← sources

A Very Strange GPU Stall Situation

source: reddit (reddit:comfyui)published 2026-10-09 12:03ingested 2026-10-10 02:37

retained text

Hardware Config:

**RTX 5090 Windforce**  
128GB DDR5 (Running at 3600MHz cause AM5 struggles with 4-slot DDR5 speeds)  
Ryzen 7 7800X3D  
Crucial Gen5 SSD

  
Model Config:

Minimax Ref2V int8 Pruned  
**Qwen VL int8**  
CK Attention  
Comfy run on WSL Ubuntu with cu13

  
Mode: Ref2VA  
**5 Reference Images (2K quality/4MP)**  
5s @ 0.9MP

  
Flags:

\--use-ck-attention  
\--disable-smart-memory  
\--disable-dynamic-vram  
**--high-ram**  
\--disable-pinned-memory  
**--fast-disk**

While running the **basic workflow** I have been experiencing this weird phenomena where before the pipeline reaches to the steps, my VRAM gets full, my GPU usage gets full but my **GPU temps stay 35** degrees Celsius (basically idle temps). As expected, no progress happens, and I eventually have to restart the server.

What has helped however is reducing the reference image resolution and adding the reserve VRAM flag (i have to reserve 10gb of VRAM since reserve VRAM isn't a hard bound limit). How does lowering VRAM limits help here?

  
What I am struggling to understand here, is why does my GPU just straight up stall when it clearly has enough RAM to offload into. I get that it can never fit all into the VRAM but with this much cumulative memory, this is a very strange behaviour I tried searching up a little myself and apparently it has to do with the speed at which the offload/load takes place from RAM to GPU to RAM, basically memory swaps, which somehow ends up bricking the GPU in that run and the kernel has to be restarted. 

**Am I the only one struggling with this? Any help/suggestions would be appreciated.** 

document fetched 2026-10-10 00:41 · canonical https://reddit.com/r/comfyui/comments/1x1jkfq/a_very_strange_gpu_stall_situation