Billy Cao
For the failing `owlvit` test, it is not because weight overflow, but implementation: https://github.com/huggingface/transformers/blob/3f93fd06949f9eae58e50fd0c9b8e60be82643bc/src/transformers/models/owlvit/modeling_owlvit.py#L1260 `-1e6` is simply out of range for fp16, min is `-65504`. Should we correct the impl?...
> Here does "all models" refer to all owlvit checkpoints? What I meant was all torch impl of models in HF transformers as more may be using a hard-coded out...
@amyeroberts the failing tf and onnx tests are due to some keras changes in https://github.com/keras-team/keras/releases/tag/v3.4.1 The failing torch pipeline test is due to network timeout
@amyeroberts CI all green now
Added the training tests and also enabled gradient checkpointing tests. I note that CLIP had issues with GC but I have used it with siglip myself and did not find...
@amyeroberts seems to need you to enable slow tests?
@amyeroberts now that the GC tests are properly skipped, shall we move forward with this?
Hi fellow TIL 2024 participant, 2 things i can point out from your provided info. 1. You set num classes to 80 which is num classes of COCO but have...
试试 https://github.com/hiyouga/LLaMA-Factory/pull/7530 能不能解决问题 在你的yaml里面加上 ``` dataset_shards: 16 dataloader_num_workers: 16 ``` 把16换成你的cpu数量除以gpu数量
你有checkout我的commit吗 用github cli 在项目目录下运行`gh pr checkout 7530`