[BUG] [Win] Installation from source fails for transformer_inference_op

[lostmsu]

Describe the bug
I am attempting to install DeepSpeed from cloned code in editable mode and the build fails with these lines present in the output:

LINK : fatal error LNK1181: cannot open input file 'aio.lib'

and

error: could not create 'deepspeed\ops\transformer_inference\transformer_inference_op.cp39-win_amd64.pyd': No such file or directory

The first one happens despite DS_BUILD_AIO=1

Full log: inst.log

To Reproduce
Steps to reproduce the behavior:

1. Clone the repo
2. Enter Python environment, that has torch installed
3. cd to the repo, and run pip install -e .
4. See error above

Expected behavior
Build should succeed.

ds_report output
N/A

System info (please complete the following information):

• OS: Windows 10
• GPU count and types: irrelevant, but 2x3090
• Interconnects (if applicable): N/A
• Python version: Python 3.9.7 (default, Sep 16 2021, 16:59:28) [MSC v.1916 64 bit (AMD64)] :: Anaconda, Inc. on win32

WORKAROUND
Disabling transformer inference using DS_BUILD_TRANSFORMER_INFERENCE=0 does the trick.

[RezaYazdaniAminabadi]

Hi @lostmsu

Thanks for looking into supporting inference through Windows.
I will look into this issue.

Best,
Reza

[lostmsu]

@RezaYazdaniAminabadi I actually did some investigation and it seems related to the location of C++ sources vs the structure of Python modules.

[jomayeri]

Please check out the latest version of DeepSpeed for Windows compatibility.