Can I downgrade DGX SPARK to cuda 12?

Accelerated Computing DGX Spark / GB10 User Forum DGX Spark / GB10
1

My current software stack requires torch and torchaudio. CUDA13 requires torch v2.9. However, it keeps producing problematic audio when using torch v2.9, both with and without GPU enabled. When I switch to torch 2.8 CPU only, the generated audio sounds great. Torch 2.8 only supports CUDA 12, therefore, I couldn’t use the GPU on DGX SPARK if I have to use torch 2.8.

So I am wondering if SPARK can be downgrade to CUDA 12? I tried to uninstall CUDA13 and install CUDA 12, then Nvidia SMI no longer shows the GPU. When I tried to install the CUDA13 back, everything is broken.

2

consider using CUDA 12.x based containers from GPU-optimized AI, Machine Learning, & HPC Software | NVIDIA NGC

this should not effect your native host software stack, including the GPU driver.

1 Like
3

Thank you for the advice, I will try out the container approach.

4

Yes – I had to do container route to get indes2-tts running:

My Key Lessons Learned for IndexTTS on DGX Spark:

1. CUDA & PyTorch Compatibility

  • Use CUDA 12.9.0 with PyTorch 2.8.0 (test/cu129 channel)

  • Critical: Match exact CUDA version with PyTorch build

  • Command: pip3 install torch==2.8.0 torchaudio==2.8.0 --index-url https://download.pytorch.org/whl/test/cu129

2. Dependency Management

  • Install in correct order: System deps → PyTorch → Core ML → Audio processing → UI

  • Essential system packages: python3 python3-pip espeak-ng ffmpeg

  • Core ML stack: transformers==4.52.1 tokenizers==0.21.0 safetensors==0.5.2 accelerate==1.8.1

  • Audio processing: librosa==0.10.2.post1 soundfile numba==0.58.1

  • Web UI: gradio==5.45.0

1 Like
5

Hi Maccormackjim3, I am doing IntexTTS on DGX too. Your input is real valuable to me. Thank you for sharing.

1 Like
7

I finally got it working with the docker image. Have to spend quite a bit of time resolving the pynini package installation issue. Basically I have to install conda in the docker to get a working version of the pynini. And getting conda to activate the enviroment in docker is also not an easy task. Other than that, things are in order now and I can produce clear audio on SPARK.

Cheers

1 Like
8

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.