Huang Fang comments

Repositories
Issues
Comments

Results 5 comments of


                                            Huang Fang

OOM and slow tokenization after upgrade LLaMA-Factory

> What do you mean by 32 A100? 32 NVIDIA A100 GPUs.

[question] Dose Muon Optimizer support deepspeed ?

> > Can you try this PR? [#7808](https://github.com/hiyouga/LLaMA-Factory/pull/7808) > > 看起来符合预期： > > * 收敛更快 > * 占显存更多，需要移植点别的进来 > > ![Image](https://github.com/user-attachments/assets/2123136d-a710-4e40-a4cf-6b8ae30a77b8) > > opencompass 精度在测，但应该问题不大。哈喽，请问一下这两个曲线的batchsize设置的不一样吧？为啥epoch结束loss突降的点不在同一个step呢？

Object Assets dataset broken?

Same issue.

Error using vllm with Qwen2.5-VL-72B-Instruct-AWQ

Only one gpu is used when deploying interactive demo

@ryansunyz We need to set the environment variables before import vLLM/Pytorch. This [solution](https://github.com/vllm-project/vllm/issues/571) should help.