Skip to content

Gemma3 layers naming convention #129

@alessiodevoto

Description

@alessiodevoto

Bug

When using AutoModelforCausaLM with Gemma3 models, the loaded version is multimodal. In this version, the LM layers are in model.language_model.layers. However, we assume they are in model.model.layers.

To Reproduce

Call any press with any Gemma3 model.

Repository version

0.2.10

Metadata

Metadata

Assignees

Labels

bugSomething isn't working

Type

No type

Projects

No projects

Milestone

No milestone

Relationships

None yet

Development

No branches or pull requests

Issue actions