🐛 Describe the bug

Hello PyTorch team,

I would like to kindly request official support for sm_120 (RTX 50-series / Blackwell GPUs, e.g. RTX 5070 Ti) in the stable PyTorch builds.

Current situation:

• CUDA 12.8/12.9 already includes support for Blackwell architectures.
• PyTorch nightly builds (e.g., 2.10.0.dev + cu12.9) can detect sm_120, but they are not yet fully stable.
• In my case, I tested the nightly build on Windows 11 with an RTX 5070 Ti. PyTorch itself launches, but DeepLabCut (DLC GUI, which relies heavily on PyTorch) still fails to start properly.
• Interestingly, Annolid GUI works fine on the same PC with RTX 5070 Ti. This suggests the underlying CUDA/NVIDIA support is there, but stable PyTorch integration is still missing.

Problem:

• DLC (and many other research tools) depend strictly on stable PyTorch releases. Without official sm_120 support in the stable channel, we cannot run these applications on RTX 50-series GPUs.
• As a researcher, I purchased RTX 5070 Ti for deep learning workloads, but currently it cannot be used productively with DLC due to this gap.

Request:

• Please prioritize adding official sm_120 support into stable PyTorch builds.
• Even partial support in an upcoming stable release (e.g., wheels with cu12.9) would greatly help researchers and developers adopt RTX 50-series hardware.
• At minimum, could you provide an ETA or roadmap for when sm_120 will be supported in stable builds?

Thank you very much for your efforts and for maintaining this essential framework.

Best regards,

Versions

RTX 5070 Ti requires CUDA 12.0+ for full support.
Multiple rebuilds of environments tested.
PyTorch, NumPy, and OpenCV work independently.
Failures appear specific to DLC’s internal module loading mechanism.

cc @seemethere @malfet @atalman @peterjc123 @mszhanyi @skyline75489 @nbcsm @iremyux @Blackhex @ptrblck @msaroufim @eqy @jerryzh168