github shootthesound/Fizgig v2.8.8
Fizgig v2.8.8 — FP8 Base hidden for Krea 2

latest releases: v9.8.5, v9.8.4, v9.8.3...
3 months ago

Fizgig v2.8.8 — FP8 Base hidden for Krea 2

FP8 Base is no longer shown on the Training tab when Krea 2 is selected. Unticking it was
never a useful option there — it was a guaranteed out-of-memory error.

The checkbox asked the trainer for a bf16 base: 25.8 GB of weights on their own, ~28 GB in
total, which no consumer card can hold. Three things made it worse than merely useless:

  • The automatic block-swap planner never saw it. The plan came out identical whether the box
    was ticked or not, so the run received a swap count sized for fp8/INT8/NF4 and then loaded a
    model twice that size.
  • Unticking it silently cancelled INT8, undoing the faster path the planner had just chosen.
  • There's no card where it's the right call. The base is frozen during LoRA training and your
    LoRA trains in bf16 either way, so the precision you'd gain is negligible.

Krea 2's base precision is chosen on the 4-bit Base control — Auto / On / Off, giving
NF4 / INT8 / fp8 as appropriate — and the planner accounts for all of those properly.

Klein is unaffected: it uses different flags and keeps the checkbox. The --no_fp8 option
remains available on the command line for anyone with the hardware to use it.

Also in this line of fixes

This is the third issue in the same pair of controls in two days, all now closed:

  • v2.8.6 — block swap planned for
    a quantisation the user had overridden (the 16 GB out-of-memory report), plus lighter,
    measured swap counts.
  • v2.8.7 — 4-bit Off
    accidentally disabling INT8 on 20 GB+ cards.
  • v2.8.8 — this one.

If you train Krea 2, leave Blocks Swap and 4-bit Base on Auto and Fizgig will pick the
fastest configuration that fits your card.

Don't miss a new Fizgig release

NewReleases is sending notifications on new releases.