`pip install` of preview wheel uninstalls XPU PyTorch and replaces it with CUDA PyTorch

[jiqing-feng]

System Info

Environment

• OS: Linux x86_64 (Ubuntu 24.04)
• Python: 3.12
• PyTorch: 2.11+xpu (XPU build, installed from Intel index)
• bitsandbytes wheel: bitsandbytes-1.33.7.preview-py3-none-manylinux_2_24_x86_64.whl from continuous-release_main

Describe the bug

Installing the latest preview wheel from GitHub releases with --force-reinstall silently uninstalls the user's XPU PyTorch and replaces it with CUDA PyTorch from PyPI.

pip install --force-reinstall \
  https://github.com/bitsandbytes-foundation/bitsandbytes/releases/download/continuous-release_main/bitsandbytes-1.33.7.preview-py3-none-manylinux_2_24_x86_64.whl

The wheel metadata declares:

Requires-Dist: torch<3,>=2.3

When pip resolves this dependency with --force-reinstall, it pulls torch from the default PyPI index, which only hosts CUDA builds. This overwrites the XPU-flavored PyTorch (e.g. torch==2.7.0+xpu installed from https://pytorch-extension.intel.com/release-whl/stable/xpu/us/) with the CUDA variant.

Reproduction

pip install --force-reinstall
https://github.com/bitsandbytes-foundation/bitsandbytes/releases/download/continuous-release_main/bitsandbytes-1.33.7.preview-py3-none-manylinux_2_24_x86_64.whl

Expected behavior

Installing the bitsandbytes wheel should not replace an already-installed XPU PyTorch with the CUDA build. Users on XPU environments should be able to install bitsandbytes without losing their PyTorch backend.

[jiqing-feng]

Hi @matthewdouglas . Would you please take a look? Thanks!

[Titus-von-Koeller]

Thanks for the report @jiqing-feng. This is unfortunately expected behavior from pip rather than a bitsandbytes bug — it affects any package that declares torch as a dependency.

When you pass --force-reinstall, pip re-resolves all dependencies from scratch against PyPI, which only hosts CUDA builds of torch. Since your XPU torch was installed from Intel's custom index, pip has no memory of that index and replaces it with the default CUDA variant. Any package depending on torch would cause the same issue in this scenario.

Workaround: use --no-deps to avoid touching your existing torch installation:

pip install --force-reinstall --no-deps \
  https://github.com/bitsandbytes-foundation/bitsandbytes/releases/download/continuous-release_main/bitsandbytes-1.33.7.preview-py3-none-manylinux_2_24_x86_64.whl

Or, if you want pip to resolve other dependencies (numpy, packaging) but still keep your XPU torch, pass Intel's index so pip can find the right torch:

pip install --force-reinstall \
  --extra-index-url https://pytorch-extension.intel.com/release-whl/stable/xpu/us/ \
  <wheel-url>

I've updated the pre-release notes to call this out for users with custom PyTorch builds. Closing as resolved.

[Titus-von-Koeller]

@jiqing-feng feel free to reopen, if you have a better proposal what to write