Dynamo.export uses dynamic=True for symbolic tracing

[SherlockNoMad]

Stack from ghstack (oldest at bottom):

• -> Dynamo.export uses dynamic=True for symbolic tracing #91899

cc @mlazos @soumith @voznesenskym @yanboliang @penguinwu @anijain2305 @EikanWang @jgong5 @Guobing-Chen @chunyuan-w @XiaobingSuper @zhuhaozhe @blzheng @Xia-Weiwen @wenzhe-nrv @jiayisunx @desertfire

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/91899

• 📄 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 Active SEVs

There are 1 currently active SEVs. If your PR is affected, please view them below:

• win-vs2019-cpu-py3 / build workflows failing consistently with linker crash

✅ No Failures

As of commit 4f625b1:
💚 Looks good so far! There are no failures yet. 💚

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

[ezyang]

ugh, why doesn't export just take dynamic directly lol

[SherlockNoMad]

ugh, why doesn't export just take dynamic directly lol

We should, and this will be an API change for dynamo.export()... and maybe we should do this sooner than latter.

Also, is fake tracing mode still a thing after the Symbolic shape branch is merged. IIUC, we should end up with only two mode, static(real tracing mode) vs. dynamic (symbolic tracing mode) , right?

[SherlockNoMad]

@pytorchbot merge

[ezyang]

I would like to get rid of real tracing mode (where you actually run the tensor compute) entirely, but IIRC last time I tried to get rid of it there were... problems

[pytorchmergebot]

Merge started

Your change will be merged once all checks pass (ETA 0-4 Hours).

Learn more about merging in the wiki.

Questions? Feedback? Please reach out to the PyTorch DevX Team

Advanced Debugging

Check the merge workflow status
here

[SherlockNoMad]

I would like to get rid of real tracing mode (where you actually run the tensor compute) entirely, but IIRC last time I tried to get rid of it there were... problems

I think there are still merits of real tracing mode.
In the case of custom op without meta implementation, (or its meta implementation doesn't support symbolic shape), real tracing mode would still work, I am not sure what would happen if it were traced with symbolic mode?

What should be our policy on custom op being traceable?