Xiaohan Zhang
Xiaohan Zhang
@irenedea thanks for reminding, I thought I opened a draft pr. Yes, potentially this will require your reviews. I will try to make the changes very limited and confined to...
Yes it should work out of the box. Use MDSWrite to convert your preference data to MDS , and create a streaming dataset out of it. Use json as the...
Hi @ericxsun want to follow up here before closing this issue.
@VikaasVarma Is this a typo in your repro script? def __getitem__(self, at: int) -> Sample: obj = super().__getitem__(at) _input = self.transforms(obj[self.input_key]) cond = torch.tensor(obj[self.cond_key]) if torch.rand(1) < self.cond_dropout_rate: cond =...
@VikaasVarma can you clarify your plot a bit? e.g., what does pss uss shared mean? and x, y axis. Can you also provide a sample dataset so I can reproduce...
@knighton I updated the description to include some of the profiling results. PTA~
This is not a bug on the streaming side most likely. I seen it before. Can you try using a non server less compute?
Has anyone found a solution/workaround to this bug?
@zhouhao001832050 before StreamingDataset (SD), what did you try and make the model converge? I wonder if your SD configuration has a mistake. Can you post here the code snippet or...
Hey @naston I am trying to reproduce but couldnt. I got a value error instead of the runtime error. ValueError: In the absence of local dataset, path to remote dataset...