`task_rust.sh` and `task_cpp_unittest.sh` fail with updated Docker images when `USE_VITIS_AI ON`

[leandron]

It looks like there is an error happening in ./tests/scripts/task_rust.sh when we have updated images including Python 3.7/TensorFlow 2.6/h5py 3.1.0. The error only happens when USE_VITIS_AI ON in ./tests/scripts/task_config_build_cpu.sh.

This is causing me issues when testing TensorFlow 2.6 images. I tried downgrading h5py and reproducing the steps without success. Despite that TensorFlow 2.6 requires h5py>3, so it is not a viable solution anyway.

Steps to reproduce:

cd <your tvm repo>
rm -rf build
docker pull tlcpackstaging/ci_cpu:20220321-061034-bd684deb5
./docker/bash.sh tlcpackstaging/ci_cpu:20220321-061034-bd684deb5

# this happens inside the container
./tests/scripts/task_config_build_cpu.sh build
./tests/scripts/task_build.py
./tests/scripts/task_ci_setup.sh
./tests/scripts/task_rust.sh         # this is the one that crashes

Then you should see an error similar to this:

test device::tests::device ... ok

test result: ok. 2 passed; 0 failed; 0 ignored; 0 measured; 0 filtered out; finished in 0.00s

free(): invalid pointer
error: test failed, to rerun pass '--lib'

Caused by:
  process didn't exit successfully: `/workspace/rust/target/debug/deps/tvm_sys-0b86e7b6a5ad689c` (signal: 6, SIGABRT: process abort signal)
script returned exit code 101

By setting USE_VITIS_AI OFF in ./tests/scripts/task_config_build_cpu.sh and repeating the steps, the issue doesn't reproduce.

Full output in Jenkins: https://ci.tlcpack.ai/blue/organizations/jenkins/docker-images-ci%2Fdocker-image-run-tests/detail/docker-image-run-tests/69/pipeline/

cc @jtuyls @anilmartha can you help me troubleshooting this?

[leandron]

Skipping task_rust.sh I also see an issue during cpp tests. Same case, similar error message. [2022-03-22T10:56:59.483Z] free(): invalid pointer.

Full log at: https://ci.tlcpack.ai/blue/rest/organizations/jenkins/pipelines/docker-images-ci/pipelines/docker-image-run-tests/runs/71/nodes/307/steps/514/log/?start=0

[jtuyls]

@leandron I am looking at this. Not sure what is going wrong yet.

[masahi]

The CI update for ci-cpu is apparently blocked by this https://ci.tlcpack.ai/blue/organizations/jenkins/tvm/detail/ci-docker-staging/239/pipeline/60

I've hit this error twice today.

[jtuyls]

This seems to be happening when tensorflow is being loaded inside pyxir. I don't know what the exact cause is for this starting to happen with tensorflow 2.6 but the issue looks very similar to this one: triton-inference-server/server#3777 but with tensorflow instead of pytorch.

I have a workaround by loading tensorflow eagerly only when needed. I am currently verifying with the ci-cpu docker image locally and will create a PR when successful.

[leandron]

Based on results at https://ci.tlcpack.ai/job/docker-images-ci/job/docker-image-run-tests/82/, which include #10858 (up to fcdf463 in the repo), we can update the Docker images now, as the bug reported here is fixed.