An older torchvision version is used in the nightly tpu image

[hosseinsarshar]

🐛 Bug - old torchvision is used in the nightly tpu image

I noticed an older torchvision being used in the nightly image - Here is what I tried:

gcloud compute tpus tpu-vm ssh $TPU_NAME --project $PROJECT --zone=$ZONE --worker=all --command='
DOCKER_IMAGE=us-central1-docker.pkg.dev/tpu-pytorch-releases/docker/xla:nightly_3.10_tpuvm

worker_id=1

cat >> /dev/null <<EOF
EOF

stdbuf -oL bash <<-PIPE_EOF 2>&1 | sed "s/^/[worker $worker_id] /g" | tee runlog
  set -o xtrace
  # Configure docker
  sudo groupadd docker
  sudo usermod -aG docker $USER
  # newgrp applies updated group permissions
  newgrp - docker
  gcloud auth configure-docker us-central1-docker.pkg.dev --quiet
  # Kill any running benchmarks
  docker kill $USER-test || true

  docker pull $DOCKER_IMAGE

docker run \
    --name $USER-test \
    --privileged \
    -v /home/$USER:/tmp/home \
    --shm-size=16G \
    --net host \
    -u root \
     --rm $DOCKER_IMAGE  python -c "import torch_xla; import torch; import torchvision; print(f'{torch_xla.devices()=}'); print(f'{torch.__version__=}'); print(f'{torchvision.__version__=}'); print(f'{torch_xla.__version__=}')"

PIPE_EOF
'

This is what I get:

[worker ] torch_xla.devices()=[device(type='xla', index=0)]
[worker ] torch.__version__='2.7.0'
[worker ] torchvision.__version__='0.19.0a0+d23a6e1'
[worker ] torch_xla.__version__='2.7.0+gita295f7d'

By checking the commit ID on the upstream torchvision repo - I get to a commit submitted in May, 2024 - wonder if this is intended or it needs to be updated to a more recent version.

The git commit on torchvision: pytorch/vision@d23a6e1

Thanks

[miladm]

cc @ysiraichi and @tengyifei to align 2.6 whl readiness + nightly whl alignments - it seems torchvision and torchxla nightlies have been incompatible since ~20241216