Avoiding importing too many dependency libraries when importing hlib

cornell-zhang / heterocl

HeteroCL: A Multi-Paradigm Programming Infrastructure for Software-Defined Heterogeneous Computing

https://cornell-zhang.github.io/heterocl/

Apache License 2.0

322 stars 92 forks source link

Avoiding importing too many dependency libraries when importing hlib #299

Closed hecmay closed 1 year ago

hecmay commented 3 years ago

When importing hlib, many other packages are imported by default (e.g. TVM, keras, e.t.c), which makes the importing process really slow (around 5s).

I do not know whether this is related to Debjit's issue. He tried to install TVM separately in the OS, but pip always complains TVM not found. Maybe there is some potential conflicts between TVM and HCL.TVM in pip?

seanlatias commented 3 years ago

Although the conclusion is not true (tvm conflicts with hcl.tvm), I agree we can reduce the number of dependency libraries needed. If some package is not found, usually that just means the environment issue.

hecmay commented 3 years ago

This is kind of an unrelated issue. Debjit said he installed TVM using the .circleci/install_tvm.sh, but in the site-packages folder, he can only find an empty tvm.egg folder. That's wired. we can check later.

chhzh123 commented 1 year ago

As we have released a new version of HeteroCL that entirely moves the compilation stack to MLIR, issues related to the original HalideIR flow will be closed.