TotallyNotOla

4 ms·

1. 1.▲
   by TotallyNotOla 8mo ago

   ML training workflows store checkpoints as full model files, even when most of the model hasn’t changed.I wanted to see what happens if you treat checkpoints as structured objects instead of opaque blobs, and deduplicate at the tensor level
2. 2.▲
   Show HN: Content Addressable Storage for ML Checkpoints(olamyy.github.io)

   2 points by TotallyNotOla 8mo ago|1 comments
3. 3.▲
   Latency Is a Curve(iop.systems)

   3 points by TotallyNotOla 3y ago|0 comments
4. 4.▲
   by TotallyNotOla 5y ago

   I see it's time to update xkcd 1987. https://xkcd.com/1987/