Bag of scripts to faster training on TPU VM
-
Create new TPUVM-V38 with name 'learnjax' at zone
europe-west-4a:./bin_local/v38.sh learnjax -
Setup conda environment
tpuvm&git-lfs:./bin_tpuvm/setup_conda.sh -
Append default python path to current virtual env:
./bin_tpuvm/path.env- Usecase: You created a conda environment and fresh install pytorch, but somehow pytorch couldn't recognize tpu. A work around is just simply append default python3 path to the current env
-
Mount 300GB of ram to
$HOME/hfcache:./bin_tpuvm/mount_tmpfs.sh- Usecase: You want to prepare a dataset but 90GB disk is not enough, you
can consider using
tmpfs - You can run
sudo umount -f $HOME/hfcacheto unmount
- Usecase: You want to prepare a dataset but 90GB disk is not enough, you
can consider using
-
Set XRT config variable for
torch_xla:sh ./bin_tpuvm/xrt.env