Setup pre-commit github action and apply pre-commit to all files (#849)

* adding pre-commit files

* applying pre-commit to all files

* adding no-strict-optional argument to mypy in circle ci config

* fix typo

* updating python versions

* [skip ci] remove extra args

* adding python 3.9

* [skip ci] set pre-commit version in requirements-dev.txt

* set CACHE_VERSION

* move linters from circleci to github actions

* update python version

* update python version in benchmarks_2

* moving to python 3.9.7

.circleci/config.yml

@@ -6,6 +6,14 @@
 # https://github.com/facebookresearch/detectron2/blob/main/.circleci/config.yml
 #
 # Pro tip: download circle ci cli to validate the config locally during development.
+#
+# To reset/clean the cache update the CACHE_VERSION variable in project settings
+# in the fairscale project in CircleCI. The CACHE_VERSION follows the convention
+# v$(FAIRSCALE_VERSION)-${CACHE_NUMBER}. E.g. v0.4.2-1. CACHE_NUMBER must start
+# at 1 and increase in whole numbers. When changing the CACHE_VERSION manually
+# always set the FAIRSCALE_VERSION value to the fairscale version being tested.
+# To reset the cache when not updating the fairscale version, only update the
+# CACHE_NUMBER value.
 
 version: 2.1
 orbs:
@@ -15,23 +23,26 @@ orbs:
 # -------------------------------------------------------------------------------------
 cpu_py37: &cpu_py37
   docker:
+    # python version 3.7.12
     - image: circleci/python:3.7
   resource_class: large
 
 cpu_py38: &cpu_py38
   docker:
+    # python version 3.8.12
     - image: circleci/python:3.8
   resource_class: large
 
 cpu_py39: &cpu_py39
   docker:
+    # python version 3.9.7
     - image: circleci/python:3.9
   resource_class: large
 
-# Here are list of GPU images:
+# Here is the list of GPU images:
 # https://circleci.com/docs/2.0/configuration-reference/#available-linux-gpu-images
-# We need to use multiple gpus for several jobs. the resource_class values are
-# available here T101565170
+# We need to use multiple gpus for several jobs. The resource_class
+# values are available here T101565170
 # gpu.nvidia.small.multi = 2 gpus with 16 GB ram each
 # gpu.nvidia.medium.multi = 4 gpus with 16 GB ram each
 
@@ -122,30 +133,6 @@ install_repo: &install_repo
         # Test import.
         python -c 'import sys; sys.path = sys.path[1:]; import fairscale'
 
-run_isort: &run_isort
-   - run:
-       name: Run Linter (isort)
-       command: |
-         isort . --check
-
-run_black: &run_black
-   - run:
-       name: Run Linter (black)
-       command: |
-         black --check .
-
-run_mypy: &run_mypy
-   - run:
-       name: Run type-checking (mypy)
-       command: |
-         mypy --ignore-missing-imports --scripts-are-modules --pretty .
-
-run_flake8: &run_flake8
-  - run:
-      name: Run Linter (flake8)
-      command: |
-        flake8 --show-source --statistics
-
 check_test_list: &check_test_list
   - run:
       name: Verify that unit test list files are correct
@@ -260,21 +247,16 @@ jobs:
       # Cache the venv directory that contains dependencies
       - restore_cache:
           keys:
-            - cache-key-cpu-py37-torch-1-10-0-{{ checksum "setup.py"}}-{{ checksum "requirements-test.txt"}}
+            - cache-key-cpu-py37-torch-1-10-0-{{.Environment.CACHE_VERSION }}-{{checksum "setup.py"}}-{{checksum "requirements-test.txt"}}
 
       - <<: *install_dep_1_10_0
 
       - save_cache:
           paths:
             - ~/venv
-          key: cache-key-cpu-py37-torch-1-10-0-{{ checksum "setup.py"}}-{{ checksum "requirements-test.txt"}}
+          key: cache-key-cpu-py37-torch-1-10-0-{{.Environment.CACHE_VERSION }}-{{checksum "setup.py"}}-{{checksum "requirements-test.txt"}}
 
       - <<: *install_repo
-
-      - <<: *run_isort
-      - <<: *run_black
-      - <<: *run_mypy
-      - <<: *run_flake8
       - <<: *run_unittests
       - <<: *run_doc_build
 
@@ -294,20 +276,15 @@ jobs:
       # Cache the venv directory that contains dependencies
       - restore_cache:
           keys:
-            - cache-key-cpu-py38-torch-1-10-0-{{ checksum "setup.py"}}-{{ checksum "requirements-test.txt"}}
+            - cache-key-cpu-py38-torch-1-10-0-{{.Environment.CACHE_VERSION }}-{{checksum "setup.py"}}-{{checksum "requirements-test.txt"}}
       - <<: *install_dep_1_10_0
 
       - save_cache:
           paths:
             - ~/venv
-          key: cache-key-cpu-py38-torch-1-10-0-{{ checksum "setup.py"}}-{{ checksum "requirements-test.txt"}}
+          key: cache-key-cpu-py38-torch-1-10-0-{{.Environment.CACHE_VERSION }}-{{checksum "setup.py"}}-{{checksum "requirements-test.txt"}}
 
       - <<: *install_repo
-
-      - <<: *run_isort
-      - <<: *run_black
-      - <<: *run_mypy
-      - <<: *run_flake8
       - <<: *run_unittests
       - <<: *run_doc_build
 
@@ -327,21 +304,16 @@ jobs:
       # Cache the venv directory that contains dependencies
       - restore_cache:
           keys:
-            - cache-key-cpu-py39-torch-1-10-0-{{ checksum "setup.py"}}-{{ checksum "requirements-test.txt"}}
+            - cache-key-cpu-py39-torch-1-10-0-{{.Environment.CACHE_VERSION }}-{{checksum "setup.py"}}-{{checksum "requirements-test.txt"}}
 
