Add setup for TPU CI to run every hour. (#6219)

* Add setup for TPU CI to run every hour.

* Re-organize config.yml
Co-authored-by: Lysandre <lysandre.debut@reseau.eseo.fr>

.circleci/config.yml

-version: 2
+version: 2.1
+orbs:
+    gcp-gke: circleci/gcp-gke@1.0.4
+    go: circleci/go@1.3.0
+
+# TPU REFERENCES
+references:
+    checkout_ml_testing: &checkout_ml_testing
+        run:
+            name: Checkout ml-testing-accelerators
+            command: |
+                git clone https://github.com/GoogleCloudPlatform/ml-testing-accelerators.git
+                cd ml-testing-accelerators
+                git fetch origin 5e88ac24f631c27045e62f0e8d5dfcf34e425e25:stable
+                git checkout stable
+    build_push_docker: &build_push_docker
+        run:
+            name: Configure Docker
+            command: |
+                gcloud --quiet auth configure-docker
+                cd docker/transformers-pytorch-tpu
+                if [ -z "$CIRCLE_PR_NUMBER" ]; then docker build --tag "$GCR_IMAGE_PATH:$CIRCLE_WORKFLOW_JOB_ID" -f Dockerfile --build-arg "TEST_IMAGE=1"; else docker build --tag "$GCR_IMAGE_PATH:$CIRCLE_WORKFLOW_JOB_ID" -f Dockerfile --build-arg "TEST_IMAGE=1" --build-arg "GITHUB_REF=pull/$CIRCLE_PR_NUMBER/head" . ; fi
+                docker push "$GCR_IMAGE_PATH:$CIRCLE_WORKFLOW_JOB_ID"
+    deploy_cluster: &deploy_cluster
+        run:
+            name: Deploy the job on the kubernetes cluster
+            command: |
+                go get github.com/google/go-jsonnet/cmd/jsonnet && \
+                export PATH=$PATH:$HOME/go/bin && \
+                kubectl create -f docker/transformers-pytorch-tpu/dataset.yaml || true && \
+                job_name=$(jsonnet -J ml-testing-accelerators/ docker/transformers-pytorch-tpu/bert-base-cased.jsonnet --ext-str image=$GCR_IMAGE_PATH --ext-str image-tag=$CIRCLE_WORKFLOW_JOB_ID | kubectl create -f -) && \
+                job_name=${job_name#job.batch/} && \
+                job_name=${job_name% created} && \
+                echo "Waiting on kubernetes job: $job_name" && \
+                i=0 && \
+                # 30 checks spaced 30s apart = 900s total.
+                max_checks=30 && \
+                status_code=2 && \
+                # Check on the job periodically. Set the status code depending on what
+                # happened to the job in Kubernetes. If we try max_checks times and
+                # still the job hasn't finished, give up and return the starting
+                # non-zero status code.
+                while [ $i -lt $max_checks ]; do ((i++)); if kubectl get jobs $job_name -o jsonpath='Failed:{.status.failed}' | grep "Failed:1"; then status_code=1 && break; elif kubectl get jobs $job_name -o jsonpath='Succeeded:{.status.succeeded}' | grep "Succeeded:1" ; then status_code=0 && break; else echo "Job not finished yet"; fi; sleep 30; done && \
+                echo "Done waiting. Job status code: $status_code" && \
+                # Allow time for logs to flush.
+                sleep 60 && \
+                echo "JOB_NAME: $job_name" && \
+                gcloud logging read "resource.type=k8s_container resource.labels.project_id=$GOOGLE_PROJECT_ID resource.labels.location=$GOOGLE_COMPUTE_ZONE resource.labels.cluster_name=$GKE_CLUSTER resource.labels.namespace_name=default resource.labels.pod_name:$job_name" --limit 10000000 --order asc --format 'value(textPayload)' --project=$GOOGLE_PROJECT_ID && \
+                echo "Done with log retrieval attempt." && \
+                gcloud container images delete "$GCR_IMAGE_PATH:$CIRCLE_WORKFLOW_JOB_ID" --force-delete-tags && \
+                exit $status_code
+    delete_gke_jobs: &delete_gke_jobs
+        run:
+            name: Delete GKE Jobs
+            command: |
+                # Match jobs whose age matches patterns like '1h' or '1d', i.e. any job
+                # that has been around longer than 1hr. First print all columns for
+                # matches, then execute the delete.
+                kubectl get job | awk 'match($4,/[0-9]+[dh]/) {print $0}'
+                kubectl delete job $(kubectl get job | awk 'match($4,/[0-9]+[dh]/) {print $1}')
+
+
+
+
 jobs:
     run_tests_torch_and_tf:
         working_directory: ~/transformers
@@ -50,7 +113,6 @@ jobs:
             - store_artifacts:
                   path: ~/transformers/output.txt
                   destination: test_output.txt
-
     run_tests_tf:
         working_directory: ~/transformers
         docker:
@@ -193,6 +255,35 @@ jobs:
             - checkout
             - run: pip install requests
             - run: python ./utils/link_tester.py
+
+# TPU JOBS
+    run_examples_tpu:
+        docker:
+            - image: circleci/python:3.6
+        environment:
+            OMP_NUM_THREADS: 1
+        resource_class: xlarge
+        parallelism: 1
+        steps:
+            - checkout
+            - go/install
+            - *checkout_ml_testing
+            - gcp-gke/install
+            - gcp-gke/update-kubeconfig-with-credentials:
+                  cluster: $GKE_CLUSTER
+                  perform-login: true
+            - setup_remote_docker
+            - *build_push_docker
+            - *deploy_cluster
+    cleanup-gke-jobs:
+        docker:
+            - image: circleci/python:3.6
+        steps:
+            - gcp-gke/install
+            - gcp-gke/update-kubeconfig-with-credentials:
+                  cluster: $GKE_CLUSTER
+                  perform-login: true
+            - *delete_gke_jobs
 workflow_filters: &workflow_filters
     filters:
         branches:
@@ -211,3 +302,15 @@ workflows:
             - run_tests_tf
             - build_doc
             - deploy_doc: *workflow_filters
+    tpu_testing_jobs:
+        triggers:
+            - schedule:
+                # Set to run at the first minute of every hour.
+                cron: "0 8 * * *"
+                filters:
+                    branches:
+                        only:
+                            - master
+        jobs:
+            - cleanup-gke-jobs
+            - run_examples_tpu

