Repository navigation
Match a cached llama.cpp context only to its live session - #423
Merged
Merged
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
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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 PR fixes a case where a new session could reuse the KV cache of a session that was already freed.
LlamaLanguageModelkeeps one llama.cpp context for a session, keyed by the session'sObjectIdentifier. When a session is freed, the next one can get the same address, so it matched the freed session's context and reported the reused prompt prefix as cached input tokens on its first request.The cached context now holds its session weakly and matches only that session while it's alive, as the MLX backend's KV store already does. A new session's first request decodes its full prompt and reports no cached tokens.