423A35C7/LLaMA-Factory
1
0
Fork 0
mirror of https://github.com/hiyouga/LLaMA-Factory.git synced 2025-08-23 06:12:50 +08:00
Code Issues Packages Projects Releases Wiki Activity

add forbidden modules

Browse Source 
Former-commit-id: df477370dc67315effac1a8f48068c4e9c4067a5
This commit is contained in:
hiyouga 2024-11-23 18:34:15 +00:00
parent e99031daa4
commit 253752ccca
2 changed files with 6 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
src/llamafactory/model/model_utils/misc.py
View File

@ -34,13 +34,15 @@ def find_all_linear_modules(model: "PreTrainedModel", freeze_vision_tower: bool)
forbidden_modules.add("output_layer") forbidden_modules.add("output_layer")
elif model_type == "internlm2": elif model_type == "internlm2":
forbidden_modules.add("output") forbidden_modules.add("output")
elif model_type in ["llava", "llava_next", "llava_next_video", "paligemma", "video_llava"]: elif model_type in ["llava", "llava_next", "llava_next_video", "mllama", "paligemma", "video_llava"]:
forbidden_modules.add("multi_modal_projector") forbidden_modules.add("multi_modal_projector")
elif model_type == "qwen2_vl": elif model_type == "qwen2_vl":
forbidden_modules.add("merger") forbidden_modules.add("merger")
if freeze_vision_tower: if freeze_vision_tower:
if model_type == "qwen2_vl": if model_type == "mllama":
forbidden_modules.add("vision_model")
elif model_type == "qwen2_vl":
forbidden_modules.add("visual") forbidden_modules.add("visual")
else: else:
forbidden_modules.add("vision_tower") forbidden_modules.add("vision_tower")

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

@ -191,6 +191,8 @@ def patch_target_modules(
if finetuning_args.freeze_vision_tower: if finetuning_args.freeze_vision_tower:
if model_type in ["llava", "llava_next", "llava_next_video", "paligemma", "pixtral", "video_llava"]: if model_type in ["llava", "llava_next", "llava_next_video", "paligemma", "pixtral", "video_llava"]:
return "^(?!.*vision_tower).*(?:{}).*".format("|".join(target_modules)) return "^(?!.*vision_tower).*(?:{}).*".format("|".join(target_modules))
elif model_type == "mllama":
return "^(?!.*vision_model).*(?:{}).*".format("|".join(target_modules))
elif model_type == "qwen2_vl": elif model_type == "qwen2_vl":
return "^(?!.*visual).*(?:{}).*".format("|".join(target_modules)) return "^(?!.*visual).*(?:{}).*".format("|".join(target_modules))
else: else: