[Feature Request]: want disable --no-half on gfx1010 series.

[AiPs1717]

Is there an existing issue for this?

• [o] I have searched the existing issues and checked the recent builds/commits

What would your feature do ?

I'm using a translator, what I'm saying may not be accurate.

I am a gfx1010 user using the --no-half setting.

If I don't use the --no-half command, only gray screen appears. then, when I use --no-half setting, VRAM usage increases twice, preventing high-resolution paintings from being created and 'slowing down' the time they are created. Even the gfx1010 series is mush slower than the previous generation of gfx803 series. Can you investigate why only the gfx1010 series should use the --no-half command?

Proposed workflow

-using RX5700, ubuntu 22.04 LTS.

1. sudo apt update
2. sudo apt upgrade
3. sudo apt install amdgpu-install
4. sudo amdgpu-install --usecase=rocm --no-dkms
5. sudo rocminfo (for checking that ROCm installed)
6. sudo apt install git python3 python3-venv
7. git clone https://github.com/AUTOMATIC1111/stable-diffusion-webui
8. cd stable-diffusion-webui
9. python3 -m venv venv
10. source venv/bin/activate
11. python -m pip install --upgrade pip wheel
12. HSA_OVERRIDE_GFX_VERSION=10.3.0 TORCH_COMMAND='pip install torch torchvision --extra-index-url https://download.pytorch.org/whl/rocm5.1.1' python launch.py --no-half

Additional information

->using : only launch.py

-> using : --no-half

-> using : --no-half -medvram

[ClashSAN]

https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Command-Line-Arguments-and-Settings

from amd guide in wiki:

It's possible that you don't need "--precision full", dropping "--no-half" however crashes my drivers

What are currently wanting to do is to generate images at fp16. But only certain cards can use fp16: Nividia 3XXX and 2XXX series. It looks like your card only makes images at fp32, since you already tried to run at fp16 and got blank images.

[AiPs1717]

https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Command-Line-Arguments-and-Settings

from amd guide in wiki:

It's possible that you don't need "--precision full", dropping "--no-half" however crashes my drivers

What are currently wanting to do is to generate images at fp16. But only certain cards can use fp16: Nividia 3XXX and 2XXX series. It looks like your card only makes images at fp32, since you already tried to run at fp16 and got blank images.

The previous and next generation graphics cards of the gfx1010 support fp16. Also, graphics cards have long been designed to compute fp16. However, it seems that only the RX5000 series does not support fp16.

[jpentland]

This seems to be a bug in ROCM: https://github.com/RadeonOpenCompute/ROCm/issues/1857