Sagemaker test (#10925)

* init

* first working test

* added todo for setup.py

* working test for single node multi node ddp and smd

* added tensorflow single node test

* added directory for pytorch and tensorflow due to different requirements.txt

* added directory for pytorch and tensorflow

* added comment for run_glue until it is available

* added output_dir to it

* smaller dataset to make test running faster

* adjust HP and script

* adjusted parameter for tensorflow

* refactored test scripts

* adjusted make file

* init

* first working test

* added todo for setup.py

* working test for single node multi node ddp and smd

* added tensorflow single node test

* added directory for pytorch and tensorflow due to different requirements.txt

* added directory for pytorch and tensorflow

* added comment for run_glue until it is available

* added output_dir to it

* smaller dataset to make test running faster

* adjust HP and script

* adjusted parameter for tensorflow

* refactored test scripts

* adjusted make file

* updated dlc container

* commented in all tests

* added both ecr images

* added new master branches

* debug

* added new datasets version

* init

* strange rebase bug

* removed changes

* changed min version for tests to work

* updated DLC

* added model parallel test

* removed test files

* removed test files

* tested with ned dlc

* added correct sagemaker sdk version

* adjust DLCs for official one

* reworked tests

* quality

* removed default profile added documentation to it

* added step in release for sagemaker tests

* reverted version for example script removed duplicated script and added install from master to requirements.txt

* removed mistaken .DS_Stores from mac

* fixed tests

* added Sylvains feedback

* make style

* added lysandre's feedback

Makefile

@@ -74,6 +74,12 @@ test:
 test-examples:
 	python -m pytest -n auto --dist=loadfile -s -v ./examples/
 
+# Run tests for SageMaker DLC release
+
+test-sagemaker: # install sagemaker dependencies in advance with pip install .[sagemaker]
+	TEST_SAGEMAKER=True python -m pytest -n auto  -s -v ./tests/sagemaker
+
+
 # Check that docs can build
 
 docs:

setup.py

@@ -19,15 +19,17 @@ To create the package for pypi.
 
 1. Run `make pre-release` (or `make pre-patch` for a patch release) then run `make fix-copies` to fix the index of the
    documentation.
+   
+2. Run Tests for Amazon Sagemaker. The documentation is located in `./tests/sagemaker/README.md`, otherwise @philschmid.
 
-2. Unpin specific versions from setup.py that use a git install.
+3. Unpin specific versions from setup.py that use a git install.
 
-3. Commit these changes with the message: "Release: VERSION"
+4. Commit these changes with the message: "Release: VERSION"
 
-4. Add a tag in git to mark the release: "git tag VERSION -m 'Adds tag VERSION for pypi' "
+5. Add a tag in git to mark the release: "git tag VERSION -m 'Adds tag VERSION for pypi' "
    Push the tag to git: git push --tags origin master
 
-5. Build both the sources and the wheel. Do not change anything in setup.py between
+6. Build both the sources and the wheel. Do not change anything in setup.py between
    creating the wheel and the source distribution (obviously).
 
    For the wheel, run: "python setup.py bdist_wheel" in the top level directory.
@@ -36,7 +38,7 @@ To create the package for pypi.
    For the sources, run: "python setup.py sdist"
    You should now have a /dist directory with both .whl and .tar.gz source versions.
 
-6. Check that everything looks correct by uploading the package to the pypi test server:
+7. Check that everything looks correct by uploading the package to the pypi test server:
 
    twine upload dist/* -r pypitest
    (pypi suggest using twine as other methods upload files via plaintext.)
@@ -46,12 +48,12 @@ To create the package for pypi.
    Check that you can install it in a virtualenv by running:
    pip install -i https://testpypi.python.org/pypi transformers
 
-7. Upload the final version to actual pypi:
+8. Upload the final version to actual pypi:
    twine upload dist/* -r pypi
 
-8. Copy the release notes from RELEASE.md to the tag in github once everything is looking hunky-dory.
+9. Copy the release notes from RELEASE.md to the tag in github once everything is looking hunky-dory.
 
-9. Run `make post-release` (or `make post-patch` for a patch release).
+10. Run `make post-release` (or `make post-patch` for a patch release).
 """
 
 import os
@@ -134,6 +136,7 @@ _deps = [
     "unidic>=1.0.2",
     "unidic_lite>=1.0.7",
     "uvicorn",
+    "sagemaker>=2.31.0",
 ]
 
 
@@ -223,12 +226,16 @@ extras["onnxruntime"] = deps_list("onnxruntime", "onnxruntime-tools")
 extras["onnx"] = deps_list("onnxconverter-common", "keras2onnx") + extras["onnxruntime"]
 extras["modelcreation"] = deps_list("cookiecutter")
 
+extras["sagemaker"] = deps_list("sagemaker")
+
 extras["serving"] = deps_list("pydantic", "uvicorn", "fastapi", "starlette")
 extras["speech"] = deps_list("soundfile", "torchaudio")
 
 extras["sentencepiece"] = deps_list("sentencepiece", "protobuf")
 extras["testing"] = (
-    deps_list("pytest", "pytest-xdist", "timeout-decorator", "parameterized", "psutil", "datasets", "pytest-sugar", "black")
+    deps_list(
+        "pytest", "pytest-xdist", "timeout-decorator", "parameterized", "psutil", "datasets", "pytest-sugar", "black"
+    )
     + extras["retrieval"]
     + extras["modelcreation"]
 )

src/transformers/dependency_versions_table.py

@@ -53,4 +53,5 @@ deps = {
     "unidic": "unidic>=1.0.2",
     "unidic_lite": "unidic_lite>=1.0.7",
     "uvicorn": "uvicorn",
+    "sagemaker": "sagemaker>=2.31.0",
 }

