.gitignore magic_wand; updated install instructions to include make -…

[afeldman-nm]

Fixes to make the neuralmagic-vllm install process smoother:

• Added README instruction to build the magic_wand .so (unless we plan to roll make -C build/ into magic_wand setup.py)
• .gitignore magic_wand
• reminder to login to huggingface

[afeldman-nm]

Oh also - added advice to use python>=3.10; vllm docs recommend installing 3.9 via conda but this breaks magic_wand

[mgoin]

Oh also - added advice to use python>=3.10; vllm docs recommend installing 3.9 via conda but this breaks magic_wand

why does 3.9 not work with the lib?

[afeldman-nm]

Oh also - added advice to use python>=3.10; vllm docs recommend installing 3.9 via conda but this breaks magic_wand

why does 3.9 not work with the lib?

3.9 works fine with neuralmagic-vllm. However, neuralmagic-vllm takes a dependency on magic_wand. @LucasWilkinson In magic_wand one or more of the sparsity related classes call "dataclasses(slot=True)"; according to

https://www.trueblade.com/blogs/news/python-3-10-new-dataclass-features

the "slot" constructor argument is new in python 3.10.

I tried installing neuralmagic-vllm twice, with Python versions 3.9 and 3.10 respectively, and confirmed that under 3.9 the "slot" dataclasses argument is not recognized.

[afeldman-nm]

Oh also - added advice to use python>=3.10; vllm docs recommend installing 3.9 via conda but this breaks magic_wand

why does 3.9 not work with the lib?

3.9 works fine with neuralmagic-vllm. However, neuralmagic-vllm takes a dependency on magic_wand. @LucasWilkinson In magic_wand one or more of the sparsity related classes call "dataclasses(slot=True)"; according to

https://www.trueblade.com/blogs/news/python-3-10-new-dataclass-features

the "slot" constructor argument is new in python 3.10.

I tried installing neuralmagic-vllm twice, with Python versions 3.9 and 3.10 respectively, and confirmed that under 3.9 the "slot" dataclasses argument is not recognized.

@mgoin