Minwoo (Dylan) Byeon

Reproduce V2-S,M and Worse Accuracy

3

I tried reproduce V2-S,M training on TPUs. (Run main.py with default configurations) But it was worse accuracy than [pre-trained checkpoints](https://github.com/google/automl/tree/master/efficientnetv2#2-pretrained-efficientnetv2-checkpoints). ## EfficientnetV2-S Evaluation Result ``` Saving dict for global step...

Initialization failed when action with `mlvm.sh`

1

### Command to create a GPU cluster ```bash $ gcloud dataproc clusters create cluster-b9b8-gpu \ --autoscaling-policy policy-d005-gpu \ --enable-component-gateway \ --region europe-west4 \ --zone europe-west4-a \ --master-machine-type n1-standard-8 \ --master-boot-disk-type...

Supports H100

1

Do you have plans to support the H100(sm90)? https://github.com/NetEase-FuXi/EETQ/blob/1657b1504faa359e2ce0ac02999439d7ac8c74c0/csrc/cutlass_kernels/cutlass_preprocessors.cc#L113-L128