I had to stay on CUDA 10.0 for personal projects.
Rather than installing Pytorch with versions appropriate for CUDA 10.0, I accidentally installed Pytorch 1.7 supported with CUDA 10.1. In particular, I installed by
pip install torch==1.7.1+cu101 torchvision==0.8.2+cu101 torchaudio==0.7.2 -f https://download.pytorch.org/whl/torch_stable.html
Surprisingly, everything works fine so far although the CUDA versions do not match.
To verify my installation, I've run the code given in collect_env.py, and it was fine.
I am just wondering few things.
Did Pytorch team officially comment Pytorch 1.7 is compatible with CUDA 10.0?
Would there be more rigorous ways to verify my Pytorch installation?
Surprisingly, everything works fine so far although the CUDA versions
do not match.
Changes between minor versions should work (mismatch like this worked in my case), although there is no promise of compatibility in 10.x release (source), only since 11.x there will be binary compatibility.
Did Pytorch team officially commented Pytorch 1.7 is compatible with
CUDA 10.0?
Not that I'm aware of, but listed wheels do not include 10.0 CUDA and PyTorch 1.7.0 (latest with 10.0 support seems to be 1.4.0).
Would there be more rigorous way to verify my Pytorch installation?
As above, maybe cloning PyTorch's github repo, reverting to tagged release and running tests (folder here, one of cuda test files here), but for personal projects might be excessive.
The Pytorch Previous Version Log offers installations for CUDA versions 9.2, 10.1, 10.2 and 11. Therefore, CUDA 10 is probably not officially supported.
Related
Has anybody tried and succeeded with installing flutter on win x86?
(no VM of course)
My question is not about what Google SAYs, but why do YOU think it is possible or not possible. I do not see anything in the source at github that may prevent it from running on win x86 apart that google does not want to provide support for x86, because it is not a priority.
If you think it is not possible, can you please provide some idea why before saying No or voting down
====BACKGROUND====
prerequisites for the flutter sdk on windows are:
dart 2.0
PowerShell 5.0 or newer
Git for Windows
inside there are also some java and libcurl executables
all of the above exist in win x86 versions
plus the rest seems to be just dart source code.
I run flutter on Ubuntu.
I have an older laptop with win 7 x86 pro which I do not want to upgrade to x64
I tried to use x64 win installation
replacing dart 2.0 x64->x86
and then using flutter doctor to update.
(flutter uses dart pub with "update" function changed to "upgrade", but this can be fixed)
the update using flutter doctor in fact runs just to the point of updating flutter_tool
then trying to update some flutter_tool related packages it comes up with an ERROR:
cannot resolve the library URL
The Windows installation page says:
To install and run Flutter, your development environment must meet
these minimum requirements:
Operating Systems: Windows 7 SP1 or later (64-bit)
And there is a recent comment (20 Aug 2018) from one of the Flutter developers that also states:
We don't have any plans to support 32-bit Windows. That said, if
anyone is willing to send pull requests to get Flutter running on
32-bit Windows, we'd gladly review the PRs.
A further response from the Flutter devs outlines some of the reasons why this is the case:
Someone would need to author 32-bit build rules in the
engine/buildroot repos to build a 32-bit SDK -- specifically the
Dart VM and gen_snapshot. Ideally, the rules would also emit the
target architecture Android artifacts as well.
The design of gen_snapshot (our AOT compiler) assumes identical host
and target architecture bitness. Only a 64-bit build of gen_snapshot
can output arm64 target binaries. Fixing this would involve a
significant amount of work.
I need to do fast image processing so I decide to go for OpenCV with CUDA support. Now I'm trying to build OpenCV as dll to use it in my project, but I am not even able to generate the VS project since CMake keeps giving me this error:
CMakeError. I am on Windows 10, CMake 3.11 with VS 15 2017 x64 generator, OpenCV 3.4 (the last release).
I have installed CUDA 9.2, 9.1 and 8.0 and set CUDA_HOME environment variable as the CUDA toolkit directory. Nothing seems to work.
Then I try to set the CUDA_SDK_ROOT_DIR through the CMake GUI, but the CUDA toolkit directory seems to be not the right one and I'm not able to find the CUDA sample directory anywhere.
I have googled a lot, but I found nothing that had been able to help me. Did someone face this very same issue? How can I solve it?
PS: I have successfully created the dll with TBB and IPP (without CUDA) and it is still not fast enough, so I really need to use Cuda.
I have solved using Cuda 9.1 and VS 15.45. To install CUDA in a correct way you need to use custom installation and delete the VS integration, since it seems to have some problems. Once installation is complete, CMake should detect it.
This is a guide that could help.
I just downloaded and built octave 4.0 from their website. There are a lot of packages already installed from the 3.6 version. I was wondering if is it possible to transfer that installation so that octave 4.0 would recognize them.
Thanks in advance.
Recent days we are trying to use Amazon GPU instances.But their graphic card NVIDIA GRID K520 supports only cuda 5.5. From installation documentation of arrayfire minimum version of cuda is 6.5.Need help to compile arrayfire with cuda 5.5.
You should be able to compile ArrayFire using CUDA 5.5. In fact I was just able to do so on my MacBook Pro. All you have to do is change the CUDA_VERSION to 5.5 in the CMake options.
You should also be able to update the CUDA Toolkit on the EC2 instance. I can't think of a reason why you wouldn't be able to update the driver and install the newest CUDA Toolkit on the machine.
If you have additional questions you can find us at:
https://groups.google.com/forum/#!forum/arrayfire-users
I have read several related posts about installing numpy for python version 2.7 on a 64 bit windows7 OS. Before I try these, does anybody know if the 32bit version will work on a 64bit system?
It should work if you're using 32-bit Python. If you're using 64-bit Python you'll need 64-bit Numpy.
If you are getting it from pip and you want a 64 bit version of NumPy, you need MSVS 2008. pip needs to compile NumPy module with the same compiler that Python binary was compiled with.
The last I checked (this Summer), python's build.py on Windows only supported up to that version of MSVS. Probably because build.py isn't updated for compilers which are not clearly available for free as compile-only versions. There is an "Express" version of MSVS 2010, 2012 and 2013 (which would satisfy that requirement). But I am not sure if there is a dedicated repository for them and if they have a redistribution license. If there is, then the only problem is that no one got around to upgrading build.py to support the newer vertsions of MSVS.