Skip to content

Support batch-size-greater-than-one Transformers generation #9

Description

@DaBestCode

Extend the Transformers adapter beyond its current batch-size-one contract without mixing sequence state or random samples across requests.

Acceptance criteria:

  • Define prompt-length, request-ID, and absolute-position state per batch element.
  • Support unequal prompt lengths represented by an attention mask.
  • Preserve independent deterministic sampling per request, layer, and KV head.
  • Test two requests with different prompt lengths through multiple eviction rounds.
  • Test finished-sequence handling and explicit failures for unsupported beam expansion.
  • Preserve current batch-size-one behavior and public call site.
  • Update the README supported-scope table and adapter documentation.

Please outline the intended tensor/state representation in a comment before coding.

Activity

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

Metadata

Metadata

Assignees

No one assigned

    Labels

    area: transformersHugging Face Transformers adapter workenhancementNew feature or requesthelp wantedExtra attention is needed

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions