Sitelet https://github.com/mlx-node/mlx-node/pull/74/commits

feat(gemma4): Gemma 4 12B unified (gemma4_unified) — encoder-free text + vision + audio - #74

Merged

Brooooooklyn merged 41 commits into

mainfrom

feat/gemma4-unified-12b

Jun 24, 2026

Commits on Jun 23, 2026

feat(gemma4): detect gemma4_unified config + skip vision_embedder weights

Brooooooklyn
and
claude
committed
021118c
feat(lm): dispatch gemma4_unified to the gemma4 loader (text-only)

Brooooooklyn
and
claude
committed
cf4b357
fix(gemma4): route arch-only unified checkpoints to gemma4 in TS detection

Brooooooklyn
and
claude
committed
a3526b6
feat(gemma4): unified encoder-free vision config + embedder modules

Brooooooklyn
and
claude
committed
aa49471
feat(gemma4): wire unified vision into load + merge + image turn

Brooooooklyn
and
claude
committed
95781ef
fix(gemma4): fail closed on missing unified vision weights

Brooooooklyn
and
claude
committed
ed10e6e
feat(gemma4): blockwise bidirectional vision mask overlay helper

Brooooooklyn
and
claude
committed
9f9abe3
feat(gemma4): apply bidirectional vision overlay in unified paged prefill

Brooooooklyn
and
claude
committed
0555240
fix(gemma4): guard unified-vision overlay against KV-shared layers

Brooooooklyn
and
claude
committed
cf1d33e
feat(gemma4): unified audio config + un-drop and load embed_audio weights

Brooooooklyn
and
claude
committed
ef9dcac
feat(gemma4): encoder-free unified audio embed pipeline + audio-aware overlay gate

Brooooooklyn
and
claude
committed
128be28
fix(gemma4): gate unified audio on non-null audio_config; guard empty audio

Brooooooklyn
and
claude
committed
5cba33d
test(gemma4): add 16kHz mono audio fixture for the unified audio e2e

Brooooooklyn
and
claude
committed
b865f76
feat(gemma4): hand-rolled WAV→PCM decode for unified audio input

Brooooooklyn
and
claude
committed
44cc919
feat(gemma4): wire unified audio input through NAPI + engine + multimodal turn

Brooooooklyn
and
claude
committed
18fd8d0
feat(lm): ChatSession audio input surface (mirrors images)

Brooooooklyn
and
claude
committed
71c354b
test(gemma4): unified audio e2e (audio-only + mixed image+audio)

Brooooooklyn
and
claude
committed
01cae1f
fix(qianfan): align continue ABI with shared audio chat surface

Brooooooklyn
and
claude
committed
3e99f6d
fix(qianfan): reject unsupported audio on start path (mirror continue guard)

Brooooooklyn
and
claude
committed
0b4c51c
fix(gemma4): replay media→text continuation through cold start path

Brooooooklyn
and
claude
committed
2d9043d
feat(gemma4): warm media→text continuation for audio + non-unified image (Phase 2)

Brooooooklyn
and
claude
committed
aaae2f9
fix(gemma4): gate warm media->text continuation on faithful sliding restore + hoist marker reset

Brooooooklyn
and
claude
committed
9dd9814
fix(gemma4): also gate media continuation on live-for-continue + assert cold-restart in e2e

Brooooooklyn
and
claude
committed
08d862b

Commits on Jun 24, 2026

fix(chat-session): commit visible streamed text for gemma4 empty-done-chunk turns

Brooooooklyn
and
claude
committed
a35a727
feat(gemma4): warm media→text continuation for unified image (12B)

Brooooooklyn
and
claude
committed
28c6cb3
docs(gemma4): correct finalize_vision_turn_media_state header to match the stored-gate code

Brooooooklyn
and
claude
committed
5e97676
chore: gitignore .superpowers/ SDD scratch and untrack leaked report

Brooooooklyn
and
claude
committed
318718d
fix(gemma4): isolate media paged blocks by content in shared prefix cache

Brooooooklyn
and
claude
committed
ef379e1
Revert "fix(gemma4): isolate media paged blocks by content in shared prefix cache"
Brooooooklyn
committed
eab0e06
fix(gemma4): require a live paged request to warm-continue media→text

Brooooooklyn
and
claude
committed
ff22ab8
fix(gemma4): skip paged prefix lookup when prompt holds media placeholders

Brooooooklyn
and
claude
committed
384ee03
feat(gemma4): extend media-held cold-restart floor to tool-result entry points

Brooooooklyn
and
claude
committed
0625db0
fix(chat-session): self-heal tool-result entry points on a cold native session

Brooooooklyn
and
claude
committed
79f13d7
fix(chat-session): derive sticky media keys from replayed history on start-path restart

Brooooooklyn
and
claude
committed
4beeb26
fix(gemma4): insert audio spans after BOS when no <|audio|> placeholder

Brooooooklyn
and
claude
committed
53f21c3
fix(ci): disable CARGO_INCREMENTAL + orphan poisoned rust-cache (v2)

Brooooooklyn
and
claude
committed
10d3fe0
fix(gemma4): harden PLE media mask and not-live media-restart guard

Brooooooklyn
and
claude
committed
e274bac
fix(gemma4): clear stale audio key on text-only paged save (paged/flat parity)

Brooooooklyn
and
claude
committed
40b5bf9
fix(gemma4): mirror image cache-miss guard for audio in verify_cache_prefix

Brooooooklyn
and
claude
committed
befca77
fix(gemma4): keep image-then-audio order in no-placeholder media fallback

Brooooooklyn
and
claude
committed
70a7f2d
fix(ci): don't create an MLX Metal stream during addon module load

Brooooooklyn
and
claude
committed
ea9c562