Transformers 436 gpu

[timkuo-amazon]

GitHub Issue #, if available:

Note:

• If merging this PR should also close the associated Issue, please also add that Issue # to the Linked Issues section on the right.

• All PR's are checked weekly for staleness. This PR will be closed if not updated in 30 days.

Description

Tests run

NOTE: By default, docker builds are disabled. In order to build your container, please update dlc_developer_config.toml and specify the framework to build in "build_frameworks"

• I have run builds/tests on commit for my changes.

Confused on how to run tests? Try using the helper utility...

Assuming your remote is called origin (you can find out more with git remote -v)...

• Run default builds and tests for a particular buildspec - also commits and pushes changes to remote; Example:

python src/prepare_dlc_dev_environment.py -b </path/to/buildspec.yml> -cp origin

• Enable specific tests for a buildspec or set of buildspecs - also commits and pushes changes to remote; Example:

python src/prepare_dlc_dev_environment.py -b </path/to/buildspec.yml> -t sanity_tests -cp origin

• Restore TOML file when ready to merge

python src/prepare_dlc_dev_environment.py -rcp origin

NOTE: If you are creating a PR for a new framework version, please ensure success of the standard, rc, and efa sagemaker remote tests by updating the dlc_developer_config.toml file:

Expand

• sagemaker_remote_tests = true
• sagemaker_efa_tests = true
• sagemaker_rc_tests = true

Additionally, please run the sagemaker local tests in at least one revision:

• sagemaker_local_tests = true

Formatting

• I have run black -l 100 on my code (formatting tool: https://black.readthedocs.io/en/stable/getting_started.html)

DLC image/dockerfile

Builds to Execute

Expand

Fill out the template and click the checkbox of the builds you'd like to execute

Note: Replace with <X.Y> with the major.minor framework version (i.e. 2.2) you would like to start.

• build_pytorch_training_<X.Y>_sm

• build_pytorch_training_<X.Y>_ec2

• build_pytorch_inference_<X.Y>_sm

• build_pytorch_inference_<X.Y>_ec2

• build_pytorch_inference_<X.Y>_graviton

• build_tensorflow_training_<X.Y>_sm

• build_tensorflow_training_<X.Y>_ec2

• build_tensorflow_inference_<X.Y>_sm

• build_tensorflow_inference_<X.Y>_ec2

• build_tensorflow_inference_<X.Y>_graviton

Additional context

PR Checklist

Expand

• I've prepended PR tag with frameworks/job this applies to : [mxnet, tensorflow, pytorch] | [ei/neuron/graviton] | [build] | [test] | [benchmark] | [ec2, ecs, eks, sagemaker]
• If the PR changes affects SM test, I've modified dlc_developer_config.toml in my PR branch by setting sagemaker_tests = true and efa_tests = true
• If this PR changes existing code, the change fully backward compatible with pre-existing code. (Non backward-compatible changes need special approval.)
• (If applicable) I've documented below the DLC image/dockerfile this relates to
• (If applicable) I've documented below the tests I've run on the DLC image
• (If applicable) I've reviewed the licenses of updated and new binaries and their dependencies to make sure all licenses are on the Apache Software Foundation Third Party License Policy Category A or Category B license list. See https://www.apache.org/legal/resolved.html.
• (If applicable) I've scanned the updated and new binaries to make sure they do not have vulnerabilities associated with them.

NEURON/GRAVITON Testing Checklist

• When creating a PR:

• I've modified dlc_developer_config.toml in my PR branch by setting neuron_mode = true or graviton_mode = true

Benchmark Testing Checklist

• When creating a PR:

• I've modified dlc_developer_config.toml in my PR branch by setting ec2_benchmark_tests = true or sagemaker_benchmark_tests = true

Pytest Marker Checklist

Expand

• (If applicable) I have added the marker @pytest.mark.model("<model-type>") to the new tests which I have added, to specify the Deep Learning model that is used in the test (use "N/A" if the test doesn't use a model)
• (If applicable) I have added the marker @pytest.mark.integration("<feature-being-tested>") to the new tests which I have added, to specify the feature that will be tested
• (If applicable) I have added the marker @pytest.mark.multinode(<integer-num-nodes>) to the new tests which I have added, to specify the number of nodes used on a multi-node test
• (If applicable) I have added the marker @pytest.mark.processor(<"cpu"/"gpu"/"eia"/"neuron">) to the new tests which I have added, if a test is specifically applicable to only one processor type

By submitting this pull request, I confirm that my contribution is made under the terms of the Apache 2.0 license. I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.

[timkuo-amazon]

It looks like it’s getting hung up at various points in the runs, sometimes at a random logging, and sometimes on a random huggingface function (map/load/set_format). Things I’ve tried:

• Upgrading the instance type
• Upgrading the various huggingface packages (datasets in particular seemed to be an offender of other “hanging” issues)
• Use only one process so mutex/lock doesn’t get stuck

Other observations

• CPU usage drops basically to 0 after some time, so I think it’s getting stuck on a step but not exiting

From various huggingface/github issue threads, it seems like the next step is to get the error from a Keyboard Interrupt, but I’m not able to reproduce it locally. So, I implemented a function to timeout and then send a KeyboardInterrupt after that to simulate it. For some reason, the dlc-pr-quick-checks action isn’t running though so need to get that fixed first.

[timkuo-amazon]

/rerun

[github-actions]

This PR has been marked stale as a result of being open for 30 days without activity or updates. Please remove the stale label or comment in order to keep this open, otherwise the PR will be closed in 5 days.