New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Install error with both pip/conda (system MKL installed) #63
Comments
Could you use
That is, install
Are you able to find
If yes, please use
and then re-run
Please describe how you installed cudatoolkit and pytorch. It looks to me that it is using a version of PyTorch with cuda 10.2
Sorry for the installation issues. Please ask for help as early as possible. |
Using your install command now it is installed correctly and I can configure sherpa successfully. With regarding to your remaining comment, I've provided the steps that I used - almost is conda install commands except kaldifeat, mkl do exists... So the problem is really weird for me Anyway, thank you for your hard work in open source ASR frameworks. |
Hi, it's me again. Error message:
|
Are you using the latest sherpa, i.e., the master branch of sherpa? |
Yes,
|
An addition information: my computer has two version of gcc 12.2.0 (default) and 10.4.0 (custom installed). |
Here is reproducible Dockerfile
|
By the way, how did you install sherpa? Have you installed sherpa before? (If yes, have you uninstalled it? I suspect that the error is caused by a previous version of sherpa). |
I've provided Dockerfile that has same problem that I encountered even if it is clean install. |
Could you run
in your terminal and post the output? |
Here is the sherpa-version output from
Here is the error message
|
Are you able to run
|
My local (the failed one):
My server (succeed one)
|
That is strange. Did you use the same method to install |
I don't remember correctly, just using the history of the shell (~/.zsh_history). |
Could you try to also install kaldifeat from source, like what you are doing for sherpa? |
I've tried docker with Rocky Linux 8 (gcc 8.5.0), Fedora 35 (gcc 11.3.1), ubuntu 20.04 (gcc 9.4.0)
The Dockerfile using ubuntu:
Ubuntu 20.04 failed that is very weird since I'm using sherpa in Ubuntu 20.04 server (but with different installing procedure). I will retry more. |
Update 2: the following steps is work in my server:
Sample:
|
Update 3: if I build previous steps using docker, it complains about cuda:
Here is the content of Dockerfile that I used:
|
Using the procedure in #63 (comment) ( If building sherpa using gcc-10 (set priority via PATH) can run it without any further steps. The steps is basically same as what I've posted in original issue. Not sure what change I've made to my computer/conda now I even can not reproduce the issue @@~. -> Update: My local computer has oneAPI/MKL installed but not in library search path. If I move /opt/intel to /opt/intel.BAK I can install successfully as this comment, otherwise with oneAPI installed using intel tarball install process is failed as report in original issue. My server also has system mkl installed but it has library path entry for it (seem like I installed it using kaldi tools).
|
Thanks for your effort in reproducing. I am not sure what causes the problem. Sorry, I don't know how to fix it. It looks to me that if you use |
I've updated my last comment, it seem like there is a problem with linking system MKL that not in library search path |
Does it mean when you switch from GCC 11.2.0 to 12.2.0, it works without reinstalling sherpa and kaldifeat? |
No, Gcc version is just addition infomation. |
I think I found the cause:
So the problem seem cleared, mainly because my MKL installation is not follows "standard" as CMake and Pytorch expected (Even if it is installed from official intel tarball) |
Great to hear that you find out the reason. I am surprised that the link issues with MKL will lead to such an error when Would you mind creating a pull-request (both in kaldifeat and in sherpa) to fix it? Thanks ! |
I can not install kaldifeat alongside with k2. Steps:
conda create --name k2 -c k2-fsa -c pytorch -c nvidia k2 pytorch pytorch-cuda=11.7 torchaudio
=> OKconda install -c kaldifeat kaldifeat
=> Conflicting/ERROR: kaldifeat-cuda-install-error.logConda failed so I've tried to install via pip. Steps:
pip install --verbose kaldifeat
=> Failed with message saying that it can't find cuDNN: kaldifeat-pip-install-error.logconda install -c conda-forge cudnn
=> OK withcuDNN-conda-install-message
:pip install --verbose kaldifeat
: .......annd- it failed with message saying that can not find mkl: kaldifeat-pip-install-error-cudnn-installed.logpip install kaldifeat
can't find MKL in a conda env (fix: set LDFLAGS) #10, and it failed with problem of cuda: k2-pip-install-error-mkl-provided.logcuDNN-conda-install-message
:The text was updated successfully, but these errors were encountered: