A method to enable the GTX1650 to generate in fp16.

[pinyangcong]

I discovered a strange feature/bug. My laptop has a GTX1650 graphics card and runs on Windows 10. According to common sense, when generating images in SD without adding "--no-half" and "--precision full", it will report NaN errors. However, when I added "set http_proxy=http://127.0.0.1:xxxx" and "set https_proxy=http://127.0.0.1:xxxx" at the beginning of the webui.bat file, which are the proxy ports I set, after running SD, I was able to generate images without reporting NaN errors and they were working normally.But NaN errors are still reported when I change the resolution.

I think this is an opportunity for 16xx graphics cards to no longer use "--no-half".

OS:windows-10 22H2
CPU:i5-9300H
GPU:GTX 1650(4GB)
stable-diffusion-webui(22bcc7b)
torch==2.0.0_cu118
cuda11.8
COMMANDLINE_ARGS=--medvram --theme dark --opt-sdp-attention --autolaunch
model=AnythingV5_v5PrtRE.safetensors(7F96A1A9CA9B3A3242A9AE95D19284F0D2DA8D5282B42D2D974398BF7663A252)
The model is Full Model fp16.

[Sakura-Luna]

The 1650 natively supports half-precision, so there is no need for no-half.

[Sakura-Luna]

It is for the GTX10xx series.

[cgessai]

Without the --no-half command line argument included, a 16xx will render steps half as fast as everyone else. Specifically the 1650 but I believe the 1660 as well.

[pinyangcong]

Without the --no-half command line argument included, a 16xx will render steps half as fast as everyone else. Specifically the 1650 but I believe the 1660 as well.

Your perspective is completely wrong.Without the --no-half command line argument included, A 16xx will generate an all-black or all-green image.

[pinyangcong]

It is for the GTX10xx series.

Other installation tutorials on various websites will prompt that 16 series graphics cards require the command line argument "--no-half", but why?
Since the 1650 graphics card has built-in support for half-precision floating-point numbers, why do tutorials prompt us to add the command line argument"--no-half" when running SD?

[Hansynily]

Have you found other workaround on this? Adding --disable-nan-check seems to bypass that error but all I get is a black image unless I added --no-half

Using the same card as well

[cgessai]

Try: --medvram --opt-split-attention --xformers --no-half --listen

Does that double your speed? It does with the pytorch under 2.0 and I'm pretty sure it would with 2.0. Maximum resolution for images like this is either 512 or 576, FWIW. If you try it would really love to know whether it works for you or not.

[pinyangcong]

You may not understand the meaning of my article, and you need to know why the 16xx requires the command line argument "--no-half".

[SharkAccount]

How did generation speed and vram usage using your method compare to running a 1650 using --precision full --no-half?

[pinyangcong]

OK,I think it is not a issue, if the benchmark is open and benchmark_limit is 0, my work is OK.

[GawainProdigy]

I have 1650, and I used all of the above arguments; then I discovered Sd FORGE.

This amazing branch automatically optimizes A111 to match your card. The results are incredible speed increases and reliability.

And, it was designed for increasing older card's performance specifically.

The project's results have been subject to peer review. It also comes preinstalled with a few good extensions AND some features designed to SDXL, but tweaked for older card's.

FORGE is faster stable diffusion without the hassle of messing with tweaks and fixes.

https://github.com/lllyasviel/stable-diffusion-webui-forge

[lj4o7]

My 2080ti needs "--no-half"or"--all-in-fp32" to not generate black image in both webui and forge,while i read about no need to use"no half" from other 2080ti users.
i am confused,how?