Swapnopam

↗ HN profile [ 129 ms ] full profile
Karma
0
Created
()
Submissions
0
  1. Not approximately the same. Bit-identical. Across different hardware vendors, kernel versions, or driver stacks. Logs and weight checksums don't answer this. Looking for anyone who has a working production answer or has…

  2. Hey, I was wondering whether the tech industries require deterministic ML training or is non-determinism acceptable? As in, I know in case of LLMs, small float differences while training are generally tolerated as the…

  3. Was wondering if float divergence/drift a very real issue which affects industries or is just a minor inconvenience at best.