[Question][PyTorch] Status of PyTorch 2.0 integration

Questions
#1

Hi community,

We have some questions about PyTorch 2.0 supporting. May you kindly give us some references please?

We are interested in the integration between TVM and PyTorch 2.0. May we know the information, like:

  1. What is the road map of integration between TVM and PyTorch.
  2. Where to check integration status.
  3. How many new features will be introduced.
  4. Any change on the current frontend conversion of PyTorch.

Recently PyTorch has demonstrated some optimized results based on the compile function in 2.0 version. It seems to be very promising, but perhaps the existing PyTorch IR might be changed or enhanced.

Currently we only found these two materials about PyTorch 2.0 integration:

  1. PyTorch 2.0 + Apache TVM: Better Together
  2. Git docker update for PyTorch 2.0

Because PyTorch IR conversion in TVM is a relatively complex conversion, it would be great if we could have more information about it, or maybe a preview about what is going to change in the current PyTorch conversion.

Thanks for reading. We are looking forward to hearing from you. :smiley:

Best
Joey

1 Like
#2

I’m not aware of existing effort / plan, but if something is going to happen probably it will be done on the unity branch. An exception is https://github.com/apache/tvm/pull/14886

The situation of PT 2.0 compile is very premature and messy right now - personally I’d wait until their dynamo.export path becomes mature enough.

2 Likes
#3

There is a PT2 support in the unity branch which we would love to continue expand to our usecases. atm we are using it for certain use-cases like stable diffussion

2 Likes
#4

Hi masa and tqchen,

Sorry for the late reply. Thank a lot you for these information! :smiley:

We haven’t read the unity branch detailedly. It might take us more time to have better understanding about the interaction between PyTorch 2.0 and unity branch. We will get back to you if anything unclear to us.

Thank you!