXiv 2025

  41. [41]

    William Won, Midhilesh Elavazhagan, Sudarshan Srinivasan, Swati Gupta, and Tushar Krishna. 2024. TACOS: Topology-Aware Collec- tive Algorithm Synthesizer for Distributed Machine Learning. InPro- ceedings of the 2024 57th IEEE/ACM International Symposium on Mi- croarchitecture(Austin, TX, USA)(MICRO ’24). IEEE Press, 856–870. doi:10.1109/MICRO61859.2024.00068

    work page doi:10.1109/micro61859.2024.00068 2024

  42. [42]

    Chenfei Wu, Jiahao Li, Jingren Zhou, Junyang Lin, Kaiyuan Gao, Kun Yan, Sheng ming Yin, Shuai Bai, Xiao Xu, Yilei Chen, Yuxi- ang Chen, Zecheng Tang, Zekai Zhang, Zhengyi Wang, An Yang, Bowen Yu, Chen Cheng, Dayiheng Liu, Deqing Li, Hang Zhang, Hao Meng, Hu Wei, Jingyuan Ni, Kai Chen, Kuan Cao, Liang Peng, Lin Qu, Minggang Wu, Peng Wang, Shuting Yu, Tingk...

    Pith/arXiv arXiv 2025

  43. [43]

    Yifei Xia, Fangcheng Fu, Hao Yuan, Hanke Zhang, Xupeng Miao, Yijun Liu, Suhan Ling, Jie Jiang, and Bin Cui. 2025. TridentServe: A Stage- level Serving System for Diffusion Pipelines. arXiv:2510.02838 [cs.DC] https://arxiv.org/abs/2510.02838

    arXiv 2025

  44. [44]

    Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, Da Yin, Yuxuan Zhang, Weihan Wang, Yean Cheng, Bin Xu, Xiaotao Gu, Yuxiao Dong, and Jie Tang. 2025. CogVideoX: Text-to-Video Diffusion Models with An Expert Transformer. arXiv:2408.06072 [cs.CV]https://arxiv.org/abs/2408.06072

    Pith/arXiv arXiv 2025

  45. [45]

    Fanjiang Ye, Zhangke Li, Xinrui Zhong, Ethan Ma, Russell Chen, Kai- jian Wang, Jingwei Zuo, Desen Sun, Ye Cao, Triston Cao, Myungjin Lee, Arvind Krishnamurthy, and Yuke Wang. 2026. GENSERVE: Efficient Co-Serving of Heterogeneous Diffusion Model Workloads. arXiv:2604.04335 [cs.DC]https://arxiv.org/abs/2604.04335

    Pith/arXiv arXiv 2026

  46. [46]

    Peiqi Yin, Jiangyun Zhu, Han Gao, Chenguang Zheng, Yongxiang Huang, Taichang Zhou, Ruirui Yang, Weizhi Liu, Weiqing Chen, Canlin Guo, Didan Deng, Zifeng Mo, Cong Wang, James Cheng, Roger Wang, and Hongsheng Liu. 2026. vLLM-Omni: Fully Disaggregated Serving for Any-to-Any Multimodal Models. arXiv:2602.02204 [cs.DC]https: //arxiv.org/abs/2602.02204

    arXiv 2026

  47. [47]

    Zangwei Zheng, Xiangyu Peng, Yuxuan Lou, Chenhui Shen, Tom Young, Xinying Guo, Binluo Wang, Hang Xu, Hongxin Liu, Mingyan Jiang, Wenjun Li, Yuhui Wang, Anbang Ye, Gang Ren, Qianran Ma, Wanying Liang, Xiang Lian, Xiwen Wu, Yuting Zhong, Zhuangyan Li, Chaoyu Gong, Guojun Lei, Leijun Cheng, Limin Zhang, Ming- hao Li, Ruijie Zhang, Silan Hu, Shijie Huang, Xia...

    Pith/arXiv arXiv 2026

  48. [48]

    Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Sheng- gui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You

  49. [49]

    arXiv:2412.20404 [cs.CV]https://arxiv.org/abs/2412.20404

    Open-Sora: Democratizing Efficient Video Production for All. arXiv:2412.20404 [cs.CV]https://arxiv.org/abs/2412.20404

    Pith/arXiv arXiv

  50. [50]

    Yang Zhou, Zhongjie Chen, Ziming Mao, ChonLam Lao, Shuo Yang, Pravein Govindan Kannan, Jiaqi Gao, Yilong Zhao, Yongji Wu, Kaichao You, Fengyuan Ren, Zhiying Xu, Costin Raiciu, and Ion Stoica

  51. [51]

    arXiv:2504.17307 [cs.NI]https://arxiv.org/abs/2504.17307

    An Extensible Software Transport Layer for GPU Network- ing. arXiv:2504.17307 [cs.NI]https://arxiv.org/abs/2504.17307

    arXiv

  52. [52]

    Yuan Zhou, Qiuyue Wang, Yuxuan Cai, and Huan Yang. 2024. Allegro: Open the Black Box of Commercial-Level Video Generation Model. arXiv:2410.15458 [cs.CV]https://arxiv.org/abs/2410.15458 15

    arXiv 2024