fix parameter placement #3
Closed
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
Summary: fix tensor placement where the remote device should receive {rank, local_rank}
Differential Revision: D31072120
fbshipit-source-id: 13de19a31a5cafeef280ed7b38a8372a4038fe89
facebook-github-bot
added
CLA Signed
|
This pull request was exported from Phabricator. Differential Revision: D31072120 |
levythu
pushed a commit
to levythu/torchrec
that referenced
this pull request
Sep 13, 2023
Summary: Some bug fixes during the integration test in PyPER O3: ### fix pytorch#1 `_embedding_bag_collection` (`ShardedEmbeddingBagCollection`) is not really called by input_dist (because the same thing is already distributed by ShardedManagedCollisionCollection) . So it never get a chance to initiate `_input_dist`. As a result, TREC pipelining thinks it's not ready for input distribution. This is not expected, since the module is not used in the stage anyway, nor should it be put in fused a2a communication. With this change, https://fburl.com/code/ud8lnixv it'll satisfy the assertion, meanwhile doesn't carry _input_dists so won't be put into fused a2a. ### fix pytorch#2 ManagedCollisionCollection.forward is not traceable because it uses unwarpped `KeyedJaggedTensor.from_jt_dict`. We don't care about its internal detail so just keep it atomic. ### fix pytorch#3 Due to how remap table is set, `MCHManagedCollisionModule` doesn't support i32 id list for now. An easy fix is to convert to i64 regardless. A more memory efficient fix is probably change the remapper to i32 if necessary Differential Revision: D48804332
levythu
pushed a commit
to levythu/torchrec
that referenced
this pull request
Sep 18, 2023
Summary: Some bug fixes during the integration test in PyPER O3: ### fix pytorch#1 `_embedding_bag_collection` (`ShardedEmbeddingBagCollection`) is not really called by input_dist (because the same thing is already distributed by ShardedManagedCollisionCollection) . So it never get a chance to initiate `_input_dist`. As a result, TREC pipelining thinks it's not ready for input distribution. This is not expected, since the module is not used in the stage anyway, nor should it be put in fused a2a communication. With this change, https://fburl.com/code/ud8lnixv it'll satisfy the assertion, meanwhile doesn't carry _input_dists so won't be put into fused a2a. ### fix pytorch#2 ManagedCollisionCollection.forward is not traceable because it uses unwarpped `KeyedJaggedTensor.from_jt_dict`. We don't care about its internal detail so just keep it atomic. ### fix pytorch#3 Due to how remap table is set, `MCHManagedCollisionModule` doesn't support i32 id list for now. An easy fix is to convert to i64 regardless. A more memory efficient fix is probably change the remapper to i32 if necessary Differential Revision: D48804332
levythu
pushed a commit
to levythu/torchrec
that referenced
this pull request
Sep 19, 2023
Summary: Some bug fixes during the integration test in PyPER O3: ### fix pytorch#1 `_embedding_bag_collection` (`ShardedEmbeddingBagCollection`) is not really called by input_dist (because the same thing is already distributed by ShardedManagedCollisionCollection) . So it never get a chance to initiate `_input_dist`. As a result, TREC pipelining thinks it's not ready for input distribution. This is not expected, since the module is not used in the stage anyway, nor should it be put in fused a2a communication. With this change, https://fburl.com/code/ud8lnixv it'll satisfy the assertion, meanwhile doesn't carry _input_dists so won't be put into fused a2a. ### fix pytorch#2 ManagedCollisionCollection.forward is not traceable because it uses unwarpped `KeyedJaggedTensor.from_jt_dict`. We don't care about its internal detail so just keep it atomic. ### fix pytorch#3 Due to how remap table is set, `MCHManagedCollisionModule` doesn't support i32 id list for now. An easy fix is to convert to i64 regardless. A more memory efficient fix is probably change the remapper to i32 if necessary Differential Revision: D48804332
levythu
pushed a commit
to levythu/torchrec
that referenced
this pull request
Sep 21, 2023
Summary: Some bug fixes during the integration test in PyPER O3: ### fix pytorch#1 `_embedding_bag_collection` (`ShardedEmbeddingBagCollection`) is not really called by input_dist (because the same thing is already distributed by ShardedManagedCollisionCollection) . So it never get a chance to initiate `_input_dist`. As a result, TREC pipelining thinks it's not ready for input distribution. This is not expected, since the module is not used in the stage anyway, nor should it be put in fused a2a communication. With this change, https://fburl.com/code/ud8lnixv it'll satisfy the assertion, meanwhile doesn't carry _input_dists so won't be put into fused a2a. ### fix pytorch#2 ManagedCollisionCollection.forward is not traceable because it uses unwarpped `KeyedJaggedTensor.from_jt_dict`. We don't care about its internal detail so just keep it atomic. ### fix pytorch#3 Due to how remap table is set, `MCHManagedCollisionModule` doesn't support i32 id list for now. An easy fix is to convert to i64 regardless. A more memory efficient fix is probably change the remapper to i32 if necessary Differential Revision: D48804332
duduyi2013
pushed a commit
to duduyi2013/torchrec
that referenced
this pull request
Nov 27, 2023
Summary: Some bug fixes during the integration test in PyPER O3: fix pytorch#1 _embedding_bag_collection (ShardedEmbeddingBagCollection) is not really called by input_dist (because the same thing is already distributed by ShardedManagedCollisionCollection) . So it never get a chance to initiate _input_dist. As a result, TREC pipelining thinks it's not ready for input distribution. This is not expected, since the module is not used in the stage anyway, nor should it be put in fused a2a communication. With this change, https://fburl.com/code/ud8lnixv it'll satisfy the assertion, meanwhile doesn't carry _input_dists so won't be put into fused a2a. fix pytorch#2 ManagedCollisionCollection.forward is not traceable because it uses unwarpped KeyedJaggedTensor.from_jt_dict. We don't care about its internal detail so just keep it atomic. fix pytorch#3 Due to how remap table is set, MCHManagedCollisionModule doesn't support i32 id list for now. An easy fix is to convert to i64 regardless. A more memory efficient fix is probably change the remapper to i32 if necessary Differential Revision: D51601041
facebook-github-bot
pushed a commit
that referenced
this pull request
Dec 1, 2023
Summary: Pull Request resolved: #1541 Some bug fixes during the integration test in PyPER O3: # fix #1 _embedding_bag_collection (ShardedEmbeddingBagCollection) is not really called by input_dist (because the same thing is already distributed by ShardedManagedCollisionCollection) . So it never get a chance to initiate _input_dist. As a result, TREC pipelining thinks it's not ready for input distribution. This is not expected, since the module is not used in the stage anyway, nor should it be put in fused a2a communication. With this change, https://fburl.com/code/ud8lnixv it'll satisfy the assertion, meanwhile doesn't carry _input_dists so won't be put into fused a2a. # fix #2 ManagedCollisionCollection.forward is not traceable because it uses unwarpped KeyedJaggedTensor.from_jt_dict. We don't care about its internal detail so just keep it atomic. # fix #3 Due to how remap table is set, MCHManagedCollisionModule doesn't support i32 id list for now. An easy fix is to convert to i64 regardless. A more memory efficient fix is probably change the remapper to i32 if necessary Reviewed By: dstaay-fb Differential Revision: D51601041 fbshipit-source-id: 95cf346b5247f1d5afb6643ecfd7dca4b3c4d575
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
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.
Summary: fix tensor placement where the remote device should receive {rank, local_rank}
Differential Revision: D31072120