       - <<: *install_dep_1_10_0
 
       - save_cache:
           paths:
             - ~/venv
-          key: cache-key-cpu-py39-torch-1-10-0-{{ checksum "setup.py"}}-{{ checksum "requirements-test.txt"}}
+          key: cache-key-cpu-py39-torch-1-10-0-{{.Environment.CACHE_VERSION }}-{{checksum "setup.py"}}-{{checksum "requirements-test.txt"}}
 
       - <<: *install_repo
-
-      - <<: *run_isort
-      - <<: *run_black
-      - <<: *run_mypy
-      - <<: *run_flake8
       - <<: *run_unittests
       - <<: *run_doc_build
 
@@ -365,21 +337,21 @@ jobs:
 
       # Run this to make sure we use python3 from the system.
       - setup_pyenv:
-          version: 3.7.0
+          version: 3.9.7
 
       - <<: *setup_venv
 
       # Cache the venv directory that contains dependencies
       - restore_cache:
           keys:
-            - cache-key-py37-gpu-torch-1-8-1-cuda-11-2-{{ checksum "setup.py"}}-{{ checksum "requirements-test.txt"}}
+            - cache-key-py-3-9-7-gpu-torch-1-8-1-cuda-11-2-{{.Environment.CACHE_VERSION }}-{{checksum "setup.py"}}-{{checksum "requirements-test.txt"}}
 
       - <<: *install_dep_1_8_1
 
       - save_cache:
           paths:
             - ~/venv
-          key: cache-key-py37-gpu-torch-1-8-1-cuda-11-2-{{ checksum "setup.py"}}-{{ checksum "requirements-test.txt"}}
+          key: cache-key-py-3-9-7-gpu-torch-1-8-1-cuda-11-2-{{.Environment.CACHE_VERSION }}-{{checksum "setup.py"}}-{{checksum "requirements-test.txt"}}
 
       - <<: *install_repo
 
@@ -408,21 +380,21 @@ jobs:
 
       # Run this to make sure we use python3 from the system.
       - setup_pyenv:
-          version: 3.8.6
+          version: 3.9.7
 
       - <<: *setup_venv
 
       # Cache the venv directory that contains dependencies
       - restore_cache:
           keys:
-            - cache-key-py38-gpu-torch-1-10-0-cuda-11-2-{{ checksum "setup.py"}}-{{ checksum "requirements-test.txt"}}
+            - cache-key-py-3-9-7-gpu-torch-1-10-0-cuda-11-2-{{.Environment.CACHE_VERSION }}-{{checksum "setup.py"}}-{{checksum "requirements-test.txt"}}
 
       - <<: *install_dep_1_10_0
 
       - save_cache:
           paths:
             - ~/venv
-          key: cache-key-py38-gpu-torch-1-10-0-cuda-11-2-{{ checksum "setup.py"}}-{{ checksum "requirements-test.txt"}}
+          key: cache-key-py-3-9-7-gpu-torch-1-10-0-cuda-11-2-{{.Environment.CACHE_VERSION }}-{{checksum "setup.py"}}-{{checksum "requirements-test.txt"}}
 
       - <<: *install_repo
 
@@ -449,21 +421,21 @@ jobs:
 
       # Run this to make sure we use python3 from the system.
       - setup_pyenv:
-          version: 3.8.6
+          version: 3.9.7
 
       - <<: *setup_venv
 
       # Cache the venv directory that contains dependencies
       - restore_cache:
           keys:
-            - cache-key-py38-gpu-pytorch-nightly-cuda-11-2-{{ checksum "setup.py"}}-{{ checksum "requirements-test.txt"}}
+            - cache-key-py-3-9-7-gpu-pytorch-nightly-cuda-11-2-{{.Environment.CACHE_VERSION }}-{{checksum "setup.py"}}-{{checksum "requirements-test.txt"}}
 
       - <<: *install_dep_pytorch_nightly
 
       - save_cache:
           paths:
             - ~/venv
-          key: cache-key-py38-gpu-pytorch-nightly-cuda-11-2-{{ checksum "setup.py"}}-{{ checksum "requirements-test.txt"}}
+          key: cache-key-py-3-9-7-gpu-pytorch-nightly-cuda-11-2-{{.Environment.CACHE_VERSION }}-{{checksum "setup.py"}}-{{checksum "requirements-test.txt"}}
 
       - <<: *install_repo
 
@@ -484,26 +456,26 @@ jobs:
       - run: nvidia-smi
 
       - setup_pyenv:
-          version: 3.7.0
+          version: 3.9.7
 
       - <<: *setup_venv
 
       # Cache the venv directory that contains dependencies
       - restore_cache:
           keys:
-            - cache-key-py37-benchmarks-torch-1-10-0-cuda-11-2-{{ checksum "setup.py"}}-{{ checksum "requirements-test.txt"}}
+            - cache-key-py-3-9-7-benchmarks-torch-1-10-0-cuda-11-2-{{.Environment.CACHE_VERSION }}-{{checksum "setup.py"}}-{{checksum "requirements-test.txt"}}
 
       # Cache the MNIST directory that contains benchmark data
       - restore_cache:
           keys:
-            - cache-key-benchmark-MNIST-{{ checksum "benchmarks/datasets/mnist.py"}}
+            - cache-key-benchmark-MNIST-{{.Environment.CACHE_VERSION }}-{{checksum "benchmarks/datasets/mnist.py"}}
 
       - <<: *install_dep_1_10_0
 
       - save_cache:
           paths:
             - ~/venv
-          key: cache-key-py37-benchmarks-torch-1-10-0-cuda-11-2-{{ checksum "setup.py"}}-{{ checksum "requirements-test.txt"}}
+          key: cache-key-py-3-9-7-benchmarks-torch-1-10-0-cuda-11-2-{{.Environment.CACHE_VERSION }}-{{checksum "setup.py"}}-{{checksum "requirements-test.txt"}}
 
