[Bug] Studio setup messed up install of torchcodec on aarch64 #4446
Description
- Did you update? This was fresh install
- Local -- DGX Spark (aarch64)
- Number GPUs used: 1
- Which notebook? None-- Unsloth Studio
- Which Unsloth version, TRL version, transformers version, PyTorch version?
unsloth==2026.3.5,trl==0.24.0,transformers==5.3.0, andtorch==2.10.0+cu130 - Which trainer?
What I ran
uv pip install unsloth --torch-backend=auto
unsloth studio setupResult
╔══════════════════════════════════════╗
║ Unsloth Studio Setup Script ║
╚══════════════════════════════════════╝
✅ Node v24.14.0 and npm 11.11.1 already meet requirements. Skipping nvm install.
✅ Node v24.14.0 | npm 11.11.1
✅ Frontend built to frontend/dist
finished finding best python
✅ Using python3 (3.12.13) — compatible (3.11.x – 3.13.x)
[=======-------------] 4/11 extra codecs uv failed, falling back to pip...
Using Python 3.12.13 environment at: /home/bjester/.unsloth/studio/.venv
× No solution found when resolving dependencies:
╰─▶ Because all versions of torchcodec have no wheels with a matching platform tag (e.g., `manylinux_2_39_aarch64`) and you require torchcodec,
we can conclude that your requirements are unsatisfiable.
hint: Pre-releases are available for `torchcodec` in the requested range (e.g., 0.0.0.dev0), but pre-releases weren't enabled (try:
`--prerelease=allow`)
hint: Wheels are available for `torchcodec` (v0.10.0) on the following platforms: `manylinux_2_28_x86_64`, `macosx_11_0_arm64`, `win_amd64`
[====================] 11/11 finalizing
✅ Python dependencies installed
Pre-installing transformers 5.x for newer model support...
✅ Transformers 5.x pre-installed to /home/bjester/.unsloth/studio/.venv_t5/
Building llama-server for GGUF inference...
Building with CUDA support (nvcc: /usr/local/cuda/bin/nvcc)...
GPU compute capabilities: 121 -- limiting build to detected archs
✅ llama-server built at /home/bjester/.unsloth/llama.cpp/build/bin/llama-server
✅ llama-quantize available for GGUF export
╔══════════════════════════════════════╗
║ Setup Complete! ║
╠══════════════════════════════════════╣
║ Launch with: ║
║ ║
║ unsloth studio -H 0.0.0.0 -p 8000 ║
╚══════════════════════════════════════╝
I didn't think much about the torchcodec error since the whole process completed. Later on, when I went to fine tune a model, it threw an error about it couldn't import from torchcodec.decoders.
Training error: Custom format mapping failed: No module named 'torchcodec.decoders'
As this issue notes, the ARM (aarch64) wheels for torchcodec aren't available on PyPi, but are on the pytorch index.
How I fixed it
First, I stoped Unsloth Studio. Then I activated the virtualenv Unsloth Studio was using:
cd ~/.unsloth/studio
source .venv/bin/activate
I took a look at what was installed:
torchcodec==0.0.0.dev0
Seeing that the current version is 0.10, this seemed wrong. Within the virtualenv, I installed it manually:
pip install torchcodec --index-url=https://download.pytorch.org/whl/cu130
I restarted Unsloth Studio, and I didn't hit the error again.