Inexistant CUDA support for last version ?

[TheHellTower]

Click to expand!

Issue Type

Feature Request

Have you reproduced the bug with TF nightly?

Yes

Source

source

Tensorflow Version

tf 2.11

Custom Code

No

OS Platform and Distribution

Windows 10 (Build 19045.2673)

Mobile device

No response

Python version

3.11

Bazel version

No response

GCC/Compiler version

No response

CUDA/cuDNN version

11.0.2

GPU model and memory

No response

Current Behaviour?

In the last version of Tensorflow it is saying the support for CUDA on Windows is kinda dropped ?

Is it possible to put it back ? I can't make a code work on a older version of Tensorflow and I want to use GPU processing to loose lowest amount of time possible..

Standalone code to reproduce the issue

It's on configuration.

Relevant log output

None

[tiruk007]

@TheHellTower
TensorFlow 2.10 was the last TensorFlow release that supported GPU on native-Windows. Starting with TensorFlow 2.11, you will need to install TensorFlow in WSL2 for GPU support. Could you please refer to the doc for reference.

Thank you !

[TheHellTower]

@TheHellTower TensorFlow 2.10 was the last TensorFlow release that supported GPU on native-Windows. Starting with TensorFlow 2.11, you will need to install TensorFlow in WSL2 for GPU support. Could you please refer to the doc for reference.

Thank you !

The problem is I'm building the project on Windows with Visual Studio for Windows so it's not possible anymore ? Can I even build the project from Windows for Linux ?
This is not practical at all, why the support got dropped ?

[tiruk007]

@TheHellTower
Sorry for the late reply, Unfortunately from TF 2.11 is not supported for GPU pip installation. If you want GPU support for windows try to build from source. Please refer to Build from source on Windows for reference.

Thank you!

[TheHellTower]

@TheHellTower Sorry for the late reply, Unfortunately from TF 2.11 is not supported for GPU pip installation. If you want GPU support for windows try to build from source. Please refer to Build from source on Windows for reference.

Thank you!

Hello,

First of all I'm not using Python I want to use C++, then you said GPU support is not in 2.11 "TensorFlow 2.10 was the last TensorFlow release that supported GPU on native-Windows" so it would be useless for me to build from source ? I was enthusiastic and excited by the idea of using Tensorflow until I see that I can't even make my project with the 2.10 version it doesn't seem to work as intented (it was crashing).

If in C++ the GPU support is still there then I woud love to know how to get it on my side because I really need it I can't waste time with CPU it's way too long next to GPU...

And like I can see here "Caution: TensorFlow 2.10 was the last TensorFlow release that supported GPU on native-Windows."

Regards.

[TheHellTower]

Well hopefully you will not avoid the issue just like this and will give a answer at some point.

[gaikwadrahul8]

Hi, @SuryanarayanaY

Could you please take look into this issue? Thank you!

[SuryanarayanaY]

Hi @TheHellTower ,

With Windows if you want to enable GPU support for Tf>2.10 versions, you can try DirectML Plugin option. Please refer the source here and follow the instructions mentioned there.

The requirement should be:

- Windows 10 Version 1709, 64-bit (Build 16299 or higher)
- Python x86-64 3.8, 3.9, 3.10 or 3.111
- One of the following supported GPUs:
          AMD Radeon R5/R7/R9 2xx series or newer
          Intel HD Graphics 5xx or newer
          NVIDIA GeForce GTX 9xx series GPU or newer
   

If you want to go with TF=2.10v only then you might need to install tensorflow first and then reinstall again with TF==2.10v.

Incase of any problem with DirectML plugin please let us know.

Thanks!

[TheHellTower]

Hi @TheHellTower ,

With Windows if you want to enable GPU support for Tf>2.10 versions, you can try DirectML Plugin option. Please refer the source here and follow the instructions mentioned there.

The requirement should be:

- Windows 10 Version 1709, 64-bit (Build 16299 or higher)
- Python x86-64 3.8, 3.9, 3.10 or 3.111
- One of the following supported GPUs:
          AMD Radeon R5/R7/R9 2xx series or newer
          Intel HD Graphics 5xx or newer
          NVIDIA GeForce GTX 9xx series GPU or newer
   

If you want to go with TF=2.10v only then you might need to install tensorflow first and then reinstall again with TF==2.10v.

Incase of any problem with DirectML plugin please let us know.

Thanks!

Isn't it for the Python bindings ? I need GPU support using C++ I'm writing my project in C++.

[SuryanarayanaY]

Hi @TheHellTower ,

If you want GPU support with C++ then you need to go with tensorflow version 2.10 only using c-lang bindings here. For more details on this please refer this source.

Please note that there are some limitations in this bindings like mentioned below.

[TheHellTower]

Hi @TheHellTower ,

If you want GPU support with C++ then you need to go with tensorflow version 2.10 only using c-lang bindings here. For more details on this please refer this source.

Please note that there are some limitations in this bindings like mentioned below.

I think the support should be added back. See Issue #59918 that list 7 issues including this one about GPU support on Windows, It's very important for my project to be on the last version and I can't even use it and in 2.10 it doesn't work at all for some strange reasons.

[TheHellTower]

Bump.

[TheHellTower]

Hi @TheHellTower ,

With the inclusion of WSL,it will be comparatively easy to maintain the framework for both Linux and Windows. Here is the link to the announcement blog which talks about official build collaborators.

Please refer to the Developer comment related to drop of GPU support.

#59918 (comment)

[AsakusaRinne]

@TheHellTower Sorry for the late reply, Unfortunately from TF 2.11 is not supported for GPU pip installation. If you want GPU support for windows try to build from source. Please refer to Build from source on Windows for reference.

Thank you!

@tiruk007 Hello, I tried building tf2.11 from source on Windows and it failed. I can't find the solution for the error in existing issues. May I ask that whether tensorflow developers already tried building from source on Windows since tf2.11 and succeeded? I'd like to build from source to use GPU for native-windows on tf2.11 and higher. However, if that means I need to modify the C++ code to make the build successful, the price is just too high.

The main part of error is listed below:

.\tensorflow/core/kernels/segment_reduction_ops_gpu.cu.h(206): error: no suitable constructor exists to convert from "int" to "tensorflow::AlignedVector<int, 4>"
          detected during:
            instantiation of "void tensorflow::SegmentOffsetsKernel(Tindex, Tsegmentids, const Tsegmentids *, Tindex *) [with Tindex=int, Tsegmentids=int]"
(233): here
            instantiation of "tsl::Status tensorflow::LaunchSegmentOffsetsKernel(const tensorflow::GPUDevice &, Tindex, Tsegmentids, const Tsegmentids *, Tindex *) [with Tindex=int, Tsegmentids=int]"
(570): here
            instantiation of "tsl::Status tensorflow::SegmentReduceGPUImpl<Treducevec,Tvec,Tindex,Tsegmentids,ReduceOp,Tinit>(tensorflow::OpKernelContext *, Tindex, Tindex, Tsegmentids, ReduceOp, Tinit, Tinit, __nv_bool, __nv_bool, const Tvec *, const Tsegmentids *, const Tindex *, const Tinit *, Tvec *) [with Treducevec=tensorflow::AlignedVector<int, 4>, Tvec=tensorflow::AlignedVector<int, 4>, Tindex=int, Tsegmentids=int, ReduceOp=tensorflow::functor::Max, Tinit=int]"
(617): here
            instantiation of "tsl::Status tensorflow::SegmentReduceGPUVectorized<Treduce>::Impl<vec_size>::operator()(tensorflow::OpKernelContext *, Tindex, Tindex, Tsegmentids, ReduceOp, T, T, __nv_bool, __nv_bool, const T *, const Tsegmentids *, const Tindex *, const T *, T *) [with Treduce=tsl::int32, vec_size=4, T=int, Tindex=int, Tsegmentids=int, ReduceOp=tensorflow::functor::Max]"
.\tensorflow/core/util/gpu_kernel_helper.h(328): here
            instantiation of "tsl::Status tensorflow::detail::DispatchToVectorizedHelper<VecSize, Functor>::operator()(int64_t, Args &&...) const [with VecSize=4LL, Functor=tensorflow::SegmentReduceGPUVectorized<tsl::int32>::Impl, Args=<tensorflow::OpKernelContext *&, int &, int &, int &, tensorflow::functor::Max &, int &, int &, __nv_bool &, __nv_bool &, const tsl::int32 *&, const tsl::int32 *&, const tsl::int32 *&, const tsl::int32 *&, tsl::int32 *&>]"

[learning-to-play]

@mraunak Could you please take a look?

[intel-tf-windows]

Hey @learning-to-play this is the Intel TF Windows account. Is it possible to tag this account or does it need to be added? - Rajeev & Mayank

[learning-to-play]

@intel-tf-windows It seems I can now.
@sachinprasadhs Thank you for assigning the issue!

[mraunak]

Hi @learning-to-play, I think TensorFlow GPU or Nvidia Team would be able to resolve this issue as the errors involve CUDA files

[learning-to-play]

Thank you @mraunak!

[TheHellTower]

News ?

[learning-to-play]

Some references that may be helpful:

• See section "Expanded GPU support on Windows" in this TensorFlow blog post.
• See TensorFlow install page and this page.

[TheHellTower]

Some references that may be helpful:

• See section "Expanded GPU support on Windows" in this TensorFlow blog post.
• See TensorFlow install page and this page.

No it's not okay for me because:

• Inexistant CUDA support for last version ? #59905 (comment)
• Inexistant CUDA support for last version ? #59905 (comment)

[TheHellTower]

Bump

[TheHellTower]

Bump

[TheHellTower]

Bump

[AsakusaRinne]

You can never wake up a bunch of people pretending to be asleep

[TheHellTower]

You can never wake up a bunch of people pretending to be asleep

Doesn't matter I have the time lol

[TheHellTower]

Bump

[learning-to-play]

@TheHellTower Unfortunately as mentioned in this comment GPU on native-Windows isn't supported.