Update-ComfyUI overwrites rocm packages with cuda packages on AMD system

### 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

the "Update ComfyUI" button keeps the ROCm packages (torch, triton, ...)

### Actual Behavior

Torch, triton, ... packages for ROCm where replaced with Cuda pacakges during "Update ComfyUI"

### Steps to Reproduce

Linux Mint computer with AMD GPU (in my case, 7900 XTX) and ComfyUI 27.0 installed, and running without any issues. Using rocm 7.2.4.
Open the Manager and click "Update ComfyUI"
After this operation, ComfyUI no longer starts up.
Error = Found no NVIDIA driver on your system (see debug log below)

### Debug Logs

```powershell
## ComfyUI-Manager: installing dependencies done.
[2026-07-12 12:13:16.804] ** ComfyUI startup time: 2026-07-12 12:13:16.804
[2026-07-12 12:13:16.804] ** Platform: Linux
[2026-07-12 12:13:16.804] ** Python version: 3.12.3 (main, Jun 19 2026, 12:46:00) [GCC 13.3.0]
[2026-07-12 12:13:16.804] ** Python executable: ~/comfyui-venv/bin/python3
[2026-07-12 12:13:16.804] ** ComfyUI Path: ~/ComfyUI
[2026-07-12 12:13:16.804] ** ComfyUI Base Folder Path: ~/ComfyUI
[2026-07-12 12:13:16.804] ** User directory: ~/ComfyUI/user
[2026-07-12 12:13:16.804] ** ComfyUI-Manager config path: ~/ComfyUI/user/__manager/config.ini
[2026-07-12 12:13:16.804] ** Log path: ~/ComfyUI/user/comfyui.log

Prestartup times for custom nodes:
[2026-07-12 12:13:16.879] 0.0 seconds: ~/ComfyUI/custom_nodes/rgthree-comfy
[2026-07-12 12:13:16.879] 0.0 seconds: ~/ComfyUI/custom_nodes/comfyui-easy-use
[2026-07-12 12:13:16.879] 0.4 seconds: ~/ComfyUI/custom_nodes/ComfyUI-Manager
[2026-07-12 12:13:16.879]
[2026-07-12 12:13:18.214] Found comfy_kitchen backend eager: {'available': True, 'disabled': False, 'unavailable_reason': None, 'capabilities': ['adaln', 'apply_rope', 'apply_rope1', 'apply_rope_split_half', 'apply_rope_split_half1', 'dequantize_int8_convrot_weight', 'dequantize_int8_convrot_weight_dtype', 'dequantize_int8_simple', 'dequantize_int8_simple_dtype', 'dequantize_mxfp8', 'dequantize_nvfp4', 'dequantize_per_tensor_fp8', 'gemv_awq_w4a16', 'int8_linear', 'quantize_and_rotate_rowwise', 'quantize_int8_convrot_weight', 'quantize_int8_rowwise', 'quantize_int8_tensorwise', 'quantize_mxfp8', 'quantize_nvfp4', 'quantize_per_tensor_fp8', 'quantize_svdquant_w4a4', 'scaled_mm_mxfp8', 'scaled_mm_nvfp4', 'scaled_mm_svdquant_w4a4', 'stochastic_rounding_fp8']}
[2026-07-12 12:13:18.214] Found comfy_kitchen backend triton: {'available': False, 'disabled': True, 'unavailable_reason': 'Neither CUDA nor XPU available on this system', 'capabilities': []}
[2026-07-12 12:13:18.214] Found comfy_kitchen backend cuda: {'available': False, 'disabled': False, 'unavailable_reason': 'CUDA not available on this system', 'capabilities': []}
[2026-07-12 12:13:18.216] Checkpoint files will always be loaded safely.
[2026-07-12 12:13:18.222] Traceback (most recent call last):
[2026-07-12 12:13:18.222] File "~/ComfyUI/main.py", line 227, in
[2026-07-12 12:13:18.222] import execution
[2026-07-12 12:13:18.222] File "~/ComfyUI/execution.py", line 18, in
[2026-07-12 12:13:18.222] import comfy.model_management
[2026-07-12 12:13:18.222] File "~/ComfyUI/comfy/model_management.py", line 362, in
[2026-07-12 12:13:18.222] total_vram = get_total_memory(get_torch_device()) / (1024 * 1024)
[2026-07-12 12:13:18.222] ^^^^^^^^^^^^^^^^^^
[2026-07-12 12:13:18.223] File "~/ComfyUI/comfy/model_management.py", line 211, in get_torch_device
[2026-07-12 12:13:18.223] return torch.device(torch.cuda.current_device())
[2026-07-12 12:13:18.223] ^^^^^^^^^^^^^^^^^^^^^^^^^^^
[2026-07-12 12:13:18.224] File "~/comfyui-venv/lib/python3.12/site-packages/torch/cuda/__init__.py", line 1205, in current_device
[2026-07-12 12:13:18.224] _lazy_init()
[2026-07-12 12:13:18.224] File "~/comfyui-venv/lib/python3.12/site-packages/torch/cuda/__init__.py", line 529, in _lazy_init
[2026-07-12 12:13:18.224] torch._C._cuda_init()
[2026-07-12 12:13:18.224] RuntimeError: Found no NVIDIA driver on your system. Please check that you have an NVIDIA GPU and installed a driver from http://www.nvidia.com/Download/index.aspx
```

### Other

I re-installed all ROCm specific packages (torch, triton, jax, etc...) and then ComfyUI starts up again.
So I am not blocked, but this should not happen.
After "Update ComfyUI": (notice "torch" is now the default package, not the rocm package)
```
(comfyui-venv) @linux:~$ pip list | grep -i torch
open_clip_torch 3.3.0
torch 2.13.0
torchaudio 2.9.0+rocm7.2.4.gite3c6ee2b
torchsde 0.2.6
torchvision 0.24.0+rocm7.2.4.gitb919bd0c
```
After the reinstall of the rocm packages: (I used the most the recent packages)
```
(comfyui-venv) @linux:~$ pip list | grep -i torch
open_clip_torch 3.3.0
torch 2.10.0+rocm7.2.4.lw.git3d3aa833
torchaudio 2.10.0+rocm7.2.4.git5047768f
torchsde 0.2.6
torchvision 0.25.0+rocm7.2.4.git82df5f59
```

Start with the ComfyUI-Manager "Update ComfyUI" entry point and reproduce the dependency installation on the Linux AMD/ROCm setup described. Compare the packages shown by pip list before and after updating, then verify that main.py starts without the NVIDIA-driver error in comfy/model_management.py and that ROCm packages remain installed.

Written by the indexing model from the issue text.