[Issue]: patching rpath with patchelf

[ksebaz]

I was wondering whether 'patching' the runpath in jax_rocm_plugin/pjrt/tools/build_gpu_plugin_wheel.py and jax_rocm_plugin/jaxlib_ext/tools/build_gpu_kernels_wheel.py is still necessary. Specifically with something as generic as '/opt/rocm/lib/'. To me it seems these patches happen before bazel will append any runpaths and thus it makes any attempt to properly set runpaths to rocm via bazel's linkopt during build useless. In a heterogeneous software environment, this is a real problem since it necessitates setting the build-time rocm version in LD_LIBRARY_PATH which we usually try to avoid.

[ksebaz]

I just realize that for this repo here, the runpath is actually completely reset by these scripts mentioned above whereas in ROCm/jax the rpath setting comes from the local_config_rocm that is pulled in from xla. For the latter case, with branch rocm-jaxlib-v0.7.1 (and the xla referenced therein; so no explicit local xla) I get any bazel-linkopts specifying rpaths to the specific rocm-module used at compile time prepended by '/opt/rocm/lib'. In any case the result regarding the necessity for LD_LIBRARY_PATH is the same.

[gcapodagAMD]

thank you @ksebaz for raising these