[NLPL Task Force (A)] NLPL modules python issue

Tue Jan 28 13:25:00 UTC 2020

Hi,

You can use this test code:
/cluster/home/andreku/tensorflow_test.py

(I guess you have access to my home directory on Saga)

Currently this code runs without errors when using
TensorFlow/1.13.1-fosscuda-2018b-Python-3.6.6, but fails with
nlpl-tensorflow/1.15.0/3.7.

Note that it should be tested on compute nodes, not on login nodes: it
seems that the latter lack the proper CUDA driver.

On 1/28/20 11:00 AM, Stephan Oepen wrote:
> do you have a minimum test case that i could use to validate a new
> (forthcoming) TF installation?  i use the script in
> $NLPLROOT/operation/python/test/tensorflow.py, which confirms basic
> CUDA functionality but apparently does not validate the parts that
> depend on cuDNN ... something involving convolutions, i suspect?
> 
> i expect we will see cuDNN 7.6 and CUDA 10.1 decoupled during the day
> today, so then i should be able to produce a fresh install of TF
> 1.15.2.
> 
> oe
> 
> 
> On Mon, Jan 27, 2020 at 4:05 PM Andrei Kutuzov <andreku at ifi.uio.no> wrote:
>>
>> Hi,
>>
>> This story will never end :)
>>
>> Even without the dependency on GCC 8.3, the new Saga modules
>> cuDNN/7.6.4.38-CUDA-10.1.243 and CUDA/10.1.243 conflict with
>> nlpl-tensorflow/1.15.0/3.7
>> The reason seems to be that the TF module is compiled using CUDA
>> 10.0.130, not CUDA 10.1.243. Thus, nlpl-tensorflow/1.15.0/3.7 and
>> CUDA/10.1.243 cannot be loaded together.
>>
>> 27.01.2020 11:04, Andrei Kutuzov wrote:
>>> Hi Stephan,
>>>
>>> The correct version of CuDNN is now installed on Saga, and I can load
>>> it, but there is another problem. It seems that all these three modules
>>> are dependent on GCCcore/8.2.0:
>>> nlpl-python-candy/201912/3.7 nlpl-scipy/201910/3.7
>>> nlpl-tensorflow/1.15.0/3.7
>>>
>>> As far as I can tell, this makes them incompatible with the new CUDA
>>> modules, which use GCCcore/8.3.0.
>>>
>>> Is it possible to re-compile the NLPL modules with GCC 8.3?
>>>
>>> 23.01.2020 18:10, Andrei Kutuzov wrote:
>>>> Hi Stephan,
>>>>
>>>> In the end, it was too early to celebrate. Indeed, I can run python3
>>>> from the nlpl-tensorflow/1.15.0/3.7 module now.
>>>>
>>>> But it seems that TF in this module was compiled a different CuDNN
>>>> version than the one used currently on the Saga GPU nodes.
>>>>
>>>> The result is that it is impossible to run GPU jobs with this module. TF
>>>> first produces the following warnings:
>>>>
>>>> 2020-01-23 17:55:36.543187: E
>>>> tensorflow/stream_executor/cuda/cuda_dnn.cc:319] Loaded runtime CuDNN
>>>> library: 7.4.2 but source was compiled with: 7.6.0.  CuDNN library major
>>>> and minor version needs to match or have higher minor version in case of
>>>> CuDNN 7.0 or later version. If using a binary install, upgrade your
>>>> CuDNN library.  If building from sources, make sure the library loaded
>>>> at runtime is compatible with the version specified during compile
>>>> configuration.
>>>>
>>>> ...and then it fails like this:
>>>>
>>>> tensorflow.python.framework.errors_impl.UnknownError: 2 root error(s) found.
>>>>   (0) Unknown: Failed to get convolution algorithm. This is probably
>>>> because cuDNN failed to initialize, so try looking to see if a warning
>>>> log message was printed above.
>>>>
>>>>
>>>> I guess, TF should be compiled with CuDNN 7.4.2 in order to work
>>>> properly on Saga.
>>>>
>>>>
>>>> On 1/13/20 10:10 AM, Sara Stymne wrote:
>>>>> Hi Stephan,
>>>>>
>>>>> Yes, it seems to work fine for me as well. I can load and run nlpl-uuparser/2.3.1, which also loads many of the other modules, and python seems to work fine.
>>>>>
>>>>> Thanks for resolving this so quickly!
>>>>>
>>>>> Best,
>>>>> Sara
>>>>>
>>>>>
>>>>> 12 jan 2020 kl. 20:34 skrev Andrei Kutuzov <andreku at ifi.uio.no>
>>>>> :
>>>>>
>>>>>> Hi Stephan,
>>>>>>
>>>>>> Yes, I can confirm that at least for me this works. I can now run
>>>>>> python3 from the nlpl-tensorflow/1.15.0/3.7 module.
>>>>>>
>>>>>> Thanks for resolving this!
>>>>>>
>>>>>> 12.01.2020 4:26, Stephan Oepen wrote:
>>>>>>> hi again, sara, andrey, all,
>>>>>>>
>>>>>>> i believe i managed to track down this problem and was relieved to see
>>>>>>> it is a recently introduced issue: the NLPL binaries for these Python
>>>>>>> add-on modules had inadvertently had their set-group-id bit ('g+s')
>>>>>>> set, which i am pretty sure was the result of a major recursive
>>>>>>> adjustment of file permissions right before the holidays.  this bit
>>>>>>> (probably) should be set on directories, where it will cause the group
>>>>>>> owner to be inherited onto new sub-directories or files; but on
>>>>>>> executable files (run by anyone but me or root), it actually caused a
>>>>>>> loss of privileges that prevented the search for the base shared
>>>>>>> libraries.  note to self: this was tedious to debug, because the
>>>>>>> problem goes away when running in the scope of strace(1); it turns
>>>>>>> out, strace(1) prevents setuid(2) and setgid(2) execution ...
>>>>>>>
>>>>>>> sara and andrey, please try again.  i hope the NLPL add-on modules are
>>>>>>> back to normal now?
>>>>>>>
>>>>>>> all best, oe
>>>>>>>
>>>>>>> On Fri, Jan 10, 2020 at 11:52 AM Sara Stymne <sara.stymne at lingfil.uu.se> wrote:
>>>>>>>>
>>>>>>>> No, neither of us had tried it before. I think it might have worked for Ali, but I'm not sure.
>>>>>>>>
>>>>>>>>
>>>>>>>> Best,
>>>>>>>>
>>>>>>>> Sara
>>>>>>>>
>>>>>>>>
>>>>>>>> ________________________________
>>>>>>>> Från: Stephan Oepen <oe at ifi.uio.no>
>>>>>>>> Skickat: den 10 januari 2020 11:50:14
>>>>>>>> Till: Sara Stymne
>>>>>>>> Kopia: Martin Matthiesen; Ali Basirat; infrastructure
>>>>>>>> Ämne: Re: [NLPL Task Force (A)] NLPL modules python issue
>>>>>>>>
>>>>>>>>> We tried it on both my and Artur's accounts here, and had the same issue.
>>>>>>>>
>>>>>>>>>> python: error while loading shared libraries: libpython3.7m.so.1.0: cannot open shared object file: No such file or directory
>>>>>>>>
>>>>>>>> had either of you tried before (in other words, is this a recent
>>>>>>>> problem)?  i installed most of these modules last november, but cannot
>>>>>>>> know how many people have tried using them (i believe i know for sure
>>>>>>>> several of them work for vinit and yves) ...
>>>>>>>>
>>>>>>>> oe
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>> När du har kontakt med oss på Uppsala universitet med e-post så innebär det att vi behandlar dina personuppgifter. För att läsa mer om hur vi gör det kan du läsa här: http://www.uu.se/om-uu/dataskydd-personuppgifter/
>>>>>>>>
>>>>>>>> E-mailing Uppsala University means that we will process your personal data. For more information on how this is performed, please read here: http://www.uu.se/en/about-uu/data-protection-policy
>>>>>>
>>>>>>
>>>>>> --
>>>>>> Andrei
>>>>>> PhD Candidate at Language Technology Group (LTG)
>>>>>> University of Oslo
>>>>>
>>>>
>>>>
>>>
>>>
>>
>>
>> --
>> Andrei
>> PhD Candidate at Language Technology Group (LTG)
>> University of Oslo

-- 
Andrei
PhD Candidate at Language Technology Group (LTG)
University of Oslo