tests/sagemaker/README.md

+# Testing new Hugging Face Deep Learning Container.
+
+This document explains the testing strategy for releasing the new Hugging Face Deep Learning Container. AWS maintains 14 days of currency with framework releases. Besides framework releases, AWS release train is bi-weekly on Monday. Code cutoff date for any changes is the Wednesday before release-Monday. 
+
+
+## Test Case 1: Releasing a New Version (Minor/Major) of 🤗 Transformers
+
+### Requirements: Test should run on Release Candidate for new `transformers` release to validate the new release is compatible with the DLCs. To run these tests you need credentials for the HF SageMaker AWS Account. You can ask @philschmid or @n1t0 to get access.
+
+### Run Tests:
+
+Before we can run the tests we need to adjust the `requirements.txt` for PyTorch under `/tests/sagemaker/scripts/pytorch` and for TensorFlow under `/tests/sagemaker/scripts/pytorch`. We adjust the branch to the new RC-tag.
+
+```
+git+https://github.com/huggingface/transformers.git@v4.5.0.rc0 # install master or adjust ist with vX.X.X for installing version specific-transforms
+```
+
+After we adjusted the `requirements.txt` we can run Amazon SageMaker tests with:  
+
+```bash
+AWS_PROFILE=<enter-your-profile> make sagemaker-test
+```
+These tests take around 10-15 minutes to finish. Preferably make a screenshot of the successfully ran tests.
+
+### After Transformers Release:
+
+After we have released the Release Candidate we need to create a PR at the [Deep Learning Container Repository](https://github.com/aws/deep-learning-containers).
+
+**Creating the update PR:**
+
+1. Update the two latest `buildspec.yaml` config for [PyTorch](https://github.com/aws/deep-learning-containers/tree/master/huggingface/pytorch) and [TensorFlow](https://github.com/aws/deep-learning-containers/tree/master/huggingface/tensorflow). The two latest `buildspec.yaml` are the `buildspec.yaml` without a version tag and the one with the highest framework version, e.g. `buildspec-1-7-1.yml` and not `buildspec-1-6.yml`.  
+
+To update the `buildspec.yaml` we need to adjust either the `transformers_version` or the `datasets_version` or both. Example for upgrading to `transformers 4.5.0` and `datasets 1.6.0`.
+```yaml
+account_id: &ACCOUNT_ID <set-$ACCOUNT_ID-in-environment>
+region: &REGION <set-$REGION-in-environment>
+base_framework: &BASE_FRAMEWORK pytorch
+framework: &FRAMEWORK !join [ "huggingface_", *BASE_FRAMEWORK]
+version: &VERSION 1.6.0
+short_version: &SHORT_VERSION 1.6
+
+repository_info:
+  training_repository: &TRAINING_REPOSITORY
+    image_type: &TRAINING_IMAGE_TYPE training
+    root: !join [ "huggingface/", *BASE_FRAMEWORK, "/", *TRAINING_IMAGE_TYPE ]
+    repository_name: &REPOSITORY_NAME !join ["pr", "-", "huggingface", "-", *BASE_FRAMEWORK, "-", *TRAINING_IMAGE_TYPE]
+    repository: &REPOSITORY !join [ *ACCOUNT_ID, .dkr.ecr., *REGION, .amazonaws.com/,
+      *REPOSITORY_NAME ]
+
+images:
+  BuildHuggingFacePytorchGpuPy37Cu110TrainingDockerImage:
+    <<: *TRAINING_REPOSITORY
+    build: &HUGGINGFACE_PYTORCH_GPU_TRAINING_PY3 false
+    image_size_baseline: &IMAGE_SIZE_BASELINE 15000
+    device_type: &DEVICE_TYPE gpu
+    python_version: &DOCKER_PYTHON_VERSION py3
+    tag_python_version: &TAG_PYTHON_VERSION py36
+    cuda_version: &CUDA_VERSION cu110
+    os_version: &OS_VERSION ubuntu18.04
+    transformers_version: &TRANSFORMERS_VERSION 4.5.0 # this was adjusted from 4.4.2 to 4.5.0
+    datasets_version: &DATASETS_VERSION 1.6.0 # this was adjusted from 1.5.0 to 1.6.0
+    tag: !join [ *VERSION, '-', 'transformers', *TRANSFORMERS_VERSION, '-', *DEVICE_TYPE, '-', *TAG_PYTHON_VERSION, '-',
+      *CUDA_VERSION, '-', *OS_VERSION ]
+    docker_file: !join [ docker/, *SHORT_VERSION, /, *DOCKER_PYTHON_VERSION, /, 
+      *CUDA_VERSION, /Dockerfile., *DEVICE_TYPE ]
+```
+2. In the PR comment describe what test, we ran and with which package versions. Here you can copy the table from [Current Tests](#current-tests). 
+
+TODO: Add a screenshot of PR + Text template to make it easy to open. 
+ 
+## Test Case 2: Releasing a New AWS Framework DLC
+
+
+## Execute Tests
+
+### Requirements:
+AWS is going to release new DLCs for PyTorch and/or TensorFlow. The Tests should run on the new framework versions with current `transformers` release to validate the new framework release is compatible with the `transformers` version. To run these tests you need credentials for the HF SageMaker AWS Account. You can ask @philschmid or @n1t0 to get access. AWS will notify us with a new issue in the repository pointing to their framework upgrade PR.
+
+### Run Tests:
+
+Before we can run the tests we need to adjust the `requirements.txt` for Pytorch under `/tests/sagemaker/scripts/pytorch` and for Tensorflow under `/tests/sagemaker/scripts/pytorch`. We add the new framework version to it.
+
+```
+torch==1.8.1 # for pytorch
+tensorflow-gpu==2.5.0 # for tensorflow
+```
+
+After we adjusted the `requirements.txt` we can run Amazon SageMaker tests with. 
+
+```bash
+AWS_PROFILE=<enter-your-profile> make sagemaker-test
+```
+These tests take around 10-15 minutes to finish. Preferably make a screenshot of the successfully ran tests.
+
+
+### After successful Tests:
+
+After we have successfully run tests for the new framework version we need to create a PR at the [Deep Learning Container Repository](https://github.com/aws/deep-learning-containers).
+
+**Creating the update PR:**
+
+1. Create a new `buildspec.yaml` config for [PyTorch](https://github.com/aws/deep-learning-containers/tree/master/huggingface/pytorch) and [TensorFlow](https://github.com/aws/deep-learning-containers/tree/master/huggingface/tensorflow) and rename the old `buildspec.yaml` to `buildespec-x.x.x`, where `x.x.x` is the base framework version, e.g. if pytorch 1.6.0 is the latest version in `buildspec.yaml` the file should be renamed to `buildspec-yaml-1-6.yaml`. 
+
+To create the new `buildspec.yaml` we need to adjust  the `version` and the `short_version`. Example for upgrading to `pytorch 1.7.1`. 
+
+```yaml
+account_id: &ACCOUNT_ID <set-$ACCOUNT_ID-in-environment>
+region: &REGION <set-$REGION-in-environment>
+base_framework: &BASE_FRAMEWORK pytorch
+framework: &FRAMEWORK !join [ "huggingface_", *BASE_FRAMEWORK]
+version: &VERSION 1.7.1 # this was adjusted from 1.6.0 to 1.7.1
+short_version: &SHORT_VERSION 1.7 # this was adjusted from 1.6 to 1.7
+
+repository_info:
+  training_repository: &TRAINING_REPOSITORY
+    image_type: &TRAINING_IMAGE_TYPE training
+    root: !join [ "huggingface/", *BASE_FRAMEWORK, "/", *TRAINING_IMAGE_TYPE ]
+    repository_name: &REPOSITORY_NAME !join ["pr", "-", "huggingface", "-", *BASE_FRAMEWORK, "-", *TRAINING_IMAGE_TYPE]
+    repository: &REPOSITORY !join [ *ACCOUNT_ID, .dkr.ecr., *REGION, .amazonaws.com/,
+      *REPOSITORY_NAME ]
+
+images:
+  BuildHuggingFacePytorchGpuPy37Cu110TrainingDockerImage:
+    <<: *TRAINING_REPOSITORY
+    build: &HUGGINGFACE_PYTORCH_GPU_TRAINING_PY3 false
+    image_size_baseline: &IMAGE_SIZE_BASELINE 15000
+    device_type: &DEVICE_TYPE gpu
+    python_version: &DOCKER_PYTHON_VERSION py3
+    tag_python_version: &TAG_PYTHON_VERSION py36
+    cuda_version: &CUDA_VERSION cu110
+    os_version: &OS_VERSION ubuntu18.04
+    transformers_version: &TRANSFORMERS_VERSION 4.4.2
+    datasets_version: &DATASETS_VERSION 1.5.0
+    tag: !join [ *VERSION, '-', 'transformers', *TRANSFORMERS_VERSION, '-', *DEVICE_TYPE, '-', *TAG_PYTHON_VERSION, '-',
+      *CUDA_VERSION, '-', *OS_VERSION ]
+    docker_file: !join [ docker/, *SHORT_VERSION, /, *DOCKER_PYTHON_VERSION, /, 
+      *CUDA_VERSION, /Dockerfile., *DEVICE_TYPE ]
+```
+2. In the PR comment describe what test we ran and with which framework versions. Here you can copy the table from [Current Tests](#current-tests). 
+
+TODO: Add a screenshot of PR + Text template to make it easy to open. 
+
+
+## Current Tests
+
+| ID                                  | Description                                                       | Platform                   | #GPUS | Collected & evaluated metrics            |
+|-------------------------------------|-------------------------------------------------------------------|-----------------------------|-------|------------------------------------------|
+| pytorch-transfromers-test-single    | test bert finetuning using BERT fromtransformerlib+PT             | SageMaker createTrainingJob | 1     | train_runtime, eval_accuracy & eval_loss |
+| pytorch-transfromers-test-2-ddp     | test bert finetuning using BERT from transformer lib+ PT DPP      | SageMaker createTrainingJob | 16    | train_runtime, eval_accuracy & eval_loss |
+| pytorch-transfromers-test-2-smd     | test bert finetuning using BERT from transformer lib+ PT SM DDP   | SageMaker createTrainingJob | 16    | train_runtime, eval_accuracy & eval_loss |
+| pytorch-transfromers-test-1-smp     | test roberta finetuning using BERT from transformer lib+ PT SM MP | SageMaker createTrainingJob | 8     | train_runtime, eval_accuracy & eval_loss |
+| tensorflow-transfromers-test-single | Test bert finetuning using BERT from transformer lib+TF           | SageMaker createTrainingJob | 1     | train_runtime, eval_accuracy & eval_loss |
+| tensorflow-transfromers-test-2-smd  | test bert finetuning using BERT from transformer lib+ TF SM DDP   | SageMaker createTrainingJob | 16    | train_runtime, eval_accuracy & eval_loss |
\ No newline at end of file

tests/sagemaker/__init__.py

+import importlib
+
+
+def is_sagemaker_available():
+    return importlib.util.find_spec("sagemaker") is not None

