pytorch - 💡(How to fix) Fix [Hardware Support] RTX 5080 (Blackwell) CUDA error: no kernel image is available for execution on the device on latest Nightly build [1 participants]

RuntimeError: CUDA error: no kernel image is available for execution on the device

TL;DR

The most likely fix is to wait for pre-compiled Windows binaries that natively support the SM architectures for the consumer RTX 5000 series or use a workaround such as compiling PyTorch from source with the necessary CUDA support.

Guidance

• The error message indicates a CUDA kernel error, suggesting that the current PyTorch version does not support the NVIDIA GeForce RTX 5080 GPU.
• Compiling PyTorch from source with TORCH_USE_CUDA_DSA and the necessary CUDA version (cu124) might enable support for the RTX 5080.
• Passing CUDA_LAUNCH_BLOCKING=1 can help with debugging by making CUDA kernel launches synchronous.
• Consider using a different PyTorch version or a nightly build that might include support for the RTX 5000 series.

Example

No code snippet is provided as the issue does not require a specific code change.

Notes

The issue is specific to the Windows platform and the NVIDIA GeForce RTX 5080 GPU. The solution might not apply to other platforms or GPUs.

Recommendation

Apply workaround: Compiling PyTorch from source with the necessary CUDA support is a viable workaround until pre-compiled Windows binaries that support the RTX 5000 series are available.