Yichen Yan

Mark as ready-for-review for it passes all tests.

Based on a wheel of pytorch, I test the speed and size of different compress level of zip. There's some outliner in extract time, but I haven't double check them...

With the demo in PR, the time has been reduced from 3m36.515s to 2m44.353s. Concurrent grafting seems not significent, I'll see the result on our real workload later. Only change...

Numbers: 10m59.651s (6.2.0) 10m27.363s (current version + level=6) 8m54.623s (current version) 5m14.166s (current version + level=0) The repair part still seem potential to be optimized. [audit_orig.log](https://github.com/user-attachments/files/18898120/audit_orig.log) [audit_fast.log](https://github.com/user-attachments/files/18898121/audit_fast.log)

`torch.compiler.is_compiling()` should be better for this case, however there's still issue, presumably on dynamo side (since we have faketensor we're definitely tracing). So keep it for now. ``` [rank1]: File...

I have some experience with this specific issue, the pypi mkl package would be installed to `../../libmkl*.so`, i.e. `venv/lib/libmkl*.so`. This path is usually not recognized by auditwheel, you can add...

> Sad, Z3 has a hard-encoded SONAME `libz3.so.x.xx`. The linker will make TVM dependent on only 1 version of Z3. We have to choose: `libz3.so` in `z3-solver==4.13.0` has a SONAME...

> Also we could use `patchelf --replace-needed` to modify the `DL_NEEDED` part of libtvm (`libz3.so.x.xx` -> `libz3.so`). This is rather common in python packaging. I could help if needed. Actually...

> And in MACOS, the libz3.dylib is placed in the `bin` directory. The inconsistent library path causes an error. We might be able to set different `INSTALL_RPATH` for macos to...

> I finally chose to build a static linked library in `libtvm.so`, which can create a 20% larger library size. I think the best option is to embedding a shared...