[BUG] I can't using pip install deepspeed

[Yeqishen]

Describe the bug
I can't using pip install deepspeed

To Reproduce
Steps to reproduce the behavior:
To build wheel on Windows:
1. torch 1.12.1+cu113 , CUDA 11.7
2.exactly install visual cpp build tool
3. Launch cmd console with Administrator privilege.
4. using cmd execute "pip install deepspeed"

System info (please complete the following information):

• OS: Windows10
• one machines with Nvidia 3080ti
• https://github.com/microsoft/DeepSpeed-0.7.3
• (if applicable) Hugging Face 1.4 / Transformers 4.22.2/ Accelerate need to install deepspeed(So, I can't install accelerate)
• Python version - 3.10.6

Additional context
(venv) G:\Download\DeepSpeed-0.7.3\DeepSpeed-0.7.3>python setup.py bdist_wheel
DS_BUILD_OPS=1
Traceback (most recent call last):
File "G:\Download\DeepSpeed-0.7.3\DeepSpeed-0.7.3\setup.py", line 166, in
ext_modules.append(builder.builder())
File "G:\Download\DeepSpeed-0.7.3\DeepSpeed-0.7.3\op_builder\builder.py", line 596, in builder
assert_no_cuda_mismatch()
File "G:\Download\DeepSpeed-0.7.3\DeepSpeed-0.7.3\op_builder\builder.py", line 87, in assert_no_cuda_mismatch
cuda_major, cuda_minor = installed_cuda_version()
File "G:\Download\DeepSpeed-0.7.3\DeepSpeed-0.7.3\op_builder\builder.py", line 40, in installed_cuda_version
assert cuda_home is not None, "CUDA_HOME does not exist, unable to compile CUDA op(s)"
AssertionError: CUDA_HOME does not exist, unable to compile CUDA op(s)

[MT010104]

Hi, I met the same problem. Did you solve it?

[Yeqishen]

Hi, I met the same problem. Did you solve it?

No, I still haven't solved the problem. But I no longer choose to use deepspeed when installing accelerate.

[d8ahazard]

Hi, I met the same problem. Did you solve it?

You need to install the proper Cuda Toolkit for your pytorch version, then add the CUDA_HOME environment variable to your windows path var.

Even after that, you will have limited success in getting this to work.

[eltonzheng]

Yes, please install cuda toolkit before generating DeepSpeed wheel, following is the "cuda_home" in my machine, which is returned from torch.utils.cpp_extension.CUDA_HOME.

(py3.7) D:\git\ds>python setup.py bdist_wheel
DS_BUILD_OPS=1

d:\git\ds\op_builder\builder.py(40)installed_cuda_version()
-> cuda_home = torch.utils.cpp_extension.CUDA_HOME
(Pdb) n
d:\git\ds\op_builder\builder.py(41)installed_cuda_version()
-> assert cuda_home is not None, "CUDA_HOME does not exist, unable to compile CUDA op(s)"
(Pdb) p cuda_home
'C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.6'

[eltonzheng]

I submitted a PR to fix build issues on Windows, @Yeqishen, @MT010104 and @dfyz could you help to check if it works on your Windows?

#2428

[tjruwase]

Closing issue, please (re)open as needed.

[xxllp]

I meet the same problem on linux system ~~~

[macabdul9]

@xxllp did you manage to fix it?

[46319943]

Refer to this new issue #2772 (comment).