Discussion and troubleshooting on PyPI (pip) / Scala Maven (Nexus) installation #8671
Comments
Is there a way to see the build flags used to build the pip packages? |
I will add a section for the flags |
头像换了,不好看,还是原来的猪头像好看。。。蛤蛤。。。 |
https://pypi.python.org/pypi/mxnet-cu91/ |
@adaaaaaa cu91 versions are WIP. |
cu90和cu91有没有benchmark,对比一下? |
求cu91的更一下win版本的? |
pinging @yajiedesign |
你@了一个不存在的人? |
realese date for cu91 for win? Xiexie |
@szha when can we get cu91 for win ? |
@yajiedesign is working on it. |
@szha Can we turn on the USE_LIBJPEG_TURBO flag. I find that it helps improve the speed of IO pipeline significantly? Results for Resnet50 v1 Imagenet, 480px resized data, batch size 1920, float16, symbolic, p3.16x As an example, Here's how I used LIBJPEG_TURBO on ubuntu.
Flags |
@rahul003 thanks for the suggestion. I will certainly take a look. For these dependencies, my approach have been to statically link them, so some steps will be different. |
Is mxnet-mkl 1.2.0 ready for windows? |
Are the pip distributions compiled with |
@fhieber Yes, they should have USE_DIST_KVSTORE=1 already. Did you have trouble using the distributed kvstore feature? |
@szha Oh great, thanks! I haven't tried it in a while but remembered that last time I checked the distributions were compiled without it. Glad to hear its enabled by default now! |
@fhieber no problem. I added the corresponding row in the table. |
@szha what version of opencv is used for pip? |
@eric-haibin-lin 3.3.0 |
@szha the windows mkl is compiled,is update? |
@szha What malloc are you using for pip? tcmalloc, jemalloc, or default? |
@MoritzMaxeiner default malloc so far. I intended to use jemalloc by building it from source but our makefile logic for jemalloc prevents a static link. https://github.com/apache/incubator-mxnet/blob/master/Makefile#L254-L278 (e.g. it forces to look in specific paths for libjemalloc.so). |
@szha Thank you very for the info, I suspected as much. OT: I was just making sure, since I use MXNet from a shared object (dynamically loaded at runtime by another executable that I can't change to tcmalloc or jemalloc, which links dynamically against OpenCV), and using tcmalloc in MXNet leads to these issues with the |
@szha : I'm trying to locally reproduce the
Thanks! |
@cconvey I use ubuntu 14.04 as the environment because I want the GLIBC to be old enough for portability. I statically link all the dependencies too. For verification, if I changed any dependency, I do a manual verification of the new wheel. For all wheels, I run unit-test as sanity test before any upload. |
@szha When building the wheels, do you use the provided Makefiles or CMake? edit for clarification: Do you use the script |
I'm using Makefile along with a separate set of scripts. |
PyPI for Apache MXNet installation has been available for a while and has been the most important channel for distribution. I'm opening this issue to collect feedbacks, suggestions and issue reports for releases and nightly pre-releases.
pip Status
Supported variants:
Windows
,Linux
, andMac OSX
, with CPU-only without MKL-DNN support.Linux
, supportsCUDA-7.5
. (discontinued in 1.2)Windows
andLinux
, supportsCUDA-8.0
.Windows
andLinux
, supportsCUDA-9.0
.Windows
andLinux
, supportsCUDA-9.1
. (will discontinue in 1.3)Windows
andLinux
, supportsCUDA-9.2
.Windows
andLinux
, supportsCUDA-10.0
.Windows
,Linux
, andMac OSX
, with CPU-onlyMKLDNN
support.Linux
, supportsCUDA-7.5
andMKLDNN
support. (discontinued in 1.2)Windows
andLinux
, supportsCUDA-8.0
andMKLDNN
support.Windows
andLinux
, supportsCUDA-9.0
andMKLDNN
support.Windows
andLinux
, supportsCUDA-9.1
andMKLDNN
support. (will discontinue in 1.3)Windows
andLinux
, supportsCUDA-9.2
andMKLDNN
support.Windows
andLinux
, supportsCUDA-10.0
andMKLDNN
support.Scala Maven (Nexus) Status
Packages can be found on Nexus
mxnet
variant on Linux)mxnet-cu90
variant on Linux)mxnet
variant on OSX)Compilation Flags
Notes:
* `USE_OPENMP` only applies to Linux
**
USE_NCCL
is supported since release version1.0.0.post2
, nightly since1.0.1b20171228
. Version: 2.1.2-1Usage
Release
For release, users can install it via
Be sure to update
pip
(setuptools
) to the most recent version, bypip install --upgrade pip
.Nightly
For nightly pre-release, users can install it via
pip install --pre --upgrade
. Be sure to updatepip
(setuptools
) to the most recent version, bypip install --upgrade pip
.Report issue
For bug submissions on issues found in the pre-releases, please attach:
mx.nd.random.uniform
), unless bug is only triggered with specific data value, in which case that specific value should be attached.The text was updated successfully, but these errors were encountered: