Skip to content

[Bug] RTX 5070 Ti (sm_120) not supported by default CUDA 12.4 PyTorch build #28

Description

@sjh00

Describe the bug
When running Inline Studio with an NVIDIA GeForce RTX 5070 Ti GPU, PyTorch throws a warning that the GPU's compute capability (sm_120) is not supported by the installed PyTorch version (which only supports up to sm_90). This prevents GPU acceleration and may cause the application to fall back to CPU, severely impacting performance.

To Reproduce
Steps to reproduce the behavior:

  1. Follow the Windows installation guide: run webui.bat --install (which installs PyTorch from the CUDA 12.4 index).
  2. Launch the application with webui.bat.
  3. Observe the warning in the console:
    UserWarning: NVIDIA GeForce RTX 5070 Ti 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_61 sm_70 sm_75 sm_80 sm_86 sm_90.
    

Expected behavior
The application should recognize the RTX 50-series GPU and use it for acceleration without compatibility warnings.

Screenshots / Logs
(Attach the full log or paste the warning message)

Environment

  • OS: Windows 11
  • GPU: NVIDIA GeForce RTX 5070 Ti (CUDA 13.2)
  • Python version: 3.11 (as specified in .python-version)
  • Inline Studio version: inline-core==1.2.63 (latest as of today)
  • PyTorch installed: from cu124 index

Additional context
The RTX 50 series GPUs are based on the Blackwell architecture and require PyTorch compiled with CUDA 12.5+ (or 13.x) to include sm_120 support. The current install script hardcodes the CUDA 12.4 index (https://download.pytorch.org/whl/cu124), which only provides wheels built for compute capabilities up to sm_90.

I attempted to manually modify webui.bat to use the cu132 index, but then encountered dependency conflicts (e.g., torchao and bitsandbytes did not have pre-built wheels for CUDA 13.2 on Windows). A more robust solution is needed.

Possible fix

  • Update the installation script to detect the GPU architecture and automatically select the appropriate CUDA version (e.g., use cu132 for RTX 50 series).
  • Alternatively, provide a command-line flag (e.g., --cuda-version) or environment variable (INLINE_CUDA_INDEX) to override the PyTorch index without editing the script.
  • Update the dependency constraints to ensure compatibility with newer PyTorch/CUDA builds (e.g., torchao>=0.14 may need a version that supports CUDA 13.2).

Metadata

Metadata

Assignees

Labels

bugSomething isn't working

Type

Projects

No projects

Milestone

No milestone

Relationships

None yet

Development

No branches or pull requests

Issue actions