Fix prompt cache leak between conversations#1039
Closed
kernelpool wants to merge 2 commits intoml-explore:mainfrom
Closed
Fix prompt cache leak between conversations#1039kernelpool wants to merge 2 commits intoml-explore:mainfrom
kernelpool wants to merge 2 commits intoml-explore:mainfrom
Conversation
Member
|
Thanks for flagging this. I am working on revamping the batch generation in general to support multiple checkpoints and ensure the cache key always contains all the processed tokens etc. |
Member
|
Should be fixed in #1072 . Feel free to test it and close it is fine on my testing at least. |
Contributor
Author
|
Ran some tests, LGTM! |
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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
I ran into an issue where KV caches leak data into new conversations (using GLM-5), and believe this addresses it.
After the checkpoint save, the remaining prompt tokens were fed through the model twice (once for prefill, then all of them again through _step). This inflated the KV cache offset, so trimming it for a new conversation would leave stale tokens from the previous one.
Example: