Pith. sign in

REVIEW 5 major objections 4 minor 48 references

ViFusion: In-Network Tensor Fusion for Scalable Video Feature Indexing

T0 review · 5 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read ViFusion claims an 8–22x throughput gain for distributed video feature indexing by fusing small tensors and offloading AllReduce to in-network servers.

desk verdict Batching and hierarchical aggregation ideas are sound, but the 8-22x claim is a microbenchmark artifact against a self-defined baseline. read the letter →

arxiv 2506.16258 v1 pith:C2MRLLGI submitted 2025-06-19 cs.MM

Yisu Wang , Yixiang Zhu , Xinjiao Li , Yulong Zhang , Ruilong Wu , Dirk Kutscher This is my paper
classification cs.MM
keywords videofeatureindexingin-networkcomputationtensorfusionAllReducedistributedinferenceretrievaldatacenternetworks
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

ViFusion targets a specific bottleneck in distributed video feature indexing: small feature tensors and the AllReduce communication needed to aggregate them across racks. The paper proposes a two-part fix: a tensor fusion module that batches many small tensors into larger ones, and an in-network AllReduce that offloads reduction work from GPUs to dedicated aggregation servers placed on the network path. On an 8-GPU server, the system reports AllReduce bandwidth 37–43x higher than Gloo, and 8–22x higher than a hierarchical fusion baseline at fine granularity, while cutting median query latency by 13% and 99th-percentile latency by 23%. A sympathetic reader would care because the design addresses the exact point where current video retrieval clusters saturate: the communication between GPUs and across racks.

What carries the argument

The central objects are the Tensor Fusion Module (TFM) and the In-Network AllReduce hierarchy. TFM dynamically batches small feature tensors by triggering on a 75% buffer fill or a 150 ms deadline, concatenates them through zero-copy GPU-to-GPU writes, and pads sequences to power-of-two sizes to exploit hardware-aligned kernels. The In-Network AllReduce is a two-tier reduction: GPUs inside the same rack aggregate locally first, then group leaders on dedicated CPU aggregation servers perform a cross-rack reduction, so inter-rack traffic carries only one message per group instead of one per GPU. Together these mechanisms cut the number of communication rounds and the bytes crossing congested uplinks.

What would settle it

Run ViFusion on a cluster spanning at least two racks with deliberate inter-rack congestion, measure end-to-end AllReduce time and aggregation-server CPU utilization, and compare against a host-based hierarchical baseline; if the throughput advantage shrinks to near zero as the aggregation servers saturate, the in-network mechanism is shifting the bottleneck rather than removing it.

Watch

Extended reading notes

Core claim

ViFusion's central claim is that the combination of dynamic tensor fusion and hierarchical in-network AllReduce removes the two dominant overheads in distributed video feature indexing: per-query kernel launches and cross-rack communication. The Tensor Fusion Module (TFM) groups incoming video segments into fused tensors using a dual-threshold trigger (75% GPU buffer capacity or a 150 ms deadline), concatenates them with zero-copy GPU-to-GPU writes, and pads each sequence to a power-of-two size for hardware-friendly kernels. The In-Network AllReduce partitions GPUs into proximity groups, performs a first-stage reduction inside each rack, then a second-stage reduction among group leaders on dedicated CPU aggregation servers, so only one message per group traverses inter-rack links. In experiments with four and eight GPU workers, ViFusion reports 37–43x higher AllReduce bandwidth than Gloo, 8–22x higher than a fine-grained hierarchical baseline (HFBA High), and 5–10x higher than a coarse-grained one (HFBA Low), with lower latency across the distribution than a synchronous non-fused baseline.

Load-bearing premise

The paper assumes that dedicated CPU aggregation servers placed on the network path can take over the reduction work without becoming the next bottleneck, so the hierarchical AllReduce is genuinely faster than host-based aggregation at the multi-rack scale it targets.