docker/transformers-pytorch-tpu/Dockerfile

+FROM google/cloud-sdk:slim
+
+# Build args.
+ARG GITHUB_REF=refs/heads/master
+
+# TODO: This Dockerfile installs pytorch/xla 3.6 wheels. There are also 3.7
+# wheels available; see below.
+ENV PYTHON_VERSION=3.6
+
+RUN apt-get update && apt-get install -y --no-install-recommends \
+         build-essential \
+         cmake \
+         git \
+         curl \
+         ca-certificates
+
+# Install conda and python.
+# NOTE new Conda does not forward the exit status... https://github.com/conda/conda/issues/8385
+RUN curl -o ~/miniconda.sh https://repo.anaconda.com/miniconda/Miniconda3-4.7.12-Linux-x86_64.sh  && \
+    chmod +x ~/miniconda.sh && \
+    ~/miniconda.sh -b && \
+    rm ~/miniconda.sh
+
+ENV PATH=/root/miniconda3/bin:$PATH
+
+RUN conda create -y --name container python=$PYTHON_VERSION
+
+# Run the rest of commands within the new conda env.
+# Use absolute path to appease Codefactor.
+SHELL ["/root/miniconda3/bin/conda", "run", "-n", "container", "/bin/bash", "-c"]
+RUN conda install -y python=$PYTHON_VERSION mkl
+
+RUN pip uninstall -y torch && \
+    # Python 3.7 wheels are available. Replace cp36-cp36m with cp37-cp37m
+    gsutil cp 'gs://tpu-pytorch/wheels/torch-nightly-cp${PYTHON_VERSION/./}-cp${PYTHON_VERSION/./}m-linux_x86_64.whl' . && \
+    gsutil cp 'gs://tpu-pytorch/wheels/torch_xla-nightly-cp${PYTHON_VERSION/./}-cp${PYTHON_VERSION/./}m-linux_x86_64.whl' . && \
+    gsutil cp 'gs://tpu-pytorch/wheels/torchvision-nightly-cp${PYTHON_VERSION/./}-cp${PYTHON_VERSION/./}m-linux_x86_64.whl' . && \
+    pip install 'torch-nightly-cp${PYTHON_VERSION/./}-cp${PYTHON_VERSION/./}m-linux_x86_64.whl' && \
+    pip install 'torch_xla-nightly-cp${PYTHON_VERSION/./}-cp${PYTHON_VERSION/./}m-linux_x86_64.whl' && \
+    pip install 'torchvision-nightly-cp${PYTHON_VERSION/./}-cp${PYTHON_VERSION/./}m-linux_x86_64.whl' && \
+    rm 'torch-nightly-cp${PYTHON_VERSION/./}-cp${PYTHON_VERSION/./}m-linux_x86_64.whl' && \
+    rm 'torch_xla-nightly-cp${PYTHON_VERSION/./}-cp${PYTHON_VERSION/./}m-linux_x86_64.whl' && \
+    rm 'torchvision-nightly-cp${PYTHON_VERSION/./}-cp${PYTHON_VERSION/./}m-linux_x86_64.whl' && \
+    apt-get install -y libomp5
+
+ENV LD_LIBRARY_PATH=root/miniconda3/envs/container/lib
+
+
+# Install huggingface/transformers at the current PR, plus dependencies.
+RUN git clone https://github.com/huggingface/transformers.git && \
+    cd transformers && \
+    git fetch origin $GITHUB_REF:CI && \
+    git checkout CI && \
+    cd .. && \
+    pip install ./transformers && \
+    pip install -r ./transformers/examples/requirements.txt && \
+    pip install pytest
+
+RUN python -c "import torch_xla; print(torch_xla.__version__)"
+RUN python -c "import transformers as trf; print(trf.__version__)"
+RUN conda init bash
+COPY docker-entrypoint.sh /usr/local/bin/
+RUN chmod +x /usr/local/bin/docker-entrypoint.sh
+ENTRYPOINT ["/usr/local/bin/docker-entrypoint.sh"]
+CMD ["bash"]

