ros/LLaMA-Factory
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

refactor mm training

Browse Source 
Former-commit-id: 179c0558699e287cbf38a2d73bff47e86d589c5a
This commit is contained in:
hiyouga
2024-08-30 02:14:31 +08:00
parent 77c2c7076b
commit c62a6ca59d
29 changed files with 499 additions and 312 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
src/llamafactory/model/model_utils/misc.py
View File

@@ -36,11 +36,14 @@ def find_all_linear_modules(model: "PreTrainedModel", freeze_vision_tower: bool)
forbidden_modules.add("output")
elif model.config.model_type in ["llava", "paligemma"]:
forbidden_modules.add("multi_modal_projector")
elif model.config.model_type in ["qwen2_vl"]:
elif model.config.model_type == "qwen2_vl":
forbidden_modules.add("merger")
if freeze_vision_tower:
forbidden_modules.add("vision_tower")
if model.config.model_type == "qwen2_vl":
forbidden_modules.add("visual")
else:
forbidden_modules.add("vision_tower")
module_names = set()
for name, module in model.named_modules():