Repository navigation

Fix Gemma 4 model loading and fix random outputs - #3968

Open

Zhang-JK wants to merge 3 commits into

huggingface:mainfrom

Zhang-JK:fix/gemma4

Open

Fix Gemma 4 model loading and fix random outputs#3968
Zhang-JK wants to merge 3 commits into
huggingface:mainfrom
Zhang-JK:fix/gemma4

Zhang-JK commented Sep 9, 2026 •

edited

Loading

changes:

update weight loading path to be model.language_model for language model
Apply the per-layer embedding for E2Band E4B
add shared KV for E4B model
fix RMSNorm, gemma 4 does not apply weight+1

verification:

Zhang-JK added 3 commits

September 8, 2026 22:03


          fix gemma 4 model load prefix

f540f71


          impl per layer embedding, add shared KV and fix RMSNorm

c026dd5


          Merge branch 'main' into fix/gemma4

b617ec6

This branch has not been deployed

No deployments

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet