Yaowei Zheng

icon indicating an email [email protected]

icon company Millennium Science School icon location Beijing, China icon location No code All live

Results 651 comments of Yaowei Zheng

华为910B train_full 微调 qwen2vl，解冻freeze_vision_tower，就报错OPS function error: Conv3DBackpropFilter

910B 不支持 unfreeze vision tower

提示bus error，insufficient shared memery (shm)

increase shm of the docker

Error when training

Try reinstall trl

Qwen/Qwen2.5-VL-7B-Instruct PPO 训练报错

We recommend using EasyR1 for RL, the RL implementation in LlamaFactory is temporarily bugged: https://github.com/hiyouga/EasyR1

跑open_r1_math数据集，qwen7b-instruct每次跑到53个step报错

试试关掉：`use_unsloth_gc`

Questions about qwen template modification

这个应该影响不大吧，去除会有问题

special token未能输出

训练时候没有添加 `resize_vocab` 参数，tokenizer 不识别

special token未能输出

@lifeng7777 modules_to_save 里面不是有吗

[NPU] update cann base image to py3.10 and fix the type error of args INSTALL_DEEPSPEED of npu docker-compose.yaml

Thanks for your fixing!

按照LLaMA-Factory QuickStart 进行微调时，没有正确输出损失函数图

数据太少或者训练步数不够导致没法画图