[DeepSeek] Potential memory bug for noaux_tc? #1030
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
This is more of a question since I'm not sure if what I'm doing is valid. However, I noticed that `noaux_tc` uses a lot of memory for DSV3. Specifically, the first step suceeded but the run OOMs on the second step -- this is surprising since we use a constant sequence length and if the first step succeeds, in theory the rest should as well. Adding `detach()` to the `scores` here fixed the OOM and the training curves / grad norms look ~identical for a small run I tried, but I'm sure what I'm doing is valid.
facebook-github-bot
added
the
CLA Signed
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Outdated suggestions cannot be applied.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.
This is more of a question since I'm not sure if what I'm doing is valid. However, I noticed that
noaux_tcuses a lot of memory for DSV3. Specifically, the first step suceeded but the run OOMs on the second step -- this is surprising since we use a constant sequence length and if the first step succeeds, in theory the rest should as well.Do we need some
torch.no_grad()calls when computing the topk indicies similar to what we do inmoe_forward()? Addingdetach()to thescoreslike this PR does fixed the OOM and the training curves / grad norms look ~identical for a small run I tried, but I'm sure what I'm doing is valid.cc @kwen2501 @lessw2020