[CI] Use auditwheel to generate manylinux wheels

[oraluben]

As title.

Currently pypi wheels do not provide libraries required by libtvm_runtime.so and stuffs. This causes troubles like #172 and a so missing error mentioned in wechat group.

This PR shell fix it.

[LeiWang1999]

awesome contributions :) !

@xwhzz would you mind help checking if this works?

[oraluben]

I didn't test maint/scripts/docker_pypi_distribute.sh because there seems to be some issue in my docker. But I do manually run tox -e audit_2_27 to verify it works.

[xwhzz]

Yeah, I'll check this PR ASAP.

[xwhzz]

Thanks for your contribution. I've tested your proposed method on ubuntu 18.04 after running maint/scripts/docker_pypi_distribute.sh to build the wheel, but ran into a issue during library loading.

When installing the built wheel and attempting to import tilelang, I encounter a runtime error about missing TVM libraries:

>>> import tilelang
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/root/miniconda3/envs/py310/lib/python3.10/site-packages/tilelang/__init__.py", line 86, in <module>
    import tvm
  File "/root/miniconda3/envs/py310/lib/python3.10/site-packages/tilelang/3rdparty/tvm/python/tvm/__init__.py", line 26, in <module>
    from ._ffi.base import TVMError, __version__, _RUNTIME_ONLY
  File "/root/miniconda3/envs/py310/lib/python3.10/site-packages/tilelang/3rdparty/tvm/python/tvm/_ffi/__init__.py", line 28, in <module>
    from .base import register_error
  File "/root/miniconda3/envs/py310/lib/python3.10/site-packages/tilelang/3rdparty/tvm/python/tvm/_ffi/base.py", line 78, in <module>
    _LIB, _LIB_NAME = _load_lib()
  File "/root/miniconda3/envs/py310/lib/python3.10/site-packages/tilelang/3rdparty/tvm/python/tvm/_ffi/base.py", line 58, in _load_lib
    lib_path = libinfo.find_lib_path()
  File "/root/miniconda3/envs/py310/lib/python3.10/site-packages/tilelang/3rdparty/tvm/python/tvm/_ffi/libinfo.py", line 166, in find_lib_path
    raise RuntimeError(message)
RuntimeError: Cannot find libraries: ['libtvm.so', 'libtvm_runtime.so', '3rdparty/cutlass_fpA_intB_gemm/cutlass_kernels/libfpA_intB_gemm.so', '3rdparty/libflash_attn/src/libflash_attn.so']

​Library Availability:

• Your build produces libtvm.so/libtvm_runtime.so in tilelang/lib, but the TVM Python package fails to detect them
• Our nightly build works without including libcudart.so/libnvrtc.so, but your method bundles these CUDA runtime libraries
  ​

Directory Comparison:

Your Build (new)               Nightly Build (existing)
────────────────────────────────────────────────────────
libcudart-...so.12.1.55        (not present)
libnvrtc-...so.12.1.55         (not present)
libtilelang_module.so          libtilelang_module.so
libtilelang.so                 libtilelang.so
libtvm.so                      libtvm.so
libtvm_runtime.so              libtvm_runtime.so

The wheel files generated using your proposed approach for reference can be accessed via the link. Would you mind reviewing the build process to help identify where things might be going wrong?

[LeiWang1999]

Thanks @xwhzz , and I think we should double check this commit and then we can release tilelang 0.1.3 :)

[oraluben]

Sure, thanks for your review! I'll try to fix it.

[oraluben]

previously add-apt-repository -y ppa:deadsnakes/ppa does not work on my environment. I modify the script a little bit to make it work. It's for installing python versions, right? Since we already have conda it can be safely removed. I also add a new dockerfile to make build environment more reproducible.

The auditwheel-ed wheels are located in ./wheelhouse, not ./dist. I've verified they can be imported successfully. Would you mind to test again on your side? @xwhzz

[oraluben]

Sorry for the format! I'll take a look later.

[LeiWang1999]

@oraluben run format.sh would help

[xwhzz]

Thanks for your guidance @oraluben. I found that the issue was caused by my outdated TVM submodule. After updating it, this PR works for me.

[oraluben]

• Our nightly build works without including libcudart.so/libnvrtc.so, but your method bundles these CUDA runtime libraries

As for this, is this still an issue? CUDA libs are excluded via --exclude /usr/local/cuda*, which should be the default install location of cuda runtime. How's nightly package built?