[CI] Automatic rerun of certain test failures (#36694)

.circleci/create_circleci_config.py

@@ -33,6 +33,26 @@ COMMON_ENV_VARIABLES = {
 COMMON_PYTEST_OPTIONS = {"max-worker-restart": 0, "dist": "loadfile", "vvv": None, "rsfE":None}
 DEFAULT_DOCKER_IMAGE = [{"image": "cimg/python:3.8.12"}]
 
+# Strings that commonly appear in the output of flaky tests when they fail. These are used with `pytest-rerunfailures`
+# to rerun the tests that match these patterns.
+FLAKY_TEST_FAILURE_PATTERNS = [
+    "OSError",  # Machine/connection transient error
+    "Timeout",  # Machine/connection transient error
+    "ConnectionError",  # Connection transient error
+    "FileNotFoundError",  # Raised by `datasets` on Hub failures
+    "PIL.UnidentifiedImageError",  # Raised by `PIL.Image.open` on connection issues
+    "HTTPError.*502",  # Hub-related
+    "HTTPError.*504",  # Hub-related
+    "AssertionError: Tensor-likes are not close!",  # `torch.testing.assert_close`, we might have unlucky random values
+    # TODO: error downloading tokenizer's `merged.txt` from hub can cause all the exceptions below. Throw and handle
+    # them under a single message.
+    "TypeError: expected str, bytes or os.PathLike object, not NoneType",
+    "TypeError: stat: path should be string, bytes, os.PathLike or integer, not NoneType",
+    "Converting from Tiktoken failed",
+    "KeyError: <class ",
+    "TypeError: not a string",
+]
+
 
 class EmptyJob:
     job_name = "empty"
@@ -124,7 +144,9 @@ class CircleCIJob:
                 # Examples special case: we need to download NLTK files in advance to avoid cuncurrency issues
         timeout_cmd = f"timeout {self.command_timeout} " if self.command_timeout else ""
         marker_cmd = f"-m '{self.marker}'" if self.marker is not None else ""
-        additional_flags = f" -p no:warning -o junit_family=xunit1 --junitxml=test-results/junit.xml"
+        junit_flags = f" -p no:warning -o junit_family=xunit1 --junitxml=test-results/junit.xml"
+        joined_flaky_patterns = "|".join(FLAKY_TEST_FAILURE_PATTERNS)
+        repeat_on_failure_flags = f"--reruns 5 --reruns-delay 2 --only-rerun '({joined_flaky_patterns})'"
         parallel = f' << pipeline.parameters.{self.job_name}_parallelism >> '
         steps = [
             "checkout",
@@ -152,7 +174,7 @@ class CircleCIJob:
             },
             {"run": {
                 "name": "Run tests",
-                "command": f"({timeout_cmd} python3 -m pytest {marker_cmd} -n {self.pytest_num_workers} {additional_flags} {' '.join(pytest_flags)} $(cat splitted_tests.txt) | tee tests_output.txt)"}
+                "command": f"({timeout_cmd} python3 -m pytest {marker_cmd} -n {self.pytest_num_workers} {junit_flags} {repeat_on_failure_flags} {' '.join(pytest_flags)} $(cat splitted_tests.txt) | tee tests_output.txt)"}
             },
             {"run": {"name": "Expand to show skipped tests", "when": "always", "command": f"python3 .circleci/parse_test_outputs.py --file tests_output.txt --skip"}},
             {"run": {"name": "Failed tests: show reasons",   "when": "always", "command": f"python3 .circleci/parse_test_outputs.py --file tests_output.txt --fail"}},

docker/custom-tokenizers.dockerfile

@@ -1,5 +1,6 @@
 FROM python:3.9-slim
 ENV PYTHONDONTWRITEBYTECODE=1
+ARG REF=main
 USER root
 RUN apt-get update && apt-get install -y libsndfile1-dev espeak-ng time git cmake wget xz-utils build-essential g++5 libprotobuf-dev protobuf-compiler
 ENV UV_PYTHON=/usr/local/bin/python
@@ -16,11 +17,11 @@ RUN make install -j 10
 
 
 RUN uv pip install --no-cache --upgrade 'torch' --index-url https://download.pytorch.org/whl/cpu
-RUN uv pip install --no-cache-dir  --no-deps accelerate --extra-index-url https://download.pytorch.org/whl/cpu 
-RUN uv pip install  --no-cache-dir "transformers[ja,testing,sentencepiece,jieba,spacy,ftfy,rjieba]" unidic unidic-lite
+RUN uv pip install --no-cache-dir  --no-deps accelerate --extra-index-url https://download.pytorch.org/whl/cpu
+RUN uv pip install  --no-cache-dir "git+https://github.com/huggingface/transformers.git@${REF}#egg=transformers[ja,testing,sentencepiece,jieba,spacy,ftfy,rjieba]" unidic unidic-lite
 # spacy is not used so not tested. Causes to failures. TODO fix later
 RUN python3 -m unidic download
 RUN pip uninstall -y transformers
 
 RUN apt-get clean && rm -rf /var/lib/apt/lists/*
-RUN apt remove -y g++ cmake  xz-utils libprotobuf-dev protobuf-compiler
+RUN apt remove -y g++ cmake  xz-utils libprotobuf-dev protobuf-compiler

docker/examples-tf.dockerfile

@@ -1,12 +1,13 @@
 FROM python:3.9-slim
 ENV PYTHONDONTWRITEBYTECODE=1
+ARG REF=main
 USER root
 RUN apt-get update && apt-get install -y libsndfile1-dev espeak-ng time git
 RUN apt-get install -y g++ cmake
 ENV UV_PYTHON=/usr/local/bin/python
 RUN pip --no-cache-dir install uv && uv venv
 RUN uv pip install --no-cache-dir -U pip setuptools albumentations seqeval
-RUN pip install  --upgrade --no-cache-dir "transformers[tf-cpu,sklearn,testing,sentencepiece,tf-speech,vision]"
-RUN uv pip install --no-cache-dir  "protobuf==3.20.3" 
+RUN pip install  --upgrade --no-cache-dir "git+https://github.com/huggingface/transformers.git@${REF}#egg=transformers[tf-cpu,sklearn,testing,sentencepiece,tf-speech,vision]"
+RUN uv pip install --no-cache-dir  "protobuf==3.20.3"
 RUN pip uninstall -y transformers
-RUN apt-get clean && rm -rf /var/lib/apt/lists/*
+RUN apt-get clean && rm -rf /var/lib/apt/lists/*

docker/examples-torch.dockerfile

@@ -1,11 +1,12 @@
 FROM python:3.9-slim
 ENV PYTHONDONTWRITEBYTECODE=1
+ARG REF=main
 USER root
 RUN apt-get update &&  apt-get install -y --no-install-recommends libsndfile1-dev espeak-ng time git g++ cmake pkg-config openssh-client git
 ENV UV_PYTHON=/usr/local/bin/python
 RUN pip --no-cache-dir install uv && uv venv && uv pip install --no-cache-dir -U pip setuptools
 RUN pip install --no-cache-dir 'torch' 'torchvision' 'torchaudio' --index-url https://download.pytorch.org/whl/cpu
-RUN uv pip install --no-deps timm accelerate --extra-index-url https://download.pytorch.org/whl/cpu 
-RUN uv pip install --no-cache-dir librosa "transformers[sklearn,sentencepiece,vision,testing]" seqeval albumentations jiwer
+RUN uv pip install --no-deps timm accelerate --extra-index-url https://download.pytorch.org/whl/cpu
+RUN uv pip install --no-cache-dir librosa "git+https://github.com/huggingface/transformers.git@${REF}#egg=transformers[sklearn,sentencepiece,vision,testing]" seqeval albumentations jiwer
 RUN pip uninstall -y transformers
-RUN apt-get clean && rm -rf /var/lib/apt/lists/*
+RUN apt-get clean && rm -rf /var/lib/apt/lists/*

setup.py

@@ -150,6 +150,7 @@ _deps = [
     "pydantic",
     "pytest>=7.2.0,<8.0.0",
     "pytest-asyncio",
+    "pytest-rerunfailures",
     "pytest-timeout",
     "pytest-xdist",
     "pytest-order",
@@ -326,6 +327,7 @@ extras["testing"] = (
         "pytest-rich",
         "pytest-xdist",
         "pytest-order",
+        "pytest-rerunfailures",
         "timeout-decorator",
         "parameterized",
         "psutil",

src/transformers/dependency_versions_table.py

@@ -56,6 +56,7 @@ deps = {
     "pydantic": "pydantic",
     "pytest": "pytest>=7.2.0,<8.0.0",
     "pytest-asyncio": "pytest-asyncio",
+    "pytest-rerunfailures": "pytest-rerunfailures",
     "pytest-timeout": "pytest-timeout",
     "pytest-xdist": "pytest-xdist",
     "pytest-order": "pytest-order",

src/transformers/testing_utils.py

@@ -2518,6 +2518,11 @@ def is_flaky(max_attempts: int = 5, wait_before_retry: Optional[float] = None, d
     """
     To decorate flaky tests. They will be retried on failures.
 
+    Please note that our push tests use `pytest-rerunfailures`, which prompts the CI to rerun certain types of
+    failed tests. More specifically, if the test exception contains any substring in `FLAKY_TEST_FAILURE_PATTERNS`
+    (in `.circleci/create_circleci_config.py`), it will be rerun. If you find a recurrent pattern of failures,
+    expand `FLAKY_TEST_FAILURE_PATTERNS` in our CI configuration instead of using `is_flaky`.
+
     Args:
         max_attempts (`int`, *optional*, defaults to 5):
             The maximum number of attempts to retry the flaky test.