Repository navigation

gemma4: fix 16x audio soft-token magnitude on NPU via single-source scaling - #726

Merged

HenryNdubuaku merged 6 commits into

mainfrom

gemma4-npu-scale-single-source

Jun 16, 2026

Merged

gemma4: fix 16x audio soft-token magnitude on NPU via single-source scaling#726
HenryNdubuaku merged 6 commits into
mainfrom
gemma4-npu-scale-single-source

Commits on Jun 16, 2026
- convert: make gemma4_scale_factor the single source for weight reparam
  ncylich
  committed
  ba95c37
- transpile: reparam gemma4 projection weights for NPU export
  ncylich
  committed
  cf453c4
- Merge remote-tracking branch 'origin/main' into gemma4-npu-scale-single-source
  ncylich
  committed
  3f88749
- Update inference.yml
  ncylich
  committed
  713374e
- ci: create venv for inference tests
  ncylich
  committed
  37aa0a4
- ci: skip inference tests when self-hosted runner is unavailable
  ncylich
  committed
  f4fb02e