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

[trainer] make projector trainable in freeze training (#7872)

Browse Source 
Co-authored-by: hoshi-hiyouga <hiyouga@buaa.edu.cn>
This commit is contained in:
zhaop-l 2025-04-28 13:19:37 +08:00 committed by GitHub
parent fcca3b0b0d
commit 1bd319d16c
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 6 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

8
src/llamafactory/model/adapter.py
View File

@ -23,7 +23,7 @@ from ..extras import logging
from .model_utils.misc import find_all_linear_modules, find_expanded_modules
from .model_utils.quantization import QuantizationMethod
from .model_utils.unsloth import get_unsloth_peft_model, load_unsloth_peft_model
from .model_utils.visual import get_forbidden_modules, patch_target_modules
from .model_utils.visual import COMPOSITE_MODELS, get_forbidden_modules, patch_target_modules
if TYPE_CHECKING:
@ -100,7 +100,7 @@ def _setup_freeze_tuning(
hidden_modules.add(name.split(".1.")[-1].split(".")[0])
if re.search(r"\.\d+\.", name) is None:
non_hidden_modules.add(name.split(".")[-2])
non_hidden_modules.add(name.split(".")[-2]) # remove weight/bias
trainable_layers = []
for module_name in finetuning_args.freeze_trainable_modules:
@ -121,6 +121,10 @@ def _setup_freeze_tuning(
trainable_layers.append(module_name)
model_type = getattr(model.config, "model_type", None)
if not finetuning_args.freeze_multi_modal_projector and model_type in COMPOSITE_MODELS:
trainable_layers.append(COMPOSITE_MODELS[model_type].projector_key)
forbidden_modules = get_forbidden_modules(model.config, finetuning_args)
for name, param in model.named_parameters():
if any(trainable_layer in name for trainable_layer in trainable_layers) and not any(