[decompositions] add decomposition for RNN with packed sequence

[samdow]

Stack from ghstack (oldest at bottom):

• [decompositions] GRU decompositon with and without packed sequence #91466
• [decompositions] LSTM with packed input #91465
• -> [decompositions] add decomposition for RNN with packed sequence #91281
• [decompositions] add LSTM decomp #91124
• [expanded weights] add RNN support via decomp #91807
• [decompositions] add RNN decomp and testing #91123

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/91281

• 📄 Preview Python docs built from this PR
• 📄 Preview C++ docs built from this PR
• ❓ Need help or want to give feedback on the CI? Visit the bot commands wiki or our office hours

Note: Links to docs will display an error until the docs builds have been completed.

❌ 1 Failures

As of commit 727ab58:

NEW FAILURES - The following jobs have failed:

• linux-bionic-py3_8-clang8-xla / test (xla, 1, 1, linux.4xlarge)

This comment was automatically generated by Dr. CI and updates every 15 minutes.

[zou3519]

A bit confused about what's going on here. Why is this necessary for LSTM? (It isn't in the RNN decomp). If it is necessary, then should this be in the previous PR?

[samdow]

Yeah so what's happening here is that all the decompositions are using _rnn_helper

But that was baking in an assumption that this was going to be a tensor. This basically replaces L2141 because I didn't really think ahead to PackedSequences when I was doing this

[zou3519]

Thanks for pointing that out, I missed that line

[zou3519]

LGTM, PackedSequence is wild. Would prefer some more test cases (e.g. batch_size > 2). Also It would be good to actually compute the per sample grads and do the comparison instead of the smoke test (I'm not sure how much work this is, if you don't have bandwidth then please file an issue and we'll get to it)

[ezyang]

We need someone else to shepherd this right