torchrec support on kvzch emb lookup module #4035
Open
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: change list 1. add bucket concept into ssd tbe 2. update split_embedding_weights to make it return a tuple of 3 tensors(weight, weight_id, id_cnt_per_bucket) 3. add new ut for the key value embedding cases Differential Revision: D73274786
Summary: # Change logs 1. add ZeroCollisionKeyValueEmbedding emb lookup 2. address existing unit test missing for ssd offloading 3. add new ut for kv zch embedding module 4. add a temp hack solution for calculate bucket metadata 5. embedding updates, details illustrated below ####################################################################### ########################### embedding.py updates ########################## ####################################################################### 1. keep the original idea to init shardedTensor during training init 2. for kv zch table, the shardeTensor will be init using virtual size for metadata calculation, and skip actual tensor size check for ST init, this is needed as during training init, the table has 0 rows 3. the new tensor, weight_id will not be registered in the EC becuase its shape is changing in realtime, the weight_id tensor will be generated in post_state_dict hooks 4. the new tensor, bucket could be registered and preserved, but in this diff we keep it the same way as weight_id 5. in post state dict hook, we call get_named_split_embedding_weights_snapshot to get Tuple[table_name, weight(ST), weight_id(ST), bucket(ST)], all 3 tensors are return in the format of ST, and we will update destination with the returned ST directly 6. in pre_load_state_dict_hook, which is called upon load_state_dict(), we will skip all 3 tensors update, because the tensor assignment is done [on the nn.module side](https://fburl.com/code/it5nior8), which doesn't support updating KVT through PMT. This is fine for now because, checkpoint loading will be done outside of the load_state_dict call, but we need future plans to make it work cohesively with other type of tensors Differential Revision: D73567631
✅ Deploy Preview for pytorch-fbgemm-docs ready!
To edit notification comments on pull requests, go to your Netlify site configuration. |
|
This pull request was exported from Phabricator. Differential Revision: D73567631 |
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:
Change logs
#######################################################################
########################### embedding.py updates ##########################
#######################################################################
Differential Revision: D73567631