Nerfstudio on RTX 5070 (sm_120) – PyTorch / CUDA stack and tiny-cuda-nn install issues

[bashtavenko]

I hit roadblocks for installing RTX 5070 on Ubuntu 24. This is not a bug report, but the only solution that I can see is to buy GeForce GTX 1050 on ebay. I would like to have some guidance and some documentation update (see below).

System

• GPU: NVIDIA GeForce RTX 5070 (Blackwell, compute capability sm_120)
• OS: Ubuntu 24.04
• System CUDA: 12.8 (driver + toolkit installed globally) Also have 11.8 CUDA toolkit as non-default
• Nerfstudio: poster dataset, nerfacto pipeline
• Python: 3.8 in a dedicated conda env nerfstudio
• Following the official Nerfstudio conda + cu118 install instructions

What works

1. Created conda env and installed PyTorch 2.1.2 + cu118 and torchvision as documented.
2. Attempted to install CUDA 11.8 toolkit inside the env via the NVIDIA channel.
3. Installed tiny-cuda-nn with ninja and the torch bindings from GitHub, after working through CUDA version mismatch and checksum issues.

Issues hit along the way

1. Initial CUDA version mismatch (11.8 vs 12.8)

   • PyTorch cu118 in the env reports CUDA 11.8, but the system exposes CUDA 12.8.
   • Building tiny-cuda-nn failed with the standard PyTorch extension error:
     “The detected CUDA version (12.8) mismatches the version that was used to compile PyTorch (11.8).”

2. GCC too new for CUDA 11.8

   • After adding a local CUDA 11.8 toolkit to the env, tiny-cuda-nn build then failed with:
     #error -- unsupported GNU version! gcc versions later than 11 are not supported!
   • Ubuntu 24 ships GCC ≥13; CUDA 11.8’s nvcc refuses to compile without workarounds (installing gcc-11 and exporting CC/CXX/CUDAHOSTCXX, or using -allow-unsupported-compiler).
   • added export CUDA_HOME="$CONDA_PREFIX" and export PATH="$CUDA_HOME/bin:$PATH"

3. Training still fails / degrades even after tiny-cuda-nn install attempt

   • Nerfstudio starts training the poster dataset but prints:

     NVIDIA GeForce RTX 5070 with CUDA capability sm_120 is not compatible with the current PyTorch installation.
     The current PyTorch install supports CUDA capabilities sm_50 sm_60 sm_70 sm_75 sm_80 sm_86 sm_37 sm_90.
     

   • So even if the extension situation is resolved, the core PyTorch build itself does not recognize sm_120, and training either falls back to CPU or fails.

Net result

• The current stable PyTorch + cu118 stack recommended in the Nerfstudio docs appears incompatible with RTX 5070 (sm_120) on Ubuntu 24.
• Trying to follow the older CUDA 11.8 + tiny-cuda-nn route runs into a tangle of:
  • CUDA runtime mismatch (11.8 vs 12.8)
  • GCC version incompatibility with CUDA 11.8
  • And finally, PyTorch itself lacking sm_120 support at the recommended version.
  • I tried using the latest Pytorch 2.9.1 with CUDA 12.8 with Python 3.8. The training succeed but ns-viewer failed with loading checkpoint due to the Pytorch old weight parameters. I patched it with newer syntax, the viewer began working but than it failed with out of memory for GPU....

What I’m asking

1. Recommended stack for RTX 50‑series GPUs (sm_120)

   • Is there a known‑good combination of:
     • PyTorch version (e.g., a specific nightly with CUDA 12.x and sm_120 support),
     • CUDA toolkit version (system vs env-local),
     • And Nerfstudio commit / viewer configuration
       that’s been validated on RTX 5070 / 5080 / 5090?
     • Python version

2. Guidance on tiny-cuda-nn for these GPUs

   • For Blackwell cards, should users:
     • Avoid tiny-cuda-nn entirely and prefer gsplat / other backends, or
     • Build it against a CUDA 12.x PyTorch with sm_120, and if so, what exact flags / arch list should be used?

3. Docs update request

   • The current install docs target Python 3.8 + cu118 + CUDA 11.8, which are increasingly out of sync with:
     • Ubuntu 24’s GCC toolchain, and
     • New RTX 50‑series GPUs.
   • A short “RTX 50‑series / sm_120” section describing the recommended PyTorch and CUDA versions (and whether tiny-cuda-nn is supported) would be extremely helpful.