Repository navigation

feat(gemma4): Google Gemma-4-E4B-it multimodal baseline - #3476

Open

mudler-agent wants to merge 82 commits into

mainfrom

row/MODEL-MM-gemma4-mm-gemma4-for-conditional-generation

Open

feat(gemma4): Google Gemma-4-E4B-it multimodal baseline#3476
mudler-agent wants to merge 82 commits into
mainfrom
row/MODEL-MM-gemma4-mm-gemma4-for-conditional-generation

Commits on Oct 18, 2026
- spec(MODEL-MM-gemma4-mm-gemma4-for-conditional-generation): pin Google E4B baseline
  mudler
  committed
  7301a5b
- fix(MODEL-MM-gemma4-mm-gemma4-for-conditional-generation): preserve derived READY
  mudler
  committed
  61d2286
- feat(MODEL-MM-gemma4-mm-gemma4-for-conditional-generation): validate original E4B metadata
  mudler
  committed
  2bb9dd6
- fix(MODEL-MM-gemma4-mm-gemma4-for-conditional-generation): pin metadata guard regressions
  mudler
  committed
  501b06b
- docs(MODEL-MM-gemma4-mm-gemma4-for-conditional-generation): split metadata prerequisites
  mudler
  committed
  799c98f
- fix(MODEL-MM-gemma4-mm-gemma4-for-conditional-generation): require valid untied heads
  mudler
  committed
  c2cc4da
- test(MODEL-MM-gemma4-mm-gemma4-for-conditional-generation): reject F16 output heads
  mudler
  committed
  ad0010e
- docs(MODEL-MM-gemma4-mm-gemma4-for-conditional-generation): bound modality projector loading
  mudler
  committed
  475a4f7
- fix(MODEL-MM-gemma4-mm-gemma4-for-conditional-generation): retain validated modality projectors
  mudler
  committed
  8f17dec
- docs(MODEL-MM-gemma4-mm-gemma4-for-conditional-generation): bound vision patch retention
  mudler
  committed
  a80d84d
- fix(MODEL-MM-gemma4-mm-gemma4-for-conditional-generation): retain vision patch weights
  mudler
  committed
  5a12185
- test(MODEL-MM-gemma4-mm-gemma4-for-conditional-generation): cover projection byte overflow
  mudler
  committed
  e35a24b
- docs(gemma4): define vision block retention
  mudler
  committed
  e08c7d8
- fix(gemma4): retain vision encoder block weights
  mudler
  committed
  e438b25
- test(gemma4): require head-product overflow diagnostic
  mudler
  committed
  0f96d83
- docs(gemma4): specify vision clipping and typed config retention
  mudler
  committed
  981a709
- feat(gemma4): retain vision clipping bounds and typed config
  mudler
  committed
  881fabd
- docs(gemma4): specify bounded audio frontend retention
  mudler
  committed
  6aba17f
- feat(gemma4): retain audio frontend weights in decoder loaders
  mudler
  committed
  41bea37
- test(gemma4): pin audio frontend BF16 refusal diagnostics
  mudler
  committed
  c011434
- docs(gemma4): specify complete audio tower retention
  mudler
  committed
  3f91215
- feat(gemma4): retain complete audio blocks and output projection
  mudler
  committed
  5e0c3a4
- spec(gemma4): bound loaded audio encoder execution
  mudler
  committed
  14bedc3
- feat(gemma4): execute loaded audio through the registry encoder seam
  mudler
  committed
  acab23d
- spec(gemma4): separate audio norm epsilon contracts
  mudler
  committed
  a3c007a
- fix(gemma4): preserve distinct audio norm epsilons
  mudler
  committed
  bf417a5
- spec(gemma4): pin executable audio preprocessing prerequisite
  mudler
  committed
  887afa9
- feat(gemma4): preprocess bounded audio into shared encoder inputs
  mudler
  committed
  cd586bf
- spec(gemma4): record prospective T05A process correction
  mudler
  committed
  359349f
- revert(gemma4): capture buildable T05A corrective red state
  mudler
  committed
  3b0a4a9
- fix(gemma4): restore complete T05A frontend after corrective red
  mudler
  committed
  9508d99
- spec(gemma4): define bounded shared embedding merge
  mudler
  committed
  c97b18c
- feat(gemma4): merge CPU audio embeddings through shared dispatch
  mudler
  committed
  d69a34d
- test(gemma4): detect missing BF16 text scale rounding
  mudler
  committed
  b810626
