[Issue]: Rocm Windows 11 Times out GPU Driver and returns black image

[NiklasRichter2222]

Issue Description

As soon as I click generate it starts and runs through the generation progress bar normaly, but: The preview is just grey with sometimes just some colord little artifacts in it. When generation has finished the saved image is just black. Also about 30% of the time my gpu driver runs into driver timeout. This happens 100% of the time.

The issue happens on a new installation of sd.next. It happens consistently even after reinstalling. It happens with multible models(dreamshaper_8, sd_xl_base_1.0, juggernautXL_juggXIByRundiffusion).

Version Platform Description

Using VENV: E:\AI\sdnext\venv
13:43:07-163054 INFO Starting SD.Next
13:43:07-171875 INFO Logger: file="E:\AI\sdnext\sdnext.log" level=DEBUG host="Niklas-PC" size=77 mode=create
13:43:07-176877 INFO Python: version=3.12.10 platform=Windows bin="E:\AI\sdnext\venv\Scripts\python.exe"
venv="E:\AI\sdnext\venv"
13:43:07-835736 INFO Version: app=sd.next updated=2026-01-28 commit=58351b1f5 branch=dev url=https://github.com/vladmandic/sdnext/tree/dev
kanvas=main ui=dev
13:43:08-777076 TRACE Repository branches: active=dev available=['dev', 'master', 'upstream']
13:43:09-069686 INFO Version: app=sd.next latest=2026-01-28T18:28:49Z hash=58351b1f branch=dev
13:43:09-115631 INFO Platform: arch=AMD64 cpu=Intel64 Family 6 Model 165 Stepping 5, GenuineIntel
system=Windows release=Windows-11-10.0.26200-SP0 python=3.12.10 locale=('de_DE',
'cp1252') docker=False
13:43:09-120649 DEBUG Packages: prefix=venv site=['venv', 'venv\Lib\site-packages']
13:43:09-124653 INFO Args: ['--use-rocm', '--port', '7862']
13:43:09-125651 DEBUG Setting environment tuning
13:43:09-128160 DEBUG Torch allocator: "garbage_collection_threshold:0.80,max_split_size_mb:512"
13:43:09-129666 INFO Verifying torch installation
13:43:09-129666 DEBUG Torch overrides: cuda=False rocm=True ipex=False directml=False openvino=False
zluda=False
13:43:10-873358 INFO ROCm: agents=['gfx1100']
13:43:10-875357 INFO ROCm: version=7.2, using agent gfx1100
13:43:10-876862 INFO Python: version=3.12.10 platform=Windows bin="E:\AI\sdnext\venv\Scripts\python.exe" venv="E:\AI\sdnext\venv"
13:43:44-127091 INFO Torch backend: version="2.9.1+rocm7.10.0a20251120" type=ROCm HIP=7.2.0-c0adf921a0
13:43:44-130603 INFO Torch detected: gpu="AMD Radeon RX 7900 XT" vram=20464 arch=(11, 0) cores=42
13:43:44-908732 INFO Install: verifying requirements
13:43:45-045868 DEBUG Timestamp repository update time: Wed Jan 28 19:28:49 2026
13:43:45-048376 INFO Startup: standard
13:43:45-049880 INFO Verifying submodules
13:43:51-438324 DEBUG Git submodule: extensions-builtin/sd-extension-chainner / main
13:43:51-706830 DEBUG Git submodule: extensions-builtin/sd-extension-system-info / main
13:43:51-966090 DEBUG Git submodule: extensions-builtin/sdnext-kanvas / main
13:43:52-223835 DEBUG Git submodule: extensions-builtin/sdnext-modernui / dev
13:43:52-496016 DEBUG Git submodule: extensions-builtin/stable-diffusion-webui-rembg / master
13:43:52-764334 DEBUG Git submodule: wiki / master
13:43:53-252335 DEBUG Installed packages: 251
13:43:53-256838 DEBUG Extensions all: ['sd-extension-chainner', 'sd-extension-system-info', 'sdnext-kanvas', 'sdnext-modernui',
'stable-diffusion-webui-rembg']
13:43:54-618346 DEBUG Extension installer: builtin=True file="E:\AI\sdnext\extensions-builtin\stable-diffusion-webui-rembg\install.py"
13:43:54-975346 DEBUG Extensions all: []
13:43:54-978382 INFO Extensions enabled: ['sd-extension-chainner', 'sd-extension-system-info', 'sdnext-kanvas', 'sdnext-modernui',
'stable-diffusion-webui-rembg']