docker/transformers-pytorch-tpu/bert-base-cased.jsonnet

+local base = import 'templates/base.libsonnet';
+local tpus = import 'templates/tpus.libsonnet';
+local utils = import "templates/utils.libsonnet";
+local volumes = import "templates/volumes.libsonnet";
+
+local bertBaseCased = base.BaseTest {
+  frameworkPrefix: "hf",
+  modelName: "bert-base-cased",
+  mode: "example",
+  configMaps: [],
+
+  timeout: 3600, # 1 hour, in seconds
+
+  image: std.extVar('image'),
+  imageTag: std.extVar('image-tag'),
+
+  tpuSettings+: {
+    softwareVersion: "pytorch-nightly",
+  },
+  accelerator: tpus.v3_8,
+
+  volumeMap+: {
+    datasets: volumes.PersistentVolumeSpec {
+      name: "huggingface-cluster-disk",
+      mountPath: "/datasets",
+    },
+  },
+  command: utils.scriptCommand(
+    |||
+      python -m pytest -s transformers/examples/test_xla_examples.py -v
+      test_exit_code=$?
+      echo "\nFinished running commands.\n"
+      test $test_exit_code -eq 0
+    |||
+  ),
+};
+
+bertBaseCased.oneshotJob

docker/transformers-pytorch-tpu/dataset.yaml

+apiVersion: v1
+kind: PersistentVolume
+metadata:
+  name: huggingface-cluster-disk
+spec:
+  storageClassName: ""
+  capacity:
+    storage: 500Gi
+  accessModes:
+    - ReadOnlyMany
+  claimRef:
+    namespace: default
+    name: huggingface-cluster-disk-claim
+  gcePersistentDisk:
+    pdName: huggingface-cluster-disk
+    fsType: ext4
+    readOnly: true
+---
+apiVersion: v1
+kind: PersistentVolumeClaim
+metadata:
+  name: huggingface-cluster-disk-claim
+spec:
+  # Specify "" as the storageClassName so it matches the PersistentVolume's StorageClass.
+  # A nil storageClassName value uses the default StorageClass. For details, see
+  # https://kubernetes.io/docs/concepts/storage/persistent-volumes/#class-1
+  storageClassName: ""
+  accessModes:
+    - ReadOnlyMany
+  resources:
+    requests:
+      storage: 1Ki

docker/transformers-pytorch-tpu/docker-entrypoint.sh

+#!/bin/bash
+source ~/.bashrc
+echo "running docker-entrypoint.sh"
+conda activate container
+echo $KUBE_GOOGLE_CLOUD_TPU_ENDPOINTS
+echo "printed TPU info"
+export XRT_TPU_CONFIG="tpu_worker;0;${KUBE_GOOGLE_CLOUD_TPU_ENDPOINTS:7}"
+exec "$@"#!/bin/bash

examples/test_xla_examples.py

@@ -14,7 +14,6 @@
 # limitations under the License.
 
 
-import argparse
 import logging
 import sys
 import unittest
@@ -29,13 +28,6 @@ logging.basicConfig(level=logging.DEBUG)
 logger = logging.getLogger()
 
 
-def get_setup_file():
-    parser = argparse.ArgumentParser()
-    parser.add_argument("-f")
-    args = parser.parse_args()
-    return args.f
-
-
 @require_torch_tpu
 class TorchXLAExamplesTests(unittest.TestCase):
     def test_run_glue(self):
@@ -47,13 +39,13 @@ class TorchXLAExamplesTests(unittest.TestCase):
         output_directory = "run_glue_output"
 
         testargs = f"""
-            text-classification/run_glue.py
+            transformers/examples/text-classification/run_glue.py
             --num_cores=8
-            text-classification/run_glue.py
+            transformers/examples/text-classification/run_glue.py
             --do_train
             --do_eval
             --task_name=MRPC
-            --data_dir=../glue_data/MRPC
+            --data_dir=/datasets/glue_data/MRPC
             --cache_dir=./cache_dir
             --num_train_epochs=1
             --max_seq_length=128
@@ -87,5 +79,5 @@ class TorchXLAExamplesTests(unittest.TestCase):
                 # Assert that the model trains
                 self.assertGreaterEqual(value, 0.70)
 
-            # Assert that the script takes less than 100 seconds to make sure it doesn't hang.
-            self.assertLess(end - start, 100)
+            # Assert that the script takes less than 300 seconds to make sure it doesn't hang.
+            self.assertLess(end - start, 300)