PokeLu
PokeLu
使用win server 2022H遇到相同问题，一直持续重命名，qb在host上下载，ab运行于wsl上的docker上
后续发现bug可能触发原因是重命名方式为advance，改回pn时bug消失。 目前发现bug仅在Rename Method=advance，Debug=false下触发。 > 使用win server 2022H遇到相同问题，一直持续重命名，qb在host上下载，ab运行于wsl上的docker上
same issue here, could someone take a look?
> H100 80G ? Yes, but sorry I found that I forgot to set tp and pp for the actor in my script. I will test again with the corrected...
> H100 80G ? I got it work with tp=4, pp=1 for the actor and ref model. One more question before I closed the issue. I checked the script for...
> Colocating with 2x gpus with 2x sequence length doesn't means you can train with 1x gpu and 1x seq length. The memory for training is a little complicated, see...
> It is related to the agent loop. when the mode is set to async, AgentLoopManager is used. see ppo/ray_trainer.py and main_ppo.py for example. > > The reason that dapo...
> [DeepSeek-V3/inference/fp8_cast_bf16.py](https://github.com/deepseek-ai/DeepSeek-V3/blob/2f7b80eecebf3d1c84da5a0d465f6639ea175012/inference/fp8_cast_bf16.py#L80) > > Line 80 in [2f7b80e](/deepseek-ai/DeepSeek-V3/commit/2f7b80eecebf3d1c84da5a0d465f6639ea175012) > > new_state_dict[weight_name] = weight_dequant(weight, scale_inv) > I tried changing > > new_state_dict[weight_name] = weight_dequant(weight, scale_inv) > to > > new_state_dict[weight_name] =...
上个release都2年前了