Go to the Indico Home Page Oldest event Older event Up to category Printable version
  • Compact style
  • Indico style
  • Indico style - inline minutes
  • Indico style - numbered
  • Indico style - numbered + minutes
  • Indico Weeks View
    • Switch to the management area of this event
    Login

    Birds-of-a-Feather Sessions: SONIC - Inference as a Service @Fast ML 2026

    → US/Pacific
    QI Auditorium

    QI Auditorium
    Description

    Find the ZOOM link in FastML main indico: https://indico.cern.ch/event/1654479/timetable/?view=standard#111-sonic-inference-as-a-servi

      • 16:00 → 16:05
        Intro 5m
        Speakers: Yongbin Feng (Texas Tech University (US)), Yuan-Tang Chou (National Tsing Hua University (TW))
      • 16:05 → 16:15
        IceCube SONIC 10m

        IceCube recently switched from Triton to Ray as a framework for running inference, and it has shown much better performance when testing at scale.

        Speaker: David Schultz (University of Wisconsin-Madison)
        IceCube and SuperSONIC.pdf
      • 16:15 → 16:25
        TensorFlow-to-ONNX Migration of CMS DeepTau Models for Efficient Inference in SONIC 10m
        Speaker: Pritam Palit (Carnegie-Mellon University (US))
        SONIC_BoFSession_PritamPalit.pdf
      • 16:25 → 16:35
        Inference as a Service for HENP in the Genesis Era 10m

        The DOE Genesis Mission and American Science Cloud (AmSC) aim to connect Scientific User Facilities through shared data, AI, and computing services. Building on existing Inference-as-a-Service (IaaS) efforts and experiences, we are extending this model to portable, cross-facility scientific AI.

        Speakers: Meghna Bhattacharya (Fermilab), Nhan Tran (Fermi National Accelerator Lab. (US))
        FastML_D3.pptx.pdf
      • 16:35 → 16:45
        Extending Inference as a Service to Heterogeneous Edge Fleets: Open Problems from a 41-Node Kubernetes Cluster of Phones 10m

        We operate a shared research cluster of 41 Google Pixel Fold phones — Tensor G2 system-on-chip, Mali-G710 GPU, 12 GB RAM, Debian on arm64 — run as one Kubernetes (K3s) cluster behind a single Linux gateway on a private gigabit LAN. Over the past quarter roughly twenty people used it across four course teams and several research efforts: student-run Kubernetes control planes on phones, Karmada federations of up to five phone clusters, an over-the-air A/B image-update system, distributed Llama-3-8B inference over RPC, and a 27-node fault-injection campaign executed as one Indexed Job.

        Speaker: Raymond Duenas
        fastml2026_bof_deck_v3.pdf
      • 16:45 → 16:55
        Possible development of lightweight SuperSONIC 10m
        Speakers: Miles Cochran-Branson (University of Washington (US)), Yuan-Tang Chou (National Tsing Hua University (TW))
        superSONIC_light.pdf
      • 16:55 → 17:15
        General discussion 20m
        Speakers: Yongbin Feng (Texas Tech University (US)), Yuan-Tang Chou (National Tsing Hua University (TW))