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

Merge pull request #5290 from simonJJJ/qwen2_vl

Browse Source 
support qwen2-vl

Former-commit-id: 7156f832af8505b26371559d340c0e69eb962bbc
This commit is contained in:
hoshi-hiyouga
2024-08-30 02:10:36 +08:00
committed by GitHub
parent c1369a1ec9 7466fd4387
commit 77c2c7076b
8 changed files with 185 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
src/llamafactory/model/adapter.py
View File

@@ -212,7 +212,7 @@ def _setup_lora_tuning(
target_modules = find_expanded_modules(model, target_modules, finetuning_args.freeze_trainable_layers)
if model_args.visual_inputs and finetuning_args.freeze_vision_tower:
target_modules = "^(?!.*vision_tower).*(?:{}).*".format("|".join(target_modules))
target_modules = "^(?!.*(?:vision_tower|visual)).*(?:{}).*".format("|".join(target_modules))
if (
finetuning_args.use_dora

2
src/llamafactory/model/model_utils/misc.py
View File

@@ -36,6 +36,8 @@ 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"]:
forbidden_modules.add("merger")
if freeze_vision_tower:
forbidden_modules.add("vision_tower")