bring TPU CI bakc to green by reinstall torch/torchvision

[ManfeiBai]

the TPU CI has used package_version as 2.1.0 for a long time, and we update it to 2.3.0 since PyTorch/XLA has announced 2.2 release

add short term hack pip3 install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cpu to the TPU CI script to bring TPU CI green before replaced with the new TPU CI

[ManfeiBai]

ok, according to the log of TPUCI(https://github.com/pytorch/xla/pull/6458/checks?check_run_id=21166603603), rename the package_version didn't change a lot, the chaneg would only rename the wheel, and this error might be a real error: RuntimeError: operator torchvision::nms does not exist

[ManfeiBai]

so, successed in other tests and failed in TPU CI, so there are some different when we built the wheel used in test, maybe like some env variable?

[ManfeiBai]

let's build on TPU and test to see whether we could see this failure or not

---

update: met RuntimeError: operator torchvision::nms does not exist when built on TPUVM directly

[will-cromar]

See this comment: #6439 (review)

I agree with @vanbasten23's hack in the short term: #6439 (comment)

Basically, just add pip3 install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cpu to the TPU CI script. This is going to fail some of the time when the last nightly doesn't include required changes from the current head. But, failing some of the time is better than failing all of the time. TPU CI script: https://github.com/pytorch/xla/blob/master/test/tpu/xla_test_job.yaml

A short term hack is fine, since we're trying to replace the current TPU CI anyway. cc @mbzomowski

[ManfeiBai]

Thanks, Will, make sense and TPUCI passed with this change: https://github.com/pytorch/xla/runs/21241385079

let's add this short term hack and wait for TPU CI replacement too

[will-cromar]

Thanks!

[will-cromar]

Can you add a comment with the error we're trying to work around?

[ManfeiBai]

Thanks, good point, added comment