Unsloth

[MickLesk]

Name of the Script

Unsloth

Script Type

CT (LXC Container)

Does this script support arm64?

arm64 not supported

📋 Script Details

Unsloth Studio — local web UI for running and fine-tuning LLMs. Search/download/serve GGUF, LoRA and safetensors models, side-by-side model comparison, export, and OpenAI-/Anthropic-compatible endpoints (/v1/chat/completions, /v1/responses, /v1/messages) so local models can back Claude Code or Codex.

This was initially assessed as "a Python library, not an app" and rejected — that was wrong. The repo description is now "Unsloth is a local UI for training and running…", studio/ contains a full backend/ + frontend/, and the README states Windows, Linux and macOS.

• No curl | sh: the vendor installer was read and its essentials reproduced — uv pip install --torch-backend=auto unsloth-zoo unsloth into a venv, launched via <venv>/bin/unsloth studio. Installer lines 4178/4487 for reference.
• -H 0.0.0.0 is passed explicitly; the default binds loopback only.
• UNSLOTH_STUDIO_PASSWORD is generated and written to the env file. Without it studio/backend/auth/terminal_prompt.py asks on the terminal, which a systemd unit cannot answer — the service would hang on first start.
• --torch-backend=auto lets uv pick CPU or CUDA wheels depending on what the container sees, so the same script works with and without passthrough. var_gpu="yes".
• Note: fine-tuning needs an NVIDIA GPU. On CPU-only, inference/model management/export/API work but training does not — documented as a warning.
• has_arm: false — the torch/CUDA stack makes arm64 untested at best.

Source: https://github.com/unslothai/unsloth

[flatlinebb]

It's not working for me:

root@unsloth-lxc:~# systemctl status unsloth.service
● unsloth.service - Unsloth Studio
     Loaded: loaded (/etc/systemd/system/unsloth.service; enabled; preset: enabled)
     Active: activating (auto-restart) (Result: exit-code) since Mon 2026-08-03 15:41:23 PDT; 4s ago
 Invocation: 134ea6e350ed435c81e11754bcc95872
    Process: 372 ExecStart=/opt/unsloth/.venv/bin/unsloth studio -H 0.0.0.0 -p 8888 (code=exited, status=1/FAILURE)
   Main PID: 372 (code=exited, status=1/FAILURE)
   Mem peak: 24.6M
        CPU: 693ms

Aug 03 15:41:23 unsloth-lxc systemd[1]: unsloth.service: Main process exited, code=exited, status=1/FAILURE
Aug 03 15:41:23 unsloth-lxc systemd[1]: unsloth.service: Failed with result 'exit-code'.
root@unsloth-lxc:~# /opt/unsloth/.venv/bin/unsloth studio -H 0.0.0.0 -p 8888
Unsloth Studio not set up. Run install.sh first.

Here is my system:

 ---------
 OS: Proxmox VE 9.2.6 x86_64
 Host: Precision T5610 (00)
 Kernel: Linux 7.0.12-1-pve
 Uptime: 33 days, 22 hours, 14 mins
 Shell: bash 5.2.37
 CPU: 2 x Intel(R) Xeon(R) E5-2660 0 (32) @ 3.00 GHz
 GPU: NVIDIA GeForce RTX 3060 Lite Hash Rate [Discrete]
 Memory: 66.57 GiB / 125.83 GiB (53%)
 Swap: 945.65 MiB / 980.00 MiB (96%)
 Disk (/): 316.75 GiB / 466.97 GiB (68%) - ext4
 Disk (/storage): 1.32 TiB / 3.58 TiB (37%) - ext4
 Locale: en_US.UTF-8

Let me know if you need any other info.

[MickLesk]

Its currently not for testing, we need to add the label if its Ready for Testing

[github-actions]

The Unsloth script is ready for testing:

bash -c "$(curl -fsSL https://github.com/community-scripts/ProxmoxVED/raw/main/ct/unsloth.sh)"

The generated UI password is stored as UNSLOTH_STUDIO_PASSWORD in /opt/unsloth.env. Without it Unsloth Studio would prompt for one on the terminal, which a systemd service cannot answer.
Fine-tuning requires an NVIDIA GPU. Without one, inference, model management, export and the OpenAI-compatible API work, but training does not. Pass a card through and install the NVIDIA driver in the container to enable it.
Models and Hugging Face caches land in /opt/unsloth_data, which fills up quickly - a single 7B model at Q4 is roughly 4 GB. Give the container more disk if you plan to keep several, and back that directory up rather than /opt/unsloth.
PyTorch is installed with uv's --torch-backend=auto, so a CPU-only container gets the CPU wheels and a GPU container the matching CUDA build. Re-run the update after adding a GPU so the right backend is pulled.

Note: This is not in the official repo yet—it's just a dev version! After merging into ProxmoxVE, it will need to be recreated.

Discord testing thread: https://discord.com/channels/1302816934508630047/1534101359845052477

[kintaro1981]

I just installed it.
I activated unsloth/gemma-4-E2B-it-qat-GGUF but if I try to ask for a web search doesn't work.
ping to google.com is working

[kintaro1981]

Another issue:
after rebooting the LXC unsloth is unreachable.

root@unsloth:~# systemctl status unsloth.service
● unsloth.service - Unsloth Studio
     Loaded: loaded (/etc/systemd/system/unsloth.service; enabled; preset: enabled)
     Active: activating (auto-restart) (Result: exit-code) since Thu 2026-08-06 19:33:45 CEST; 9s ago
 Invocation: f11010149240452089a3860e0b2d3ad0
    Process: 379 ExecStart=/opt/unsloth/unsloth_studio/bin/unsloth studio -H 0.0.0.0 -p 8888 (code=exited, status=1/FAILURE)
   Main PID: 379 (code=exited, status=1/FAILURE)
   Mem peak: 24.8M
        CPU: 251ms

Aug 06 19:40:38 unsloth systemd[1]: unsloth.service: Scheduled restart job, restart counter is at 57.
Aug 06 19:40:38 unsloth systemd[1]: Started unsloth.service - Unsloth Studio.
Aug 06 19:40:38 unsloth unsloth[479]: Error: an Unsloth admin password is already set; --password only sets the initial password. Change it in the UI, or run `unsloth studio reset-password` for a new one.
Aug 06 19:40:38 unsloth systemd[1]: unsloth.service: Main process exited, code=exited, status=1/FAILURE
Aug 06 19:40:38 unsloth systemd[1]: unsloth.service: Failed with result 'exit-code'

[kintaro1981]

The issue is the password in /opt/unsloth.env
AFAIK this env var is needed only during the installation.
I tried to comment the env variable UNSLOTH_STUDIO_PASSWORD and after systemctl start unsloth.service the service started without errors.

[MickLesk]

Its their official Installer 🤔

[MickLesk]

ive pushed now a little fix, but i dont have the time now to test it, feel free to redeploy the container and test it :-)

[kintaro1981]

ive pushed now a little fix, but i dont have the time now to test it, feel free to redeploy the container and test it :-)

I confirm that now after a reboot unsloth studio start correctly
right now after installing it, you need to use unsloth studio reset-password (no need to restart) because the password is unknown so a warning in the info page is needed.

I just installed it. I activated unsloth/gemma-4-E2B-it-qat-GGUF but if I try to ask for a web search doesn't work. ping to google.com is working

about web search is now working so it works with local models too.

[MickLesk]

Password is in /root/unsloth.creds

[kintaro1981]

So it's needed to update the script page with this new information.

[kintaro1981]

The tool calling issue from connected model has been addressed, see here:
unslothai/unsloth#7761 (comment)

But I haven’t tried yet.