Packs
A pack is what veizik convert makes from a checkpoint you already hold: the calibrated quality profile, prepared once and reused. The current release deliberately ships without one.
PLANNED
Two ways to prepare a model, and only one of them is finished
On the device, at load
SHIPPINGWhat the experimental release does. The checkpoint streams in, the quantiser runs on the GPU, the layers are built, nothing is written to disk. It costs a little time at every start and it is not calibrated, which is where the quality gap comes from.
Ahead of time, into a pack
PLANNEDWhat convert will do. Calibration runs once against real activations, and the result is a file the engine can map straight in. Better quality, faster start, one artifact to move between machines.
What a pack has to guarantee before it is a format
- Versioned, and refused when mismatched. An unknown version is an error, not an attempt. A misread that still produces plausible numbers is worse than a crash, because it ships as quality loss.
- Bound to the engine that can read it. A pack and a runtime that do not match is an error you see immediately, not a quality problem you find later.
- Self-describing. Everything the runtime needs to read it is in the file, so a pack that outlives the machine that made it still means what it meant.
- No silent route change. If a pack cannot be used on this device, the engine says so and stops. It does not quietly fall back to another path and keep going with settings chosen for the first one.
- Your weights stay yours. A pack is derived from a checkpoint you hold, on your machine; producing one does not send weights anywhere.
Pack entries
Loading…