Repository navigation
How do I install the CUDA thingie? #34225
Replies: 10 comments 6 replies
|
TheBigBobo You were right about the archive -- it is the 1. Check the GPU/driver first. Run 2. Install the GPU package into the same Python environment the auto-tagger runs in. Remove the CPU package first -- the CPU and GPU distributions both provide the 3. Get the CUDA/cuDNN libraries. The easiest way is to let pip bring them: That installs the CUDA and cuDNN runtime libraries alongside the package, so no separate CUDA toolkit install is needed. One thing to know: since ORT 1.27 the default 4. Verify it worked:
If the tagger has a provider/device setting, set it to CUDA there. About the archive: you only need the Two caveats from the repo's own validation doc: the CUDA 13 Linux packages ship SASS only (no PTX), so a GPU whose compute capability isn't covered fails with |
|
TheBigBobo Fair question, step 2 assumed too much. A Python environment is just one Python install plus the packages installed for it. When you run To check which one you are in, run this in the same terminal you launch the auto-tagger from: That prints the Python the auto-tagger will use. Then If you never set up an environment yourself, you are just using your system Python, and that is fine for this. Run the pip commands from the same terminal you run the auto-tagger in and you are in the right place. If you want a clean separate one later, |
|
TheBigBobo Sorry, my last two replies were wrong for your setup. I assumed a Python install with pip, but monbooru's portable bundle doesn't work that way. It uses ONNX Runtime as a library file next to the program, not a Python package. Let me redo this properly, from the monbooru docs (https://monbooru.github.io/mondocs/advanced/auto-tagger.html). The easy path first. If you are on Windows, you don't need to install anything for GPU tagging. The Windows download already includes a runtime built with DirectML. Just go to Settings -> Auto-Tagger in monbooru and set the provider to If you specifically want CUDA instead, here is the actual procedure from the docs:
One correction to my earlier reply: the pip commands ( |
|
TheBigBobo Good follow-up, I should have said where. The CUDA libraries come from NVIDIA, in two downloads:
Once both are in place, set the provider to And again, if all of this feels like too much, the |
|
TheBigBobo You are right to flag that, my last reply skipped a step. The file you want is the zip archive, named something like If your download has no Inside One shortcut: if you ran the |
|
I think I've got something working. It now no longer fails when I try to save cuda as the tagger in the settings. Instead it just skips every single image I try to auto tag. |
|
TheBigBobo That changes two things, and my last reply was written for Windows, so let me redo it for Linux. First, good news: the Second, the cuDNN instructions change. On Linux there is no
If you already installed the CUDA Toolkit on your system, the same files can go into Sorry for the Windows detour. The |
|
Following your earlier advice, my terminal shows that my computer is running CUDA 13.5. Do I need the CUDA 13 archive from https://github.com/microsoft/onnxruntime/releases#release-v1.32.0 instead of the CUDA 12 archive? |
|
TheBigBobo Yes -- take the CUDA 13 archive: Two things that make this the answer: 1. That "CUDA 13.5" is your driver, not your toolkit. The CUDA version 2. CUDA major versions don't mix. The CUDA 12 archive is built against CUDA 12.8 and needs the CUDA 12.8 (or newer 12.x) runtime libraries on the machine. A CUDA 13.5 setup doesn't provide CUDA 12 libraries, so the CUDA 12 archive can't load its CUDA code. Same the other way: the CUDA 13 archive needs the CUDA 13 runtime libraries. One correction to my earlier replies: the cuDNN has to match the CUDA major version too. I told you cuDNN 9 for CUDA 12 -- that was for the CUDA 12 archive. For the CUDA 13 archive, get the cuDNN 9 package for CUDA 13 from the cuDNN downloads page (Linux Before you swap the archive, one quick check that the CUDA 13 runtime is actually installed and not just driver-supported: If that shows On the skipping-every-image part from your last message: that is a separate symptom from the archive question. The provider now saves as |
|
TheBigBobo That That should show And pacman makes the cuDNN part much simpler than my earlier replies made it sound. You don't need NVIDIA's downloads page at all for cuDNN: That installs the cuDNN 9 libraries into the system library path, so the auto-tagger can find them without you copying anything next to the program. The So the full picture for your machine is now just this:
One more thing: the skipping-every-image symptom you mentioned earlier was most likely the old CUDA 12 archive failing against your CUDA 13 runtime. Once the CUDA 13 archive is in place, that should go away. If it still skips images after this, the next thing to look at is whatever monbooru prints in the terminal when it tries to tag, that usually says why the model failed to load. |
Uh oh!
There was an error while loading. Please reload this page.
Stupid title because I don't want technical guides, I want usable guides. It's not obvious to a lay person how to enable anything other than CPU auto tagging from https://monbooru.github.io/mondocs/advanced/auto-tagger.html
Can we have a step by step guide on what to put where and how to get the CUDA stuff?
ChatGPT indicated that I want the linux_x64 archive and not the linux_aarch_x64 archive, but I'm completely lost from there.
All reactions