Big rework of IR representation

ezyang / pytorch-unattached

Tensors and Dynamic neural networks in Python with strong GPU acceleration

http://pytorch.org

Other

20 stars 8 forks source link

Big rework of IR representation #229

Closed ezyang closed 7 years ago

ezyang commented 7 years ago

You should read the commits individually.

zdevito commented 7 years ago

Let's talk about these at the standup today because some of these commits go beyond refactoring.

ezyang commented 7 years ago

CR commits are labeled FIXUP, and are right after the commits they would be merged into.

ezyang commented 7 years ago

(If anyone is wondering, "Why am I so obsessed with IR invariants? Why can't we just have a bag of useful manipulation utility functions, and trust users to use them correctly?" just shout, I'll be more than happy to explain :)

ezyang commented 7 years ago

See new FIXUP commits.

ezyang commented 7 years ago

The two unfinished comments are:

https://github.com/ezyang/pytorch/pull/229#discussion_r139161542 (removeInputs semantics with alias information)
https://github.com/ezyang/pytorch/pull/229/files/310daead108806e34b99428d3a81b40aae4bb670#diff-c7fa0d815e308a9fc70b52c9a049a8a3 (on the cloning of stages)

ezyang commented 7 years ago

Since we're dropping aliasing now, review up to fac237eb880768d3521f102a67f743c7fc91f7f6, but there are other patches on top which need to be moved backwards.

ezyang commented 7 years ago

OK, aliasing has been excised.

zdevito commented 7 years ago

Looks good now. I am going to merge as soon as the rebase I just did passes the tests.