Skip to content

aging-LFU eviction weights, persistent weights, and warm prefixes - #518

Closed
auryn-macmillan wants to merge 1 commit into
theroyallab:mainfrom
auryn-macmillan:prompt-cache-aging-lfu
Closed

auryn-macmillan wants to merge 1 commit into
theroyallab:mainfrom
auryn-macmillan:prompt-cache-aging-lfu

Conversation

@auryn-macmillan

@auryn-macmillan auryn-macmillan commented Nov 10, 2026 •

Copy link
Copy Markdown

mistakenly opened
Superseded by #519

- Add memory.aging_lfu and the aging_* knobs to MemoryConfig and pass
  them to AsyncGenerator, so exllamav3 can score eviction candidates by
  decaying reuse credit instead of raw recency.
- Add memory.warm_prefixes / warm_prefix_token_ids and prefill them in a
  background task after load, before the first request, so their pages
  (and weights) are resident again after a restart.
- Replay prefixes captured from earlier cache hits by the engine
  (pagetable.stored_prefixes()) within memory.warm_prefix_max_tokens, in
  most-recently-written order, so a restart does not spend minutes
  re-prefilling stale chains before serving.
- Cancel a running warmup on unload so no job outlives its generator,
  and report the warmed token total.
@auryn-macmillan auryn-macmillan changed the title Prompt cache: aging-LFU eviction weights, persistent weights, and warm prefixes aging-LFU eviction weights, persistent weights, and warm prefixes Nov 10, 2026
@auryn-macmillan

auryn-macmillan commented Nov 10, 2026 •

Copy link
Copy Markdown
Author

mistakenly opened by meat proxy
Superseded by #519

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant