Binary install not working in airgapped/pip proxy scenario

[zeronewb]

Describe the bug

Binary install not working, since wheel fetch hardcoded, and wheel not published at pypi.
Pypi cached via artifactory/nexus etc.
GPU Info

To Reproduce

pip install -v gptqmodel --no-build-isolation with custom proxied indexes.

Expected behavior

Binary install working with pypi caching proxies.

[zeronewb]

Same issues:
#1634
#1634
#1633

[Qubitium]

The wheels with precompiled kernels are large and need a unique wheel for every python and torch combo. I don't think pypi allows GB sized pkgs per version.

Not sure what the solution is for an airgapped secure env with only pypi cache proxy. Allow the download urls to be configurable via env vars?

[Qubitium]

From a security standpoint, having unrestricted pypi cache is a higher security risk than wheel downloads from restricted
(whitelisted) github release urls.

BASE_WHEEL_URL = (
    "https://github.com/ModelCloud/GPTQModel/releases/download/{tag_name}/{wheel_name}"
)

Perhaps ask the WAF firewall team to allow white listed github.com release paths such as https://github.com/ModelCloud/GPTQModel/releases/download/? Or setup a proxy for whitelisted github repo downloads/artifcats? Even if we allow BASE_WHEEL_URL to be env configurable, the security teams still needs to allow the github release download proxy.

[Qubitium]

Also, our setup mimics the logic used by https://github.com/Dao-AILab/flash-attention which also builds kernel but tries download precompiled wheels from github/releaes. How does your team resolve the flash-attn issue? The same download issue for us should be also visible for the flash-attn module and the same fix can be applied.

[Qubitium]

Completed with: #1761