Kingsley comments

Results 80 comments of


                                            Kingsley

input_features values are all -1.5 in modeling_qwen2_5_omni.py when using lora sft Qwen2.5Omni

I think it is because we add a fake audio input into the batch features when the whole batch doesn't contain 'audio' input. https://github.com/hiyouga/LLaMA-Factory/blob/5817cda37ec2ede2d94347f168f2c75241dd21ad/src/llamafactory/data/collator.py#L130

input_features values are all -1.5 in modeling_qwen2_5_omni.py when using lora sft Qwen2.5Omni

> [@Kuangdd01](https://github.com/Kuangdd01) Thanks again! One follow-up question: > > Is adding the fake audio input strictly necessary? It seems like, based on the official inference code, it's also possible to...

单机双卡A100 40G训练gemma3-27b lora非常慢

deepspeed z3要把模型部分也广播了卡间通信量可能很大

how to train with vqa

Insert `` into every sample in the question column.

how to train with vqa

For example, you should have a in your question when you made up your vqa dataset. "`+`Please parse the parent-child one to one relationships from the attaches directed structure chart...

qlora微调Qwen2.5-Omni模型报错ValueError: Processor was not found, please check and update your processor config.

pip install git+https://github.com/Kuangdd01/transformers.git@qwen25omni 暂时先用这个试一下

qlora微调Qwen2.5-Omni模型报错ValueError: Processor was not found, please check and update your processor config.

参考data/mllm_audio_demo.jdon组织一下数据格式

qlora微调Qwen2.5-Omni模型报错ValueError: Processor was not found, please check and update your processor config.

> 训练参数为： > llamafactory-cli train ` > --stage sft ` > --do_train True ` > --model_name_or_path C://Users//76425//Desktop//LLaMA-Factory//models//Qwen2.5-Omni-3B ` > --preprocessing_num_workers 16 ` > --finetuning_type lora ` > --template qwen2_omni `...

qlora微调Qwen2.5-Omni模型报错ValueError: Processor was not found, please check and update your processor config.

数据请参考mllm_audio_demo，不需要自己加特殊token类似

Qwen3-14-Base 训练异常

accelerate run用的也是同一份yaml还是用的fsdp2，换成dsz3再试试。