ApolloRay

> Please try `--quantization-level 2.5` because TRT currently does not support INT8 MHA fusion for SeqLen>512 due to accuracy reason. Therefore, use `--quantization-level 2.5` instead so that the MHA part...

> I like open source, but because I am no longer a Ph.D. student, based on commercial considerations, this may not be a simple matter. I will do my best...

> I like open source, but because I am no longer a Ph.D. student, based on commercial considerations, this may not be a simple matter. I will do my best...

> Yes I am also mostly interested in the 1024 version also. > > Can you provide more detail on the 1024 holdup? I know you mentioned commercial reasons. Is...

> Yes I am also mostly interested in the 1024 version also. > > Can you provide more detail on the 1024 holdup? I know you mentioned commercial reasons. Is...

> > Yes I am also mostly interested in the 1024 version also. > > Can you provide more detail on the 1024 holdup? I know you mentioned commercial reasons....

请问进行referring grounding训练的时候，可以按以下的方案进行标注吗？`{ "from": "user", "value": "img path格子衬衫" }, { "from": "assistant", "value": "(588,499),(725,789)" }` 这么训练之后，模型的grounding能力消失了

> > 请问进行referring grounding训练的时候，可以按以下的方案进行标注吗？`{ "from": "user", "value": "img path格子衬衫" }, { "from": "assistant", "value": "(588,499),(725,789)" }` 这么训练之后，模型的grounding能力消失了 > > 这种方式我没有实验过，我认为你可以试试如下两个方法： > > 1. user的询问问题是一个名词，不符合大模型微调要求，你可以使用如：“找出`格子衬衫`的坐标”作为user提问。 > 2. ref和box是需要在一起进行对应。你可以在训练时使用 > `img path格子衬衫(588,499),(725,789)`目标+坐标的形式，然后在你实际应用中，你再询问，比如：输出格子衬衫坐标，应该是可以的。...

> > > > 请问进行referring grounding训练的时候，可以按以下的方案进行标注吗？`{ "from": "user", "value": "img path格子衬衫" }, { "from": "assistant", "value": "(588,499),(725,789)" }` 这么训练之后，模型的grounding能力消失了 > > > > > > > > > 这种方式我没有实验过，我认为你可以试试如下两个方法： >...

rm -rf *