-
Notifications
You must be signed in to change notification settings - Fork 25.6k
enable data type propagation #98065
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
enable data type propagation #98065
Conversation
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/98065
Note: Links to docs will display an error until the docs builds have been completed. ❌ 2 FailuresAs of commit dc5731a: BROKEN TRUNK - The following jobs failed but were present on the merge base 148d492:👉 Rebase onto the `viable/strict` branch to avoid these failures
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
[ghstack-poisoned]
Attach data type information inside scheduling nodes. Just leverage CI to check whether this attaching will impact UT now. Plan to do: (1) Understand and resolve data type propagation related to "masked". (2) Add debug print to check the data types after data type propagation. (3) Write UT based on (1). cc soumith voznesenskym penguinwu anijain2305 EikanWang jgong5 Guobing-Chen XiaobingSuper blzheng Xia-Weiwen wenzhe-nrv jiayisunx peterbell10 desertfire [ghstack-poisoned]
Attach data type information inside scheduling nodes. Just leverage CI to check whether this attaching will impact UT now. Plan to do: (1) Understand and resolve data type propagation related to "masked". (2) Add debug print to check the data types after data type propagation. (3) Write UT based on (1). cc soumith voznesenskym penguinwu anijain2305 EikanWang jgong5 Guobing-Chen XiaobingSuper blzheng Xia-Weiwen wenzhe-nrv jiayisunx peterbell10 desertfire [ghstack-poisoned]
@pytorchbot rebase |
@pytorchbot successfully started a rebase job. Check the current status here |
Attach data type information inside scheduling nodes. Just leverage CI to check whether this attaching will impact UT now. Plan to do: (1) Understand and resolve data type propagation related to "masked". (2) Add debug print to check the data types after data type propagation. (3) Write UT based on (2). cc soumith voznesenskym penguinwu anijain2305 EikanWang jgong5 Guobing-Chen XiaobingSuper blzheng Xia-Weiwen wenzhe-nrv jiayisunx peterbell10 desertfire [ghstack-poisoned]
Successfully rebased |
Attach data type information inside scheduling nodes. Just leverage CI to check whether this attaching will impact UT now. Plan to do: (1) Understand and resolve data type propagation related to "masked". (2) Add debug print to check the data types after data type propagation. (3) Write UT based on (2). cc soumith voznesenskym penguinwu anijain2305 EikanWang jgong5 Guobing-Chen XiaobingSuper blzheng Xia-Weiwen wenzhe-nrv jiayisunx peterbell10 desertfire [ghstack-poisoned]
Attach data type information inside scheduling nodes. Just leverage CI to check whether this attaching will impact UT now. Plan to do: (1) Understand and resolve data type propagation related to "masked". (2) Add debug print to check the data types after data type propagation. (3) Write UT based on (2). cc soumith voznesenskym penguinwu anijain2305 EikanWang jgong5 Guobing-Chen XiaobingSuper blzheng Xia-Weiwen wenzhe-nrv jiayisunx peterbell10 desertfire [ghstack-poisoned]
Have you validated this PR with E2E models? |
Enable data type propagation in schedule node level. Propagation policy: (1) ops with dtype args [constant, load, rand, randn] -> direct use dtype as node dtype (2) ops semantics decide output dtype -> using output dtype All `override_return_dtype` in https://github.com/pytorch/pytorch/blob/master/torch/_inductor/lowering.py. (3) other ops: perform promote on input nodes dtype. ADD(BF16, FP32) -> FP32 output. cc soumith voznesenskym penguinwu anijain2305 EikanWang jgong5 Guobing-Chen XiaobingSuper blzheng Xia-Weiwen wenzhe-nrv jiayisunx peterbell10 desertfire [ghstack-poisoned]
Enable data type propagation in schedule node level. Propagation policy: (1) ops with dtype args [constant, load, rand, randn] -> direct use dtype as node dtype (2) ops semantics decide output dtype -> using output dtype All `override_return_dtype` in https://github.com/pytorch/pytorch/blob/master/torch/_inductor/lowering.py. (3) other ops: perform promote on input nodes dtype. ADD(BF16, FP32) -> FP32 output. cc soumith voznesenskym penguinwu anijain2305 EikanWang jgong5 Guobing-Chen XiaobingSuper blzheng Xia-Weiwen wenzhe-nrv jiayisunx peterbell10 desertfire [ghstack-poisoned]
Enable data type propagation in schedule node level. Propagation policy: (1) ops with dtype args [constant, load, rand, randn] -> direct use dtype as node dtype (2) ops semantics decide output dtype -> using output dtype All `override_return_dtype` in https://github.com/pytorch/pytorch/blob/master/torch/_inductor/lowering.py. (3) other ops: perform promote on input nodes dtype. ADD(BF16, FP32) -> FP32 output. cc soumith voznesenskym penguinwu anijain2305 EikanWang jgong5 Guobing-Chen XiaobingSuper blzheng Xia-Weiwen wenzhe-nrv jiayisunx peterbell10 desertfire [ghstack-poisoned]
@pytorchbot merge |
Merge startedYour 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 |
Merge failedReason: Command
Details for Dev Infra teamRaised by workflow job |
Enable data type propagation in schedule node level. Propagation policy: (1) ops with dtype args [constant, load, rand, randn] -> direct use dtype as node dtype (2) ops semantics decide output dtype -> using output dtype All `override_return_dtype` in https://github.com/pytorch/pytorch/blob/master/torch/_inductor/lowering.py. (3) other ops: perform promote on input nodes dtype. ADD(BF16, FP32) -> FP32 output. cc soumith voznesenskym penguinwu anijain2305 EikanWang jgong5 Guobing-Chen XiaobingSuper blzheng Xia-Weiwen wenzhe-nrv jiayisunx peterbell10 desertfire [ghstack-poisoned]
@zhuhaozhe , please check the failed cases. |
@pytorchbot rebase |
@pytorchbot successfully started a rebase job. Check the current status here |
Enable data type propagation in schedule node level. Propagation policy: (1) ops with dtype args [constant, load, rand, randn] -> direct use dtype as node dtype (2) ops semantics decide output dtype -> using output dtype All `override_return_dtype` in https://github.com/pytorch/pytorch/blob/master/torch/_inductor/lowering.py. (3) other ops: perform promote on input nodes dtype. ADD(BF16, FP32) -> FP32 output. cc soumith voznesenskym penguinwu anijain2305 EikanWang jgong5 Guobing-Chen XiaobingSuper blzheng Xia-Weiwen wenzhe-nrv jiayisunx peterbell10 desertfire [ghstack-poisoned]
Successfully rebased |
Enable data type propagation in schedule node level. Propagation policy: (1) ops with dtype args [constant, load, rand, randn] -> direct use dtype as node dtype (2) ops semantics decide output dtype -> using output dtype All `override_return_dtype` in https://github.com/pytorch/pytorch/blob/master/torch/_inductor/lowering.py. (3) other ops: perform promote on input nodes dtype. ADD(BF16, FP32) -> FP32 output. cc soumith voznesenskym penguinwu anijain2305 EikanWang jgong5 Guobing-Chen XiaobingSuper blzheng Xia-Weiwen wenzhe-nrv jiayisunx peterbell10 desertfire [ghstack-poisoned]
@pytorchbot merge |
Merge startedYour 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 |
Stack from ghstack (oldest at bottom):
Enable data type propagation in schedule node level.
Propagation policy:
(1) ops with dtype args [constant, load, rand, randn] -> direct use dtype as node dtype
(2) ops semantics decide output dtype -> using output dtype
All
override_return_dtype
in https://github.com/pytorch/pytorch/blob/master/torch/_inductor/lowering.py.(3) other ops: perform promote on input nodes dtype. ADD(BF16, FP32) -> FP32 output.
cc @soumith @voznesenskym @penguinwu @anijain2305 @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @blzheng @Xia-Weiwen @wenzhe-nrv @jiayisunx @peterbell10 @desertfire