Open hxssgaa opened 1 year ago
I have the same issue
me too
Me too
Workaround for this:
Create a v3 or v4 TPU VM (I have tested this on a v3-32 and v4-32 TPU Pod) using the tpu-ubuntu2204-base
runtime version, e.g. with resource manager it can be created like:
$ gcloud alpha compute tpus queued-resources create byt5-experiments \
--node-id byt5 --project <project-name> \
--zone us-central2-b \
--accelerator-type v4-32 \
--runtime-version tpu-ubuntu2204-base
On each worker, perform the following steps:
Install jax
on each worker:
$ pip install jax[tpu] -f https://storage.googleapis.com/jax-releases/libtpu_releases.html
Install latest flax
(here: 8360b7c
)
$ pip3 install --upgrade git+https://github.com/google/flax.git
Install latest T5X version (here: b642f30)
$ git clone https://github.com/google-research/t5x.git
$ sed -i 's!flax @ git+https://github.com/google/flax#egg=flax!flax!g' t5x/setup.py
$ cd t5x && pip3 install -e
$ cd ..
Then T5X can be used :)
Hi,
I have been following the installation instruction here.
However I encounted the following error when installing the t5x library on cloud TPU v4-8 (tpu version: tpu-ubuntu2204-base):
May I know how to resolve the issue?