Editorial extensions

If this is right

  • If the throughput numbers hold, a video retrieval cluster could serve 8–22x more queries per second without adding GPUs, since fused tensors amortize kernel launches and in-network reduction shortens the communication critical path.
  • The latency CDF shift implies fusion-aware scheduling improves not just throughput but also tail latency, keeping query completion times stable under bursty arrivals.
  • Because the reduction hierarchy mirrors the physical topology, the inter-rack data volume stays proportional to the number of racks rather than the number of GPUs, which is the property that would let the system scale to multi-rack datacenters.
  • The design plugs into existing stacks—PyTorch for inference and a vector database like Milvus for retrieval—so the gain could be realized without changing the embedding model or the retrieval algorithm.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The same fusion-plus-in-network-reduction pattern could generalize to other embedding-heavy retrieval workloads—image, audio, or mixed-modality search—where many small feature tensors dominate communication.
  • A natural extension would be to let aggregation servers do more than summation during reduction, such as partial top-K selection or deduplication, which would shrink the data that must return to end hosts.
  • The fixed 75%/150ms dual-threshold policy is an obvious place for adaptation; a policy that learns arrival statistics per modality could push the throughput/latency tradeoff further under non-Poisson traffic.
  • The paper's evaluation stays on a single 8-GPU server plus CPU-only clusters, so the strongest confirmation of the in-network claim would be a multi-rack deployment with generated congestion—exactly the scenario the design is meant to solve.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 4 minor

Summary. The paper proposes ViFusion, a tensor fusion and hierarchical AllReduce framework for distributed video feature indexing. The design combines a Tensor Fusion Module with dynamic batching, zero-copy concatenation, power-of-two padding, and a dual-threshold trigger, together with an in-network AllReduce scheme implemented on dedicated aggregation servers. The evaluation consists of a CLIP/MSRVTT retrieval prototype with reported hit ratios, an AllReduce bandwidth microbenchmark comparing ViFusion with Gloo and two self-defined HFBA variants, and a query-latency CDF against a synchronous baseline. The abstract and conclusion claim an 8-22x throughput improvement for the video retrieval system, but the experiments measure AllReduce bandwidth rather than end-to-end retrieval throughput, and the in-network mechanism is not tested at the scale it targets.

Significance. If substantiated, an order-of-magnitude throughput gain for distributed video feature indexing without a latency penalty would be a valuable contribution. The paper addresses a genuine bottleneck in distributed multimedia systems, and the design is described clearly enough to be reproduced. Credit is due for explicitly acknowledging in Section 4.3 that no comparative retrieval evaluation was conducted, and for describing the fusion scheduler in sufficient detail. However, the central quantitative claim is not supported by the reported measurements: the 8-22x figure comes from an AllReduce microbenchmark with a self-defined baseline, and the in-network component is not evaluated under the multi-rack or congested conditions that motivate it. As presented, the contribution is a plausible system proposal whose headline performance remains unvalidated.

major comments (5)
  1. [Abstract, §4.3, §4.4] The headline claim that ViFusion 'improves the throughput of the video retrieval system by 8-22 times' is not supported by the experiments. The only quantitative evidence for this factor is Figure 7, which reports AllReduce bandwidth for 4 and 8 GPU workers, not video retrieval throughput. Section 4.3 explicitly states that 'we have not yet conducted comparative analyses against other established retrieval methods,' so there is no end-to-end throughput comparison, no queries-per-second measurement, and no retrieval latency under load. The 8-22x figure therefore cannot be attributed to the video retrieval system, and the abstract's 'deployment results' wording is not justified by the experiments presented.
  2. [§4.4, Figure 7] The speedup ratios are confounded by the baseline definition. HFBA (High) and HFBA (Low) are described as performing AllReduce on 1 KB and 4 KB tensors, while ViFusion fuses tensors into buffers up to 32 MB; the reported 8-22x and 5-10x gains are therefore largely an artifact of the number and size of reduction operations, not of in-network aggregation. No justification or sensitivity analysis is given for the 1 KB/4 KB choices, so the comparison does not establish an independent advantage over a properly tuned baseline. A fair comparison would fix the data size and vary only the aggregation mechanism, or would sweep baseline chunk sizes and report the range of speedups.
  3. [§3.2, Figures 4 and 6; §4.4] The 'in-network' computation is implemented on dedicated CPU aggregation servers, not on programmable switches or SmartNICs, and the evaluation never exercises the mechanism at the scale it targets. There are no measurements of aggregation-server CPU or network saturation, no multi-rack test, and no induced-congestion experiment; the scalability argument in Section 3.2 and in the conclusion is therefore untested. The term 'in-network' is used for server-side reduction, which is not the same as the in-network computation in the switch-based systems cited in Section 2.2, and this discrepancy should be addressed explicitly.
  4. [§4.2, §4.4] The microbenchmark setup is under-specified. The paper states that experiments use an 8-GPU server with 40 Gbps links, but it does not state whether the 4 and 8 'GPU workers' in Figure 7 are co-located on that server or spread across nodes, and no topology, switch model, or congestion information is given. Without this information, the results cannot be interpreted as evidence about cross-rack or inter-datacenter communication, which is the stated motivation for the hierarchical AllReduce design.
  5. [§4.5, Figure 8] The latency CDF compares ViFusion against an unnamed 'Sync' baseline with no description of its implementation, request rate, or query count. The reported median 13% and 99th-percentile 23% reductions in query latency are not connected to the 8-22x throughput claim, and the two claims are never reconciled in a single end-to-end experiment. The 'same level of latency as state-of-the-art systems' assertion in the abstract is likewise unsupported because no state-of-the-art retrieval system is used as a latency baseline.
minor comments (4)
  1. [Figure 7] The bandwidth plots use a logarithmic y-axis but report no error bars, repetition counts, or variance; the plotted values cannot be assessed for statistical significance.
  2. [Section 4.4] The text states that the 4-GPU experiments show 'approximately 37x to 43x higher throughput than Gloo' without specifying the data-size range over which this factor is computed; please state whether this is a peak or average over the swept sizes.
  3. [Figure 5] The caption says 'A typical inter-datacenter network,' while the surrounding text discusses inter-rack communication; the terminology should be harmonized to avoid confusion about the intended deployment scenario.
  4. [Figure 2] There is a typo in the figure ('Freature' for 'Feature'), and the labels 'Summery' and 'TSS/LLM' are unclear; the figure should be cleaned up for readability.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation found: ViFusion's claims rest on external references and direct measurements, not on self-citation or parameters that reduce to the claimed result.