- docs(gemma4): specify bounded decoded RGB prerequisite
  mudler
  committed
  2a858f7
- feat(gemma4): preprocess decoded RGB identity-size images
  mudler
  committed
  1e60ad4
- docs(gemma4): specify CPU tensor uint8 resize parity
  mudler
  committed
  d963d9e
- feat(gemma4): resize decoded RGB with tensor bicubic antialias
  mudler
  committed
  71c7952
- docs(gemma4): specify bounded CPU vision execution
  mudler
  committed
  9bc5f1f
- feat(gemma4): execute loaded vision through CPU multimodal hooks
  mudler
  committed
  d7eb067
- docs(gemma4): specify vision mutation coverage
  mudler
  committed
  2161022
- test(gemma4): detect pool rounding and excess weights
  mudler
  committed
  976a14e
- docs(gemma4): specify original Google template adapter
  mudler
  committed
  1739447
- docs(gemma4): bound Google literal syntax adaptation
  mudler
  committed
  f748918
- docs(gemma4): specify scoped empty-schema truth compatibility
  mudler
  committed
  439a513
- fix(gemma4): preserve original Google typed template content
  mudler
  committed
  fc60486
- spec(gemma4): define empty schema byte-parity repair
  mudler
  committed
  3709b57
- fix(gemma4): preserve empty schema truthiness in Google prompts
  mudler
  committed
  6abd270
- spec(gemma4): bound engine reasoning-end repair
  mudler
  committed
  f57b24f
- fix(gemma4): honor thinking and latest-turn reasoning end
  mudler
  committed
  e5bb09d
- spec(gemma4): define engine prompt-state preparation
  mudler
  committed
  293e445
- fix(gemma4): prepare engine reasoning from prompt IDs once
  mudler
  committed
  d26484e
- spec(gemma4): bound reasoning registry adapter
  mudler
  committed
  e8181e3
- fix(gemma4): register prompt-aware reasoning adapter
  mudler
  committed
  1ede2a4
- spec(gemma4): bound request reasoning settings
  mudler
  committed
  3e36dcb
- fix(gemma4): honor request reasoning settings
  mudler
  committed
  c3ce1a7
- spec(gemma4): define bounded production tokenizer bridge
  mudler
  committed
  37c39a2
- feat(gemma4): bridge the production tokenizer to parsers
  mudler
  committed
  216c1fe
- spec(gemma4): reconcile T03C3/C4/C5 evidence to accepted
  mudler
  committed
  9b230ac
- spec(gemma4): define T03B config.json special token ID parity
  mudler
  committed
  1e86755
- test(gemma4): add RED for config.json special token ID parity
  mudler
  committed
  b751567
- feat(gemma4): read special token IDs from config.json
  mudler
  committed
  46486b7
- spec(gemma4): reconcile T03B evidence to accepted
  mudler
  committed
  05b2a39
- spec(gemma4): define T03D video token replacement sequence
  mudler
  committed
  3a0aa53
- test(gemma4): add RED for video token replacement sequence
  mudler
  committed
  6845718
- feat(gemma4): build video token replacement sequence
  mudler
  committed
  a06b047
- spec(gemma4): reconcile T03D evidence to accepted
  mudler
  committed
  044c940
- spec(gemma4): define T09 mm_inputs_ready flip
  mudler
  committed
  09b0286
- feat(gemma4): wire mm_inputs_ready to the runner
  mudler
  committed
  bcac330
- spec(gemma4): reconcile T09 evidence to accepted
  mudler
  committed
  ab2fb4d
- spec(gemma4): define T10A mm chat seam registration
  mudler
  committed
  173643b
- test(gemma4): add RED for mm chat seam registration
  mudler
  committed
  ef286a0
- feat(gemma4): register mm chat seam for Gemma4
  mudler
  committed
  cfbf2c1
- spec(gemma4): reconcile T10A evidence to accepted
  mudler
  committed
  b7e055c
- spec(gemma4): define T10B image processing in chat seam
  mudler
  committed
  8c5dae8
- test(gemma4): add RED for image processing in chat seam
  mudler
  committed
  7096b10
- feat(gemma4): process images in the chat seam
  mudler
  committed
  2047fae
- spec(gemma4): reconcile T10B evidence to accepted
  mudler
  committed
  2ac067e
Commits on Oct 20, 2026

Commits

Commits on Oct 18, 2026

Commits on Oct 20, 2026