[Gemma 4] Add multimodal support (apply_liger_kernel_to_gemma4 for Gemma4ForConditionalGeneration)#1203

Draft

dvdimitrov13 wants to merge 31 commits intolinkedin:mainfrom

dvdimitrov13:feat/gemma4-multimodal

Commits on Apr 24, 2026

gemma4: implementation plan

amrtini
and
claude
committed
gemma4: narrow plan scope to 31B text-only

amrtini
and
claude
committed
gemma4: assimilate user's deployment spec into the plan

amrtini
and
claude
committed
gemma4: mirror full gemma3 test coverage, not just one test file

amrtini
and
claude
committed
gemma4: add LigerRMSNormForGemma4 (ones init, no +1 offset)

amrtini
and
claude
committed
gemma4: scaffold causal_forward for Gemma4ForCausalLM

amrtini
and
claude
committed
gemma4: add apply_liger_kernel_to_gemma4_text + model-type registration

amrtini
and
claude
committed
gemma4: align monkey_patch function style with gemma3 sibling

amrtini
and
claude
committed
gemma4: export apply_liger_kernel_to_gemma4_text from package root

amrtini
and
claude
committed
gemma4: add revert_liger_kernel_to_gemma4_text test helper

amrtini
and
claude
committed
gemma4: add mini_gemma4_text bf16 convergence test

amrtini
and
claude
committed
gemma4: add mini_gemma4_text to bf16 logits-parity convergence test

amrtini
and
claude
committed
gemma4: add test_apply_liger_kernel_to_instance_for_gemma4_text

amrtini
and
claude
committed
gemma4: polish comments from whole-branch review

amrtini
and
claude
committed
gemma4: handle with_scale=False (v_norm) in RMSNorm subclass

amrtini
and
claude
committed
gemma4: wrap LigerGEGLUMLP to absorb Gemma4TextMLP's layer_idx arg

amrtini
and
claude
committed
gemma4: skip rope kernel swap (signature incompatibility with HF Gemma 4)

amrtini
and
claude
committed
gemma4: loosen bf16 convergence tolerances for 6-layer mini model

amrtini
and
claude
committed
gemma4: also bump logprobs_atol in test_mini_models.py

amrtini
and
claude
committed
gemma4: PR description with LUMI-measured numbers

amrtini
and
claude
committed
gemma4: address pre-PR review findings

amrtini
and
claude
committed
gemma4: align PR description with maintainer AI-assisted conventions

amrtini
and
claude
committed
gemma4: drop internal planning docs (not for upstream)

amrtini
and
claude
committed
address review: double-wide MLP, gemma4 model type, rope default

amrtini
and
claude
committed
test: add Gemma 4 double-wide MLP edge case tests

amrtini
and
claude
committed
review: remove gemma4 model type mapping (defer to multimodal PR)

amrtini
and
claude
committed

Commits on Apr 25, 2026

style: fix import sort and assert formatting in test_geglu.py

amrtini
and
claude
committed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[Gemma 4] Add multimodal support (apply_liger_kernel_to_gemma4 for Gemma4ForConditionalGeneration)#1203

[Gemma 4] Add multimodal support (apply_liger_kernel_to_gemma4 for Gemma4ForConditionalGeneration)#1203
dvdimitrov13 wants to merge 31 commits intolinkedin:mainfrom
dvdimitrov13:feat/gemma4-multimodal

Commits on Apr 24, 2026

gemma4: implementation plan

gemma4: narrow plan scope to 31B text-only

gemma4: assimilate user's deployment spec into the plan

gemma4: mirror full gemma3 test coverage, not just one test file

gemma4: add LigerRMSNormForGemma4 (ones init, no +1 offset)

gemma4: scaffold causal_forward for Gemma4ForCausalLM

gemma4: add apply_liger_kernel_to_gemma4_text + model-type registration

gemma4: align monkey_patch function style with gemma3 sibling

gemma4: export apply_liger_kernel_to_gemma4_text from package root

gemma4: add revert_liger_kernel_to_gemma4_text test helper

gemma4: add mini_gemma4_text bf16 convergence test

gemma4: add mini_gemma4_text to bf16 logits-parity convergence test

gemma4: add test_apply_liger_kernel_to_instance_for_gemma4_text

gemma4: polish comments from whole-branch review

gemma4: handle with_scale=False (v_norm) in RMSNorm subclass

gemma4: wrap LigerGEGLUMLP to absorb Gemma4TextMLP's layer_idx arg

gemma4: skip rope kernel swap (signature incompatibility with HF Gemma 4)

gemma4: loosen bf16 convergence tolerances for 6-layer mini model

gemma4: also bump logprobs_atol in test_mini_models.py

gemma4: PR description with LUMI-measured numbers

gemma4: address pre-PR review findings

gemma4: align PR description with maintainer AI-assisted conventions

gemma4: drop internal planning docs (not for upstream)

address review: double-wide MLP, gemma4 model type, rope default

test: add Gemma 4 double-wide MLP edge case tests

review: remove gemma4 model type mapping (defer to multimodal PR)

Commits on Apr 25, 2026

style: fix import sort and assert formatting in test_geglu.py

Commits on Apr 26, 2026

[Gemma 4] Add multimodal support (apply_liger_kernel_to_gemma4 for Gemma4ForConditionalGeneration)

fix: filter text_classes by isinstance(cls, type) under unittest.mock.patch

style: ruff format

fix: same isinstance(cls, type) filter in apply_liger_kernel_to_gemma4_text