full rationale

ViFusion is an engineering-systems paper whose central claims are empirical performance improvements, not results derived from a model that presupposes those improvements. The paper contains no self-citations: every cited system (SwitchML, ATP, Gloo, NCCL, Milvus, etc.) is external to the authors, so the self-citation and uniqueness-imported-from-authors patterns do not apply. The 8–22x throughput figure is presented as a measured AllReduce bandwidth ratio in Section 4.4 and Figure 7, not as a quantity forced by an equation or by a fitted parameter. The HFBA baselines are locally implemented, and their choice of 1 KB and 4 KB fragments means the reported ratio largely reflects fusion granularity rather than isolating in-network aggregation, but this is a baseline-comparison validity concern, not circularity: ViFusion's measured bandwidth is not defined in terms of HFBA's fragmentation, and the paper does not fit HFBA's parameters to any data to produce the reported numbers. Section 4.3 explicitly states 'we have not yet conducted comparative analyses against other established retrieval methods,' which undermines the abstract's 'deployment results' wording as an overclaiming/evidence issue rather than a circular-derivation issue. The in-network AllReduce design is a hierarchical reduction scheme similar to known multi-level AllReduce, but the paper does not present that resemblance as a derived prediction. No step in the paper's claimed derivation chain reduces by construction to its own inputs, so no significant circularity is present.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central claims depend on four hand-set parameters and four untested domain assumptions. The two most consequential are the server-side aggregation bottleneck assumption and the accuracy-neutrality assumption, because the evaluation does not exercise either.

free parameters (4)
  • Fusion buffer capacity threshold = 75% of GPU buffer
    Hand-chosen in Section 3.1 to trigger tensor fusion. No sensitivity analysis is reported.
  • Fusion deadline = 150 ms
    Hand-chosen in Section 3.1 to bound latency under low arrival rates. No sensitivity analysis is reported.
  • HFBA (High) aggregation chunk size = 1 KB per tensor
    Baseline variable chosen in Section 4.4; small chunks inflate the relative gain of ViFusion.
  • HFBA (Low) aggregation chunk size = 4 KB per tensor
    Baseline variable chosen in Section 4.4; still small, yielding 5-10x relative gains.
assumptions (4)
  • domain assumption Inter-rack links are oversubscribed and dominate AllReduce cost.
    Motivates hierarchical design in Section 3.2 but is never measured in a multi-rack experiment.
  • domain assumption Dedicated aggregation servers in the data path perform reductions without becoming a bottleneck.
    Central to the in-network AllReduce claim in Section 3.2; not stress-tested.
  • ad hoc to paper Fusing queries into batches does not change retrieval accuracy.
    Section 4.3 reports hit ratios with no non-fused baseline comparison.
  • domain assumption An 8-GPU single-node server is representative of multi-rack datacenter inference.
    All experiments in Section 4 use one 8-GPU server or CPU clusters; no multi-rack topology is built.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ViFusion: In-Network Tensor Fusion for Scalable Video Feature Indexing." pith.science (2026). https://pith.science/paper/C2MRLLGI

@misc{pith2026250616258,
  author       = {Pith},
  title        = {Pith review of: ViFusion: In-Network Tensor Fusion for Scalable Video Feature Indexing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/C2MRLLGI}},
  note         = {Machine review of arXiv:2506.16258}
}
read the original abstract

Large-scale video feature indexing in datacenters is critically dependent on efficient data transfer. Although in-network computation has emerged as a compelling strategy for accelerating feature extraction and reducing overhead in distributed multimedia systems, harnessing advanced networking resources at both the switch and host levels remains a formidable challenge. These difficulties are compounded by heterogeneous hardware, diverse application requirements, and complex multipath topologies. Existing methods focus primarily on optimizing inference for large neural network models using specialized collective communication libraries, which often face performance degradation in network congestion scenarios. To overcome these limitations, we present ViFusion, a communication aware tensor fusion framework that streamlines distributed video indexing by merging numerous small feature tensors into consolidated and more manageable units. By integrating an in-network computation module and a dedicated tensor fusion mechanism within datacenter environments, ViFusion substantially improves the efficiency of video feature indexing workflows. The deployment results show that ViFusion improves the throughput of the video retrieval system by 8--22 times with the same level of latency as state-of-the-art systems.

Figures

Figures reproduced from arXiv: 2506.16258 by the authors.

Figure 1
Figure 1. ViFusion Processing Pipeline. Multimodal video features are processed through a combination of tensor fusion, LLM [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. ViFusion Multi-Modal Video Feature Extraction and Retrieval. The process begins with the ingestion of raw video into [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Example of Tensor Fusion Module prompts for contextual guidance, whereas in-network AllReduce efficiently coordinates distributed computations. Together, these components enable ViFusion to produce text, images, and audio outputs at scale, significantly improving indexing efficiency and retrieval speed. 3.1 Tensor Fusion Module Efficient processing of video queries in large-scale clusters requires addressing two key… view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: In-Network AllReduce Execution Flow 3.2 in-network AllReduce In ViFusion’s LLM backbone, tensor-parallel inference is adopted, in which large model layers are distributed across multiple devices. Each device computes a partial result that must be aggregated to form the…
Figure 5
Figure 5. Figure 5: A typical inter-datacenter network [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Example of ViFusion In-Network AllReduce Module [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 7
Figure 7. Figure 7: Overall comparison of AllReduce bandwidth [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]
Figure 8
Figure 8. Figure 8: Comparison of Query Latency Distribution: ViFu [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

48 extracted references · 18 canonical work pages

  1. [1]

    Barefoot Tofino Software Behavior Model

    2024. Barefoot Tofino Software Behavior Model. https://www.barefootnetworks. com/products/brief-p4-studio/

    work page 2024

  2. [2]

    Martín Abadi, Paul Barham, Jianmin Chen, Zhifeng Chen, Andy Davis, et al. 2016. TensorFlow: A system for large-scale machine learning. arXiv:1605.08695 [cs.DC] https://arxiv.org/abs/1605.08695

    arXiv 2016

  3. [3]

    Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lučić, and Cordelia Schmid. 2021. ViViT: A Video Vision Transformer. arXiv:2103.15691 [cs.CV] https://arxiv.org/abs/2103.15691

    arXiv 2021

  4. [4]

    Kirolos Ataallah, Xiaoqian Shen, Eslam Abdelrahman, Essam Sleiman, Deyao Zhu, Jian Ding, and Mohamed Elhoseiny. 2024. MiniGPT4-Video: Advancing Mul- timodal LLMs for Video Understanding with Interleaved Visual-Textual Tokens. arXiv:2404.03413 [cs.CV] https://arxiv.org/abs/2404.03413

    arXiv 2024

  5. [5]

    Kirolos Ataallah, Xiaoqian Shen, Eslam Abdelrahman, Essam Sleiman, Mingchen Zhuge, Jian Ding, Deyao Zhu, Jürgen Schmidhuber, and Mohamed Elhoseiny

    work page

  6. [6]

    Franklin, Joseph E

    Daniel Crankshaw, Xin Wang, Giulio Zhou, Michael J. Franklin, Joseph E. Gon- zalez, and Ion Stoica. 2017. Clipper: a low-latency online prediction serving system. InProceedings of the 14th USENIX Conference on Networked Systems De- sign and Implementation(Boston, MA, USA)(NSDI’17). USENIX Association, USA, 613–627

    work page 2017

  7. [7]

    DeepSeek-AI. 2025. DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning. arXiv:2501.12948 [cs.CL] https://arxiv.org/abs/2501. 12948

    arXiv 2025

  8. [8]

    Gherghescu, Vlad-Andrei Bădoiu, Alexandru Agache, Mihai- Valentin Dumitru, Iuliu Vasilescu, Radu Mantu, and Costin Raiciu

    Alexandru M. Gherghescu, Vlad-Andrei Bădoiu, Alexandru Agache, Mihai- Valentin Dumitru, Iuliu Vasilescu, Radu Mantu, and Costin Raiciu. 2024. I’ve Got 99 Problems But FLOPS Ain’t One. doi:10.1145/3696348.3696893 arXiv:2407.12819 [cs.DC]

    arXiv 2024

Show all 48 references
  1. [9]

    Runsheng Benson Guo, Utkarsh Anand, Arthur Chen, and Khuzaima Daudjee

  2. [10]

    Bo He, Hengduo Li, Young Kyun Jang, Menglin Jia, Xuefei Cao, Ashish Shah, Abhi- nav Shrivastava, and Ser-Nam Lim. 2024. MA-LMM: Memory-Augmented Large Multimodal Model for Long-Term Video Understanding. arXiv:2404.05726 [cs.CV] https://arxiv.org/abs/2404.05726

    2024 arXiv

  3. [11]

    arXiv:2411.01075 [cs.DC] https://arxiv.org/abs/2411.01075

    Cephalo: Harnessing Heterogeneous GPU Clusters for Training Trans- former Models. arXiv:2411.01075 [cs.DC] https://arxiv.org/abs/2411.01075

    arXiv

  4. [12]

    Facebook Incubator. 2023. Gloo: Collective Communication Library. https: //github.com/facebookincubator/gloo

    2023

  5. [13]

    Bo He, Xitong Yang, Le Kang, Zhiyu Cheng, Xin Zhou, and Abhinav Shrivastava

  6. [14]

    Somayeh Kianpisheh and Tarik Taleb. 2023. A Survey on In-Network Computing: Programmable Data Plane and Technology Specific Applications.IEEE Com- munications Surveys & Tutorials25, 1 (2023), 701–761. doi:10.1109/COMST.2022. 3213237

    2023 doi

  7. [15]

    Gonzalez, Hao Zhang, and Ion Stoica

    Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph E. Gonzalez, Hao Zhang, and Ion Stoica. 2023. Efficient Memory Management for Large Language Model Serving with PagedAttention. arXiv:2309.06180 [cs.LG] https://arxiv.org/abs/2309.06180

    2023 arXiv

  8. [16]

    Yimin Jiang, Yibo Zhu, Chang Lan, Bairen Yi, Yong Cui, and Chuanxiong Guo

  9. [17]

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. 2023. Visual Instruc- tion Tuning. arXiv:2304.08485 [cs.CV] https://arxiv.org/abs/2304.08485

    2023 arXiv

  10. [18]

    Yuhan Liu, Hanchen Li, Yihua Cheng, Siddhant Ray, Yuyang Huang, Qizheng Zhang, Kuntai Du, Jiayi Yao, Shan Lu, Ganesh Ananthanarayanan, Michael Maire, Henry Hoffmann, Ari Holtzman, and Junchen Jiang. 2024. CacheGen: KV Cache Compression and Streaming for Fast Large Language Mod...

    2024 arXiv

  11. [19]

    Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. 2021. CLIP4Clip: An Empirical Study of CLIP for End to End Video Clip Retrieval. arXiv:2104.08860 [cs.CV] https://arxiv.org/abs/2104.08860

    2021 arXiv

  12. [20]

    ChonLam Lao, Yanfang Le, Kshiteej Mahajan, Yixi Chen, Wenfei Wu, Aditya Akella, and Michael Swift. 2021. ATP: In-network Aggregation for Multi-tenant Learning. In18th USENIX Symposium on Networked Systems Design and Imple- mentation (NSDI 21). USENIX Association, 741–761. http...

