Open Raz0rStorm opened 1 week ago
If you're trying to run this model in full (16 bit) precision, you're probably running out of memory. You need a lot more memory than just the amount needed to store the model weights to actually generate.
For reference, I'm using a 4090 24gb, and it was still using more than 32gb of system ram in order to do all the offloading needed to make FLUX run.
I had the same problem when generating on PDXL models with lora. I didn't find a solution, I rolled back the update
If you're trying to run this model in full (16 bit) precision, you're probably running out of memory. You need a lot more memory than just the amount needed to store the model weights to actually generate.
For reference, I'm using a 4090 24gb, and it was still using more than 32gb of system ram in order to do all the offloading needed to make FLUX run.
I switched to using a NF4 model that is 6gb in size and it's still CUDA out of memory; Do the author of Forge have a guide for this?
spec: 16g vram and 32g ram, trying to load a 22gb FLUX model, every forge setting is set to default.