Hi Team,
I was following the playbook (LLaMA Factory | DGX Spark) to fine-tune the LLM with LLama factory in my dgx spark. The docker environment seems to be conflicting with the LLama installation (in step 4) and here is the error message:
and apparently, the CUDA is not activated.
Should I install all dependent packages manually? Or it is an internal bug…
Hi!
I’m having the same issue also I’m unable to use the Llama Factory WebUI interface. Is there something I’m missing?
Thank you very much.
Hey folks,
I have moved this post over to the DGX Spark section of the forums to get some better visibility on the issue.
Best,
Aharpster
Your torch version is incorrect. The default version in Docker should be 2.10.0a0+b558c986e8.nv25.11, but when I pull the latest version of Llamafactory from Git, I need torchaudio, and the container doesn’t have the corresponding version, causing an error. I don’t know how to solve this yet.
Just tried this and succesfully installed the packages by following the playbook instructions. Please make sure you are using this pytorch image: nvcr.io/nvidia/pytorch:25.11-py3
and you installed the torchaudio package without dependencies that conflicts with NVIDIA’s PyTorch build
sed -i 's/"torchaudio[^"]*",\?//' pyproject.toml
pip install -e ".[metrics]"
pip install --no-deps torchaudio