       - <<: *install_repo
 
@@ -520,7 +492,7 @@ jobs:
       - save_cache:
           paths:
             - /tmp/MNIST
-          key: cache-key-benchmark-MNIST-{{ checksum "benchmarks/datasets/mnist.py"}}
+          key: cache-key-benchmark-MNIST-{{.Environment.CACHE_VERSION }}-{{checksum "benchmarks/datasets/mnist.py"}}
 
   benchmarks_2:
     <<: *gpu_cu_11_2_medium_multi
@@ -533,27 +505,27 @@ jobs:
       - run: nvidia-smi
 
       - setup_pyenv:
-          version: 3.7.0
+          version: 3.9.7
 
       - <<: *setup_venv
 
       # Cache the venv directory that contains dependencies
       - restore_cache:
           keys:
-            - cache-key-py37-benchmarks-torch-1-10-0-cuda-11-2-{{ checksum "setup.py"}}-{{ checksum "requirements-test.txt"}}
+            - cache-key-py-3-9-7-benchmarks-torch-1-10-0-cuda-11-2-{{.Environment.CACHE_VERSION }}-{{checksum "setup.py"}}-{{checksum "requirements-test.txt"}}
 
 
       # Cache the MNIST directory that contains benchmark data
       - restore_cache:
           keys:
-            - cache-key-benchmark-MNIST-{{ checksum "benchmarks/datasets/mnist.py"}}
+            - cache-key-benchmark-MNIST-{{.Environment.CACHE_VERSION }}-{{checksum "benchmarks/datasets/mnist.py"}}
 
       - <<: *install_dep_1_10_0
 
       - save_cache:
           paths:
             - ~/venv
-          key: cache-key-py37-benchmarks-torch-1-10-0-cuda-11-2-{{ checksum "setup.py"}}-{{ checksum "requirements-test.txt"}}
+          key: cache-key-py-3-9-7-benchmarks-torch-1-10-0-cuda-11-2-{{.Environment.CACHE_VERSION }}-{{checksum "setup.py"}}-{{checksum "requirements-test.txt"}}
 
       - <<: *install_repo
 
@@ -562,7 +534,7 @@ jobs:
       - save_cache:
           paths:
             - /tmp/MNIST
-          key: cache-key-benchmark-MNIST-{{ checksum "benchmarks/datasets/mnist.py"}}
+          key: cache-key-benchmark-MNIST-{{.Environment.CACHE_VERSION}}-{{checksum "benchmarks/datasets/mnist.py"}}
 
 
 workflows:

.github/workflows/pre-commit.yml

+name: pre-commit
+
+on:
+  pull_request:
+  push:
+    branches: [main]
+
+jobs:
+  pre-commit:
+    runs-on: ubuntu-latest
+    strategy:
+      matrix:
+        # make sure python versions are consistent with those used in .circleci/config.yml
+        python-version: ['3.7.12', '3.8.12', '3.9.7']
+    steps:
+    - uses: actions/checkout@v2
+    - uses: actions/setup-python@v2
+      with:
+        python-version: ${{ matrix.python-version }}
+    - uses: pre-commit/action@v2.0.3

.pre-commit-config.yaml

@@ -8,7 +8,7 @@ default_language_version:
 
 repos:
 -   repo: https://github.com/pre-commit/pre-commit-hooks
-    rev: v2.5.0
+    rev: v4.0.1
     hooks:
     -   id: trailing-whitespace
     -   id: check-ast
@@ -20,29 +20,31 @@ repos:
     -   id: end-of-file-fixer
 
 -   repo: https://github.com/ambv/black
-    rev: 19.10b0
+    rev: 21.10b0
     hooks:
     - id: black
 
 -   repo: https://gitlab.com/pycqa/flake8
-    rev: 3.7.9
+    rev: 4.0.1
     hooks:
     -   id: flake8
+        args: [--show-source, --statistics]
 
 -   repo: https://github.com/asottile/seed-isort-config
-    rev: v2.1.0
+    rev: v2.2.0
     hooks:
     -   id: seed-isort-config
 
 -   repo: https://github.com/pycqa/isort
-    rev: 5.6.4
+    rev: 5.10.1
     hooks:
     -   id: isort
         exclude: README.md
         additional_dependencies: [toml]
 
 -   repo: https://github.com/pre-commit/mirrors-mypy
-    rev: 'v0.790'
+    rev: 'v0.910'
     hooks:
     -   id: mypy
+        args: [--no-strict-optional, --ignore-missing-imports, --scripts-are-modules, --pretty]
         additional_dependencies: [numpy]

benchmarks/experimental/benchmark_dataset.py

@@ -42,7 +42,10 @@ class BenchmarkLMDataset(Dataset):
     """
 
     def __init__(
-        self, vocab_size=10000, max_source_positions=1024, total_samples=10000,
+        self,
+        vocab_size=10000,
+        max_source_positions=1024,
+        total_samples=10000,
     ):
         self.vocab_size = vocab_size
         self.max_source_positions = max_source_positions

benchmarks/experimental/benchmark_mevo.py

@@ -35,7 +35,7 @@ KERNELS = [
 
 
 def run_on_gpu(kernel, data, repeats, no_grad, fwd_bwd):
-    """ Measure both GPU runtime and peak memory usage of a kernel. """
+    """Measure both GPU runtime and peak memory usage of a kernel."""
     tokens = data[0].shape[0]
 
     def get_cuda_data():

benchmarks/experimental/experimental_async_approaches.py

@@ -142,7 +142,7 @@ class MySGD(Optimizer):
         super(MySGD, self).__setstate__(state)
 
     def step(self, closure=None):
-        """ Performs a single optimization step.
+        """Performs a single optimization step.
         Args:
             closure (callable, optional): A closure that reevaluates the model
                 and returns the loss.
@@ -234,7 +234,7 @@ class SpectrainSGDMomentum(Optimizer):
                     p.data.sub_(param_state["momentum_buffer"].data, alpha=multiplier)
 
     def step(self, weight_prediction=True, closure=None):
-        """ Performs a single optimization step.
+        """Performs a single optimization step.
         Args:
             weight_prediction (bool, optional): Enable weight prediction based updates
             closure (callable, optional): A closure that reevaluates the model

benchmarks/experimental/offload.py

@@ -413,7 +413,10 @@ parser.add_argument(
     help="Print debugging statements which is more verbose than the default.",
 )
 parser.add_argument(
-    "--model_name", default="lm", type=str, help="Language Model(LM) used to benchmark nn.pipe.",
+    "--model_name",
+    default="lm",
+    type=str,
+    help="Language Model(LM) used to benchmark nn.pipe.",
 )
 parser.add_argument(
     "--use_synthetic_data", default=True, action="store_true", help="Uses synthetic data for running benchmarks."

benchmarks/oss.py

@@ -320,14 +320,22 @@ if __name__ == "__main__":
     if args.optim_type == OptimType.oss_ddp or args.optim_type == OptimType.everyone:
         logging.info("\n*** Benchmark OSS with DDP")
         mp.spawn(
-            train, args=(args, BACKEND, OptimType.oss_ddp, args.check_regression), nprocs=args.world_size, join=True,  # type: ignore
+            train,
+            args=(args, BACKEND, OptimType.oss_ddp, args.check_regression),
+            nprocs=args.world_size,
+            join=True,  # type: ignore
         )
 
     if args.optim_type == OptimType.oss_sharded_ddp or args.optim_type == OptimType.everyone:
         logging.info("\n*** Benchmark OSS with ShardedDDP")
         mp.spawn(
             train,  # type: ignore
-            args=(args, BACKEND, OptimType.oss_sharded_ddp, args.check_regression,),
+            args=(
+                args,
+                BACKEND,
+                OptimType.oss_sharded_ddp,
+                args.check_regression,
+            ),
             nprocs=args.world_size,
             join=True,
         )

fairscale/experimental/nn/ampnet_pipe/ampnet.py

@@ -34,13 +34,20 @@ from fairscale.nn.pipe.worker import Task
 
 
 def create_task_without_skip_trackers(
-    checkpoint_stop: int, i: int, j: int, batch: Batch, partition: nn.Sequential,
+    checkpoint_stop: int,
+    i: int,
+    j: int,
+    batch: Batch,
+    partition: nn.Sequential,
 ) -> Task:
     # Determine whether checkpointing or not.
     if i < checkpoint_stop:
 
         def function(
-            input: TensorOrTensors, partition: nn.Sequential = partition, chunk_id: int = i, part_id: int = j,
+            input: TensorOrTensors,
+            partition: nn.Sequential = partition,
+            chunk_id: int = i,
+            part_id: int = j,
         ) -> TensorOrTensors:
             with record_function("chunk%d-part%d" % (chunk_id, part_id)):
                 return partition(input)
@@ -52,7 +59,10 @@ def create_task_without_skip_trackers(
     else:
 
         def compute(
-            batch: Batch = batch, partition: nn.Sequential = partition, chunk_id: int = i, part_id: int = j,
+            batch: Batch = batch,
+            partition: nn.Sequential = partition,
+            chunk_id: int = i,
+            part_id: int = j,
         ) -> Batch:
             with record_function("chunk%d-part%d" % (chunk_id, part_id)):
                 return batch.call(partition)
@@ -93,7 +103,11 @@ class AsyncAMPnetEventLoop:
 
     def async_send_inner(self, batch: Batch, index: int) -> Tuple[Batch, PipeMessage]:
         task = create_task_without_skip_trackers(
-            self.checkpoint_stop, index, self.group.rank(), batch, self.partitions[0].module,
+            self.checkpoint_stop,
+            index,
+            self.group.rank(),
+            batch,
+            self.partitions[0].module,
         )
         result = task.compute()
         task.finalize(result)
@@ -258,7 +272,11 @@ class AsyncAMPnetEventLoop:
                 if self.weight_prediction:
                     optimizer.update_weight_using_future_predictions(cur_rank, N, count, self.chunks, forward=True)
                 task = create_task_without_skip_trackers(
-                    self.checkpoint_stop, args.microbatch_index, self.group.rank(), batch, self.partitions[0].module,
+                    self.checkpoint_stop,
+                    args.microbatch_index,
+                    self.group.rank(),
+                    batch,
+                    self.partitions[0].module,
                 )
                 output = task.compute()
                 activations[args.microbatch_index] = output

fairscale/experimental/nn/ampnet_pipe/pipe.py

@@ -46,8 +46,8 @@ class AMPnetPipe(AsyncPipe):
         assert self.group
         rank = self.group.rank()
 
-        transport = self.pipeline.transport  # type: ignore
-        checkpoint_stop = self.pipeline.checkpoint_stop  # type: ignore
+        transport = self.pipeline.transport
+        checkpoint_stop = self.pipeline.checkpoint_stop
         ampnet_event_loop = AsyncAMPnetEventLoop(
             partitions,
             self.group,

fairscale/experimental/nn/data_parallel/gossip/distributed.py

@@ -312,7 +312,7 @@ class SlowMoDistributedDataParallel(Module):
         self.logger.debug("Initialization of SlowMoDistributedDataParallel complete")
 
     def _initialize_logger(self, verbose: bool, process_rank: int) -> None:
-        """ Initializes the logger """
+        """Initializes the logger"""
         self.logger = logging.getLogger(__name__)
         if verbose:
             self.logger.setLevel(logging.DEBUG)
@@ -331,7 +331,7 @@ class SlowMoDistributedDataParallel(Module):
         master_group: Optional[torch.distributed.ProcessGroup],
         local_node_group: Optional[torch.distributed.ProcessGroup],
     ) -> Tuple[int, int]:
-        """ Creates the process groups required for the SlowMo implementation """
+        """Creates the process groups required for the SlowMo implementation"""
 
