Skip to content

Board-verify int8 FunctionGemma (numeric quality, speed, RAM) #13

Description

@michalharakal

int8 weight-only quant (GEMMA_QUANT=int8) is export-verified on CPU (transformers PR #245). On-device, confirm:

  • Numeric quality of per-row symmetric int8 from Q5_K (oracle token-for-token check).
  • Decode speed vs bf16.
  • RAM claim: irpa 831 → ~415 MiB on the 1.9 GB board.
  • iree-compile (Torq-fork) accepts the in-graph dequant (i8→f32 × scale).

Tracker: sl2610-function-calling/docs/GEMMA-KV-INT8.md.

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

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions