marcoabreu commented on issue #11616: Flaky test test_gluon.test_export
URL:
https://github.com/apache/incubator-mxnet/issues/11616#issuecomment-404238313
But yeah, we are regularly testing MXNet CPU on windows and don't seem to
encounter many problems. I also wouldn't know much reason bes
marcoabreu commented on issue #11616: Flaky test test_gluon.test_export
URL:
https://github.com/apache/incubator-mxnet/issues/11616#issuecomment-404237611
I tried, what you described was exactly my observation: you run out of cuda
memory if you run MXNet in parallel. I didn't know we had t
marcoabreu commented on issue #11616: Flaky test test_gluon.test_export
URL:
https://github.com/apache/incubator-mxnet/issues/11616#issuecomment-404230989
Every process can manage it's own GPU memory - cuda supports multi tenancy
out of the box. You're probably only going to run into out o
marcoabreu commented on issue #11616: Flaky test test_gluon.test_export
URL:
https://github.com/apache/incubator-mxnet/issues/11616#issuecomment-404226812
My stand on that would be that our tests should all be completely self
contained. What do you rgibk
--
marcoabreu commented on issue #11616: Flaky test test_gluon.test_export
URL:
https://github.com/apache/incubator-mxnet/issues/11616#issuecomment-404226482
Yes, because everything else is undefined behaviour, right? If we design a
system to be used only in a specific way, we don't "guarante
marcoabreu commented on issue #11616: Flaky test test_gluon.test_export
URL:
https://github.com/apache/incubator-mxnet/issues/11616#issuecomment-404101435
This is the assumption stated above: "Many design decisions are based on
having a single instance of mxnet in a system. "
marcoabreu commented on issue #11616: Flaky test test_gluon.test_export
URL:
https://github.com/apache/incubator-mxnet/issues/11616#issuecomment-404100607
Considering the fact that MXNet is not threadsafe and the interface (CAPI)
is single threaded (with sticky thread) together with the as
marcoabreu commented on issue #11616: Flaky test test_gluon.test_export
URL:
https://github.com/apache/incubator-mxnet/issues/11616#issuecomment-403936302
Just to clarify: The problem here seems to be function-under-test
python\mxnet\gluon\model_zoo\model_store.py:get_model_file, thus it's
marcoabreu commented on issue #11616: Flaky test test_gluon.test_export
URL:
https://github.com/apache/incubator-mxnet/issues/11616#issuecomment-403595104
To clarify: do we expect our users to be able to run multiple instances of
MXNet on the same host?
---
marcoabreu commented on issue #11616: Flaky test test_gluon.test_export
URL:
https://github.com/apache/incubator-mxnet/issues/11616#issuecomment-403579612
Thanks for elaborating on the home variable. That makes sense because we run
as administrator on windows.
marcoabreu commented on issue #11616: Flaky test test_gluon.test_export
URL:
https://github.com/apache/incubator-mxnet/issues/11616#issuecomment-403579484
On production systems while running a service, we can't necessarily assume
that every single request is handled in a sandbox. On Ubuntu
marcoabreu commented on issue #11616: Flaky test test_gluon.test_export
URL:
https://github.com/apache/incubator-mxnet/issues/11616#issuecomment-403567727
To me, it looks like a test problem. It seems like we are hardcoding paths,
causing problems if multiple runs are triggered in parallel
12 matches
Mail list logo