Make relationship between outputs and inputs of auto_functionalize static and not depending on addresses of new inputs passed #134315
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
…atic and not depending on new inputs passed [ghstack-poisoned]
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/134315
Note: Links to docs will display an error until the docs builds have been completed. ❌ 35 New Failures, 1 Unrelated FailureAs of commit eb59152 with merge base 938f37b ( NEW FAILURES - The following jobs have failed:
BROKEN TRUNK - The following job failed but were present on the merge base:👉 Rebase onto the `viable/strict` branch to avoid these failures
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
laithsakka
changed the title
Contributor
Author
|
ah why :
|
…ionalize static and not depending on new inputs passed" In the previous diff, auto_functionalize used to have the form ``` [x', y'] = auto_functionalize( foo, x, y, _arg0_base=x, _arg1_base = y, _all_bases=[x,y]) ``` Which can be read as auto_functionalize 1) has two outputs len(all_bases) 2) the first output starts with all_bases[0] `x` then it will observe the mutations each arg that have arg_base =`x` In this case the mutation on x. (which is read either from copy(x) that is passed to foo). 3) same for the second output, it will be start with y and observe the mutations of the arg y. Now one problem with the above is if a pass run and decided that y=x then we get : ``` [x', y'] = auto_functionalize( foo, x, x, _arg0_base=x, _arg1_base = x,_ all_bases=[x,x]) ``` This is problematic because 1) x' now start with x and observe both mutations of first and second arg of foo, because _arg0_base=x and _arg1_base = x which changes the semantics of the function. and make it wrong even if do no re-inplace anything! 2) same for y' not it starts with x but observe both mutations. **Solution** Make the relationships between the outputs and the args that we need to observe their mutations static and not changeable. ``` [x', y'] = auto_functionalize( foo, x, y, _all_bases=[x,y], _observe_mutation_from=[[arg0],[arg1]) ``` we use the addresses recorded at the time of the functionlization to know for each base, which args need to have their mutations observed 1) x' output starts with x and observe mutations of arg0 2) y' output starts with and observe mutations of arg1 if we do CSE we get ``` [x', y'] = auto_functionalize( foo, x, x, _all_bases=[x,x], _observe_mutation_from=[[arg0],[arg1]) ``` 1) x' output starts with x and observe mutations of arg0 2) y' output starts with x and observe mutations of arg1 this makes inductor/test_inplacing_pass.py::TestReinplacingPassCorrectness::test_multi_output_intermediate pass. cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx ipiszy yf225 chenyang78 kadeng muchulee8 ColinPeppler amjames desertfire chauhang rec [ghstack-poisoned]
…ionalize static and not depending on new inputs passed" In the previous diff, auto_functionalize used to have the form ``` [x', y'] = auto_functionalize( foo, x, y, _arg0_base=x, _arg1_base = y, _all_bases=[x,y]) ``` Which can be read as auto_functionalize 1) has two outputs len(all_bases) 2) the first output starts with all_bases[0] `x` then it will observe the mutation from each arg that have arg_base =`x` In this case the mutation on arg0. 3) same for the second output, it will be start with y and observe the mutations of the arg1. Now one problem with the above is if a pass run and decided that y=x then we get : ``` [x', y'] = auto_functionalize( foo, x, x, _arg0_base=x, _arg1_base = x,_ all_bases=[x,x]) ``` This is problematic because 1) x' now start with x and observe both mutations of first and second arg of foo, because _arg0_base=x and _arg1_base = x which changes the semantics of the function. and make it wrong even if do not re-inplace anything! 2) same for y' it starts with x but observe both mutations. **Solution** Make the relationships between the outputs and the args that we need to observe their mutations static and not changeable. ``` [x', y'] = auto_functionalize( foo, x, y, _all_bases=[x,y], _observe_mutation_from=[[arg0],[arg1]) ``` we use the addresses recorded at the time of the functionlization to know for each base, which args need to have their mutations observed 1) x' output starts with x and observe mutations of arg0 2) y' output starts with and observe mutations of arg1 if we do CSE we get ``` [x', y'] = auto_functionalize( foo, x, x, _all_bases=[x,x], _observe_mutation_from=[[arg0],[arg1]) ``` 1) x' output starts with x and observe mutations of arg0 2) y' output starts with x and observe mutations of arg1 this makes inductor/test_inplacing_pass.py::TestReinplacingPassCorrectness::test_multi_output_intermediate pass. cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx ipiszy yf225 chenyang78 kadeng muchulee8 ColinPeppler amjames desertfire chauhang rec [ghstack-poisoned]
laithsakka
changed the title
…ionalize static and not depending on addresses of new inputs passed" In the previous diff, auto_functionalize used to have the form ``` [x', y'] = auto_functionalize( foo, x, y, _arg0_base=x, _arg1_base = y, _all_bases=[x,y]) ``` Which can be read as auto_functionalize 1) has two outputs len(all_bases) 2) the first output starts with all_bases[0] `x` then it will observe the mutation from each arg that have arg_base =`x` In this case the mutation on arg0. 3) same for the second output, it will be start with y and observe the mutations of the arg1. Now one problem with the above is if a pass run and decided that y=x then we get : ``` [x', y'] = auto_functionalize( foo, x, x, _arg0_base=x, _arg1_base = x,_ all_bases=[x,x]) ``` This is problematic because 1) x' now start with x and observe both mutations of first and second arg of foo, because _arg0_base=x and _arg1_base = x which changes the semantics of the function. and make it wrong even if do not re-inplace anything! 2) same for y' it starts with x but observe both mutations. **Solution** Make the relationships between the outputs and the args that we need to observe their mutations static and not changeable. ``` [x', y'] = auto_functionalize( foo, x, y, _all_bases=[x,y], _observe_mutation_from=[[arg0],[arg1]) ``` we use the addresses recorded at the time of the functionlization to know for each base, which args need to have their mutations observed 1) x' output starts with x and observe mutations of arg0 2) y' output starts with and observe mutations of arg1 if we do CSE we get ``` [x', y'] = auto_functionalize( foo, x, x, _all_bases=[x,x], _observe_mutation_from=[[arg0],[arg1]) ``` 1) x' output starts with x and observe mutations of arg0 2) y' output starts with x and observe mutations of arg1 this makes inductor/test_inplacing_pass.py::TestReinplacingPassCorrectness::test_multi_output_intermediate pass. cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx ipiszy yf225 chenyang78 kadeng muchulee8 ColinPeppler amjames desertfire chauhang rec [ghstack-poisoned]
Contributor
Author
|
closed for the favor of #134409 |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Stack from ghstack (oldest at bottom):
In the previous diff, auto_functionalize used to have the form
Which can be read as auto_functionalize
xthen it will observe the mutation from each arg that have arg_base =xIn this case the mutation on arg0.
Now one problem with the above is if a pass run and decided that y=x
then we get :
This is problematic because
Solution
Make the relationships between the outputs and the args that we need to observe their mutations static and not
changeable.
we use the addresses recorded at the time of the functionlization to know for each base, which args need to have their mutations observed
if we do CSE we get
this makes inductor/test_inplacing_pass.py::TestReinplacingPassCorrectness::test_multi_output_intermediate
pass.
cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @yf225 @chenyang78 @kadeng @muchulee8 @ColinPeppler @amjames @desertfire @chauhang @rec