Skip to content

Clarify How Recommendation Logic Works and discuss more flexible recommendation #29

@toslali-ibm

Description

@toslali-ibm

It would be good to discuss and document and improve how the recommendation system currently works, including the matching logic and limitations.

Points to Clarify

  • Does the system require an exact match for prompt_length and output_length?
  • For QPS: is it true that it does not require exact QPS matching today?
  • Confirm that there is no interpolation between prompt/output ranges.
  • If a (prompt_len, output_len) pair doesn't exist, we simply fail or return nothing.
  • What happens when an exact (prompt_len, output_len) combination exists for multiple QPS values?
  • What is the fallback behavior when there is no match at all?

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions