🚀 The feature, motivation and pitch

🚀 The Feature

Please add official support for CUDA Compute Capability sm_120, which corresponds to NVIDIA’s Blackwell architecture (e.g., RTX 5090 GPUs).

Motivation, Pitch

Many researchers and professionals have already adopted RTX 5090 and other Blackwell-based GPUs. Currently, the latest PyTorch stable releases support only up to sm_90. Running models on these new GPUs leads to errors like:

RuntimeError: CUDA error: no kernel image is available for execution on the device
CUDA kernel errors might be asynchronously reported at some other API call, so the stacktrace below might be incorrect.
For debugging consider passing CUDA_LAUNCH_BLOCKING=1
Compile with TORCH_USE_CUDA_DSA to enable device-side assertions.

Alternatives

Users are forced to either build PyTorch from source or use experimental nightly builds with limited platform support.

Adding sm_120 support will:

• Enable full compatibility for cutting-edge hardware
• Support researchers and labs relying on high-performance GPUs
• Eliminate barriers for Windows and non-developer users

Alternatives

Some workarounds exist (e.g., building from source with TORCH_CUDA_ARCH_LIST="sm_120"), but this is not ideal for most users.

Additional context

NVIDIA has already released public drivers and CUDA 12.4+/12.5 toolkits that support sm_120. Nightly builds (as of early 2025) support sm_120 partially for Linux. However, official support in stable PyTorch builds is still missing.

A stable release with sm_120 support would be highly appreciated.

Thanks a lot for your great work!

Additional context

using PyTorch. I recently upgraded my hardware to an NVIDIA RTX 5090, which is based on the Blackwell architecture (sm_120).

Unfortunately, the current stable PyTorch release does not support this architecture, making it impossible to run or test models without building from source or switching to Linux with nightly builds. This presents a significant barrier to reproducibility and research progress, especially for Windows-based environments.

Adding official support for sm_120 in stable Windows builds would benefit a growing number of researchers, developers, and institutions already adopting RTX 50-series GPUs.

Thank you again for your continued efforts and support for the PyTorch community.

cc @seemethere @malfet @atalman