Stable Diffusion 3.5 Large CUDA OUT_OF_MEMORY on RTX 3090

[danielclough]

When I run
cargo run --example stable-diffusion-3 --release --features=cuda -- --which 3.5-large --prompt "pretty picture"
I am get
Error: DriverError(CUDA_ERROR_OUT_OF_MEMORY, "out of memory")
with Stable Diffusion 3.5 Large and Turbo.

According to this chart from stability.ai they should run on an RTX 3090.

[LaurentMazare]

That seems odd, we made a couple optimizations to memory usage following #2574 and in the end, SD 3.5 large was reported to work well on a GPU with only 20GB of memory. Maybe there are some other processes using the memory?
If not it would be good to run a nsys profile to see when the memory is being used.

[danielclough]

There are no other processes running.

How would you like me to run nsys?

Here's some system info:

cat /etc/os-release
PRETTY_NAME="Ubuntu 24.04.1 LTS"

rustc --version
rustc 1.81.0 (eeb90cda1 2024-09-04)

cargo --version
cargo 1.81.0 (2dbb1af80 2024-08-20)

NVIDIA-SMI 560.35.03              Driver Version: 560.35.03      CUDA Version: 12.6
...

[super-fun-surf]

have you tried it with cudnn in addition to cuda feafture. I found it used less ram when cudnn was enabled.

[danielclough]

have you tried it with cudnn in addition to cuda feafture. I found it used less ram when cudnn was enabled.

I have not.

@LaurentMazare
Should cudnn be required to run it properly?

[LaurentMazare]

cudnn shouldn't be necessary but might indeed help reduce gpu memory usage.
That said, I can running the command you mentioned only results in using ~20GB of memory in my case so my guess is that something else is off there.

[danielclough]

The GPU doesn't actually fill up all the memory.

Any suggestions for how to troubleshoot this would be welcome.

[LaurentMazare]

Not sure how much I would trust the memory usage reported by some external tool (especially here where it seems to only measure memory usage every 10s), it's probably safer to use nsys to get a proper memory profile.