[AOTI] Add standalone version of TORCH_CHECK #136873
Conversation
Summary: Use TORCH_CHECK_STD_ERROR which throws std::runtime_error, instead of using TORCH_CHECK which throws c10::Error, to cut dependency on libtorch. Specifically, Inductor CPU backend may generate cpp code that calls ATen vectorization ops and we need to make sure those ops don't call TORCH_CHECK. [ghstack-poisoned]
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/136873
Note: Links to docs will display an error until the docs builds have been completed. ✅ No FailuresAs of commit db42a23 with merge base d1b87e2 ( This comment was automatically generated by Dr. CI and updates every 15 minutes. |
pytorch-bot
bot
added
the
module: cpu
Summary: Use TORCH_CHECK_STD_ERROR which throws std::runtime_error, instead of using TORCH_CHECK which throws c10::Error, to cut dependency on libtorch. Specifically, Inductor CPU backend may generate cpp code that calls ATen vectorization ops and we need to make sure those ops don't call TORCH_CHECK. cc jgong5 mingfeima XiaobingSuper sanchitintel ashokei jingxu10 [ghstack-poisoned]
Summary: Use TORCH_CHECK_STD_ERROR which throws std::runtime_error, instead of using TORCH_CHECK which throws c10::Error, to cut dependency on libtorch. Specifically, Inductor CPU backend may generate cpp code that calls ATen vectorization ops and we need to make sure those ops don't call TORCH_CHECK. cc jgong5 mingfeima XiaobingSuper sanchitintel ashokei jingxu10 [ghstack-poisoned]
Summary: Use TORCH_CHECK_STD_ERROR which throws std::runtime_error, instead of using TORCH_CHECK which throws c10::Error, to cut dependency on libtorch. Specifically, Inductor CPU backend may generate cpp code that calls ATen vectorization ops and we need to make sure those ops don't call TORCH_CHECK. cc jgong5 mingfeima XiaobingSuper sanchitintel ashokei jingxu10 [ghstack-poisoned]
albanD
left a comment
albanD
left a comment
There was a problem hiding this comment.
I'm a bit surprised by this change. c10::Error gives a lot of niceties like stack traces tracking, internal logging, TORCH_SHOW_CPP_STACKTRACES, etc
Also this is going to change the error we give to the python users?
I think we have to give up some properties if we want to make Vectorized ABI-compatible. The first attempt I made was to make c10::Error header-only, but it is not easy with something like Lines 61 to 63 in 2b329d3 |
|
Making this particular call go through the capi shim should be relatively simple though given the types involved right? |
If we give up on supporting varargs for this one, then yes. |
|
Ho we were actually chatting about this one with Jane earlier, wouldn't we be able to do the conversion vararg -> string in a header-only c++ layer and make the raw c api take a single raw string? |
|
AOTI_TORCH_CHECK already does that. I could use that for now, but there will still be a need to make these utility Vectorized implementation self-contained, i.e. libtorch independent. What about if I define |
Summary: Use TORCH_CHECK_STD_ERROR which throws std::runtime_error, instead of using TORCH_CHECK which throws c10::Error, to cut dependency on libtorch. Specifically, Inductor CPU backend may generate cpp code that calls ATen vectorization ops and we need to make sure those ops don't call TORCH_CHECK. cc jgong5 mingfeima XiaobingSuper sanchitintel ashokei jingxu10 [ghstack-poisoned]
Summary: In the standalone mode, TORCH_CHECK throws std::runtime_error, instead of c10::Error. The goal is to cut dependency on libtorch. Specifically, AOTI generates CPU code which may call ATen vectorization ops and we need to make sure those ops are self-contained. [ghstack-poisoned]
desertfire
changed the title
Summary: In the standalone mode, TORCH_CHECK throws std::runtime_error, instead of c10::Error. The goal is to cut dependency on libtorch. Specifically, AOTI generates CPU code which may call ATen vectorization ops and we need to make sure those ops are self-contained. ghstack-source-id: 5707c8a Pull Request resolved: #136873
albanD
left a comment
albanD
left a comment
There was a problem hiding this comment.
Sounds ok as a temporary unblock, let's discuss offline for broader plan and how to ensure this is preserved.
|
@desertfire has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator. |
Summary: In the standalone mode, TORCH_CHECK throws std::runtime_error, instead of c10::Error. The goal is to cut dependency on libtorch. Specifically, AOTI generates CPU code which may call ATen vectorization ops and we need to make sure those ops are self-contained. Differential Revision: [D63911928](https://our.internmc.facebook.com/intern/diff/D63911928) cc jgong5 mingfeima XiaobingSuper sanchitintel ashokei jingxu10 voznesenskym penguinwu EikanWang Guobing-Chen zhuhaozhe blzheng wenzhe-nrv jiayisunx ipiszy yf225 chenyang78 kadeng muchulee8 ColinPeppler amjames chauhang [ghstack-poisoned]
Summary: In the standalone mode, TORCH_CHECK throws std::runtime_error, instead of c10::Error. The goal is to cut dependency on libtorch. Specifically, AOTI generates CPU code which may call ATen vectorization ops and we need to make sure those ops are self-contained. ghstack-source-id: 91649dd Pull Request resolved: #136873
|
@desertfire has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator. |
|
@pytorchbot merge -f 'Landed internally' (Initiating merge automatically since Phabricator Diff has merged, using force because this PR might not pass merge_rules.json but landed internally) |
Merge startedYour change will be merged immediately since you used the force (-f) flag, bypassing any CI checks (ETA: 1-5 minutes). Please use Learn more about merging in the wiki. Questions? Feedback? Please reach out to the PyTorch DevX Team |
Summary: In the standalone mode, TORCH_CHECK throws std::runtime_error, instead of c10::Error. The goal is to cut dependency on libtorch. Specifically, AOTI generates CPU code which may call ATen vectorization ops and we need to make sure those ops are self-contained. ghstack-source-id: 6dab1f7 Pull Request resolved: pytorch/pytorch#136873
Stack from ghstack (oldest at bottom):
Summary: In the standalone mode, TORCH_CHECK throws std::runtime_error, instead of c10::Error. The goal is to cut dependency on libtorch. Specifically, AOTI generates CPU code which may call ATen vectorization ops and we need to make sure those ops are self-contained.
Differential Revision: D63911928
cc @jgong5 @mingfeima @XiaobingSuper @sanchitintel @ashokei @jingxu10 @voznesenskym @penguinwu @EikanWang @Guobing-Chen @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @yf225 @chenyang78 @kadeng @muchulee8 @ColinPeppler @amjames @chauhang