Installation docs don't verify MultiScaleDeformableAttention CUDA op is built — silent CPU-only fallback on HPC

[ffhammer]

The current installation guide (https://hugobaudchon.github.io/CanopyRS/getting-started/installation/) does not verify that detrex's multi_scale_deform_attn CUDA extension is actually compiled with GPU support. On HPC systems this fails silently: the install completes, torch.cuda.is_available() returns True, but inference crashes only once the model runs:

RuntimeError: Not compiled with GPU support
  File ".../detrex/layers/multi_scale_deform_attn.py", line 56, in forward
    output = _C.ms_deform_attn_forward(...)

Root cause: on many HPC setups (e.g. ETH Euler) the conda env ships torch with cu126 but no nvcc / CUDA_HOME, so detrex's setup.py falls back to a CPU-only build without erroring. After ~2h of tile preprocessing the pipeline dies at the first forward pass.

Suggested doc additions:

1. State explicitly that nvcc matching torch.version.cuda must be available before installing detrex, and that build must happen on a GPU node.
2. Add a pre-install check block:

   nvcc --version                 # must match torch.version.cuda
   nvidia-smi                     # GPU visible
   echo $CUDA_HOME                # set
   

   With a fix for the common conda case:

   conda install -c nvidia/label/cuda-12.6.0 cuda-nvcc cuda-cudart-dev cuda-libraries-dev
   export CUDA_HOME=$CONDA_PREFIX FORCE_CUDA=1 TORCH_CUDA_ARCH_LIST="8.9"
   

3. Add a post-install verification step users can run in seconds:

   python -c "
   import torch
   from detrex.layers.multi_scale_deform_attn import MultiScaleDeformableAttnFunction
   from detrex.layers import _C
   assert hasattr(_C, 'ms_deform_attn_forward')
   print('detrex CUDA op OK')
   "
   

Environment where this occurred:

• ETH Euler, RTX 4090 node
• conda env, torch 2.7.1+cu126, driver 580 / CUDA 13.0
• nvcc missing until installed manually into the env