Memory allocating error with new amd driver 25.12.1

[mks812]

Custom Node Testing

• I have tried disabling custom nodes and the issue persists (see how to disable custom nodes if you need help)

Expected Behavior

worky worky

Actual Behavior

HIP out of memory. Tried to allocate 333.38 GiB. GPU 0 has a total capacity of 15.92 GiB of which 7.28 GiB is free. Of the allocated memory 8.22 GiB is allocated by PyTorch, and 19.88 MiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid fragmentation. See documentation for Memory Management (https://pytorch.org/docs/stable/notes/cuda.html#environment-variables)
This error means you ran out of memory on your GPU.

Steps to Reproduce

Install amd preview driver 25.12.1

Debug Logs

C:\ComfyUI_windows_portable>.\python_embeded\python.exe -s ComfyUI\main.py --windows-standalone-build --disable-all-custom-nodes
[WARNING] failed to run amdgpu-arch: binary not found.
Checkpoint files will always be loaded safely.
Total VRAM 16304 MB, total RAM 49063 MB
pytorch version: 2.9.0+rocmsdk20251116
Set: torch.backends.cudnn.enabled = False for better AMD performance.
AMD arch: gfx1201
ROCm version: (7, 1)
Set vram state to: NORMAL_VRAM
Device: cuda:0 AMD Radeon RX 9070 XT : native
Enabled pinned memory 22078.0
Using sub quadratic optimization for attention, if you have memory or speed issues try using: --use-split-cross-attention
Python version: 3.12.10 (tags/v3.12.10:0cc8128, Apr  8 2025, 12:21:36) [MSC v.1943 64 bit (AMD64)]
ComfyUI version: 0.4.0
________________________________________________________________________
WARNING WARNING WARNING WARNING WARNING

Installed frontend version 1.33.13 is lower than the recommended version 1.34.8.

Please install the updated requirements.txt file by running:
C:\ComfyUI_windows_portable\python_embeded\python.exe -s -m pip install -r C:\ComfyUI_windows_portable\ComfyUI\requirements.txt
If you are on the portable package you can run: update\update_comfyui.bat to solve this problem.

This error is happening because the ComfyUI frontend is no longer shipped as part of the main repo but as a pip package instead.
________________________________________________________________________
[Prompt Server] web root: C:\ComfyUI_windows_portable\python_embeded\Lib\site-packages\comfyui_frontend_package\static
Total VRAM 16304 MB, total RAM 49063 MB
pytorch version: 2.9.0+rocmsdk20251116
Set: torch.backends.cudnn.enabled = False for better AMD performance.
AMD arch: gfx1201
ROCm version: (7, 1)
Set vram state to: NORMAL_VRAM
Device: cuda:0 AMD Radeon RX 9070 XT : native
Enabled pinned memory 22078.0
Skipping loading of custom nodes
Context impl SQLiteImpl.
Will assume non-transactional DDL.
No target revision found.
________________________________________________________________________
WARNING WARNING WARNING WARNING WARNING

Installed frontend version 1.33.13 is lower than the recommended version 1.34.8.

Please install the updated requirements.txt file by running:
C:\ComfyUI_windows_portable\python_embeded\python.exe -s -m pip install -r C:\ComfyUI_windows_portable\ComfyUI\requirements.txt
If you are on the portable package you can run: update\update_comfyui.bat to solve this problem.

This error is happening because the ComfyUI frontend is no longer shipped as part of the main repo but as a pip package instead.
________________________________________________________________________
Starting server

To see the GUI go to: http://127.0.0.1:8188
got prompt
invalid prompt: {'type': 'invalid_prompt', 'message': 'Cannot execute because a node is missing the class_type property.', 'details': "Node ID '#28'", 'extra_info': {}}
got prompt
model weight dtype torch.bfloat16, manual cast: None
model_type FLOW
Using split attention in VAE
Using split attention in VAE
VAE load device: cuda:0, offload device: cpu, dtype: torch.bfloat16
Requested to load LuminaTEModel_
loaded completely; 95367431640625005117571072.00 MB usable, 4986.46 MB loaded, full load: True
CLIP/text encoder model load device: cuda:0, offload device: cpu, current: cuda:0, dtype: torch.float16
Requested to load Lumina2
Unloaded partially: 2160.46 MB freed, 2826.00 MB remains loaded, 121.50 MB buffer reserved, lowvram patches: 0
loaded completely; 9045.30 MB usable, 4977.74 MB loaded, full load: True
100%|██████████████████████████████████████████████████████████████████████████████████| 30/30 [00:03<00:00,  8.10it/s]
Requested to load AutoencodingEngine
Unloaded partially: 1940.24 MB freed, 3037.50 MB remains loaded, 40.50 MB buffer reserved, lowvram patches: 0
loaded completely; 5120.57 MB usable, 159.87 MB loaded, full load: True
Prompt executed in 12.93 seconds

Other

No response

[lostdisc]

The new 25.12.1 driver may not be intended for use with ROCm; Note the lower version number compared to 25.20.01.17 which was released together with Windows ROCm. I saw someone else having ROCm trouble with 25.12.1 too, and reverting to 25.20.01.17 fixed it for him. (In which case, yes it's dumb that the Adrenalin software is offering to update people who are on a higher version number.)

[mks812]

right, I didnt even think about it. thats stupid there is older rocm preview driver and one week newer redstone stable driver without rocm. I just took it for granted that it has the rocm in it

[SWQDswqd]

经过我四天的努力，用尽各种方法找到了解决方案，我的配置是9950X 9070 ，1.用AMD驱动清理工具卸载25.12.1版本2.安装25.20.01.17，安装时禁用AMD的自动更新，不要打勾！！因为这是预览版，即便比25.12新，AMD也会以稳定版的最新版当作自动更新的最新版3.在设备管理器或者BIOS里禁用核显（如果有的话）4.每次启动comfyUI确认核显没有打开，且不要一开就打开comfyUI，等待系统完全准备好

[lostdisc]

26.1.1 came out today with updated AI support, FYI. The installer can also optionally install ComfyUI, PyTorch, etc, but they seem to be outdated versions, so installng separately may still be preferred.