Uploaded image for project: 'AI Platform Core Components'
  1. AI Platform Core Components
  2. AIPCC-4643

transition torch build ownership from AE to torch team

    • True
    • Hide

      None

      Show
      None
    • True
    • 50% To Do, 50% In Progress, 0% Done
    • M
    • Hide

      Jan 5, 2026 YELLOW

      • Created tickets for the versioned package dependency for torch and tagged to EPIC AIPCC-6999
      • 9/12 dependency packages tickets are closed. We should be able to build torch once these packages are available.
      • AIPCC-7777: Built PyTorch with the latest dependencies to see if we can nudge upstream torch to move to latest dependency versions. However observing test failures. Failures are seen compared to pinned version. Issues are logged in AIPCC-8018, AIPCC-7994, AIPCC-8017
      Show
      Jan 5, 2026 YELLOW Created tickets for the versioned package dependency for torch and tagged to EPIC AIPCC-6999 .  9/12 dependency packages tickets are closed. We should be able to build torch once these packages are available. AIPCC-7777 : Built PyTorch with the latest dependencies to see if we can nudge upstream torch to move to latest dependency versions. However observing test failures. Failures are seen compared to pinned version. Issues are logged in AIPCC-8018 , AIPCC-7994, AIPCC-8017 . 

      Feature Overview:

      We are moving the responsibility for building “torch wheels” from the Accelerator Enablement team to the PyTorch team to remove some friction in that build process and speed up delivery.

      Product(s) associated:

      All uses of torch.

      Goals:

      • Place responsibility for building the set of wheels from the torch community on the team with the most expertise.
      • Enable that team to use the best tool for the job.
      • Continue to build for all accelerator variants supported in the product.

      Requirements:

      1. Builds must be compatible with other wheels built by AE and DP using the wheel builder image. (Build for RHEL, link against library versions in the base images, etc.)
      2. We establish a process for the PyTorch team to recognize updated dependencies and make the AE and DP teams aware so they can perform updates.

      Done - Acceptance Criteria:

      1. Builds for tagged releases of torch are published to a package index visible to the rest of the wheel build pipelines for use in image builds.
      2. Nightly build jobs try to build using the same tools and the HEAD of the most recent development branch of torch, with monitored notifications for failures.

      Use Cases - i.e. User Experience & Workflow:
      Include use case diagrams, main success scenarios, alternative flow scenarios.

      Out of Scope:

      • We will not move all accelerator-enabled builds to the torch team.
      • We will work on making the upstream build tools provide hermetic over time.

      Documentation Considerations :

      N/A

              sdharane Sudhir Dharanendraiah
              dhellman@redhat.com Doug Hellmann
              Andre Lustosa Cabral de Paula Motta, Emilien Macchi, Joseph Groenenboom, Rohan Devasthale
              Votes:
              0 Vote for this issue
              Watchers:
              13 Start watching this issue

                Created:
                Updated: