Best parameteres for AMD users?

[dawid3294443]

I'm using RX6600 and my parameteres are:
--medvram --precision full --no-half --opt-sub-quad-attention --sub-quad-q-chunk-size 512 --sub-quad-kv-chunk-size 512 --sub-quad-chunk-threshold 80

Is there something I can do to improve speed? Maybe add or delete something? Also I noticed that img2img (mostly inpainting) is much slower than txt2img.

[SenaKaito]

Maybe you can try mine, i'm using 5500XT 4GB and I can say this the best settings for my card.
--no-half --always-batch-cond-uncond --opt-sub-quad-attention --lowvram --disable-nan-check
You can change lowvram to medvram.

[Zweieckiger]

thank you it workt on my RX6800XT as well

[slowlearnerLspeedup]

DON'T use --lowvram, it's not proper for 6800XT

[gigi798]

I'm using RX6600 and my parameteres are: --medvram --precision full --no-half --opt-sub-quad-attention --sub-quad-q-chunk-size 512 --sub-quad-kv-chunk-size 512 --sub-quad-chunk-threshold 80

Is there something I can do to improve speed? Maybe add or delete something? Also I noticed that img2img (mostly inpainting) is much slower than txt2img.

how many it/s do you get ?

[dawid3294443]

512x512 20 steps euler a about 1,4 it/s, DPM+SDE Karras about 1,5s/it

[gigi798]

linux or windows? I get 3 it/s on linux with rocm

[dawid3294443]

Windows

[FrakerKill]

512x512 20 steps euler a about 1,4 it/s, DPM+SDE Karras about 1,5s/it

I have the same GPU but I don't take this speed I'm between 4-6s/it with this config:

set COMMANDLINE_ARGS=--opt-split-attention --no-half-vae --disable-nan-check --autolaunch --listen
set SAFETENSORS_FAST_GPU=1

[elen07zz]

Im running it with

export HSA_OVERRIDE_GFX_VERSION=10.3.0
export LD_PRELOAD=libtcmalloc.so.4
export PYTORCH_CUDA_ALLOC_CONF=garbage_collection_threshold:0.6,max_split_size_mb:128
python3 launch.py --opt-split-attention-v1 --upcast-sampling --opt-sub-quad-attention --medvram

[dawid3294443]

how many it/s ?

[elen07zz]

that depends on which model, steps, etc.. resolution..

for example sd 1.5 model with 512x512 cfg 6 and 20 steps. I can get 3.9 it/s ? with a rx 6700

[Sakura-Luna]

Graphics cards with RDNA2 architecture do not need --precision full --no-half, which will improve a lot of speed and VRAM usage.

[qwerkilo]

set HSA_OVERRIDE_GFX_VERSION=10.3.0
set LD_PRELOAD=libtcmalloc.so.4
set PYTORCH_CUDA_ALLOC_CONF=garbage_collection_threshold:0.6,max_split_size_mb:128
set COMMANDLINE_ARGS=--opt-split-attention --medvram --no-half --precision full --no-half-vae --opt-sub-quad-attention --sub-quad-q-chunk-size 512 --sub-quad-kv-chunk-size 512 --sub-quad-chunk-threshold 80 --upcast-sampling --always-batch-cond-uncond --disable-nan-check --use-cpu interrogate gfpgan scunet codeformer --autolaunch

512x768,1lora，25step，cfg7，1.48it/s,2M karras,6800xt

[elen07zz]

its slower on windows.

[Soulreaver90]

You do not need half those arguments for a 6800xt. I get double the speed doing 768x768 with a 6700xt.
Remove --no-half --precision full, keep --no-half-vae. I did a test with the sub-quad arcs and saw my it/s drop. You only need those if you need to run larger sizes, but honestly can just use hires fix to get large resolution images.
For reference, I only use --no-half-vae on my setup.

[qwerkilo]

You do not need half those arguments for a 6800xt. I get double the speed doing 768x768 with a 6700xt.
Remove --no-half --precision full, keep --no-half-vae. I did a test with the sub-quad arcs and saw my it/s drop. You only need those if you need to run larger sizes, but honestly can just use hires fix to get large resolution images.
For reference, I only use --no-half-vae on my setup.

thx，i will try it later

[qwerkilo]

I tested it, and the ARGS I ended up using were :--opt-split-attention --no-half-vae --disable-nan-check --autolaunch, which was actually much faster than before

[Soulreaver90]

That is great to hear! The old settings were better for cards with less vram, which the 6700xt-6900xt do not have a problem with.

[ghost]

Just learned about Stable Diffusion today, and learning how to OPTIMIZE my settings. Making me wish I had gone with Nvidia haha!

7900XT 20GB VRAM

--opt-sub-quad-attention --precision full --no-half --sub-quad-q-chunk-size 512 --sub-quad-kv-chunk-size 512 --sub-quad-chunk-threshold 80 --upcast-sampling --always-batch-cond-uncond --disable-nan-check --use-cpu interrogate gfpgan scunet codeformer --autolaunch --allow-code --skip-torch-cuda-test

[gigi798]

linux or windwos ?

[ghost]

Windows 11
7900XT 20GB

[ericek111]

Guys... Linux is free. Why torture yourself?

[ghost]

I tried using WSL2, and it was a headache.

Are you referring to a pure Linux system?

[Hamguy21]

I'm using RX6600 and my parameteres are: --medvram --precision full --no-half --opt-sub-quad-attention --sub-quad-q-chunk-size 512 --sub-quad-kv-chunk-size 512 --sub-quad-chunk-threshold 80

Is there something I can do to improve speed? Maybe add or delete something? Also I noticed that img2img (mostly inpainting) is much slower than txt2img.

someone achieved 860x faster performance using microsoft olive or 10x increase but you will not be able to use other models, you need to convert them to olive. this is faster than a Nvidia RTX 4080
https://community.amd.com/t5/ai/how-to-running-optimized-automatic1111-stable-diffusion-webui-on/ba-p/625585

• Running on the default PyTorch path, the AMD Radeon RX 7900 XTX delivers 1.87 iterations/second.
• Running on the optimized model with Microsoft Olive, the AMD Radeon RX 7900 XTX delivers 18.59 iterations/second.

[ghost]

I am using other models for my work so it's important that I can achieve different styles.

Any other alternatives we should consider?

[Hamguy21]

as far as i know, you'll have to convert it to olive.

How to convert civitai models to ONNX! AMD GPU's on windows can use tons of SD models!

AMD GPU's are screaming fast at stable diffusion! How to install Automatic1111 on windows with AMD

[RominaMinka]

Hey, I have a problem and don't know how to solve it. I hope someone can help me. I have an RX 6950 XT 16GB. My parameters are:

set PYTHON=
set GIT=
set VENV_DIR=
set COMMANDLINE_ARGS= --use-directml --opt-sub-quad-attention --autolaunch --medvram --no-half
git pull

I can generate images with low resolution, but it stops at 800. It loads up to 90% and then an error occurs:

RuntimeError: Could not allocate tensor with 2066743296 bytes. There is not enough GPU video memory available!

[Infamousfish]

I don't know if you still have this problem but I had the same issue with my 6700 xt I removed --no-half and it fixed it for me

[pinea00]

IF your GPU is higher than RX6600,Zluda is your best choices.VRAM save more 30% on windows.
Even you can train lora on local PC.

[Infamousfish]

I tried it just doesn't want to work it gave a bunch of errors and I couldn't solve it for a couple of days so I just left it

[pinea00]

follow this step
https://www.youtube.com/watch?v=n8RhNoAenvM