MultiScaleDeformableAttention CUDA Compilation Fails on PyTorch 2.6+

[LadyAfrik]

I encountered a compilation issue while setting up LLaVA-Grounding. When running:
CUDA_VISIBLE_DEVICES=0 python gradio_demo/LLaVA_G_Demo.py
--path_vision_cfg configs/openseed/openseed_swint_lang_joint_2st_visual_prompt.yaml
--path_inter_cfg configs/semsam/visual_prompt_encoder.yaml
--model_path checkpoints/llava_grounding

The program suggests running:
cd mask2former/modeling/pixel_decoder/ops
sh make.sh

However, when executing sh make.sh, compilation fails with the following error:
error: no suitable conversion function from "const at::DeprecatedTypeProperties" to "c10::ScalarType" exists

This appears to be an API change in PyTorch starting from version 2.6, causing issues with Deformable DETR’s custom CUDA kernel.

Reproduction Steps:
Install dependencies (Python 3.9+, PyTorch 2.6+, BitsandBytes 0.46.0).
Run the LLaVA-Grounding setup.
Attempt CUDA compilation (sh make.sh).
Observe error related to DeprecatedTypeProperties.

Workarounds Tried:
Installed a patched version of Transformers:
pip install -U git+https://github.com/qubvel/transformers@fix-custom-kernels
Still fails.

Checked for .type() occurrences in source files:
grep -rnw ~/LLaVA-Grounding/Mask2Former/ -e 'tensor.type()'
No relevant matches found.

Ensured all dependencies are correctly installed.

Additional Context: This issue may also impact other projects relying on Mask2Former + PyTorch 2.6+. Would appreciate guidance on an official fix or required code modifications.