Trying to increase speed by using CUDAExecutionprovider from onnxruntime-gpu instead of cpu onnxruntime, met with warning about cpu-gpu transfer bottleneck

[andrewtvuong]

I want to use CUDA instead of CPU to increase the speed on tag inference.

My machine Ubuntu 22.04.3 LTS (GNU/Linux 6.5.0-35-generic x86_64), CUDA 12.2

I learned from https://onnxruntime.ai/docs/install/ that if you have cuda 12 must install using pip install onnxruntime-gpu --extra-index-url https://aiinfra.pkgs.visualstudio.com/PublicPackages/_packaging/onnxruntime-cuda-12/pypi/simple/ as of time of writing, instead of simply pip install onnxruntime-gpu which is for cuda 11. This took me a while to figure out. Kept getting errors that didn't make sense:

[E:onnxruntime
, provider_bridge_ort.cc:1744 TryGetProviderInfo_CUDA] /onnxruntime_src/onnxruntime/core/session/provider_bridge_ort.cc:1426 onnxruntime::Provider& onnxruntime::ProviderLibrary::Get() [ONNXRuntimeError] : 1 : FAIL : Failed to load library libonnxruntime_providers_cuda.so with error: libcublasLt.so.11: cannot open shared object file: No such file or directory

[W:onnxruntime, onnxruntime_pybind_state.cc:870 CreateExecutionProviderInstance] Failed to create CUDAExecutionProvider. Please reference https://onnxruntime.ai/docs/execution-providers/CUDA-ExecutionProvider.html#requirementsto ensure all dependencies are met.

I had those objects. but after reading carefully and reinstalling based on the above for cuda 12 it worked. Using CUDAExecutionprovider instead of CPUExecutionprovider however did cause a new warning:

[W:onnxruntime:, transformer_memcpy.cc:74 ApplyImpl] 12 Memcpy nodes are added to the graph main_graph for CUDAExecutionProvider. It might have negative impact on performance (including unable to run CUDA graph). Set session_options.log_severity_level=1 to see the detail logs before this message.

Basically bottlenecked by CPU/GPU data transfer. Trying to figure out but have not been able to successfully.

[andrewtvuong]

Opened this pull to show what I tried #63

[stanleyftf1005]

any luck fixing? having the same issue. WD14 tagger been a pain to run

[andrewtvuong]

I just started looking into this yesterday, will try to fix it when I have more time. Just starting discussions here in case I miss something.

[KurtCocain]

was working fine and then it started doing this between 2 generation, no idea what i did. Anyone fixed it yet?

[sddiky]

When I have to switch to CUDAExecutionProvider (because other custom node requiring onnxruntime-gpu), the speed of WD14 tagging became very slow.
As long as I can't uninstall onnxruntime-gpu, I just modify ortProviders in pysssss.json from ["CUDAExecutionProvider","CPUExecutionProvider"] to ["CPUExecutionProvider","CUDAExecutionProvider"], and the problem solved.
Hope this helps.

[smallersoup]

When I have to switch to CUDAExecutionProvider (because other custom node requiring onnxruntime-gpu), the speed of WD14 tagging became very slow. As long as I can't uninstall onnxruntime-gpu, I just modify ortProviders in pysssss.json from ["CUDAExecutionProvider","CPUExecutionProvider"] to ["CPUExecutionProvider","CUDAExecutionProvider"], and the problem solved. Hope this helps.

I'm extremely grateful, the solution you provided has helped me a lot.

[ctrlz526]

May I ask if your problem has been solved? I have encountered this issue as well