[Bug]: Windows NVIDIA build setup loop: Fails to install torch==2.6.0+cu121 and forces CPU rollback #644

Activity

1. imbalon

   added

   bugSomething isn't workingSomething isn't working

   on Sep 20, 2026
2. thcp

   self-assigned this

   on Sep 20, 2026
3. thcp

   added a commit that references this issue on Sep 21, 2026

   Install a CUDA wheel that exists on a pre-12.4 driver (#650)

   ...

   2687fde
4. thcp

   closed this as completedin #650on Sep 21, 2026
5. thcp commented on Sep 21, 2026

   

   thcp

   on Sep 21, 2026

   Collaborator

   More actions

   Thanks for the setup.log. That is what made this findable. The exact spec pip was asked for was sitting right there in the error.

   Here is what was going on. StemDeck picks a CUDA wheel index from what your driver reports, then installs a fixed torch line, 2.6.0, from it. Your driver reports CUDA below 12.4, so setup chose the cu121 index. cu121 never published torch 2.6.0. It stops at 2.5.1, and the same is true for torchaudio and torchvision. So pip was asked for a wheel that does not exist, the install failed, CPU torch went back in, and your 4060 sat there unused.

   Fixed in #650, merged as 2687fde. It ships in the next release.

   What changed:

   • Every CUDA 12 driver now gets cu124, not only 12.4 and newer. CUDA 12 is minor version compatible, so a cu124 build runs on your driver, and cu124 does publish the 2.6.0 line.
   • cu118 follows as a second attempt if cu124 installs but cannot launch a kernel. Before this, one failed tag meant CPU with nothing else tried.
   • A test now checks every wheel tag setup can offer against what that index actually publishes, so a pairing like this one fails on my machine instead of yours.

   If you want the GPU before the release, either of these works today.

   The simpler one: update your NVIDIA driver to a version reporting CUDA 12.4 or newer, then start StemDeck again. A CPU result that came from a failure is re-probed on the next launch, so setup will retry and pick cu124 on its own.

   If you would rather not touch the driver, install the working wheels into the build you already have. From the extracted StemDeck folder:

   python\Scripts\python.exe -m pip install torch==2.6.0+cu124 torchaudio==2.6.0+cu124 torchvision==0.21.0+cu124 --index-url https://download.pytorch.org/whl/cu124 --ignore-installed --no-deps
   

   Then start StemDeck again. Setup sees working CUDA torch and keeps it.

   One thing I cannot check from here is whether cu124 launches kernels on your exact driver, since that is the part relying on minor version compatibility. If it does not, the new fallback puts you on cu118 instead. Either way logs\setup.log names the tag that verified. I would be glad to know which one you land on, if you get a chance.