Post #1913838
2026-04-06 20:27 UTC
@mdekstrand yea, once you get past "simple serialization of numpy(ish) weights" things turn into a total hash
DIY on top of HDF5 ain't bad at all IMO.
Replies (1)
-
@mdekstrand@hci.social 2026-04-06 20:30
@SnoopJ it's a little frustrating how the safetensors folks seem to think all the world is an LLM. Compression was dismissed b/c model weights are “incompressible by design” and… that is just not true outside of a gradient descent dense matrix. Glad to have your +1 to HDF5 not being a terrible idea.