Ubuntu 22.04 pip install kaldifeat: cannot find CUDNN

[ravi-mr]

@csukuangfj

Despite any flags, compiling kaldifeat could not find CUDNN,

` -- Caffe2: CUDA detected: 11.7

-- Caffe2: CUDA nvcc is: /usr/local/cuda-11.7/bin/nvcc

-- Caffe2: CUDA toolkit directory: /usr/local/cuda-11.7

-- Caffe2: Header version is: 11.7

-- /usr/local/cuda-11.7/lib64/libnvrtc.so shorthash is d833c4f3

-- USE_CUDNN is set to 0. Compiling without cuDNN support ` Following https://github.com/csukuangfj/kaldifeat/issues/12, I tried passing in

export KALDIFEAT_CMAKE_ARGS="-DCMAKE_BUILD_TYPE=Release -DCUDNN_LIBRARY_PATH=/usr/local/cuda-11.7/lib64/libcudnn.so -DCUDNN_INCLUDE_PATH=/usr/local/cuda-11.7/include -DUSE_CUDNN=1"

what is the impact of not having kaldifeat compiled with cudnn? any other suggestion for compiling kaldifeat with cudnn?

[csukuangfj]

There are two fixes:

1. If you only want a CPU version of kaldifeat, please install a CPU version of PyTorch, which does not need cudnn at all.

2. Please follow

https://k2-fsa.github.io/k2/installation/cuda-cudnn.html

to install cudnn.

[ravi-mr]

@csukuangfj adapted the same as given at https://k2-fsa.github.io/k2/installation/cuda-cudnn.html#cuda-11-7,

-- /usr/local/cuda-11.7.1/lib64/libnvrtc.so shorthash is 581f1f99
-- USE_CUDNN is set to 0. Compiling without cuDNN support
-- Autodetected CUDA architecture(s):  8.6
-- Added CUDA NVCC flags for: -gencode;arch=compute_86,code=sm_86
CMake Warning at /venv/lib/python3.8/site-packages/torch/share/cmake/Torch/TorchConfig.cmake:22 (message):
  static library kineto_LIBRARY-NOTFOUND not found.
Call Stack (most recent call first):
  /venv/lib/python3.8/site-packages/torch/share/cmake/Torch/TorchConfig.cmake:127 (append_torchlib_if_found)
  cmake/torch.cmake:14 (find_package)
  CMakeLists.txt:59 (include)

-- Found Torch: /venv/lib/python3.8/site-packages/torch/lib/libtorch.so  
-- PyTorch version: 2.0.0+cu117

still USE_CUDNN is not getting set to 1?

[csukuangfj]

That's a warning, not an error.

Is there still anything stopping you from installing kaldifeat?

[ravi-mr]

I could install kaldifeat. Was thinking if USE_CUDNN sets to 1, kaldifeat extraction can be faster in respective contexts. Does linking kaldifeat with cuda_dnn increase any speed?

[csukuangfj]

I could install kaldifeat.

Was thinking if USE_CUDNN sets to 1, kaldifeat extraction can be faster in respective contexts.

Does linking kaldifeat with cuda_dnn increase any speed?

USE_CUDNN exists only in torch 2.0.0

We don't have much experience about it now.

Maybe you could try torch < 2.0.0

[ravi-mr]

True.