Dear Community!
Is there a specific version of cudnn that is compiled for cuda/cudatoolkit 11.4?
During the execution of a tutorial skript that teaches tensorflow I get an exception hinting that the cudnn version installed is incompatible:
2025-01-30 11:53:20.995020: E tensorflow[/compiler/xla/stream_executor/cuda/cuda_dnn.cc:437](http://localhost:8888/compiler/xla/stream_executor/cuda/cuda_dnn.cc#line=436)] Could not create cudnn handle: CUDNN_STATUS_ARCH_MISMATCH 2025-01-30 11:53:20.995524: E tensorflow[/compiler/xla/stream_executor/cuda/cuda_dnn.cc:441](http://localhost:8888/compiler/xla/stream_executor/cuda/cuda_dnn.cc#line=440)] Memory usage: 454033408 bytes free, 4238606336 bytes total. 2025-01-30 11:53:20.997039: W tensorflow[/core/framework/op_kernel.cc:1828](http://localhost:8888/core/framework/op_kernel.cc#line=1827)] OP_REQUIRES failed at xla_ops.cc:503 : FAILED_PRECONDITION: DNN library initialization failed. Look at the errors above for more details. 2025-01-30 11:53:21.019085: E tensorflow[/compiler/xla/stream_executor/cuda/cuda_dnn.cc:437](http://localhost:8888/compiler/xla/stream_executor/cuda/cuda_dnn.cc#line=436)] Could not create cudnn handle: CUDNN_STATUS_ARCH_MISMATCH 2025-01-30 11:53:21.019174: E tensorflow[/compiler/xla/stream_executor/cuda/cuda_dnn.cc:441](http://localhost:8888/compiler/xla/stream_executor/cuda/cuda_dnn.cc#line=440)] Memory usage: 454033408 bytes free, 4238606336 bytes total. 2025-01-30 11:53:21.020953: W tensorflow[/core/framework/op_kernel.cc:1828](http://localhost:8888/core/framework/op_kernel.cc#line=1827)] OP_REQUIRES failed at xla_ops.cc:503 : FAILED_PRECONDITION: DNN library initialization failed. Look at the errors above for more details.
Now from the cudnn’s documentation it seems that cudnn needs to be compiled/built against/for a specific version for the cuda/cudatoolkit and that this is not just a question of the version of the cudnn used. I am bound to cuda 11.4 as the nvidia driver only supports cuda up to that version and I cannot upgrade the drive as my graphics board is not supported by the latest versions of the driver.
As a consequence I need to find a cudnn version for cuda 11.4 … now the cudnn packages in the anaconda repo do not mention the cuda version they work with.
So my question is: How to find a cudnn package working with cuda 11.4? Any helpful piece of information is highly appreciated.
Cheers