         self.local_rank = process_rank % self.nprocs_per_node
         assert (
@@ -392,7 +392,12 @@ class SlowMoDistributedDataParallel(Module):
 
         self.logger.debug("Initializing local process groups")
         for node in range(logical_world_size):
-            node_processes_ranks = list(range(node * self.nprocs_per_node, (node + 1) * self.nprocs_per_node,))
+            node_processes_ranks = list(
+                range(
+                    node * self.nprocs_per_node,
+                    (node + 1) * self.nprocs_per_node,
+                )
+            )
             # Process group to communicate between processes on this machine
             new_local_group = create_process_group(node_processes_ranks)
             if process_rank in node_processes_ranks:
@@ -401,24 +406,26 @@ class SlowMoDistributedDataParallel(Module):
         self.logger.debug("Initialization of local groups complete")
 
     def forward(self, *inputs: Any, **kwargs: Any) -> Union[torch.Tensor, List[torch.Tensor]]:
-        """ Forward pass performed in parallel across all devices on node """
+        """Forward pass performed in parallel across all devices on node"""
         return self.module(*inputs, **kwargs)
 
     def _sync_params(self) -> None:
-        """ Synchronize parameters across devices (intra-node) """
+        """Synchronize parameters across devices (intra-node)"""
         if self.local_node_group is None:
             return
 
         # intra-node parameter sync
         params = cast(List[torch.Tensor], list(self.module.parameters()))
         communication_op = functools.partial(
-            dist.broadcast, src=self.logical_rank * self.nprocs_per_node, group=self.local_node_group,
+            dist.broadcast,
+            src=self.logical_rank * self.nprocs_per_node,
+            group=self.local_node_group,
         )
         communicate(params, communication_op)
         self.logger.debug("Intra-node param sync complete")
 
     def _sync_buffers(self) -> None:
-        """ Synchronize buffers across nodes """
+        """Synchronize buffers across nodes"""
         # module buffer sync
         if self.broadcast_buffers and len(self.module_buffers) > 0:
             # Synchronize buffers across processes.
@@ -432,17 +439,18 @@ class SlowMoDistributedDataParallel(Module):
         dist._broadcast_coalesced(process_group, tensors, buffer_size)
 
     def _create_event_recorder(self, event_name: str) -> EventRecorder:
-        """ Creates an cuda event recorder which helps in profiling """
+        """Creates an cuda event recorder which helps in profiling"""
         return create_event_recorder(event_name, dummy=not self.profile_mode)
 
     def _fp16_fp32_iterator(
         self, optimizer: torch.optim.Optimizer, fp32_params: Optional[torch.Tensor]
     ) -> Iterable[Tuple[torch.Tensor, torch.Tensor]]:
-        """ Iterator for those fp16 parameters which have a fp32 copy """
+        """Iterator for those fp16 parameters which have a fp32 copy"""
         # Handle apex fp16 optimizer
         if hasattr(optimizer, "_amp_stash") and hasattr(optimizer._amp_stash, "fp16_groups"):
             for p_fp16_group, p_fp32_group in zip(
-                optimizer._amp_stash.fp16_groups, optimizer._amp_stash.fp32_from_fp16_groups,
+                optimizer._amp_stash.fp16_groups,
+                optimizer._amp_stash.fp32_from_fp16_groups,
             ):
                 for p_fp16, p_fp32 in zip(p_fp16_group, p_fp32_group):
                     yield p_fp16, p_fp32
@@ -594,7 +602,7 @@ class SlowMoDistributedDataParallel(Module):
                     ef1.copy_(p_fp32 - p_fp16.float())
 
     def perform_slowmo(self, optimizer: torch.optim.Optimizer, fp32_params: Optional[torch.Tensor] = None) -> None:
-        """ This is to be called after optimizer.step(). It performs the approximate averaging using
+        """This is to be called after optimizer.step(). It performs the approximate averaging using
         the base algorithm (SGP/ LocalSGD) and the slow momentum step. Since LocalSGD and the slow
         momentum step are not performed every iteration, it only performs those when needed.
 
@@ -645,7 +653,7 @@ class SlowMoDistributedDataParallel(Module):
         self.num_updates += 1
 
     def _init_global_momentum_buffers(self, optimizer: torch.optim.Optimizer) -> None:
-        """ Initializes the slow momentum buffers """
+        """Initializes the slow momentum buffers"""
         self.global_momentum_buffers_initialized = True
 
         if not self.slowmo:
@@ -707,7 +715,7 @@ class SlowMoDistributedDataParallel(Module):
         self.global_momentum_buffer = torch.zeros_like(self.old_params).detach()
 
     def _distributed_comm(self, optimizer: torch.optim.Optimizer, mode: str) -> None:
-        """ Performs the communication needed for the efficient SlowMo implementation """
+        """Performs the communication needed for the efficient SlowMo implementation"""
         offset = 0
         slowmo_comm_lists: List[List[torch.Tensor]] = [[] for _ in range(self.slowmo_num_shards)]
         with torch.no_grad():
@@ -743,7 +751,7 @@ class SlowMoDistributedDataParallel(Module):
                 communicate(slowmo_comm_list, communication_op)
 
     def _global_momentum_step(self, optimizer: torch.optim.Optimizer) -> None:
-        """ Performs the slow momentum step """
+        """Performs the slow momentum step"""
         if not self.slowmo:
             return
 
