Gangin Park comments

Results 5 comments of


                                            Gangin Park

dequantization offload accounting (fixes Flux2 OOMs - incl TEs)

Confirmed that it resolves https://github.com/comfyanonymous/ComfyUI/issues/10891#issuecomment-3621595666.

Flux2: torch.OutOfMemoryError: Allocation on device

Another OOM logs here @rattus128. Tested on RTX4090, using official workflow [here](https://docs.comfy.org/tutorials/flux/flux-2-dev) with reference images disabled. ``` got prompt Using pytorch attention in VAE Using pytorch attention in VAE VAE...

Fix OOM by reserving extra VRAM for cached memory from CudaMallocAsync allocator

cc. @rattus128 as related to the recent fixes.

Fix OOM by reserving extra VRAM for cached memory from CudaMallocAsync allocator

> Hey. So 32MB is very small in the context of flux 2. Comfy defines minimum inference VRAM and headrooms on the order of 100s of MB. If this is...

Fix OOM by reserving extra VRAM for cached memory from CudaMallocAsync allocator

Thanks for looking in detail @rattus128. I have checked that work and can confirm that it also resolved the issue. Will close this if it's merged. To leave notes for...