tests/sagemaker/conftest.py

+# we define a fixture function below and it will be "used" by
+# referencing its name from tests
+
+import os
+
+import pytest
+
+from attr import dataclass
+
+
+os.environ["AWS_DEFAULT_REGION"] = "us-east-1"  # defaults region
+
+
+@dataclass
+class SageMakerTestEnvironment:
+    framework: str
+    role = "arn:aws:iam::558105141721:role/sagemaker_execution_role"
+    hyperparameters = {
+        "task_name": "mnli",
+        "per_device_train_batch_size": 32,
+        "per_device_eval_batch_size": 32,
+        "do_train": True,
+        "do_eval": True,
+        "do_predict": True,
+        "output_dir": "/opt/ml/model",
+        "overwrite_output_dir": True,
+        "max_steps": 500,
+        "save_steps": 5500,
+    }
+    distributed_hyperparameters = {**hyperparameters, "max_steps": 1000}
+
+    @property
+    def metric_definitions(self) -> str:
+        if self.framework == "pytorch":
+            return [
+                {"Name": "train_runtime", "Regex": "train_runtime.*=\D*(.*?)$"},
+                {"Name": "eval_accuracy", "Regex": "eval_accuracy.*=\D*(.*?)$"},
+                {"Name": "eval_loss", "Regex": "eval_loss.*=\D*(.*?)$"},
+            ]
+        else:
+            return [
+                {"Name": "train_runtime", "Regex": "train_runtime.*=\D*(.*?)$"},
+                {"Name": "eval_accuracy", "Regex": "loss.*=\D*(.*?)]?$"},
+                {"Name": "eval_loss", "Regex": "sparse_categorical_accuracy.*=\D*(.*?)]?$"},
+            ]
+
+    @property
+    def base_job_name(self) -> str:
+        return f"{self.framework}-transfromers-test"
+
+    @property
+    def test_path(self) -> str:
+        return f"./tests/sagemaker/scripts/{self.framework}"
+
+    @property
+    def image_uri(self) -> str:
+        if self.framework == "pytorch":
+            return "763104351884.dkr.ecr.us-east-1.amazonaws.com/huggingface-pytorch-training:1.6.0-transformers4.4.2-gpu-py36-cu110-ubuntu18.04"
+        else:
+            return "763104351884.dkr.ecr.us-east-1.amazonaws.com/huggingface-tensorflow-training:2.4.1-transformers4.4.2-gpu-py37-cu110-ubuntu18.04"
+
+
+@pytest.fixture(scope="class")
+def sm_env(request):
+    request.cls.env = SageMakerTestEnvironment(framework=request.cls.framework)

tests/sagemaker/scripts/pytorch/requirements.txt

+git+https://github.com/huggingface/transformers.git@master # install master or adjust ist with vX.X.X for installing version specific transforms
\ No newline at end of file

tests/sagemaker/scripts/pytorch/run_ddp.py

+import json
+import logging
+import os
+import subprocess
+from argparse import ArgumentParser
+
+
+logger = logging.getLogger(__name__)
+
+
+def parse_args():
+    parser = ArgumentParser()
+    parsed, unknown = parser.parse_known_args()
+    for arg in unknown:
+        if arg.startswith(("-", "--")):
+            parser.add_argument(arg.split("=")[0])
+
+    return parser.parse_args()
+
+
+def main():
+    args = parse_args()
+    port = 8888
+    num_gpus = int(os.environ["SM_NUM_GPUS"])
+    hosts = json.loads(os.environ["SM_HOSTS"])
+    num_nodes = len(hosts)
+    current_host = os.environ["SM_CURRENT_HOST"]
+    rank = hosts.index(current_host)
+    os.environ["NCCL_DEBUG"] = "INFO"
+
+    if num_nodes > 1:
+        cmd = f"""python -m torch.distributed.launch \
+                --nnodes={num_nodes}  \
+                --node_rank={rank}  \
+                --nproc_per_node={num_gpus}  \
+                --master_addr={hosts[0]}  \
+                --master_port={port} \
+                ./run_glue.py \
+                {"".join([f" --{parameter} {value}" for parameter,value in args.__dict__.items()])}"""
+    else:
+        cmd = f"""python -m torch.distributed.launch \
+            --nproc_per_node={num_gpus}  \
+            ./run_glue.py \
+            {"".join([f" --{parameter} {value}" for parameter,value in args.__dict__.items()])}"""
+    try:
+        subprocess.run(cmd, shell=True)
+    except Exception as e:
+        logger.info(e)
+
+
+if __name__ == "__main__":
+    main()

tests/sagemaker/scripts/tensorflow/requirements.txt

+git+https://github.com/huggingface/transformers.git@master # install master or adjust ist with vX.X.X for installing version specific transforms
\ No newline at end of file

tests/sagemaker/scripts/tensorflow/run_tf.py

+import argparse
+import logging
+import sys
+import time
+
+import tensorflow as tf
+from datasets import load_dataset
+
+from transformers import AutoTokenizer, TFAutoModelForSequenceClassification
+
+
+if __name__ == "__main__":
+
+    parser = argparse.ArgumentParser()
+
+    # Hyperparameters sent by the client are passed as command-line arguments to the script.
+    parser.add_argument("--epochs", type=int, default=1)
+    parser.add_argument("--per_device_train_batch_size", type=int, default=16)
+    parser.add_argument("--per_device_eval_batch_size", type=int, default=8)
+    parser.add_argument("--model_name_or_path", type=str)
+    parser.add_argument("--learning_rate", type=str, default=5e-5)
+    parser.add_argument("--do_train", type=bool, default=True)
+    parser.add_argument("--do_eval", type=bool, default=True)
+    parser.add_argument("--output_dir", type=str)
+
+    args, _ = parser.parse_known_args()
+
+    # overwrite batch size until we have tf_glue.py
+    args.per_device_train_batch_size = 16
+    args.per_device_eval_batch_size = 16
+
+    # Set up logging
+    logger = logging.getLogger(__name__)
+
+    logging.basicConfig(
+        level=logging.getLevelName("INFO"),
+        handlers=[logging.StreamHandler(sys.stdout)],
+        format="%(asctime)s - %(name)s - %(levelname)s - %(message)s",
+    )
+
+    # Load model and tokenizer
+    model = TFAutoModelForSequenceClassification.from_pretrained(args.model_name_or_path)
+    tokenizer = AutoTokenizer.from_pretrained(args.model_name_or_path)
+
+    # Load dataset
+    train_dataset, test_dataset = load_dataset("imdb", split=["train", "test"])
+    train_dataset = train_dataset.shuffle().select(range(5000))  # smaller the size for train dataset to 5k
+    test_dataset = test_dataset.shuffle().select(range(500))  # smaller the size for test dataset to 500
+
+    # Preprocess train dataset
+    train_dataset = train_dataset.map(
+        lambda e: tokenizer(e["text"], truncation=True, padding="max_length"), batched=True
+    )
+    train_dataset.set_format(type="tensorflow", columns=["input_ids", "attention_mask", "label"])
+
+    train_features = {
+        x: train_dataset[x].to_tensor(default_value=0, shape=[None, tokenizer.model_max_length])
+        for x in ["input_ids", "attention_mask"]
+    }
+    tf_train_dataset = tf.data.Dataset.from_tensor_slices((train_features, train_dataset["label"])).batch(
+        args.per_device_train_batch_size
+    )
+
+    # Preprocess test dataset
+    test_dataset = test_dataset.map(
+        lambda e: tokenizer(e["text"], truncation=True, padding="max_length"), batched=True
+    )
+    test_dataset.set_format(type="tensorflow", columns=["input_ids", "attention_mask", "label"])
+
+    test_features = {
+        x: test_dataset[x].to_tensor(default_value=0, shape=[None, tokenizer.model_max_length])
+        for x in ["input_ids", "attention_mask"]
+    }
+    tf_test_dataset = tf.data.Dataset.from_tensor_slices((test_features, test_dataset["label"])).batch(
+        args.per_device_eval_batch_size
+    )
+
+    # fine optimizer and loss
+    optimizer = tf.keras.optimizers.Adam(learning_rate=args.learning_rate)
+    loss = tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True)
+    metrics = [tf.keras.metrics.SparseCategoricalAccuracy()]
+    model.compile(optimizer=optimizer, loss=loss, metrics=metrics)
+
+    start_train_time = time.time()
+    train_results = model.fit(tf_train_dataset, epochs=args.epochs, batch_size=args.per_device_train_batch_size)
+    end_train_time = time.time() - start_train_time
+
+    logger.info("*** Train ***")
+    logger.info("train_runtime = %s", end_train_time)
+    for key, value in train_results.history.items():
+        logger.info("  %s = %s", key, value)

tests/sagemaker/scripts/tensorflow/run_tf_dist.py

+import argparse
+import logging
+import os
+import sys
+import time
+
+import tensorflow as tf
+from datasets import load_dataset
+from tqdm import tqdm
+
+from transformers import AutoTokenizer, TFAutoModelForSequenceClassification
+from transformers.file_utils import is_sagemaker_distributed_available
+
+
+if os.environ.get("SDP_ENABLED") or is_sagemaker_distributed_available():
+    SDP_ENABLED = True
+    os.environ["SAGEMAKER_INSTANCE_TYPE"] = "p3dn.24xlarge"
+    import smdistributed.dataparallel.tensorflow as sdp
+else:
+    SDP_ENABLED = False
+
+
+def fit(model, loss, opt, train_dataset, epochs, train_batch_size, max_steps=None):
+    pbar = tqdm(train_dataset)
+    for i, batch in enumerate(pbar):
+        with tf.GradientTape() as tape:
+            inputs, targets = batch
+            outputs = model(batch)
+            loss_value = loss(targets, outputs.logits)
+
+        if SDP_ENABLED:
+            tape = sdp.DistributedGradientTape(tape, sparse_as_dense=True)
+
+        grads = tape.gradient(loss_value, model.trainable_variables)
+        opt.apply_gradients(zip(grads, model.trainable_variables))
+
+        pbar.set_description(f"Loss: {loss_value:.4f}")
+
+        if SDP_ENABLED and i == 0:
+            sdp.broadcast_variables(model.variables, root_rank=0)
+            sdp.broadcast_variables(opt.variables(), root_rank=0)
+
+        if max_steps and i >= max_steps:
+            break
+
+    train_results = {"loss": loss_value.numpy()}
+    return train_results
+
+
+def get_datasets(tokenizer, train_batch_size, eval_batch_size):
+    # Load dataset
+    train_dataset, test_dataset = load_dataset("imdb", split=["train", "test"])
+
+    # Preprocess train dataset
+    train_dataset = train_dataset.map(
+        lambda e: tokenizer(e["text"], truncation=True, padding="max_length"), batched=True
+    )
+    train_dataset.set_format(type="tensorflow", columns=["input_ids", "attention_mask", "label"])
+
+    train_features = {
+        x: train_dataset[x].to_tensor(default_value=0, shape=[None, tokenizer.model_max_length])
+        for x in ["input_ids", "attention_mask"]
+    }
+    tf_train_dataset = tf.data.Dataset.from_tensor_slices((train_features, train_dataset["label"]))
+
+    # Preprocess test dataset
+    test_dataset = test_dataset.map(
+        lambda e: tokenizer(e["text"], truncation=True, padding="max_length"), batched=True
+    )
+    test_dataset.set_format(type="tensorflow", columns=["input_ids", "attention_mask", "label"])
+
+    test_features = {
+        x: test_dataset[x].to_tensor(default_value=0, shape=[None, tokenizer.model_max_length])
+        for x in ["input_ids", "attention_mask"]
+    }
+    tf_test_dataset = tf.data.Dataset.from_tensor_slices((test_features, test_dataset["label"]))
+
+    if SDP_ENABLED:
+        tf_train_dataset = tf_train_dataset.shard(sdp.size(), sdp.rank())
+        tf_test_dataset = tf_test_dataset.shard(sdp.size(), sdp.rank())
+    tf_train_dataset = tf_train_dataset.batch(train_batch_size, drop_remainder=True)
+    tf_test_dataset = tf_test_dataset.batch(eval_batch_size, drop_remainder=True)
+
+    return tf_train_dataset, tf_test_dataset
+
+
+if __name__ == "__main__":
+
+    parser = argparse.ArgumentParser()
+
+    # Hyperparameters sent by the client are passed as command-line arguments to the script.
+    parser.add_argument("--epochs", type=int, default=3)
+    parser.add_argument("--per_device_train_batch_size", type=int, default=16)
+    parser.add_argument("--per_device_eval_batch_size", type=int, default=8)
+    parser.add_argument("--model_name_or_path", type=str)
+    parser.add_argument("--learning_rate", type=str, default=5e-5)
+    parser.add_argument("--do_train", type=bool, default=True)
+    parser.add_argument("--do_eval", type=bool, default=True)
+    parser.add_argument("--output_dir", type=str)
+    parser.add_argument("--max_steps", type=int, default=None)
+
+    # Data, model, and output directories
+    parser.add_argument("--output_data_dir", type=str, default=os.environ["SM_OUTPUT_DATA_DIR"])
+    parser.add_argument("--model_dir", type=str, default=os.environ["SM_MODEL_DIR"])
+    parser.add_argument("--n_gpus", type=str, default=os.environ["SM_NUM_GPUS"])
+
+    args, _ = parser.parse_known_args()
+
+    # Set up logging
+    logger = logging.getLogger(__name__)
+
+    logging.basicConfig(
+        level=logging.getLevelName("INFO"),
+        handlers=[logging.StreamHandler(sys.stdout)],
+        format="%(asctime)s - %(name)s - %(levelname)s - %(message)s",
+    )
+
+    if SDP_ENABLED:
+        sdp.init()
+
+        gpus = tf.config.experimental.list_physical_devices("GPU")
+        for gpu in gpus:
+            tf.config.experimental.set_memory_growth(gpu, True)
+        if gpus:
+            tf.config.experimental.set_visible_devices(gpus[sdp.local_rank()], "GPU")
+
+    # Load model and tokenizer
+    model = TFAutoModelForSequenceClassification.from_pretrained(args.model_name_or_path)
+    tokenizer = AutoTokenizer.from_pretrained(args.model_name_or_path)
+
+    # get datasets
+    tf_train_dataset, tf_test_dataset = get_datasets(
+        tokenizer=tokenizer,
+        train_batch_size=args.per_device_train_batch_size,
+        eval_batch_size=args.per_device_eval_batch_size,
+    )
+
+    # fine optimizer and loss
+    optimizer = tf.keras.optimizers.Adam(learning_rate=args.learning_rate)
+    loss = tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True)
+    metrics = [tf.keras.metrics.SparseCategoricalAccuracy()]
+    model.compile(optimizer=optimizer, loss=loss, metrics=metrics)
+
+    # Training
+    if args.do_train:
+
+        # train_results = model.fit(tf_train_dataset, epochs=args.epochs, batch_size=args.train_batch_size)
+        start_train_time = time.time()
+        train_results = fit(
+            model,
+            loss,
+            optimizer,
+            tf_train_dataset,
+            args.epochs,
+            args.per_device_train_batch_size,
+            max_steps=args.max_steps,
+        )
+        end_train_time = time.time() - start_train_time
+        logger.info("*** Train ***")
+        logger.info("train_runtime = %s", end_train_time)
+
+        output_eval_file = os.path.join(args.output_dir, "train_results.txt")
+
+        if not SDP_ENABLED or sdp.rank() == 0:
+            with open(output_eval_file, "w") as writer:
+                logger.info("***** Train results *****")
+                logger.info(train_results)
+                for key, value in train_results.items():
+                    logger.info("  %s = %s", key, value)
+                    writer.write("%s = %s\n" % (key, value))
+
+    # Evaluation
+    if args.do_eval and (not SDP_ENABLED or sdp.rank() == 0):
+
+        result = model.evaluate(tf_test_dataset, batch_size=args.per_device_eval_batch_size, return_dict=True)
+        logger.info("*** Evaluate ***")
+
+        output_eval_file = os.path.join(args.output_dir, "eval_results.txt")
+
+        with open(output_eval_file, "w") as writer:
+            logger.info("***** Eval results *****")
+            logger.info(result)
+            for key, value in result.items():
+                logger.info("  %s = %s", key, value)
+                writer.write("%s = %s\n" % (key, value))
+
+    # Save result
+    if SDP_ENABLED:
+        if sdp.rank() == 0:
+            model.save_pretrained(args.output_dir)
+            tokenizer.save_pretrained(args.output_dir)
+    else:
+        model.save_pretrained(args.output_dir)
+        tokenizer.save_pretrained(args.output_dir)

tests/sagemaker/test_multi_node_data_parallel.py

+import os
+import subprocess
+import unittest
+from ast import literal_eval
+
+import pytest
+
+from parameterized import parameterized, parameterized_class
+
+from . import is_sagemaker_available
+
+
+if is_sagemaker_available():
+    from sagemaker import TrainingJobAnalytics
+    from sagemaker.huggingface import HuggingFace
+
+
+@pytest.mark.skipif(
+    literal_eval(os.getenv("TEST_SAGEMAKER", "False")) is not True,
+    reason="Skipping test because should only be run when releasing minor transformers version",
+)
+@pytest.mark.usefixtures("sm_env")
+@parameterized_class(
+    [
+        {
+            "framework": "pytorch",
+            "script": "run_glue.py",
+            "model_name_or_path": "distilbert-base-cased",
+            "instance_type": "ml.p3dn.24xlarge",
+            "results": {"train_runtime": 300, "eval_accuracy": 0.7, "eval_loss": 0.6},
+        },
+        {
+            "framework": "pytorch",
+            "script": "run_ddp.py",
+            "model_name_or_path": "distilbert-base-cased",
+            "instance_type": "ml.p3dn.24xlarge",
+            "results": {"train_runtime": 300, "eval_accuracy": 0.7, "eval_loss": 0.6},
+        },
+        {
+            "framework": "tensorflow",
+            "script": "run_tf_dist.py",
+            "model_name_or_path": "distilbert-base-cased",
+            "instance_type": "ml.p3dn.24xlarge",
+            "results": {"train_runtime": 500, "eval_accuracy": 0.6, "eval_loss": 0.7},
+        },
+    ]
+)
+class MultiNodeTest(unittest.TestCase):
+    def setUp(self):
+        if self.framework == "pytorch":
+            subprocess.run(
+                f"cp ./examples/text-classification/run_glue.py {self.env.test_path}/run_glue.py".split(),
+                encoding="utf-8",
+                check=True,
+            )
+        assert hasattr(self, "env")
+
+    def create_estimator(self, instance_count):
+        job_name = f"{self.env.base_job_name}-{instance_count}-{'ddp' if 'ddp' in self.script else 'smd'}"
+        # distributed data settings
+        distribution = {"smdistributed": {"dataparallel": {"enabled": True}}}
+
+        # creates estimator
+        return HuggingFace(
+            entry_point=self.script,
+            source_dir=self.env.test_path,
+            role=self.env.role,
+            image_uri=self.env.image_uri,
+            base_job_name=job_name,
+            instance_count=instance_count,
+            instance_type=self.instance_type,
+            debugger_hook_config=False,
+            hyperparameters={**self.env.distributed_hyperparameters, "model_name_or_path": self.model_name_or_path},
+            metric_definitions=self.env.metric_definitions,
+            distribution=distribution,
+            py_version="py36",
+        )
+
+    def save_results_as_csv(self, job_name):
+        TrainingJobAnalytics(job_name).export_csv(f"{self.env.test_path}/{job_name}_metrics.csv")
+
+    # @parameterized.expand([(2,), (4,),])
+    @parameterized.expand([(2,)])
+    def test_script(self, instance_count):
+        # create estimator
+        estimator = self.create_estimator(instance_count)
+
+        # run training
+        estimator.fit()
+
+        # save csv
+        self.save_results_as_csv(estimator.latest_training_job.name)
+        # result dataframe
+        result_metrics_df = TrainingJobAnalytics(estimator.latest_training_job.name).dataframe()
+
+        # extract kpis
+        train_runtime = list(result_metrics_df[result_metrics_df.metric_name == "train_runtime"]["value"])
+        eval_accuracy = list(result_metrics_df[result_metrics_df.metric_name == "eval_accuracy"]["value"])
+        eval_loss = list(result_metrics_df[result_metrics_df.metric_name == "eval_loss"]["value"])
+
+        # assert kpis
+        assert all(t <= self.results["train_runtime"] for t in train_runtime)
+        assert any(t >= self.results["eval_accuracy"] for t in eval_accuracy)
+        assert all(t <= self.results["eval_loss"] for t in eval_loss)