@@ -760,7 +768,7 @@ class SlowMoDistributedDataParallel(Module):
             self._distributed_comm(optimizer, mode="scatter")
 
     def _perform_local_optimization(self, optimizer: torch.optim.Optimizer) -> None:
-        """ Performs the slow momentum on the local shard """
+        """Performs the slow momentum on the local shard"""
         assert self.portion_start is not None
 
         with torch.no_grad():
@@ -838,7 +846,7 @@ class SlowMoDistributedDataParallel(Module):
         self.logger.debug("making forward pre-hook")
 
         def hook(*unused: Any) -> None:
-            """ Query gossip queue and de-bias during forward pass """
+            """Query gossip queue and de-bias during forward pass"""
             # sync buffers before the forward pass
             self._sync_buffers()
 
@@ -869,7 +877,7 @@ class SlowMoDistributedDataParallel(Module):
         use_streams: bool = True,
         slowmo_sgp_average_params: bool = False,
     ) -> None:
-        """ Perform initialization for Stochastic Gradient Push base algorithm """
+        """Perform initialization for Stochastic Gradient Push base algorithm"""
 
         if graph is None:
             graph = NPDDEGraph(logical_rank, logical_world_size, self.nprocs_per_node, self.local_rank)
@@ -959,7 +967,7 @@ class SlowMoDistributedDataParallel(Module):
         super(SlowMoDistributedDataParallel, self).load_state_dict(cast(Dict[str, torch.Tensor], state_dict))
 
     def _sgp_ps_numerator(self) -> None:
-        """ Convert model params to ps-numerator """
+        """Convert model params to ps-numerator"""
         if not self.is_sgp_ps_numerator:
             if not self.lazy_mixing:
                 ps_weight = self.ps_weight
@@ -969,7 +977,7 @@ class SlowMoDistributedDataParallel(Module):
             self.is_sgp_ps_numerator = True
 
     def _sgp_unbias(self) -> None:
-        """ Convert model params to de-biased estimate """
+        """Convert model params to de-biased estimate"""
         if self.is_sgp_ps_numerator:
             if not self.lazy_mixing:
                 ps_weight = self.ps_weight
@@ -992,7 +1000,7 @@ class SlowMoDistributedDataParallel(Module):
         return self
 
     def _sgp_query_gossip_queue(self, non_blocking: bool = False) -> bool:
-        """ Check gossip-queue for push-sum residuals and update model """
+        """Check gossip-queue for push-sum residuals and update model"""
         if not self.gossip_enable:
             return False
 
@@ -1046,7 +1054,7 @@ class SlowMoDistributedDataParallel(Module):
         return False
 
     def _sgp_transfer_params(self, mix: bool = True) -> bool:
-        """ Transfers COPY of model parameters to gossip queue """
+        """Transfers COPY of model parameters to gossip queue"""
         if not self.gossip_enable or self.process_rank % self.nprocs_per_node != 0:
             return False
 
@@ -1130,7 +1138,7 @@ class SlowMoDistributedDataParallel(Module):
         gossip_ps_factor: torch.Tensor,
         gossip_stream: torch.cuda.Stream,
     ) -> None:
-        """ Gossip thread, which performs push-sum on model params """
+        """Gossip thread, which performs push-sum on model params"""
         logger = make_logger(dist_config["logical_rank"], dist_config["verbose"])
 
         gossip_params_by_dtype = group_by_dtype(gossip_params)

fairscale/experimental/nn/data_parallel/gossip/gossiper.py

@@ -30,7 +30,7 @@ class dist_backend(str, Enum):
 
 
 class Gossiper(object):
-    """ Generic gossip averaging object for multi-peer communication
+    """Generic gossip averaging object for multi-peer communication
 
     Args:
         msg (torch.Tensor): message used to initialize recv buffer
@@ -121,7 +121,7 @@ class Gossiper(object):
         self._graph_manager.peers_per_itr = v
 
     def refresh_peers_(self, rotate: Optional[bool] = None) -> None:
-        """ Update in- and out-peers """
+        """Update in- and out-peers"""
         if rotate is None:
             rotate = self._graph_manager.is_dynamic_graph()
         # cannot cycle peers in a static graph
@@ -129,11 +129,11 @@ class Gossiper(object):
         self.out_edges, self.in_edges = self._graph_manager.get_edges(rotate)
 
     def refresh_mixing_weights_(self, residual_adjusted: bool = False) -> None:
-        """ Update mixing-matrix weights """
+        """Update mixing-matrix weights"""
         self.mixing_weights = self._mixing_manager.get_mixing_weights(residual_adjusted)
 
     def mix_out_msg_(self, out_msg: torch.Tensor, ps_weight: torch.Tensor) -> Iterator[torch.Tensor]:
-        """ Returns a generator mixing messages on the fly """
+        """Returns a generator mixing messages on the fly"""
         self.refresh_mixing_weights_(residual_adjusted=True)
         self.ps_weight = ps_weight
 
@@ -153,14 +153,14 @@ class Gossiper(object):
                 yield out_msg.mul(weight.type(out_msg.dtype))  # type: ignore
 
     def clean_msg_buffers_(self) -> None:
-        """ Clean outgoing message buffer """
+        """Clean outgoing message buffer"""
         while len(self.out_msg_buffer) > 0:
             req, msg = self.out_msg_buffer.pop()
             req.wait()
             msg.set_()
 
     def parse_in_msg_buffer(self) -> Tuple[torch.Tensor, torch.Tensor]:
-        """ Parse in-msg buffer and return msg and ps-weight separately """
+        """Parse in-msg buffer and return msg and ps-weight separately"""
         msg = self.in_msg_buffer
         if not self.regular:
             return msg.narrow(0, 0, len(msg) - 1), msg[-1]
@@ -168,15 +168,15 @@ class Gossiper(object):
             return msg, self.ps_weight * self.peers_per_itr_device
 
     def mix(self, out_msg: torch.Tensor, ps_weight: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
-        """ Single gossip step """
+        """Single gossip step"""
         raise NotImplementedError
 
