Skip to content

[Feature Request] Need for ability to save partial results #227

@thameem-abbas

Description

@thameem-abbas

As it stands, when run with multiple values to the rate parameter, there is nothing written to disk until the end of all the rates. This causes a potential for loss of all results when the higher concurrencies might potentially run into some error. Some checkpointing to the disk would greatly improve the usability of the tool when running into potential errors.

The following is another scenario that we regularly encounter

  1. We run a list of concurrencies.
  2. We observe KV Cache depletion on the server side.
  3. The latter(higher) concurrencies are of no real use. The user cannot choose to end the run. The runs have to complete to get any usable results.

Metadata

Metadata

Assignees

No one assigned

    Labels

    internalfiled by core contributor or associate

    Projects

    No projects

    Milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions