Update KJT stride calculation logic to be based off of inverse_indices for VBE KJTs. #2949
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
facebook-github-bot
added
the
CLA Signed
|
This pull request was exported from Phabricator. Differential Revision: D74273083 |
jd7-tr
added a commit
to jd7-tr/torchrec
that referenced
this pull request
May 6, 2025
…s for VBE KJTs. (pytorch#2949) Summary: Update the `_maybe_compute_stride_kjt` logic to calculate stride based off of `inverse_indices` for VBE KJTs. Currently, stride of VBE KJT with `stride_per_key_per_rank` is calculated as the max "stride per key". This is different from the batch size of the EBC output KeyedTensor which is based off of inverse_indices. This causes issues in IR module serialization: debug doc. Differential Revision: D74273083
jd7-tr
added a commit
to jd7-tr/torchrec
that referenced
this pull request
May 7, 2025
…s for VBE KJTs. (pytorch#2949) Summary: Update the `_maybe_compute_stride_kjt` logic to calculate stride based off of `inverse_indices` for VBE KJTs. Currently, stride of VBE KJT with `stride_per_key_per_rank` is calculated as the max "stride per key". This is different from the batch size of the EBC output KeyedTensor which is based off of inverse_indices. This causes issues in IR module serialization: debug doc. Differential Revision: D74273083
|
This pull request was exported from Phabricator. Differential Revision: D74273083 |
jd7-tr
added a commit
to jd7-tr/torchrec
that referenced
this pull request
May 7, 2025
…s for VBE KJTs. (pytorch#2949) Summary: Update the `_maybe_compute_stride_kjt` logic to calculate stride based off of `inverse_indices` for VBE KJTs. Currently, stride of VBE KJT with `stride_per_key_per_rank` is calculated as the max "stride per key". This is different from the batch size of the EBC output KeyedTensor which is based off of inverse_indices. This causes issues in IR module serialization: debug doc. Differential Revision: D74273083
jd7-tr
added a commit
to jd7-tr/torchrec
that referenced
this pull request
May 7, 2025
…s for VBE KJTs. (pytorch#2949) Summary: Update the `_maybe_compute_stride_kjt` logic to calculate stride based off of `inverse_indices` for VBE KJTs. Currently, stride of VBE KJT with `stride_per_key_per_rank` is calculated as the max "stride per key". This is different from the batch size of the EBC output KeyedTensor which is based off of inverse_indices. This causes issues in IR module serialization: debug doc. Differential Revision: D74273083
Summary: Pull Request resolved: pytorch#2959 # context * this diff is part of the "variable-batch KJT refactoring" project ([doc](https://fburl.com/gdoc/svfysfai)) * previously the `stride_per_key_per_rank` variable is `List[List[int]] | None` which can't be handled correctly in PT2 IR (torch.export) * this change makes the KJT class variable `_stride_per_key_per_rank` as `torch.IntTensor | None` so it would be compatible with PT2 IR. # equivalency * to check if `self._stride_per_key_per_rank` is `None` this logic is used to differentiate variable_batch case, and should have the same behavior after this diff * to use `self._stride_per_key_per_rank` as `List[List[int]]` most of the callsite use the function to get the list: `def stride_per_key_per_rank(self) -> List[List[int]]:`, and this function is modified to covert the `torch.IntTensor` to list as ` _stride_per_key_per_rank.tolist()`, the results should be the same NOTE: this `self. _stride_per_key_per_rank.tolist()` tensor should always be on CPU since it's effective the meta data of a KJT. For generic torch APIs like `.to(...)`, `record_stream()`, etc. should in general avoid altering this variable. Differential Revision: D74366343 Reviewed By: jd7-tr
…s for VBE KJTs. (pytorch#2949) Summary: Pull Request resolved: pytorch#2949 Update the `_maybe_compute_stride_kjt` logic to calculate stride based off of `inverse_indices` for VBE KJTs. Currently, stride of VBE KJT with `stride_per_key_per_rank` is calculated as the max "stride per key". This is different from the batch size of the EBC output KeyedTensor which is based off of inverse_indices. This causes issues in IR module serialization: debug doc. Differential Revision: D74273083
|
This pull request was exported from Phabricator. Differential Revision: D74273083 |
# for free
to join this conversation on GitHub.
Already have an account?
# 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:
Update the
_maybe_compute_stride_kjtlogic to calculate stride based off ofinverse_indicesfor VBE KJTs.Currently, stride of VBE KJT with
stride_per_key_per_rankis calculated as the max "stride per key". This is different from the batch size of the EBC output KeyedTensor which is based off of inverse_indices. This causes issues in IR module serialization: debug doc.Differential Revision: D74273083