Anyone wanting to run UnSloth Studio on you Spark you have likely run into the issue with their script claiming that there isn’t an arm64 version of torchcodec @v0.10.0.
No idea why that error is happening or why the Unsloth crew claim incompatibility, but if you let the script run you can then manually install torchcodec from http://download.pytorch.org/whl/cu130/torchcodec-0.10.0%2Bcu130-cp313-cp313-manylinux_2_28_aarch64.whl
I got it up and running without further issue.
thats my guide
Installing Unsloth Studio on DGX Spark
A clean install guide that avoids the aarch64 dependency cascade.
Prerequisites
Verify your environment:
bashnvidia-smi # confirms GB10 visible
nvcc --version # CUDA 13.0
python3 --version # 3.11.x – 3.13.x
Install
Step 1 — Run the official installer:
curl -fsSL https://unsloth.ai/install.sh | sh
It will fail at “extra codecs” with a torchcodec==0.10.0 error. This is expected on aarch64 — keep going.
Step 2 — Activate the venv:
source ~/.unsloth/studio/unsloth_studio/bin/activate
Step 3 — Install the runtime deps that got skipped:
pip install structlog diceware fastapi aiofiles python-jose passlib bcrypt argon2-cffi
Step 4 — (Optional) Install aarch64 torchcodec for audio/TTS support:
pip install torchcodec --index-url=https://download.pytorch.org/whl/cu130
Skip this if you only need LLM training/chat.
Launch
unsloth studio -H 0.0.0.0 -p 8888
Open http://localhost:8888 in your browser. Set a password on first load. Studio should detect your GB10 with CUDA. Why this works
The official installer pins torchcodec==0.10.0, which has no aarch64 wheel on PyPI. When that pin fails, pip rolls back the entire extras batch, leaving Studio’s runtime deps (structlog, diceware, etc.) uninstalled. Step 3 installs them directly; step 4 grabs torchcodec from PyTorch’s own index where the ARM wheel exists.
Next launches
After the first install, you only need:
source ~/.unsloth/studio/unsloth_studio/bin/activate
unsloth studio -H 0.0.0.0 -p 8888
Or skip the activate line if you symlink unsloth into your PATH.
Hi @zezo_khaled, is this still valid?
You had a guide somewhere?
I did a Unsloth Studio installation today (6/5/26) and the script run without any errors:
curl -fsSL https://unsloth.ai/install.sh | sh
and this is full installation log:
elsaco@spark1:~$ curl -fsSL https://unsloth.ai/install.sh | sh
🦥 Unsloth Studio Installer
────────────────────────────────────────────────────
platform linux
deps all system dependencies found
venv creating Python 3.13 virtual environment
/home/elsaco/.unsloth/studio/unsloth_studio
venv using environment
/home/elsaco/.unsloth/studio/unsloth_studio
gpu NVIDIA GPU detected
installing PyTorch (https://download.pytorch.org/whl/cu130)...
installing unsloth (this may take a few minutes)...
setup running unsloth studio update...
🦥 Unsloth Studio Setup
────────────────────────────────────────────────────
frontend up to date
deps [==========----------] 7/13 flash-attn warning Skipping flash-attn: Blackwell GPU detected (sm_100+); no compatible prebuilt wheel
deps [====================] 13/13 ROCm torch (final)
deps installed
transformers 5.3.0 pre-installed
transformers 5.5.0 pre-installed
gpu NVIDIA GPU detected
installing prebuilt llama.cpp...
llama.cpp prebuilt install failed (continuing)
llama.cpp diagnostics (last 120 lines):
| [llama-prebuilt] no existing llama.cpp install detected at /home/elsaco/.unsloth/llama.cpp; performing fresh prebuilt install
| [llama-prebuilt] prebuilt install path failed; falling back to source build
| [llama-prebuilt] prebuilt fallback reason: unslothai/llama.cpp ships only linux-x64 prebuilts; aarch64 Linux falls back to source build
| platform=Linux machine=aarch64
| driver_cuda_version=(13, 0)
| compute_caps=121
| cuda_visible_devices=unset
| has_physical_nvidia=True
| has_usable_nvidia=True
| chosen_asset=none
| asset_source=none
| linux_runtime_lines=cuda13
| linux_runtime_dirs_cuda13=/usr/local/cuda-13.0/targets/sbsa-linux/lib,/home/elsaco/.unsloth/studio/unsloth_studio/lib/python3.13/site-packages/nvidia/cu13/lib
| linux_runtime_dirs_cuda12=none
| nvidia-smi:
| Fri Jun 5 13:53:12 2026
| +-----------------------------------------------------------------------------------------+
| | NVIDIA-SMI 580.159.03 Driver Version: 580.159.03 CUDA Version: 13.0 |
| +-----------------------------------------+------------------------+----------------------+
| | GPU Name Persistence-M | Bus-Id Disp.A | Volatile Uncorr. ECC |
| | Fan Temp Perf Pwr:Usage/Cap | Memory-Usage | GPU-Util Compute M. |
| | | | MIG M. |
| |=========================================+========================+======================|
| | 0 NVIDIA GB10 On | 0000000F:01:00.0 On | N/A |
| | N/A 38C P0 11W / N/A | Not Supported | 0% Default |
| | | | N/A |
| +-----------------------------------------+------------------------+----------------------+
|
| +-----------------------------------------------------------------------------------------+
| | Processes: |
| | GPU GI CI PID Type Process name GPU Memory |
| | ID ID Usage |
| |=========================================================================================|
| | 0 N/A N/A 2555 G /usr/lib/xorg/Xorg 559MiB |
| | 0 N/A N/A 2852 G /usr/bin/gnome-shell 371MiB |
falling back to source build
building (CUDA, sm_121)...
llama.cpp built
llama-quantize built
────────────────────────────────────────────────────
Unsloth Studio Installed
────────────────────────────────────────────────────
launch unsloth studio -p 8888
(add -H 0.0.0.0 to allow network / cloud access)
Created Unsloth Studio shortcut
Unsloth Studio installed!
────────────────────────────────────────────────────
Start Unsloth Studio now? [Y/n]
launch starting Unsloth Studio...
Starting Unsloth Studio on http://127.0.0.1:8888
[OK] Frontend loaded from /home/elsaco/.unsloth/studio/unsloth_studio/lib/python3.13/site-packages/studio/frontend/dist
INFO: Started server process [102822]
INFO: Waiting for application startup.
Hardware detected: CUDA -- NVIDIA GB10
{"timestamp": "2026-06-05T20:56:48.930525Z", "level": "warning", "event": "llama.cpp prebuilt lacks MTP support (--spec-type mtp/draft-mtp). Run `unsloth studio update`. MTP GGUFs will load without speculative decoding."}
WARNING: llama.cpp prebuilt lacks MTP support (--spec-type mtp/draft-mtp). Run `unsloth studio update`. MTP GGUFs will load without speculative decoding.
============================================================
DEFAULT ADMIN ACCOUNT CREATED
username: unsloth
password saved to: /home/elsaco/.unsloth/studio/auth/.bootstrap_password
Open the Studio UI to sign in and change it.
============================================================
INFO: Application startup complete.
INFO: Unsloth Studio running on http://127.0.0.1:8888 (To stop: press Ctrl+C -- on macOS, Control+C not Command+C)
🦥 Unsloth Studio is running
────────────────────────────────────────────────────
On this machine -- open this in your browser:
http://127.0.0.1:8888
(same as http://localhost:8888)
API & health:
http://127.0.0.1:8888/api
http://127.0.0.1:8888/api/health
────────────────────────────────────────────────────
Tip: if you are on this computer, open http://localhost:8888/ in your browser.
Studio is only reachable on this machine (bound to 127.0.0.1).
To deploy and access globally:
1. press Ctrl+C to stop Studio
2. relaunch with: unsloth studio -H 0.0.0.0 -p 8888
Only do this on trusted networks -- it exposes the API on every interface.
To stop Unsloth Studio: press Ctrl+C in this terminal.
(On macOS this is Control+C, not Command+C.)
────────────────────────────────────────────────────