]> git.djapps.eu Git - pkg/ggml/sources/llama.cpp/commit
convert : handle per_layer_config in Gemma4 (transformers 5.15) (#26882)
authorWang Zhiyu <redacted>
Wed, 12 Aug 2026 05:05:13 +0000 (13:05 +0800)
committerGitHub <redacted>
Wed, 12 Aug 2026 05:05:13 +0000 (08:05 +0300)
commit6eff59326205854cf4b9edd45e67ee3436c16d58
tree1d0950babdc7c3960b82f51a3d854c9f5ce6ffc6
parentcb27fe9c35d953b49b49a0048e73decdac0338c6
convert : handle per_layer_config in Gemma4 (transformers 5.15) (#26882)

* fix: handle nested global_head_dim in Gemma4 config

Gemma-4 E4B models have global_head_dim inside text_config
rather than at the top level. Add fallback to support both layouts.

* fix: add fallback for global_head_dim to support per_layer_config format

* fix: read head_dim only from full_attention layers in per_layer_config and num_global_key_value_heads compatibility

* fix: added fallback for num_global_key_value_heads

* fix: read per_layer_config from root hparams

* fix: delete unused text_config

* cleanup and fixes

---------

Co-authored-by: Sigbjørn Skjæret <redacted>
conversion/gemma.py