Hello,
has anyone encountered Flux RNNs invoking cudnn code, although none of the models from Flux.CUDA
are used? My model is built on top of Flux.LSTM
(not CuLSTM
), but I get the following error:
ERROR: LoadError: CUDNNError(code 3, CUDNN_STATUS_BAD_PARAM)
Stacktrace:
[1] macro expansion at /data/.julia/packages/CuArrays/PwSdF/src/dnn/error.jl:19 [inlined]
[2] cudnnRNNBackwardData(::Flux.CUDA.RNNDesc{Float32}, ::Int64, ::Array{CuArrays.CUDNN.TensorDesc,1}, ::CuArr
ays.CuArray{Float32,2}, ::Array{CuArrays.CUDNN.TensorDesc,1}, ::CuArrays.CuArray{Float32,2}, ::CuArrays.CUDNN.TensorDesc, ::CuArrays.CuArray{Float32,2}, ::CuArrays.CUDNN.TensorDesc, ::CuArrays.CuArray{Float32,2}, ::CuArrays.CUDNN.FilterDesc, ::CuArrays.CuArray{Float32,1}, ::CuArrays.CUDNN.TensorDesc, ::CuArrays.CuArray{Float32,2}, ::CuArrays.CUDNN.TensorDesc, ::CuArrays.CuArray{Float32,2}, ::Array{CuArrays.CUDNN.TensorDesc,1}, ::CuArrays.CuArray{Float32,2}, ::CuArrays.CUDNN.TensorDesc, ::CuArrays.CuArray{Float32,2}, ::CuArrays.CUDNN.TensorDesc, ::CuArrays.CuArray{Float32,2}, ::CuArrays.CuArray{UInt8,1}, ::CuArrays.CuArray{UInt8,1}) at /data/.julia/packages/Flux/qXNjB/src/cuda/curnn.jl:170
...
The error itself might be due to driver version mismatch, but I was not expecting cudnn
to be invoked at all. Am I missing something?
Thanks in advance.