• flere-imsaho@awful.systems
    link
    fedilink
    English
    arrow-up
    4
    ·
    2 days ago

    i would swear that in an earlier version of this message the optimal batch size was estimated to be as large as twenty.

    • BB84
      link
      fedilink
      English
      arrow-up
      1
      ·
      2 days ago

      Yes it was <20. That was my guess given how inefficient their inference is. People were downvoting me so I thought maybe the statement was too assertive (because to be fair we don’t know their model size) and relaxed it to <100. Got me more downvotes 🤷‍♂️.