tests/sagemaker/test_multi_node_model_parallel.py

+import os
+import unittest
+from ast import literal_eval
+
+import pytest
+
+from parameterized import parameterized, parameterized_class
+
+from . import is_sagemaker_available
+
+
+if is_sagemaker_available():
+    from sagemaker import TrainingJobAnalytics
+    from sagemaker.huggingface import HuggingFace
+
+
+@pytest.mark.skipif(
+    literal_eval(os.getenv("TEST_SAGEMAKER", "False")) is not True,
+    reason="Skipping test because should only be run when releasing minor transformers version",
+)
+@pytest.mark.usefixtures("sm_env")
+@parameterized_class(
+    [
+        {
+            "framework": "pytorch",
+            "script": "run_glue_model_parallelism.py",
+            "model_name_or_path": "roberta-large",
+            "instance_type": "ml.p3dn.24xlarge",
+            "results": {"train_runtime": 700, "eval_accuracy": 0.3, "eval_loss": 1.2},
+        },
+    ]
+)
+class MultiNodeTest(unittest.TestCase):
+    def setUp(self):
+        assert hasattr(self, "env")
+
+    def create_estimator(self, instance_count):
+
+        # configuration for running training on smdistributed Model Parallel
+        mpi_options = {
+            "enabled": True,
+            "processes_per_host": 8,
+        }
+        smp_options = {
+            "enabled": True,
+            "parameters": {
+                "microbatches": 4,
+                "placement_strategy": "spread",
+                "pipeline": "interleaved",
+                "optimize": "speed",
+                "partitions": 4,
+                "ddp": True,
+            },
+        }
+
+        distribution = {"smdistributed": {"modelparallel": smp_options}, "mpi": mpi_options}
+
+        # creates estimator
+        return HuggingFace(
+            entry_point=self.script,
+            source_dir=self.env.test_path,
+            role=self.env.role,
+            image_uri=self.env.image_uri,
+            base_job_name=f"{self.env.base_job_name}-{instance_count}-smp",
+            instance_count=instance_count,
+            instance_type=self.instance_type,
+            debugger_hook_config=False,
+            hyperparameters={
+                **self.env.hyperparameters,
+                "model_name_or_path": self.model_name_or_path,
+                "max_steps": 500,
+            },
+            metric_definitions=self.env.metric_definitions,
+            distribution=distribution,
+            py_version="py36",
+        )
+
+    def save_results_as_csv(self, job_name):
+        TrainingJobAnalytics(job_name).export_csv(f"{self.env.test_path}/{job_name}_metrics.csv")
+
+    # @parameterized.expand([(2,), (4,),])
+    @parameterized.expand([(1,)])
+    def test_scripz(self, instance_count):
+        # create estimator
+        estimator = self.create_estimator(instance_count)
+
+        # run training
+        estimator.fit()
+
+        # save csv
+        self.save_results_as_csv(estimator.latest_training_job.name)
+        # result dataframe
+        result_metrics_df = TrainingJobAnalytics(estimator.latest_training_job.name).dataframe()
+
+        # extract kpis
+        train_runtime = list(result_metrics_df[result_metrics_df.metric_name == "train_runtime"]["value"])
+        eval_accuracy = list(result_metrics_df[result_metrics_df.metric_name == "eval_accuracy"]["value"])
+        eval_loss = list(result_metrics_df[result_metrics_df.metric_name == "eval_loss"]["value"])
+
+        # assert kpis
+        assert all(t <= self.results["train_runtime"] for t in train_runtime)
+        assert all(t >= self.results["eval_accuracy"] for t in eval_accuracy)
+        assert all(t <= self.results["eval_loss"] for t in eval_loss)

tests/sagemaker/test_single_node_gpu.py

+import os
+import subprocess
+import unittest
+from ast import literal_eval
+
+import pytest
+
+from parameterized import parameterized_class
+
+from . import is_sagemaker_available
+
+
+if is_sagemaker_available():
+    from sagemaker import TrainingJobAnalytics
+    from sagemaker.huggingface import HuggingFace
+
+
+@pytest.mark.skipif(
+    literal_eval(os.getenv("TEST_SAGEMAKER", "False")) is not True,
+    reason="Skipping test because should only be run when releasing minor transformers version",
+)
+@pytest.mark.usefixtures("sm_env")
+@parameterized_class(
+    [
+        {
+            "framework": "pytorch",
+            "script": "run_glue.py",
+            "model_name_or_path": "distilbert-base-cased",
+            "instance_type": "ml.g4dn.xlarge",
+            "results": {"train_runtime": 200, "eval_accuracy": 0.6, "eval_loss": 0.9},
+        },
+        {
+            "framework": "tensorflow",
+            "script": "run_tf.py",
+            "model_name_or_path": "distilbert-base-cased",
+            "instance_type": "ml.g4dn.xlarge",
+            "results": {"train_runtime": 350, "eval_accuracy": 0.3, "eval_loss": 0.9},
+        },
+    ]
+)
+class SingleNodeTest(unittest.TestCase):
+    def setUp(self):
+        if self.framework == "pytorch":
+            subprocess.run(
+                f"cp ./examples/text-classification/run_glue.py {self.env.test_path}/run_glue.py".split(),
+                encoding="utf-8",
+                check=True,
+            )
+        assert hasattr(self, "env")
+
+    def create_estimator(self, instance_count=1):
+        # creates estimator
+        return HuggingFace(
+            entry_point=self.script,
+            source_dir=self.env.test_path,
+            role=self.env.role,
+            image_uri=self.env.image_uri,
+            base_job_name=f"{self.env.base_job_name}-single",
+            instance_count=instance_count,
+            instance_type=self.instance_type,
+            debugger_hook_config=False,
+            hyperparameters={**self.env.hyperparameters, "model_name_or_path": self.model_name_or_path},
+            metric_definitions=self.env.metric_definitions,
+            py_version="py36",
+        )
+
+    def save_results_as_csv(self, job_name):
+        TrainingJobAnalytics(job_name).export_csv(f"{self.env.test_path}/{job_name}_metrics.csv")
+
+    def test_glue(self):
+        # create estimator
+        estimator = self.create_estimator()
+
+        # run training
+        estimator.fit()
+
+        # save csv
+        self.save_results_as_csv(estimator.latest_training_job.name)
+        # result dataframe
+        result_metrics_df = TrainingJobAnalytics(estimator.latest_training_job.name).dataframe()
+
+        # extract kpis
+        train_runtime = list(result_metrics_df[result_metrics_df.metric_name == "train_runtime"]["value"])
+        eval_accuracy = list(result_metrics_df[result_metrics_df.metric_name == "eval_accuracy"]["value"])
+        eval_loss = list(result_metrics_df[result_metrics_df.metric_name == "eval_loss"]["value"])
+
+        # assert kpis
+        assert all(t <= self.results["train_runtime"] for t in train_runtime)
+        assert all(t >= self.results["eval_accuracy"] for t in eval_accuracy)
+        assert all(t <= self.results["eval_loss"] for t in eval_loss)