Repository navigation

Add EmbeddingGemma 2 export: coreai.embedding.export and registry entry - #369

Open

stikves wants to merge 40 commits into

apple:mainfrom

stikves:sukru/embedding-gemma2-export

Open

Add EmbeddingGemma 2 export: coreai.embedding.export and registry entry#369
stikves wants to merge 40 commits into
apple:mainfrom
stikves:sukru/embedding-gemma2-export

Commits on Oct 23, 2026
- RoPE: public rotate_half in primitives/macos/rope
  stikves
  committed
  374db33
- EmbeddingGemma 2: text and image encoders for GPU and Neural Engine
  stikves
  committed
  6184ee9
- coreai.embedding.export: embedding bundle exporter
  stikves
  committed
  03bb974
- EmbeddingGemma 2: model README and catalog entry
  stikves
  committed
  85db172
Commits on Oct 27, 2026
- Merge branch 'main' into sukru/embedding-gemma2-export
  stikves
  authored
  90e08e1
- Keep ScaledRMSNorm finite in float16 for rows with max-abs below 1/256
  stikves
  committed
  3b15173
Commits on Oct 28, 2026
- Clamp the squared norm in the Neural Engine L2 normalize
  stikves
  committed
  5b5e7f1
- Store the Neural Engine token table in float16
  stikves
  committed
  0ea9996
- Say which encoder the family precision fields describe
  stikves
  committed
  18b2ce6
- Check the float16 Neural Engine text encoder against real weights
  stikves
  committed
  6b00b9c
- Store the macOS text weights once when the bundle has an image tower
  stikves
  committed
  ce3b018
- Merge branch 'main' into sukru/embedding-gemma2-export
  stikves
  authored
  561f019
Commits on Oct 29, 2026
Commits on Oct 30, 2026
- Revert the Muse Glimmer rotate_half move
  stikves
  committed
  a8364dc
- Share the RoPE frequency and 1x1 conv helpers publicly
  stikves
  committed
  188d4bc
- Read the EmbeddingGemma 2 checkpoint in one place and load strictly
  stikves
  committed
  531775f
- Call add_function directly in the embedding exporters
  stikves
  committed
  a57aacd
- Look up the RoPE head dims through the text config
  stikves
  committed
  55ae6ad
- Promote the Gemma 4 max-abs RMSNorm to primitives/ios
  stikves
  committed
  b1f93de
- Reshape each iOS text projection to per-head layout once
  stikves
  committed
  ad81f5e
- Share the macOS attention and sandwich layer between text and image
  stikves
  committed
  2f2e086
- Drop a rationale from a section comment in the macOS image encoder
  stikves
  committed
  377f503
- Fold overlapping embedding schema tests and reuse the shared tiny configs
  stikves
  committed
  d7a56f2
- Trim the EmbeddingGemma 2 model tests to one path per check
  stikves
  committed
  3559d62
- Clarify embedding helper names and derive bundle assets from the pipelines
  stikves
  committed
  5b3db09
- Make the base.py safetensors helpers public
  stikves
  committed
  ac7fe2a
- Read EmbeddingGemma 2 checkpoints through the base.py safetensors helpers
  stikves
  committed
  d979627
- Restore the EmbeddingGemma 2 test guards dropped by the trims
  stikves
  committed
  6f29e93
- Test the Neural Engine embedding conversion on tiny configs
  stikves
  committed
  49f7336
- Record each embedding asset once and the component output dtypes
  stikves
  committed
  1915334
- Tighten the EmbeddingGemma 2 comments and docstrings
  stikves
  committed
  41b294b

Commits

Commits on Oct 23, 2026

Commits on Oct 27, 2026

Commits on Oct 28, 2026

Commits on Oct 29, 2026

Commits on Oct 30, 2026