[RFC] Tracking Issue for TECompiler

[jroesch]

Tracking issue for te_compiler.

About tracking issues

Tracking issues are used to record the overall progress of implementation and relevant issues.
A tracking issue is however not meant for large scale discussion, questions, or bug reports about a feature.
For open ended discussion, check out: https://discuss.tvm.apache.org/.

Implementation

• Introduce TECompiler (see [RFC]TECompiler: Staged refactor and removal of compile engine #7518)
• Unify tvm::lower into a single API
  • Remove lower wrapper inside of te_compiler_cache
• Rewrite GraphRuntimeCodegen
• Rewrite Interpreter
• Rewrite VM
• Handle BYOC Modules

Unresolved Questions

How to unify passes across backends?

[Mousius]

@jroesch are there plans to include AOT in this now?

[jroesch]

cc @mbs-octoml @denise-k is it possible for you guys to take over this one as you now have the most context?

[mbs-octoml]

Here's where we've gotten to:

1. TECompiler has fully replaced CompileEngine :-)
2. te_complier_cache.cc still in pretty messy state, the cache datastructures themselves can be simplified however we still have customers of TECompiler::Lower etc via te_compiler.py.
3. After [Relay] Switch the VM to use the LowerTE pass instead of TECompiler::{Lower,LowerShapeFunc}. #9483 all backends will use the TELowerPass instead of the one-function-at-a-time lowering API.
4. IRModule could do with better support for external functions, runtime::Modules and generated AOT source code.

At some point I'd lie to tackle 2 head-on, that will require discussion to figure out how to move folks off of depending on the CCache* structures and one-function-at-a-time.

And 4 can also be split out.

Close this one?

[jroesch]

Ok Mark will split out into follow up issues that go on the roadmap and we will close this one. cc @Mousius