423A35C7/LLaMA-Factory
1
0
Fork 0
mirror of https://github.com/hiyouga/LLaMA-Factory.git synced 2025-07-31 10:42:50 +08:00
Code Issues Packages Projects Releases Wiki Activity

[script] add Script description for qwen_omni_merge (#8293)

Browse Source
This commit is contained in:
Kingsley 2025-06-05 13:22:01 +08:00 committed by GitHub
parent 81c4d9bee6
commit 5308424705
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 12 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

12
scripts/qwen_omni_merge.py
View File

@ -12,6 +12,18 @@
# See the License for the specific language governing permissions and # See the License for the specific language governing permissions and
# limitations under the License. # limitations under the License.
"""Why we need this script for qwen_omni?
Because the qwen_omni model is constructed by two parts:
1. [Thinker]:[audio_encoder, vision_encoder, LLM backbone], which our repository does support to post-training.
2. [Talker]: [audio_decoder, wave_model], which is not supported to post-training without specific tokenizer.
When we post-training the model, we exactly train the [Thinker] part, and the [Talker] part is dropped.
So, to get the complete model, we need to merge the [Talker] part back to the [Thinker] part.
LoRA mode: [Thinker + LoRA weights] + [Original Talker] -> [Omni model]
Full mode: [Thinker] + [Original Talker] -> [Omni model]
For Processor, we do saved the processor from trained model instead of the original model.
"""
import os import os
import shutil import shutil