Training not starting on v3-128

I'm trying to train a T5 model with C4 data. This is the command that I'm using to train the model

```
python ${T5X_DIR}/t5x/train.py \
--gin_file="t5x/examples/t5/t5_1_1/examples/small_c4_pretrain.gin" \
--gin.MODEL_DIR=\"${MODEL_DIR}\" \
--alsologtostderr
```

UPDATE, even this is stuck
```
import jax
jax.device_count()
```

It's stuck at this
> UserWarning: TPU backend initialization is taking more than 60.0 seconds. Did you run your code on all TPU hosts? See https://jax.readthedocs.io/en/latest/multi_process.html for more information.

Can someone please point me what I'm doing wrong?

Start with t5x/train.py and reproduce the reported command, then run the supplied jax.device_count() check while reviewing the TPU multi-process initialization warning. Done means TPU initialization completes and the T5X training command proceeds beyond startup.

Written by the indexing model from the issue text.