cuda-core Linux wheels ship unstripped .so files, causing 6x size bloat #1881

Summary

Linux wheels for cuda-core have grown steadily across releases (11 MB in 0.4.0 → 21 MB in 0.6.0 → 30 MB in 0.7.0), while Windows wheels have stayed at ~4-5 MB. The root cause is that Linux .so files ship with debug symbols and are not stripped.

Details

Running file on any .so in the 0.7.0 Linux wheel confirms:

ELF 64-bit LSB shared object, x86-64, ... with debug_info, not stripped

Size comparison for the same module (0.7.0):

Stripping all .so files reduces the total extracted size from 103.7 MB → 11.2 MB (89% reduction), which would bring the compressed wheel from ~30 MB down to ~4-5 MB, matching Windows.

Growth across releases is compounded by more Cython modules per release (each duplicated for cu12 + cu13):

Root cause

cuda_bindings already strips debug symbols for wheel builds via -Wl,--strip-all in its build_hooks.py (link):

if strip and sys.platform == "linux":
    extra_link_args += ["-Wl,--strip-all"]

with strip=True passed only for build_wheel().

However, cuda_core/build_hooks.py has no stripping logic at all — its Extension objects have no extra_link_args. The same pattern from cuda_bindings should be applied.

Proposed fix

Add extra_link_args=["-Wl,--strip-all"] to the Extension objects in cuda_core/build_hooks.py when building wheels on Linux, following the same strip=True/False pattern as cuda_bindings.

-- Leo's bot