Unable to install DeepSpeed using "pip install deepspeed" command in Windows 11

[S-Kathirvel]

My system configs (if needed) - Single 3060 gpu with vram of 6gb Python 3.12 (primary) and python 3.11 (for virtual environments) torch 2.2.2+cu121 CUDA - nvcc: NVIDIA (R) Cuda compiler driver Copyright (c) 2005-2024 NVIDIA Corporation Built on Wed_Apr_17_19:36:51_Pacific_Daylight_Time_2024 Cuda compilation tools, release 12.5, V12.5.40 Build cuda_12.5.r12.5/compiler.34177558_0

tried setting set DS_BUILD_OPS=0 but that didnt work either.

[loadams]

Hi @S-Kathirvel - with the latest 0.14.5 release we have published a Windows whl as well. Could you please give installing it a try?

[S-Kathirvel]

Is it compatible with python 3.12 and cuda 12.5, if not can you tell me the version of them that works so that i can downgrade to try again. And what are the packages i need to preinstall for this, i have seen that torch need to be preinstalled for this

And for the current status the error persists and I tried to download on a virtual environment again but its also a failure

[S-Kathirvel]

Hello @loadams , So it worked on my python311 which i use to create virtual environments, so currently deepspeed is not available for python 3.12?

[loadams]

@S-Kathirvel - we haven't built the whl with python 3.12 yet, we're working on getting support added to our python workflows.

With 3.11 were you able to see DeepSpeed installed on Windows and able to run ds_report?

[S-Kathirvel]

@loadams i can see the package being downloaded thru pip list in my virtual environment but when i run ds_report

(deepspeed) E:\Deepspeed>ds_report Traceback (most recent call last): File "E:\Deepspeed\deepspeed\Scripts\ds_report", line 3, in from deepspeed.env_report import cli_main File "E:\Deepspeed\deepspeed\Lib\site-packages\deepspeed_init_.py", line 10, in import torch File "E:\Deepspeed\deepspeed\Lib\site-packages\torch_init_.py", line 148, in raise err OSError: [WinError 126] The specified module could not be found. Error loading "E:\Deepspeed\deepspeed\Lib\site-packages\torch\lib\fbgemm.dll" or one of its dependencies.

I get this error and i think its due to torch, i faced the similar error when dealing with some other environment in the system that there is no active torch versions with compatibility for cuda 12.5 and the in build binaries torch comes with both 11.8 and 12.4 doesnt work and only 12.1 works in the python 312 environment

Lemme try installing the version that works with cuda in the 3.12 env to this 3.11 environment and try again

[S-Kathirvel]

@loadams I think it worked but can you help me with these warnings explaining what and why are they appearing

And is this the ds_report?

[loadams]

Hi @S-Kathirvel - thanks, that is the ds_report I was looking for. It looks like DeepSpeed is installed properly then via pip, which is great to see. The warnings are for ops that are currently not supported on Windows due to incompatibilities in other libraries they use (cutlass or async_io for example). So you can ignore those warnings for now if you're just trying to use DeepSpeed on Windows.

[loadams]

Hi @S-Kathirvel - could you also run with ds_report --hide_errors_and_warnings? That should remove these warnings, we are working on getting them cleaned up.

[S-Kathirvel]

what these dependcies for and based on what you said, does it affect any functionalities that deepspeed provide and how the windows and linux libraries differ or do they diifer

[loadams]

They are built from the same git tag, the windows whl is just built for windows, and contains the ops pre-built. More information on the Windows whl and Windows support can be found here: https://github.com/microsoft/DeepSpeed/blob/master/blogs/windows/08-2024/README.md

[S-Kathirvel]

@loadams Thanks for the support man, I hope good things of this project.