Wrong device allocation?

### Custom Node Testing

- [x] 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

I normally startup ComfyUI and once I start run (after I setted up a workflow), I expect the model to be loaded onto VRAM and executed on the same GPU.

### Actual Behavior

I have two CUDA GPUs inside of my setup. For whatever reason the model is loaded onto CUDA1, while the execution is running on CUDA0, producing a noticable performance loss, as the data has to be transfered from CUDA1 VRAM to CUDA0 compute.

### Steps to Reproduce

Run ComfyUI from a successful setted up workflow with two CUDA devices plugged into the computer.

### Debug Logs

```powershell
.\python_embeded\python.exe -s ComfyUI\main.py --windows-standalone-build
Found comfy_kitchen backend eager: {'available': True, 'disabled': False, 'unavailable_reason': None, 'capabilities': ['apply_rope', 'apply_rope1', 'dequantize_mxfp8', 'dequantize_nvfp4', 'dequantize_per_tensor_fp8', 'quantize_mxfp8', 'quantize_nvfp4', 'quantize_per_tensor_fp8', 'scaled_mm_mxfp8', 'scaled_mm_nvfp4']}
Found comfy_kitchen backend cuda: {'available': True, 'disabled': False, 'unavailable_reason': None, 'capabilities': ['apply_rope', 'apply_rope1', 'dequantize_nvfp4', 'dequantize_per_tensor_fp8', 'quantize_mxfp8', 'quantize_nvfp4', 'quantize_per_tensor_fp8', 'scaled_mm_nvfp4']}
Found comfy_kitchen backend triton: {'available': False, 'disabled': True, 'unavailable_reason': "ImportError: No module named 'triton'", 'capabilities': []}
Checkpoint files will always be loaded safely.
Total VRAM 49140 MB, total RAM 130199 MB
pytorch version: 2.11.0+cu130
Set vram state to: NORMAL_VRAM
Device: cuda:0 NVIDIA GeForce RTX 4090 : cudaMallocAsync
Using async weight offloading with 2 streams
Enabled pinned memory 52079.0
Using pytorch attention
aimdo: src-win/cuda-detour.c:77:INFO:aimdo_setup_hooks: found driver at 00007FFF87410000, installing 4 hooks
aimdo: src-win/cuda-detour.c:61:DEBUG:install_hook_entrys: hooks successfully installed
aimdo: src/control.c:69:INFO:comfy-aimdo inited for GPU: NVIDIA GeForce RTX 4090 (VRAM: 49139 MB)
DynamicVRAM support detected and enabled
Python version: 3.13.12 (tags/v3.13.12:1cbe481, Feb 3 2026, 18:22:25) [MSC v.1944 64 bit (AMD64)]
ComfyUI version: 0.19.3
comfy-aimdo version: 0.2.12
comfy-kitchen version: 0.2.8
ComfyUI frontend version: 1.42.14
[Prompt Server] web root: E:\AI\ComfyUI_windows_portable\python_embeded\Lib\site-packages\comfyui_frontend_package\static
Asset seeder disabled

Import times for custom nodes:
0.0 seconds: E:\AI\ComfyUI_windows_portable\ComfyUI\custom_nodes\websocket_image_save.py

Context impl SQLiteImpl.
Will assume non-transactional DDL.
Starting server

To see the GUI go to: http://127.0.0.1:8188
got prompt
Found quantization metadata version 1
Detected mixed precision quantization
Using mixed precision operations
model weight dtype torch.bfloat16, manual cast: torch.bfloat16
model_type FLUX
VAE load device: cuda:0, offload device: cpu, dtype: torch.bfloat16
no CLIP/text encoder weights in checkpoint, the text encoder model will not be loaded.
Found quantization metadata version 1
Using MixedPrecisionOps for text encoder
CLIP/text encoder model load device: cuda:0, offload device: cpu, current: cpu, dtype: torch.float16
Requested to load LTXAVTEModel_
Model LTXAVTEModel_ prepared for dynamic VRAM loading. 11200MB Staged. 0 patches attached. Force pre-loaded 290 weights: 1497 KB.
Model LTXAVTEModel_ prepared for dynamic VRAM loading. 11200MB Staged. 0 patches attached. Force pre-loaded 290 weights: 1497 KB.
Requested to load VideoVAE
Model VideoVAE prepared for dynamic VRAM loading. 1384MB Staged. 0 patches attached.
Model VideoVAE prepared for dynamic VRAM loading. 1384MB Staged. 0 patches attached.
Model VideoVAE prepared for dynamic VRAM loading. 1384MB Staged. 0 patches attached.
Model VideoVAE prepared for dynamic VRAM loading. 1384MB Staged. 0 patches attached.
Model VideoVAE prepared for dynamic VRAM loading. 1384MB Staged. 0 patches attached.
Requested to load LTXAV
Model LTXAV prepared for dynamic VRAM loading. 24206MB Staged. 0 patches attached.
12%|██████▏ | 1/8 [00:11<01:20, 11.56s/it, Model Initialization complete! ]Interrupting prompt 0eba4d07-e645-469c-ac1f-0f762a77c64a
12%|██████▏ | 1/8 [00:12<01:28, 12.61s/it, Model Initialization complete! ]
Processing interrupted
Prompt executed in 23.52 seconds
got prompt
Model VideoVAE prepared for dynamic VRAM loading. 1384MB Staged. 0 patches attached.
Model VideoVAE prepared for dynamic VRAM loading. 1384MB Staged. 0 patches attached.
Model VideoVAE prepared for dynamic VRAM loading. 1384MB Staged. 0 patches attached.
Model VideoVAE prepared for dynamic VRAM loading. 1384MB Staged. 0 patches attached.
Model VideoVAE prepared for dynamic VRAM loading. 1384MB Staged. 0 patches attached.
0 models unloaded.
Model LTXAV prepared for dynamic VRAM loading. 24206MB Staged. 0 patches attached.
100%|████████████████████████████████████████████████████████████████████████████████████| 8/8 [09:29<00:00, 71.17s/it]
Requested to load AudioVAE
loaded completely; 14017.40 MB usable, 693.46 MB loaded, full load: True
0 models unloaded.
Model VideoVAE prepared for dynamic VRAM loading. 1384MB Staged. 0 patches attached.
Prompt executed in 00:10:12
```

### Other

- ComfyUI does not recognize both CUDA devices
- The VRAM allocation at all seems awkward
- Other LLM backends work fine (e.g. koboldcpp, ollama) for single or dual GPU usage, including VRAM allocation