Skip to content

Ensure prefixes cannot be cached for synthetic datasets #186

@markurtz

Description

@markurtz

For the SyntheticTextItemsGenerator, we need to add a controllably different prefix or start to each request to ensure automatic prefix caching within vLLM is never hit. Specifically for the create_prompt pathway, we need to set the start text as an auto-incrementing number, for example, for the requested number of items.

The validating test is that for each request made out of the total requests generated, the string match starting from the first character will always fail across requests and will not be equal. The end to end test is running a benchmark with the latest vLLM version and checking the prefix caching hit rate, which should evaluate to 0%

Metadata

Metadata

Assignees

Labels

internalfiled by core contributor or associate

Projects

No projects

Milestone

Relationships

None yet

Development

No branches or pull requests

Issue actions