mlcommons / mlcommons/endpoints
[Feature]: cache salting for multi-turn
Nobody has claimed this yet.
- Dominant language
- Python
- Stars
- 21
- Forks
- 28
- Avg merge
- 3d 17h
- Merged PRs (30d)
- 13
Description
Motivation
Multi-turn conversation relies on kv reuse/prefix-caching/radix attention to reduce the cost of long prefill from increasing context length, but allowing full kv reuse between the exact same sample (e.g.: between dataset copies) is not the intended behavior.
Proposed Solution
Add cache salt to identical samples.
Alternatives Considered
No response
Additional Context
No response
Contributor guide
First steps
- Read the whole issue, then the project's contributing guide.
- Comment on the issue to say you are picking it up — it saves two people doing the same work.
- Fork the repository and make your change on a branch.
- Open a pull request that references the issue number.
Research direction
The issue names no files or tests; start by locating the multi-turn KV reuse, prefix-caching, or radix-attention entry points and their existing tests. Clarify how identical samples are identified and define done as preventing full KV reuse between dataset copies while preserving intended multi-turn reuse.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- python
- Domain
- ai, performance
- Issue type
- Feature
- Difficulty
- 5/5
- Estimated time
- Over a week
- Activity status
- Quiet
- Clarity
- Needs clarification
- Newbie friendliness
- 35/100