Request for Official CUDA and PyTorch Support for RTX 5080/5090 (sm_120)

Gaming and Visualization Technologies General Topics and Other SDKs GPU - Hardware
1

Hello NVIDIA Dev Team,

I’m an AI engineer and systems architect currently working on a next-generation AI assistant and local inference platform. This project is designed for real-time, on-device LLMs, computer vision, and multimodal processing—all running 100% locally on consumer-grade hardware with no reliance on external APIs or cloud backends.

My current build environment includes:

  • GPU: NVIDIA RTX 5080 (Ada Gen3)
  • CUDA Version: 12.8
  • Linux: WSL2 + Ubuntu
  • Build Target: PyTorch v2.5.0 from source
  • Arch Flags: sm_120, which is currently unsupported by upstream select_compute_arch.cmake

After patching select_compute_arch.cmake manually to support sm_120, I’ve run into constant compatibility and build-breaking issues. It seems sm_120 is not yet integrated as a known architecture, despite this GPU family being commercially available for over a month.

My questions:

  1. When can we expect official support for sm_120 in CUDA and PyTorch?
  2. Is there a preferred patch or integration pathway for developers building experimental CUDA/PyTorch environments for next-gen hardware?
  3. Can you confirm if sm_120 is aligned with an upcoming compute capability update (e.g., 9.x, 10.x)?

I am happy to provide build logs, system diagnostics, and project goals if it helps expedite integration or beta access to upcoming CUDA revisions.

Thank you for your time and commitment to developers like me pushing the limits of local AI compute.

— The Architect

1 Like
2

The current nighly build of pytorch 2.8.0 works better with cuda 12.8. I dont think cuda 12.8 is supported at all in previous builds. Hopefully installing that will make some of your issues go away.

3

SUCCESS: PyTorch with CUDA 12.8 on RTX 5080 (sm_120) — Working

Following up with an important update:

I was able to successfully build PyTorch from source with full sm_120 support for my RTX 5080 using CUDA 12.8. This required a few custom modifications, but the result is stable and fully functional.

Summary of What Worked:

  • PyTorch version: 2.5.0 (built from source)
  • CUDA version: 12.8
  • GPU architecture: sm_120 (manually patched)
  • Platform: WSL2 + Ubuntu
  • Compiler: gcc / g++ installed manually
  • Build system: Ninja

Key Steps:

  1. Cloned PyTorch and checked out the v2.5.0 tag.
  2. Manually patched select_compute_arch.cmake to recognize sm_120.
  3. Set TORCH_CUDA_ARCH_LIST="12.0;12.1;12.2;12.3+PTX".
  4. Configured CMake with:

bash

CopyEdit

cmake .. -G Ninja \
  -DCMAKE_BUILD_TYPE=Release \
  -DUSE_CUDA=1 \
  -DCUDA_TOOLKIT_ROOT_DIR=/usr/local/cuda \
  -DTORCH_CUDA_ARCH_LIST="12.0;12.1;12.2;12.3+PTX" \
  -DCMAKE_C_COMPILER=/usr/bin/gcc \
  -DCMAKE_CXX_COMPILER=/usr/bin/g++ \
  -DCMAKE_EXPORT_COMPILE_COMMANDS=ON
  1. Ran python setup.py install inside the PyTorch directory.

Result:

Confirmed success with:

python

CopyEdit

import torch
print(torch.cuda.get_device_properties(0))

Output:

scss

CopyEdit

_CudaDeviceProperties(name='NVIDIA GeForce RTX 5080', major=12, minor=0, total_memory=16302MB, multi_processor_count=84)

Let me know if anyone wants the patch or full build script. I’m happy to share.

— The Architect

2 Likes
4

Hi - yes please share the full build script. I’m hoping to use my 5080 to fine-tune using unsloth but running into issues. thanks in advance!