Replies: 2 comments
-
|
I'm not sure I fully trust Dependencies when it shows the list of DLLs that ONNX Runtime CUDA depends on. I'm wondering if it's possible that some libraries I removed are actually required, but just not visible in the dependency list. For example, some CUDA or cuDNN libraries might be loaded dynamically at runtime, which is why they don't appear in the static dependency analysis. |
Beta Was this translation helpful? Give feedback.
-
|
I managed to check which libraries are loaded dynamically. Here’s the full list but I’m not sure how to verify it for certain. Is there anywhere on their website or in the documentation that lists the libraries requiring OnnxRuntime CUDA? |
Beta Was this translation helpful? Give feedback.
Uh oh!
There was an error while loading. Please reload this page.
-
Hi, I am trying to understand which CUDA libraries I actually need to ship with ONNX Runtime CUDA(version 1.22).
Right now, the full set of CUDA-related libraries I have is about 3 GB, and CuDNN itself adds another 1 GB. That is quite a lot, so I am wondering whether I really need to include the entire CUDA / cuDNN package, or only a smaller subset.
When I inspect ONNX Runtime CUDA with Dependency Walker, it does not show all CUDA-related dependencies. It only shows a few libraries. Because of that, I am not sure how complete or reliable that dependency list is for deployment purposes.
So my question is: how safe is it to ship only the libraries that appear as direct dependencies of ONNX Runtime CUDA? Or do I still need to include the full CUDA and cuDNN runtime package?
In other words, instead of shipping more than 4 GB of dependencies, I would like to include only the minimum required set for running ONNX Runtime CUDA, if that is supported and safe. Ideally, that would reduce the package size to something closer to 1–1.5 GB instead of 4+ GB.
List of all CUDA/CuDNN DLLs
Beta Was this translation helpful? Give feedback.
All reactions