PyTorch problem

Fri Sep 7 17:12:48 EDT 2018

Thank you very much Emre! That’s so clever, and (almost) completely resolves my issue. I only have a small error when using h5py:
  File "h5py/_objects.pyx", line 54, in h5py._objects.with_phil.wrapper
  File "h5py/_objects.pyx", line 55, in h5py._objects.with_phil.wrapper
  File "h5py/h5f.pyx", line 78, in h5py.h5f.open
OSError: Unable to open file (unable to lock file, errno = 5, error message = 'Input/output error')

I resolved the problem by moving the file to scratch space. I think the new disk possibly have some small problems (either performance/permission), and that’s causing the problem.

Thanks,
Yichong

On Sep 6, 2018, at 8:46 PM, Emre Yolcu <eyolcu at cs.cmu.edu<mailto:eyolcu at cs.cmu.edu>> wrote:

I think I got it. If I'm not mistaken NFS is the root of all our problems in this thread. Can anyone having problems try doing the equivalent of `export CUDA_CACHE_PATH=/home/scratch/eyolcu/computecache` (replacing eyolcu with your andrew id) and try everything again? This seems to fix it for me.

On Thu, Sep 6, 2018 at 3:28 PM, Yichong Xu <yichongx at cs.cmu.edu<mailto:yichongx at cs.cmu.edu>> wrote:
Hi Predrag,
I just tested the simpleCUBLAS sample in cuda library. It still does not work for me with the same error:
GPU Device 0: "TITAN Xp" with compute capability 6.1

simpleCUBLAS test running..
!!!! CUBLAS initialization error

I’m not sure where exactly the access problem is, but here is what I get from ls -all:
yichongx at gpu8$ ls -all
total 2136200
drwxr-xr-x. 3 sheath sheath      8192 May 31 15:16 .
drwxr-xr-x. 4 root   root          32 Sep  2  2017 ..
lrwxrwxrwx. 1 root   root          18 Mar 13 13:05 libaccinj64.so -> libaccinj64.so.9.0
lrwxrwxrwx. 1 root   root          22 Mar 13 13:05 libaccinj64.so.9.0 -> libaccinj64.so.9.0.176
-rwxr-xr-x. 1 root   root     6858944 Sep  2  2017 libaccinj64.so.9.0.176
-rw-r--r--. 1 root   root    71952010 Dec 19  2017 libcublas_device.a
lrwxrwxrwx. 1 root   root          16 Mar 13 13:04 libcublas.so -> libcublas.so.9.0
lrwxrwxrwx. 1 root   root          20 Mar 13 13:04 libcublas.so.9.0 -> libcublas.so.9.0.282
-rwxr-xr-x. 1 root   root    52590576 Dec 19  2017 libcublas.so.9.0.176
-rwxr-xr-x. 1 root   root    55781312 Dec 19  2017 libcublas.so.9.0.282
-rw-r--r--. 1 root   root    62813620 Dec 19  2017 libcublas_static.a

Thanks,
Yichong

On Sep 6, 2018, at 3:14 PM, Yichong Xu <yichongx at cs.cmu.edu<mailto:yichongx at cs.cmu.edu>> wrote:

1. I think yes. Biswajit and I cannot use the system cuda libraries.
2. I think yes as well. Predrag said he can run matlab with cuda well (probably with root access), so I think there should be some problem with the privilege setting of system libraries. We do not have root access on our accounts.
3. Not yet so far.
4. That can be a solution. Maybe we have a public access library as Jay-Yoon did and that can work for us.

Also for gpu8 - I just reinstalled pytorch again on scratch of gpu8 and it still does not work. I’m making the cuda libraries right now and trying to see if it works.

Thanks,
Yichong

On Sep 6, 2018, at 11:20 AM, Barnabas Poczos <bapoczos at cs.cmu.edu<mailto:bapoczos at cs.cmu.edu>> wrote:

Hi All,

I'm somewhat confused:

* Do I understand correctly that Manzil actually is using the CUDA
libraries installed by himself
(/zfsauton/home/manzilz/local/cuda-9.0/) and not the system libraries
(/usr/local/cuda/lib64) ?
* Since he is using different CUDA libraries is that the reason that
pytorch is working for him and not for the other users? If so, should
we double check the system libraries?
* Do we know anyone who can use pytorch now with the CUDA system
libraries ? If so, those users please let us know your system env
variables.
* As a quick solution, should we ask Manzil to copy his cuda libraries
to a public place where others could access them?

Best,
Barnabas

======================
Barnabas Poczos, PhD
Associate Professor
Co-Director of PhD Program
Machine Learning Department
Carnegie Mellon University
On Wed, Sep 5, 2018 at 5:33 PM Manzil Zaheer <manzil at cmu.edu<mailto:manzil at cmu.edu>> wrote:

Here is my related env variables:

CUDA_HOME=/zfsauton/home/manzilz/local/cuda-9.0/

LD_LIBRARY_PATH=/zfsauton/home/manzilz/local/lib64:/zfsauton/home/manzilz/local/lib:/zfsauton/home/manzilz/local/cuda-9.0/lib64:/usr/local/cuda/lib64:

PATH=/zfsauton/home/manzilz/local/bin:/zfsauton/home/manzilz/.local/bin:/zfsauton/home/manzilz/local/cuda-9.0/bin:/usr/local/cuda/bin:/usr/lib64/qt-3.3/bin:/usr/local/bin:/usr/bin:/usr/local/sbin:/usr/sbin

C_INCLUDE_PATH=/zfsauton/home/manzilz/local/include:

From: Biswajit Paria <bparia at cs.cmu.edu<mailto:bparia at cs.cmu.edu>>
Sent: Wednesday, September 05, 2018 5:29 PM
To: Yichong Xu <yichongx at cs.cmu.edu<mailto:yichongx at cs.cmu.edu>>
Cc: Biswajit Paria <bparia at cs.cmu.edu<mailto:bparia at cs.cmu.edu>>; eyolcu at cs.cmu.edu<mailto:eyolcu at cs.cmu.edu>; Predrag Punosevac <predragp at andrew.cmu.edu<mailto:predragp at andrew.cmu.edu>>; Manzil Zaheer <manzil at cmu.edu<mailto:manzil at cmu.edu>>; users at autonlab.org<mailto:users at autonlab.org>
Subject: Re: PyTorch problem

If the CUDA examples work for anyone, can they share their PATH and LD_LIBRARY_PATH variables?

Thanks

On Wed, Sep 5, 2018 at 5:27 PM Yichong Xu <yichongx at cs.cmu.edu<mailto:yichongx at cs.cmu.edu>> wrote:

I think with Biswajit’s and my problem with cuda, we should isolate the problem with just CUDA (and drivers) instead of wandering around python or pytorch.

Predrag can you test the CUDA examples? I sort of agree with Manzil that this might be a user account problem.

Thanks,

Yichong

On Sep 5, 2018, at 5:14 PM, Biswajit Paria <bparia at cs.cmu.edu<mailto:bparia at cs.cmu.edu>> wrote:

I just tried Yichong's way of testing cuBLAS, and get the same error as earlier:

[Matrix Multiply CUBLAS] - Starting...

GPU Device 0: "TITAN Xp" with compute capability 6.1

MatrixA(640,480), MatrixB(480,320), MatrixC(640,320)

CUDA error at matrixMulCUBLAS.cpp:275 code=1(CUBLAS_STATUS_NOT_INITIALIZED) "cublasCreate(&handle)"

So I believe it is not a conda error. I also tried removing .nv, doesn't help either. Maybe someone can share the PATH env variable?

On Wed, Sep 5, 2018 at 5:08 PM Emre Yolcu <eyolcu at cs.cmu.edu<mailto:eyolcu at cs.cmu.edu>> wrote:

Manzil, could you share your `conda env export` (or equivalent) output for the environment you use for pytorch? It's still not working for me after reboot, maybe I can try replicating your exact setup and try with that.

Thanks,

Emre

On Wed, Sep 5, 2018 at 4:56 PM, Predrag Punosevac <predragp at andrew.cmu.edu<mailto:predragp at andrew.cmu.edu>> wrote:

Manzil Zaheer <manzil at cmu.edu<mailto:manzil at cmu.edu>> wrote:

It was working me before reboot as well. PyTorch does work on all
nodes for me.

Aha! Gotcha.

I am trying to say is that i think it is not issue at system level but
at user account level. I might be wrong though.

That was my hunch as well. They were trying to convince me in a 150
e-mails chain over the weekend that pytorch was broken when I replaced a
failed HDD on the main file server. That didn't make any sense.

Could you please share your binaries and setup with outher pytorch
users?

Cheers,
Predrag

-------- Original message --------
From: Predrag Punosevac <predragp at andrew.cmu.edu<mailto:predragp at andrew.cmu.edu>>
Date: 9/5/18 4:44 PM (GMT-05:00)
To: Manzil Zaheer <manzil at cmu.edu<mailto:manzil at cmu.edu>>
Cc: Biswajit Paria <bparia at cs.cmu.edu<mailto:bparia at cs.cmu.edu>>, Yichong Xu <yichongx at cs.cmu.edu<mailto:yichongx at cs.cmu.edu>>, Emre Yolcu <eyolcu at cs.cmu.edu<mailto:eyolcu at cs.cmu.edu>>, users at autonlab.org<mailto:users at autonlab.org>
Subject: Re: PyTorch problem

Should I go ahead and reboot all GPU computing nodes? Can somebody else confirm that a reboot fixes the issue?

Predrag

On Wed, Sep 5, 2018 at 4:42 PM, Manzil Zaheer <manzil at cmu.edu<mailto:manzil at cmu.edu><mailto:manzil at cmu.edu>> wrote:
It does work for me and my friends

-------- Original message --------
From: Predrag Punosevac <predragp at andrew.cmu.edu<mailto:predragp at andrew.cmu.edu><mailto:predragp at andrew.cmu.edu>>
Date: 9/5/18 4:40 PM (GMT-05:00)
To: Biswajit Paria <bparia at cs.cmu.edu<mailto:bparia at cs.cmu.edu><mailto:bparia at cs.cmu.edu>>
Cc: Manzil Zaheer <manzil at cmu.edu<mailto:manzil at cmu.edu><mailto:manzil at cmu.edu>>, Yichong Xu <yichongx at cs.cmu.edu<mailto:yichongx at cs.cmu.edu><mailto:yichongx at cs.cmu.edu>>, Emre Yolcu <eyolcu at cs.cmu.edu<mailto:eyolcu at cs.cmu.edu><mailto:eyolcu at cs.cmu.edu>>, users at autonlab.org<mailto:users at autonlab.org><mailto:users at autonlab.org>
Subject: Re: PyTorch problem

I just rebooted GPU8. All packages are up to date. NVidia driver appears to be working properly and I can do GPU computations from MATLAB. Let's try now to get pytorch working on GPU8.

Predrag

On Wed, Sep 5, 2018 at 12:19 AM, Biswajit Paria <bparia at cs.cmu.edu<mailto:bparia at cs.cmu.edu><mailto:bparia at cs.cmu.edu>> wrote:
I am facing a similar error on all GPU machines. Did someone find a solution yet?

2018-09-05 00:27:41.546064: E tensorflow/stream_executor/cuda/cuda_blas.cc:459<http://cuda_blas.cc:459/>] failed to create cublas handle: CUBLAS_STATUS_NOT_INITIALIZED

On Tue, Sep 4, 2018 at 10:03 PM Manzil Zaheer <manzil at cmu.edu<mailto:manzil at cmu.edu><mailto:manzil at cmu.edu>> wrote:
Hi Yichong

Yes I am able to run TF and PyTorch on these machines. Recently someone else also had similar issue, but it got fixed by reinstalling some local packages.

Thanks,
Manzil

-------- Original message --------
From: Yichong Xu <yichongx at cs.cmu.edu<mailto:yichongx at cs.cmu.edu><mailto:yichongx at cs.cmu.edu>>
Date: 9/4/18 9:58 PM (GMT-05:00)
To: Emre Yolcu <eyolcu at cs.cmu.edu<mailto:eyolcu at cs.cmu.edu><mailto:eyolcu at cs.cmu.edu>>, Predrag Punosevac <predragp at andrew.cmu.edu<mailto:predragp at andrew.cmu.edu><mailto:predragp at andrew.cmu.edu>>
Cc: users at autonlab.org<mailto:users at autonlab.org><mailto:users at autonlab.org>
Subject: Re: PyTorch problem

Just wondering - can Tensorflow run well on these machines? I hope someone to confirm about this so that we can isolate the problem.
OK so here?s a further test: I tried running the cuda examples from the cuda installation (in /usr/local/cuda/sample), on gpu2 in my scratch directory. Simple jobs like deviceQuery succeeds, but simpleCUBLAS failed:
yichongx at gpu2$ cd /home/scratch/yichongx/
yichongx at gpu2$ cd
0_Simple/        2_Graphics/      4_Finance/       6_Advanced/      bin/             conda/
1_Utilities/     3_Imaging/       5_Simulations/   7_CUDALibraries/ common/          miniconda3/
yichongx at gpu2$ cd 7_CUDALibraries/
yichongx at gpu2$ cd simpleCUBLAS
yichongx at gpu2$ CUDA_VISIBLE_DEVICES=3 ./simpleCUBLAS
GPU Device 0: "TITAN X (Pascal)" with compute capability 6.1

simpleCUBLAS test running..
!!!! CUBLAS initialization error
yichongx at gpu2$

This is also consistent with our previous errors from pytorch, which say cublas library not initialized.

So this means at least there is some problem with CUBLAS on gpu2. This post suggests that using sudo can resolve this problem, and this is probably because of some permission problems on CUBLAS libraries:
https://devtalk.nvidia.com/default/topic/1027602/cuda-setup-and-installation/cublas-libraries-with-incorrect-permissions/
@Predrag: Can you try running the simpleCUBLAS example from the CUDA library, with and without root privilege? I think that might be something that you are more familiar with. Thank you very much!

Thanks,
Yichong

On Sep 4, 2018, at 3:18 PM, Emre Yolcu <eyolcu at cs.cmu.edu<mailto:eyolcu at cs.cmu.edu><mailto:eyolcu at cs.cmu.edu>> wrote:

Hi,

We are trying to troubleshoot the PyTorch issue with Predrag and were wondering:

Is anybody able to run PyTorch GPU models on gpu1-9? If you can, we would appreciate if you can respond.

Also, is it a problem for anyone if gpu8 is rebooted today?

Thanks,

Emre

--
Biswajit Paria
PhD in ML @ CMU

--

Biswajit Paria

PhD in ML @ CMU

--

Biswajit Paria

PhD in ML @ CMU

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.srv.cs.cmu.edu/pipermail/autonlab-users/attachments/20180907/0aef616c/attachment-0001.html>