ComfyUI Windows Version AMD 6900XT - Still instant crashing after model load

### Custom Node Testing

- [ ] I have tried disabling custom nodes and the issue persists (see [how to disable custom nodes](https://docs.comfy.org/troubleshooting/custom-node-issues#step-1%3A-test-with-all-custom-nodes-disabled) if you need help)

### Expected Behavior

Ability to generate anything without ComfyUI crashing after loading/attempting to load the data to the GPU RAM

### Actual Behavior

ComfyUI backend immediately crashes after loading/attempting to load the model to the GPU RAM

### Steps to Reproduce

1. Download ComfyUI for windows.
2. Install on a up to date Windows 11 system with a AMD 6900XT installed. (unknown if the RX560 is causing any issue by existing)
3. Run ComfyUI and try to generate a simple image model capable of fitting in VRAM.
4. Watch in task manager as the data loading into the GPU RAM and then crashes.

### Debug Logs

```powershell
Adding extra search path custom_nodes C:\ComfyUI_Win\custom_nodes
Adding extra search path download_model_base C:\ComfyUI_Win\models
Adding extra search path custom_nodes C:\Comfyui_installer\resources\ComfyUI\custom_nodes
Setting output directory to: C:\ComfyUI_Win\output
Setting input directory to: C:\ComfyUI_Win\input
Setting user directory to: C:\ComfyUI_Win\user
[START] Security scan
[DONE] Security scan
** ComfyUI startup time: 2026-01-27 11:35:56.235
** Platform: Windows
** Python version: 3.12.11 (main, Aug 18 2025, 19:17:54) [MSC v.1944 64 bit (AMD64)]
** Python executable: C:\ComfyUI_Win\.venv\Scripts\python.exe
** ComfyUI Path: C:\Comfyui_installer\resources\ComfyUI
** ComfyUI Base Folder Path: C:\Comfyui_installer\resources\ComfyUI
** User directory: C:\ComfyUI_Win\user
** ComfyUI-Manager config path: C:\ComfyUI_Win\user\__manager\config.ini
** Log path: C:\ComfyUI_Win\user\comfyui.log
[ComfyUI-Manager] Skipped fixing the 'comfyui-frontend-package' dependency because the ComfyUI is outdated.
[PRE] ComfyUI-Manager
Checkpoint files will always be loaded safely.
Total VRAM 16368 MB, total RAM 65416 MB
pytorch version: 2.9.0+rocmsdk20251116
AMD arch: gfx1030
ROCm version: (7, 1)
Set vram state to: NORMAL_VRAM
Device: cuda:0 AMD Radeon RX 6900 XT : native
Using async weight offloading with 2 streams
Enabled pinned memory 29437.0
Found comfy_kitchen backend triton: {'available': False, 'disabled': True, 'unavailable_reason': "ImportError: No module named 'triton'", 'capabilities': []}
Found comfy_kitchen backend eager: {'available': True, 'disabled': False, 'unavailable_reason': None, 'capabilities': ['apply_rope', 'apply_rope1', 'dequantize_nvfp4', 'dequantize_per_tensor_fp8', 'quantize_nvfp4', 'quantize_per_tensor_fp8', 'scaled_mm_nvfp4']}
Found comfy_kitchen backend cuda: {'available': True, 'disabled': True, 'unavailable_reason': None, 'capabilities': ['apply_rope', 'apply_rope1', 'dequantize_nvfp4', 'dequantize_per_tensor_fp8', 'quantize_nvfp4', 'quantize_per_tensor_fp8', 'scaled_mm_nvfp4']}
Using sub quadratic optimization for attention, if you have memory or speed issues try using: --use-split-cross-attention
Python version: 3.12.11 (main, Aug 18 2025, 19:17:54) [MSC v.1944 64 bit (AMD64)]
ComfyUI version: 0.9.1
[Prompt Server] web root: C:\Comfyui_installer\resources\ComfyUI\web_custom_versions\desktop_app
[START] ComfyUI-Manager
[ComfyUI-Manager] network_mode: public
[ComfyUI-Manager] The matrix sharing feature has been disabled because the `matrix-nio` dependency is not installed.
To use this feature, please run the following command:
C:\ComfyUI_Win\.venv\Scripts\python.exe -m pip install matrix-nio

Import times for custom nodes:
0.0 seconds: C:\Comfyui_installer\resources\ComfyUI\custom_nodes\websocket_image_save.py

Context impl SQLiteImpl.
Will assume non-transactional DDL.
Assets scan(roots=['models']) completed in 0.094s (created=58, skipped_existing=0, total_seen=58)
Starting server

To see the GUI go to: http://127.0.0.1:8001
comfyui-frontend-package not found in requirements.txt
got prompt
Using split attention in VAE
Using split attention in VAE
VAE load device: cuda:0, offload device: cpu, dtype: torch.float32
CLIP/text encoder model load device: cuda:0, offload device: cpu, current: cpu, dtype: torch.float16
Requested to load ZImageTEModel_
loaded completely; 14707.67 MB usable, 7672.25 MB loaded, full load: True
```

### Other

The only thing I can think of note is that the 6900XT is the primary graphics card. I run 6 displays so I use the RX560 to offload processing for the other 3. The 6900XT is primary for gaming but I can swap the primary boot choice in bios so I will test if there is any difference, thought i suspect not as I have previously been able to get only image generation occasionally working using patientx's manual setup method here: https://github.com/patientx/ComfyUI-Zluda/issues/170#issuecomment-2972793016

But that is not very stable or can do anything more than images when others report doing so much more on this card and it's family models.

Anything info on anything else i can try would be appreciated.

Start by reproducing the crash on Windows 11 with the AMD RX 6900 XT and all custom nodes disabled, using the model-loading sequence described in the issue. Inspect the startup log around the ComfyUI 0.9.1 model-load entry point and compare behavior with the RX560 or alternate primary-GPU setup; done means a reproducible cause, traceback, or confirmed environment-specific failure is identified.

Written by the indexing model from the issue text.