Throwing OOM with NVFP4
Hey Joey, we have already established that I am not the bright. Anything obvious to you why I would throw a OOM on the second sampler pass using the Take Plus flow with NVFP4?
Figure I will leave this up here in case anyone else is as big of a moron as I am. Updated Comfy Kitchen to the cublas version and seem to be plowing along fine after a few test runs.
Hope all is well!
Hey man! I've been hit with this a few times myself, which is why I created the "auto" option for upscale.
It's usually the second pass being asked to render a bigger picture than the card can hold.
The quant shrinks the WEIGHTS. Pass 2 does not fail on weights, it fails on activations, and those scale with the OUTPUT size, not the file size. Pass 2 runs on the upscaled grid: x2 is four times the pixels of pass 1, x1.5 is 2.25 times. So going to a smaller quant buys you almost nothing on pass 2 - NVFP4 and int8 fall over at nearly the same place.
What to do:
Set
upscaleon MASTER CONTROLS back toauto. It reads your card and picks the biggest output that fits - x2, then x1.5, then a single pass. If you have forced x2 or x1.5, that is almost certainly the OOM.Look at the line the panel prints in the console when you queue. It looks like this:
[JoyLTX TakeControls] LTX TAKE: 8.0 s -> 193 frames in ONE generation | render 960x544 | two passes x2 -> 1920x1088 (51k pass-2 tokens, of ~53k this 24 GB card fits)
If it says FORCED, or the token count is above what it says fits, that is your answer before you've waited for the render.
3. If you want the bigger output anyway, shorten the take. Length and upscale spend the same budget.
Rough pass-2 cost, in the same tokens the panel prints. A card fits about 2200 per GB: 26k on 12 GB, 35k on 16 GB, 53k on 24 GB, 70k on 32 GB.
render seconds one pass x1.5 x2
960x544 8 13k 29k 51k
960x544 14 22k 50k 88k
960x544 20 31k 71k 124k
1280x704 8 22k 50k 88k
1280x704 14 38k 85k 151k
So on a 24 GB card, 960x544 at 8 s upscales x2 and lands at 1920x1088. Fourteen seconds does not - it wants 88k. Drop to x1.5 there and you get 1440x832 at 50k, comfortably. Past about 20 seconds you want a single pass at the render size, and you're better off raising the render size than upscaling a long take.
The defaults in the shipped canvases (960x544, 8 s, auto) are chosen to fit a 24 GB card with room, so if you have not moved those three, tell me your card and what the console line says and we'll work it out from there.