[dynamo] warn instead of error on fullgraph=True fallback to eager#181940
Conversation
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/181940
Note: Links to docs will display an error until the docs builds have been completed. ❗ 1 Active SEVsThere are 1 currently active SEVs. If your PR is affected, please view them below: ✅ You can merge normally! (1 Unrelated Failure)As of commit b1fdd17 with merge base 58e3fb0 ( FLAKY - The following job failed but was likely due to flakiness present on trunk:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
Lucaskabela
left a comment
There was a problem hiding this comment.
For more context, this change surfaced an error in vLLM - we have patched it, but it is very likely it surfaces errors in other frameworks too so I think that underscores the criticality of including in this release
|
#181247 for reference |
malfet
left a comment
There was a problem hiding this comment.
This is probably looks fine, but we should reject if similar cherry-pick is proposed for 2.13 release
|
The run is successful with 2.12.0 rc: |
Original issue: #181817
We introduced a new behavior for fullgraph=True in #177809 where an error is thrown when no compiled code is run. However, an internal issue revealed that this new behavior is incompatible with the "eager_on_recompile" stance.
Our solution is that for 2.12, we will issue a warning instead of erroring out if no compiled code is ran under fullgraph=True. We will keep the erroring behavior on trunk as we seek to improve the behavior - i.e. steps 2 and 3 in the original issue.
Future plans: we will error on 2.13 after addressing the other points (2 and 3) in the issue.
cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @kadeng @chauhang @amjames @Lucaskabela @jataylo @azahed98 @mlazos