Relevant log output

13:45:00-005839 INFO     Processing modifiers: apply
13:45:00-072160 DEBUG    Tokenizer: words=49408
                         file="models\huggingface\models--stable-diffusion-v1-5--stable-diffusion-v1-5\snapshots\451f4fe16113bff5a5d2269ed5ad43b05
                         92e9a14\tokenizer\vocab.json"
13:45:01-002996 INFO     Base: pipeline=StableDiffusionPipeline task=TEXT_2_IMAGE batch=1/1x1 set={'prompt_embeds': torch.Size([1, 77, 768]),
                         'negative_prompt_embeds': torch.Size([1, 77, 768]), 'guidance_scale': 6, 'generator': 'cuda:[902592423]',
                         'num_inference_steps': 20, 'eta': 1.0, 'guidance_rescale': 0, 'output_type': 'latent', 'width': 1024, 'height': 1024,
                         'parser': 'native'}
100%|████████████████████████████████████████████████████████████████████████████████████████████████████████| 4.69M/4.69M [00:00<00:00, 23.4MB/s]
 96%|███████████████████████████████████████████████████████████████████████████████████████████████████▉    | 4.50M/4.69M [00:00<00:00, 24.2MB/s]
13:45:04-721893 INFO     Decode: type="taesd" variant="TAESD 1.3 Mocha Croissant":
                         uri="https://github.com/madebyollin/taesd/raw/7f572ca629c9b0d3c9f71140e5f501e09f9ea280/taesd_decoder.pth"
                         fn="models\TAESD\taesd_13_decoder_sd.pth" download

13:45:04-727909 DEBUG    Decode: type="taesd" variant="TAESD 1.3 Mocha Croissant" fn="models\TAESD\taesd_13_decoder_sd.pth" layers=3 load
Progress  3.10it/s ███████████████████████████████▎   95% 19/20 00:08 00:00 BaseMIOpen(HIP): Warning [IsEnoughWorkspace] [GetSolutionsFallback WTI] Solver <GemmFwdRest>, workspace required: 75497472, provided ptr: 0000000000000000 size: 0
MIOpen(HIP): Warning [IsEnoughWorkspace] [GetSolutionsFallback WTI] Solver <GemmFwdRest>, workspace required: 301989888, provided ptr: 0000000000000000 size: 0
Progress  2.29it/s █████████████████████████████████ 100% 20/20 00:08 00:00 Base
13:45:09-796800 INFO     Processing modifiers: unapply
MIOpen(HIP): Warning [IsEnoughWorkspace] [GetSolutionsFallback WTI] Solver <GemmFwdRest>, workspace required: 1207959552, provided ptr: 0000000000000000 size: 0
MIOpen(HIP): Warning [IsEnoughWorkspace] [GetSolutionsFallback WTI] Solver <GemmFwdRest>, workspace required: 75497472, provided ptr: 0000000000000000 size: 0
MIOpen(HIP): Warning [IsEnoughWorkspace] [GetSolutionsFallback WTI] Solver <GemmFwdRest>, workspace required: 301989888, provided ptr: 0000000000000000 size: 0
MIOpen(HIP): Warning [IsEnoughWorkspace] [GetSolutionsFallback WTI] Solver <GemmFwdRest>, workspace required: 1207959552, provided ptr: 0000000000000000 size: 0
13:45:10-158140 DEBUG    Setting decode: component=vae {'slicing': True, 'tiling': False}
13:45:16-114515 DEBUG    Decode: vae="default" scale=8 upcast=False slicing=True tiling=False latents=[1, 4, 128, 128]:cuda:0 dtype=torch.bfloat16
                         time=5.956
13:45:16-179595 INFO     Save: image="outputs\text\00001-2026-01-29-dreamshaper_8.jpg" type=JPEG width=1024 height=1024 size=6926
13:45:16-185111 INFO     Processed: images=1 its=1.22 ops=['txt2img']
13:45:16-188114 DEBUG    Processed: timers={'total': 29.38, 'pipeline': 9.75, 'decode': 6.33, 'preview': 6.26, 'callback': 4.57, 'onload': 1.04,
                         'prompt': 0.92, 'pre': 0.25, 'move': 0.12}
13:45:16-194247 DEBUG    Processed: memory={'ram': {'total': 31.86, 'rss': 4.17, 'used': 23.55, 'free': 8.31, 'avail': 8.31, 'buffers': 0,
                         'cached': 0}, 'gpu': {'used': 5.11, 'total': 19.98, 'active': 2.09, 'peak': 7.96, 'retries': 0, 'oom': 0, 'swap': 0},
                         'job': ''}



sometimes (I think when I run again after the error occurd before):
13:57:39-859941 ERROR    Prompt parser encode: couldn't find start of token sequence for fragment at index 0 'brick wall'

Backend

Diffusers

Compute

AMD ROCm

Interface

Standard

Branch

Dev

Model

Other image model

Acknowledgements

• I have read the above and searched for existing issues
• I confirm that this is classified correctly and its not an extension issue

[vladmandic]

rocm-for-windows is still not official and definitely not a polished product.
what i suggest is to engage with community on sdnext discord #rocm channel as things are moving daily.
i really cant offer much assistance via normal github issue.

[NiklasRichter2222]

ok

[youday-gh]

I had same issue with different models and find a workaround.
Advanced - VAE type : Tiny works fine on Radeon RX 6600 using ZULDA.

[astrelsky]

@NiklasRichter2222 I see you are using 2.9.1+rocm7.10.0a20251120, perhaps try updating to a recent version. The latest is 7.12.0a20260206. https://rocm.nightlies.amd.com/v2/gfx110X-all.

You can try manually activating the sd.next virtual environment and then running

pip install --index-url https://rocm.nightlies.amd.com/v2/gfx110X-all --upgrade rocm[devel,libraries] torch torchaudio torchvision
rocm-sdk init
rocm-sdk test

Then try launching sd.next again with python launch.py --quick --skip-torch --skip-all --debug --use-rocm so that it doesn't tamper with the installed torch.