search

comfyanonymous / ComfyUI

The most powerful and modular diffusion model GUI, api and backend with a graph/nodes interface.
https://www.comfy.org/
GNU General Public License v3.0
56.41k stars 5.98k forks source link

--`reserve-vram` is reserving too much VRAM #5117

Open JorgeR81 opened 1 month ago

JorgeR81 commented 1 month ago

Expected Behavior

The amount of reserved VRAM stays constant during generations, and proportional to the amount specified in the cmd.

Actual Behavior

With --reserve-vram 1.2 I usually have 6.5 ~ 7 GB VRAM usage, during the KSampler stage ( I have an 8 GB card ).

Now, after some generations the VRAM usage decreased to 3.4 GB

The generation speed did not decrease significantly, but I never noticed this before.

vram

Steps to Reproduce

I need --reserve-vram to use Flux ControlNets.

I was using --reserve-vram 1.2, with Flux Q4_K_S, some loras, and Flux Depth ControlNets ( Shaker AI and Jasper AI ).

I only used one ControlNet at the time, but I was switching between both of them. I also canceled some generations, just to check some settings faster, because my GPU is slow.

After some generations, I noticed I was at only 3.4 GB usage.

Debug Logs

.

Other

Windows 10 Portable Comfy Install

Total VRAM 8192 MB, total RAM 32727 MB
pytorch version: 2.3.1+cu121
Set vram state to: NORMAL_VRAM
Device: cuda:0 NVIDIA GeForce GTX 1070 : cudaMallocAsync
ltdrdata commented 1 month ago

Please first check whether this phenomenon occurs with custom nodes like gguf, and also whether it occurs when all custom nodes are disabled and only core nodes are used in the configuration.