Gemma-4 E2B: enable image (multimodal) E2E parity — vision clipped-linears + padded-patch masking + PLE substitution (E4B vision-tower validated; E4B E2E pending) - #4790

Open

lokic233 wants to merge 14 commits into

AI-Hypercomputer:mainfrom

lokic233:gemma4-e2b-e4b-vision-clipped-linears

Commits on Aug 26, 2026

Gemma-4 vision: add per-projection clipped-linears (E2B/E4B parity prerequisite)
lokic233
committed
bb0c359
Gemma-4 E2B/E4B: close image parity (padded-patch masking + PLE image-row substitution)
lokic233
committed
7734eee
Gemma-4 E2B/E4B: set image-contract defaults in the model configs
lokic233
committed
1e9d2b4
M1: fail-closed guard — legacy vision path rejects pre-patchified input without image_position_ids
lokic233
committed
8bb801b
M3: make 448 clip bounds immutable — stop_gradient at use-site + optimizer freeze integration
lokic233
committed
375a36e
M4+M5: real post-load clip validation (exact 112/448, min<=max, fail-closed) + fused_qkv guard
lokic233
committed
8a77804
M1-M5 in-tree tests: tests/unit/gemma4_clip_linears_test.py (13 tests, all pass)
lokic233
committed
181aa50
M1/M2/M6: stock-trainer position threading + PLE enum + MM contract gate
lokic233
committed
8ab0931
M2/M5/M6 in-tree test: Gemma-4 MM static contract gate (enum/causal/fused/clipped)
lokic233
committed
0c67c77
M7: Option-S sentinel integrity + placeholder==pooled validators + 5 unit tests
lokic233
committed
250f6c6
M7.4: sanitize padded-patch rows to zero before attention (isolate non-finite pad from valid outputs)
lokic233
committed
99d6c28
M8: fail-close packing for E2B/E4B multimodal (any mode) + packing-guard test
lokic233
committed
c16da94
E2B/E4B ymls: default packing=false (MM does not support packed image spans)
lokic233
committed
da07c52
Gemma4-small ViT review: consolidate clip tests into gemma4_small_layers_test.py; optimizer clip-freeze comment; E2B multimodal e2e script
lokic233
committed
85f817f

Navigation Menu

Uh oh!

Gemma-4 E2B: enable image (multimodal) E2E parity — vision clipped-linears + padded-patch masking + PLE substitution (E4B vision-tower validated; E4B E2E pending) - #4790

Gemma-4 E2B: enable image (multimodal) E2E parity — vision clipped-linears + padded-patch masking + PLE substitution (E4B vision-tower validated; E4B E2E pending)#4790
lokic233 wants to merge 14 commits into
AI-Hypercomputer:mainfrom
lokic233:gemma4-e2b-e4b-vision-clipped-linears

Commits on Aug 26, 2026

Gemma-4 vision: add per-projection clipped-linears (E2B/E4B parity prerequisite)

Gemma-4 E2B/E4B: close image parity (padded-patch masking + PLE image-row substitution)

Gemma-4 E2B/E4B: set image-contract defaults in the model configs

M1: fail-closed guard — legacy vision path rejects pre-patchified input without image_position_ids

M3: make 448 clip bounds immutable — stop_gradient at use-site + optimizer freeze integration

M4+M5: real post-load clip validation (exact 112/448, min<=max, fail-closed) + fused_qkv guard

M1-M5 in-tree tests: tests/unit/gemma4_clip_linears_test.py (13 tests, all pass)

M1/M2/M6: stock-trainer position threading + PLE enum + MM contract gate

M2/M5/M6 in-tree test: Gemma-4 MM static contract gate (enum/causal/fused/clipped)

M7: Option-S sentinel integrity + placeholder==pooled validators + 5 unit tests

M7.4: sanitize padded-patch rows to zero before attention (isolate non-finite pad from valid outputs)

M8: fail-close packing for E2B/E4B multimodal (any mode) + packing-guard test

E2B/E4B ymls: default packing=false (MM does not support packed image spans)

Gemma4-small ViT review: consolidate clip tests into gemma4_small_layers_test.py; optimizer clip-freeze comment; E2B multimodal e2e script