fix: don't double-count cached tokens when deciding to compact context - #28593
Open
Classic298 wants to merge 1 commit into
Open
fix: don't double-count cached tokens when deciding to compact context#28593Classic298 wants to merge 1 commit into
Classic298 wants to merge 1 commit into
Conversation
Context compaction fired at roughly half the configured threshold on backends that report prompt caching. With a 70k threshold, a chat showing 39k tokens in the UI was already being compacted, and the more the cache helped, the earlier it happened. llama.cpp's `timings` are merged into `usage`, so the same dict carries both `prompt_tokens` and `cache_n`, where `prompt_tokens` already includes `cache_n` (38153 cached + 43 new = 38196 prompt). The threshold check added `cache_n` on top of that, counting every cached token twice. Dropping the extra `cache_n` term makes the backend agree with the number the UI shows, which is what the threshold is configured against. Fixes open-webui#28590
11 tasks
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
1 participant
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Context compaction fired at roughly half the configured threshold on backends that report prompt caching. With a 70k threshold, a chat showing 39k tokens in the UI was already being compacted, and the more the cache helped, the earlier it happened.
llama.cpp's
timingsare merged intousage, so the same dict carries bothprompt_tokensandcache_n, whereprompt_tokensalready includescache_n(38153 cached + 43 new = 38196 prompt). The threshold check addedcache_non top of that, counting every cached token twice.Dropping the extra
cache_nterm makes the backend agree with the number the UI shows, which is what the threshold is configured against.Fixes #28590
Contributor License Agreement
Note
Deleting the CLA section will lead to immediate closure of your PR and it will not be merged in.