    2021

  13. [21]

    OpenAI. 2024. GPT-4 Technical Report. arXiv:2303.08774 [cs.CL] https://arxiv. org/abs/2303.08774

    2024 arXiv

  14. [22]

    OpenAI. 2024. Sora: Creating video from text. https://openai.com/sora. Accessed on <month> <day>, <year>

    2024

  15. [23]

    Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Des- maison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, L...

    2019

  16. [24]

    NVIDIA. 2024. NVIDIA Collective Communications Library (NCCL). https: //developer.nvidia.com/nccl

    2024

  17. [25]

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. 2021. Learning Transferable Visual Models From Natural Language Supervision. arXiv:2103.000...

    2021 arXiv

  18. [26]

    Samyam Rajbhandari, Conglong Li, Zhewei Yao, Minjia Zhang, Reza Yazdani Aminabadi, Ammar Ahmad Awan, Jeff Rasley, and Yuxiong He. 2022. DeepSpeed- MoE: Advancing Mixture-of-Experts Inference and Training to Power Next- Generation AI Scale. arXiv:2201.05596 [cs.LG] https://arxi...

    2022 arXiv

  19. [27]

    Amedeo Sapio, Marco Canini, Chen-Yu Ho, Jacob Nelson, Panos Kalnis, Changhoon Kim, Arvind Krishnamurthy, Masoud Moshref, Dan Ports, and Peter Richtarik. 2021. Scaling Distributed Machine Learning with In-Network Aggrega- tion. In18th USENIX Symposium on Networked Systems Desig...

    2021

  20. [28]

    William Peebles and Saining Xie. 2023. Scalable Diffusion Models with Trans- formers. InProceedings of the IEEE/CVF International Conference on Computer Vision. 4195–4205

    2023

  21. [29]

    Yang Song and Stefano Ermon. 2020. Generative Modeling by Estimating Gra- dients of the Data Distribution. arXiv:1907.05600 [cs.LG] https://arxiv.org/abs/ 1907.05600

    2020 arXiv

  22. [30]

    Yunzhuo Sun, Yifang Xu, Zien Xie, Yukun Shu, and Sidan Du. 2024. GPTSee: Enhancing Moment Retrieval and Highlight Detection via Description-Based Similarity Features.IEEE Signal Processing Letters31 (2024), 521–525. doi:10. 1109/lsp.2023.3340103

    2024

  23. [31]

    Wadekar, Abhishek Chaurasia, Aman Chadha, and Eugenio Culurciello

    Shakti N. Wadekar, Abhishek Chaurasia, Aman Chadha, and Eugenio Culurciello

  24. [32]

    Alexander Sergeev and Mike Del Balso. 2018. Horovod: fast and easy distributed deep learning in TensorFlow.CoRRabs/1802.05799 (2018). arXiv:1802.05799 http://arxiv.org/abs/1802.05799

    2018 arXiv

  25. [33]

    Jianguo Wang, Xiaomeng Yi, Rentong Guo, Hai Jin, Peng Xu, Shengjun Li, Xi- angyu Wang, Xiangzhou Guo, Chengming Li, Xiaohai Xu, Kun Yu, Yuxing Yuan, Yinghao Zou, Jiquan Long, Yudong Cai, Zhenxiang Li, Zhifeng Zhang, Yihua Mo, Jun Gu, Ruiyi Jiang, Yi Wei, and Charles Xie. 2021....

    2021

  26. [34]

    Jiaxin Wu, Chong-Wah Ngo, and Wing-Kwong Chan. 2024. Improving Inter- pretable Embeddings for Ad-hoc Video Search with Generative Captions and Multi-word Concept Bank. arXiv:2404.06173 [cs.CV] https://arxiv.org/abs/2404. 06173

    2024 arXiv

  27. [35]

    Jun Xu, Tao Mei, Ting Yao, and Yong Rui. 2016. MSR-VTT: A Large Video Description Dataset for Bridging Video and Language. In2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). 5288–5296. doi:10.1109/CVPR. 2016.571

    2016 doi

  28. [36]

    arXiv:2405.17927 [cs.AI] https://arxiv.org/abs/2405.17927

    The Evolution of Multimodal Model Architectures. arXiv:2405.17927 [cs.AI] https://arxiv.org/abs/2405.17927

    arXiv

  29. [37]

    Xinchen Wan, Hong Zhang, Hao Wang, Shuihai Hu, Junxue Zhang, and Kai Chen. 2020. RAT - Resilient Allreduce Tree for Distributed Machine Learning. In Proceedings of the 4th Asia-Pacific Workshop on Networking(Seoul, Republic of Korea)(APNet ’20). Association for Computing Machi...

    2020

  30. [38]

    Zhengyuan Yang, Linjie Li, Kevin Lin, Jianfeng Wang, Chung-Ching Lin, Zicheng Liu, and Lijuan Wang. 2023. The Dawn of LMMs: Preliminary Explorations with GPT-4V(ision). arXiv:2309.17421 [cs.CV] https://arxiv.org/abs/2309.17421

    2023 arXiv

  31. [39]

    Qiang Zhang, Zhipeng Teng, Disheng Wu, and Jiayin Wang. 2024. An Enhanced Batch Query Architecture in Real-time Recommendation. InProceedings of the 33rd ACM International Conference on Information and Knowledge Management (Boise, ID, USA)(CIKM ’24). Association for Computing ...

    2024

  32. [40]

    Bohan Zhao, Wenfei Wu, and Wei Xu. 2023. NetRPC: Enabling In-Network Com- putation in Remote Procedure Calls. In20th USENIX Symposium on Networked Systems Design and Implementation (NSDI 23). USENIX Association, Boston, MA, 199–217. https://www.usenix.org/conference/nsdi23/pre...

    2023

  33. [41]

    Wenquan Xu, Zijian Zhang, Yong Feng, Haoyu Song, Zhikang Chen, Wenfei Wu, Guyue Liu, Yinchao Zhang, Shuxin Liu, Zerui Tian, and Bin Liu. 2023. ClickINC: In-network Computing as a Service in Heterogeneous Programmable Data-center Networks. InProceedings of the ACM SIGCOMM 2023 ...

    2023

  34. [42]

    Mingran Yang, Alex Baban, Valery Kugel, Jeff Libby, Scott Mackie, Swamy Sadashivaiah Renu Kananda, Chang-Hong Wu, and Manya Ghobadi. 2022. Us- ing trio: juniper networks’ programmable chipset - for emerging in-network applications. InProceedings of the ACM SIGCOMM 2022 Confere...

    2022

  35. [43]

    Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. 2023. MiniGPT-4: Enhancing Vision-Language Understanding with Advanced Large Language Models. arXiv:2304.10592 [cs.CV] https://arxiv.org/abs/2304.10592

    2023 arXiv

  36. [46]

    Changgang Zheng, Mingyuan Zang, Xinpeng Hong, Liam Perreault, Riyad Ben- soussane, Shay Vargaftik, Yaniv Ben-Itzhak, and Noa Zilberman. 2024. Planter: Rapid Prototyping of In-Network Machine Learning Inference.SIGCOMM Com- put. Commun. Rev.54, 1 (Aug. 2024), 2–21. doi:10.1145/...

    2024

  37. [47]

    Yinmin Zhong, Shengyu Liu, Junda Chen, Jianbo Hu, Yibo Zhu, Xuanzhe Liu, Xin Jin, and Hao Zhang. 2024. DistServe: Disaggregating Prefill and Decoding for Goodput-optimized Large Language Model Serving. arXiv:2401.09670 [cs.DC] https://arxiv.org/abs/2401.09670

    2024 arXiv

  38. [2020]

    In14th USENIX Symposium on Operating Systems Design and Implementation (OSDI 20)

    A Unified Architecture for Accelerating Distributed DNN Training in Heterogeneous GPU/CPU Clusters. In14th USENIX Symposium on Operating Systems Design and Implementation (OSDI 20). USENIX Association, 463–479. https://www.usenix.org/conference/osdi20/presentation/jiang ViFusi...

    2025

  39. [2022]

    arXiv:2203.15187 [cs.CV] https://arxiv.org/abs/2203

    ASM-Loc: Action-aware Segment Modeling for Weakly-Supervised Tem- poral Action Localization. arXiv:2203.15187 [cs.CV] https://arxiv.org/abs/2203. 15187

    arXiv

  40. [2024]

    arXiv:2407.12679 [cs.CV] https://arxiv.org/abs/2407.12679

    Goldfish: Vision-Language Understanding of Arbitrarily Long Videos. arXiv:2407.12679 [cs.CV] https://arxiv.org/abs/2407.12679

    arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.