 
 class PushSum(Gossiper):
-    """ 1-peer Push-Sum consensus averaging module """
+    """1-peer Push-Sum consensus averaging module"""
 
     def mix(self, out_msg: torch.Tensor, ps_weight: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
-        """ Consensus averaging step """
+        """Consensus averaging step"""
         # out_msg must be on the correct device
         assert out_msg.device.type == self.device.type
         if self.logger is not None:
@@ -189,7 +189,12 @@ class PushSum(Gossiper):
         for out_edge in self.out_edges:
             msg = next(mixed_out_msgs)
             assert self.rank == out_edge.src
-            req = dist.broadcast(tensor=msg, src=out_edge.src, group=out_edge.process_group, async_op=True,)
+            req = dist.broadcast(
+                tensor=msg,
+                src=out_edge.src,
+                group=out_edge.process_group,
+                async_op=True,
+            )
             self.out_msg_buffer.append((req, msg))
 
         # blocking recv w/ some code optimization to avoid buffer prep overhead
@@ -204,7 +209,9 @@ class PushSum(Gossiper):
 
             for in_edge in self.in_edges:
                 dist.broadcast(
-                    tensor=self.placeholder, src=in_edge.src, group=in_edge.process_group,
+                    tensor=self.placeholder,
+                    src=in_edge.src,
+                    group=in_edge.process_group,
                 )
                 self.in_msg_buffer.add_(self.placeholder)  # type: ignore
 
@@ -214,7 +221,7 @@ class PushSum(Gossiper):
 
 
 class PushPull(Gossiper):
-    """ Doubly-stochastic consensus averaging module """
+    """Doubly-stochastic consensus averaging module"""
 
     def mix(self, out_msg: torch.Tensor, ps_weight: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
         # out_msg must be on the correct device
@@ -232,11 +239,15 @@ class PushPull(Gossiper):
             if not self.passive:
                 dist.broadcast(tensor=msg, src=out_edge.src, group=out_edge.process_group)
                 dist.broadcast(
-                    tensor=self.in_msg_buffer, src=in_edge.src, group=in_edge.process_group,
+                    tensor=self.in_msg_buffer,
+                    src=in_edge.src,
+                    group=in_edge.process_group,
                 )
             else:
                 dist.broadcast(
-                    tensor=self.in_msg_buffer, src=in_edge.src, group=in_edge.process_group,
+                    tensor=self.in_msg_buffer,
+                    src=in_edge.src,
+                    group=in_edge.process_group,
                 )
                 dist.broadcast(tensor=msg, src=out_edge.src, group=out_edge.process_group)
 
@@ -251,11 +262,15 @@ class PushPull(Gossiper):
                 if not self.passive:
                     dist.broadcast(tensor=msg, src=out_edge.src, group=out_edge.process_group)
                     dist.broadcast(
-                        tensor=self.placeholder, src=in_edge.src, group=in_edge.process_group,
+                        tensor=self.placeholder,
+                        src=in_edge.src,
+                        group=in_edge.process_group,
                     )
                 else:
                     dist.broadcast(
-                        tensor=self.placeholder, src=in_edge.src, group=in_edge.process_group,
+                        tensor=self.placeholder,
+                        src=in_edge.src,
+                        group=in_edge.process_group,
                     )
                     dist.broadcast(tensor=msg, src=out_edge.src, group=out_edge.process_group)
                 self.in_msg_buffer.add_(self.placeholder)  # type: ignore

fairscale/experimental/nn/data_parallel/gossip/graph_manager.py

@@ -77,26 +77,26 @@ class GraphManager(ABC):
 
     @abstractmethod
     def is_regular_graph(self) -> bool:
-        """ Whether each node has the same number of in-peers as out-peers """
+        """Whether each node has the same number of in-peers as out-peers"""
         raise NotImplementedError
 
     @abstractmethod
     def is_bipartite_graph(self) -> bool:
-        """ Whether graph is bipartite or not """
+        """Whether graph is bipartite or not"""
         raise NotImplementedError
 
     @abstractmethod
     def is_passive(self, rank: Optional[int] = None) -> bool:
-        """ Whether 'rank' is a passive node or not """
+        """Whether 'rank' is a passive node or not"""
         raise NotImplementedError
 
     @abstractmethod
     def is_dynamic_graph(self) -> bool:
-        """ Whether the graph-type is dynamic (as opposed to static) """
+        """Whether the graph-type is dynamic (as opposed to static)"""
         raise NotImplementedError
 
     def get_peers(self, rotate: bool = False) -> Tuple[List[int], List[int]]:
-        """ Returns the out and in-peers corresponding to 'self.rank' """
+        """Returns the out and in-peers corresponding to 'self.rank'"""
         # cycle through in- and out-peers by updating group-index
         if rotate:
             self._rotate_group_indices()
@@ -113,8 +113,8 @@ class GraphManager(ABC):
         return out_peers, in_peers
 
     def get_edges(self, rotate: bool = False) -> Tuple[List[Edge], List[Edge]]:
-        """ Returns the pairwise process groups between rank and the out and
-        in-peers corresponding to 'self.rank' """
+        """Returns the pairwise process groups between rank and the out and
+        in-peers corresponding to 'self.rank'"""
         # cycle through in- and out-peers by updating group-index
         if rotate:
             self._rotate_group_indices()
@@ -131,17 +131,17 @@ class GraphManager(ABC):
         return out_edges, in_edges
 
     def _rotate_group_indices(self) -> None:
-        """ Incerement group indices to point to the next out-peer """
+        """Incerement group indices to point to the next out-peer"""
         increment = self.peers_per_itr
         for i, group_index in enumerate(self._group_indices):
             self._group_indices[i] = int((group_index + increment) % len(self.phone_book[self.rank]))
 
     def _rotate_forward(self, r: int, p: int) -> int:
-        """ Helper function returns peer that is p hops ahead of r """
+        """Helper function returns peer that is p hops ahead of r"""
         return (r + p) % self.world_size
 
     def _rotate_backward(self, r: int, p: int) -> int:
-        """ Helper function returns peer that is p hops behind r """
+        """Helper function returns peer that is p hops behind r"""
         return (r - p) % self.world_size
 
 

fairscale/experimental/nn/data_parallel/gossip/mixing_manager.py

@@ -32,18 +32,18 @@ class MixingManager(ABC):
 
     @abstractmethod
     def is_uniform(self) -> bool:
-        """ Whether mixing weights are distributed uniformly over peers """
+        """Whether mixing weights are distributed uniformly over peers"""
         raise NotImplementedError
 
     @abstractmethod
     def get_mixing_weights(self, residual_adjusted: bool = True) -> Dict[Union[str, int], torch.Tensor]:
-        """ Create mixing weight dictionary using uniform allocation """
+        """Create mixing weight dictionary using uniform allocation"""
         raise NotImplementedError
 
 
 class UniformMixing(MixingManager):
     def get_mixing_weights(self, residual_adjusted: bool = True) -> Dict[Union[str, int], torch.Tensor]:
-        """ Create mixing weight dictionary using uniform allocation """
+        """Create mixing weight dictionary using uniform allocation"""
         mixing_weights: Dict[Union[str, int], torch.Tensor] = {}
         out_peers, _ = self.graph_manager.get_peers()
 

fairscale/experimental/nn/data_parallel/gossip/utils/cuda_metering.py

@@ -36,7 +36,7 @@ def create_event_recorder(event_name: str, dummy: bool = False) -> EventRecorder
 
 
 class CudaEventRecorder(EventRecorder):
-    """ Allows profiling in an easy-to-use manner. CudaEventRecorder can be used
+    """Allows profiling in an easy-to-use manner. CudaEventRecorder can be used
     in a loop. When it is used in a loop (or when an event recorder is created
     multiple times with the same name), get_timings returns the statistics of the
     timings since the last reset. Note: in case the number of timings is greater than
@@ -92,19 +92,22 @@ class CudaEventRecorder(EventRecorder):
             time_taken_list = [event_recorder.find_time_elapsed() for event_recorder in event_recorder_list]
 
             all_timings_str += ("{}: Time taken: avg: {}, std: {}, count: " "{}\n").format(
-                event_name, statistics.mean(time_taken_list), statistics.pstdev(time_taken_list), len(time_taken_list),
+                event_name,
+                statistics.mean(time_taken_list),
+                statistics.pstdev(time_taken_list),
+                len(time_taken_list),
             )
 
         return all_timings_str
 
     @classmethod
     def get_timings(cls) -> str:
-        """ Returns the timings since last reset was called """
+        """Returns the timings since last reset was called"""
         return cls.get_common_timings(cls.event_recorders, "Timings since last reset")
 
     @classmethod
     def get_all_timings(cls) -> str:
-        """ Returns the statistics of all the timings """
+        """Returns the statistics of all the timings"""
         return cls.get_common_timings(cls.all_event_recorders, "All timings")
 
 

fairscale/experimental/nn/data_parallel/gossip/utils/helpers.py

@@ -86,7 +86,10 @@ def communicate(tensors: List[torch.Tensor], communication_op: Any, logger: logg
         if logger is not None:
             logger.debug("Commmunication completed")
         with torch.no_grad():
-            for f, t in zip(unflatten_tensors(flat_tensor, tensors_with_same_dtype), tensors_with_same_dtype,):
+            for f, t in zip(
+                unflatten_tensors(flat_tensor, tensors_with_same_dtype),
+                tensors_with_same_dtype,
+            ):
                 t.copy_(f)
         if logger is not None:
             logger.debug("Unflatten completed")

fairscale/experimental/nn/distributed_pipeline/graph.py

@@ -198,7 +198,9 @@ class PipelineModulesGraph(nn.Module):
                 remote_module = partition[0].module.get_module_rref()
             else:
                 remote_module = rpc.remote(
-                    partition[0].module.on, RemoteSequential, args=([p.module.get_module_rref() for p in partition],),
+                    partition[0].module.on,
+                    RemoteSequential,
+                    args=([p.module.get_module_rref() for p in partition],),
                 )
             partitions.append((partition, remote_module))
 

fairscale/experimental/nn/distributed_pipeline/partition_handler.py

@@ -25,7 +25,7 @@ ExcInfo = Tuple[Type[BaseException], BaseException, TracebackType]
 
 
 class DistributedPipelineRecord:
-    """ A class for storing a single mini-batch (consisting of multiple micro-batches) as input to
+    """A class for storing a single mini-batch (consisting of multiple micro-batches) as input to
     a single partition.
     Args:
         device: the local device that runs the partition.
@@ -73,7 +73,7 @@ class DistributedPipelineRecord:
         return {}
 
     def feed(self, chunk: int, input_idx: int, input: Tensor) -> Tensor:
-        """ This function is called remotely to provide individual tensors of a given chunk."""
+        """This function is called remotely to provide individual tensors of a given chunk."""
         if input.device.type == "cpu":
             input = input.to(self.device)
         cuda_stream = torch.cuda.current_stream(input.device) if input.device.type == "cuda" else None

fairscale/experimental/nn/distributed_pipeline/pipeline.py

@@ -70,7 +70,12 @@ class DistributedPipeline(nn.Module):
 
     DataConsumer = DataConsumer[Partition]
 
-    def __init__(self, graph: PipelineModulesGraph, chunks: int = 1, checkpoint: str = "except_last",) -> None:
+    def __init__(
+        self,
+        graph: PipelineModulesGraph,
+        chunks: int = 1,
+        checkpoint: str = "except_last",
+    ) -> None:
         super().__init__()
 
         check_pytorch_version()