Update quality tools to 2023 (#1046)

* Setup 2023 tooling for quality

* Result of styling

* Simplify inits and remove isort and flake8 from doc

* Puts back isort skip flag

.gitignore

@@ -139,3 +139,6 @@ dmypy.json
 
 # More test things
 wandb
+
+# ruff
+.ruff_cache

CONTRIBUTING.md

@@ -152,7 +152,7 @@ Follow these steps to start contributing:
    $ make test
    ```
 
-   `accelerate` relies on `black` and `isort` to format its source code
+   `accelerate` relies on `black` and `ruff` to format its source code
    consistently. After you make changes, apply automatic style corrections and code verifications
    that can't be automated in one go with:
 
@@ -165,7 +165,7 @@ Follow these steps to start contributing:
    $ make style
    ```
 
-   `accelerate` also uses `flake8` and a few custom scripts to check for coding mistakes. Quality
+   `accelerate` also uses a few custom scripts to check for coding mistakes. Quality
    control runs in CI, however you can also run the same checks with:
 
    ```bash

Makefile

@@ -13,14 +13,13 @@ extra_quality_checks:
 # this target runs checks on all files
 quality:
 	black --check $(check_dirs)
-	isort --check-only $(check_dirs)
-	flake8 $(check_dirs)
+	ruff $(check_dirs)
 	doc-builder style src/accelerate docs/source --max_len 119 --check_only
 
 # Format source code automatically and check is there are any problems left that need manual fixing
 style:
 	black $(check_dirs)
-	isort $(check_dirs)
+	ruff $(check_dirs) --fix
 	doc-builder style src/accelerate docs/source --max_len 119
 	
 # Run tests for the library

benchmarks/big_model_inference.py

@@ -16,12 +16,12 @@ import argparse
 import time
 
 import torch
-
 import transformers
-from accelerate.utils import compute_module_sizes
 from measures_util import end_measure, log_measures, start_measure
 from transformers import AutoConfig, AutoModelForCausalLM, AutoModelForSeq2SeqLM, AutoTokenizer
 
+from accelerate.utils import compute_module_sizes
+
 
 DEFAULT_MODELS = {
     "gpt-j-6b": {"is_causal": True, "model": "sgugger/sharded-gpt-j-6B", "tokenizer": "EleutherAI/gpt-j-6B"},

benchmarks/measures_util.py

@@ -2,9 +2,8 @@ import gc
 import threading
 import time
 
-import torch
-
 import psutil
+import torch
 
 
 class PeakCPUMemory:

docs/source/usage_guides/megatron_lm.mdx

@@ -290,6 +290,7 @@ You will implement the `accelerate.utils.AbstractTrainStep` or inherit from thei
 ```python
 from accelerate.utils import MegatronLMDummyScheduler, GPTTrainStep, avg_losses_across_data_parallel_group
 
+
 # Custom loss function for the Megatron model
 class GPTTrainStepWithCustomLoss(GPTTrainStep):
     def __init__(self, megatron_args, **kwargs):

examples/by_feature/automatic_gradient_accumulation.py

@@ -14,16 +14,16 @@
 import argparse
 import os
 
-import torch
-from torch.optim import AdamW
-from torch.utils.data import DataLoader
-
 # New Code #
 import evaluate
+import torch
+from datasets import load_dataset
+from torch.optim import AdamW
+from torch.utils.data import DataLoader
+from transformers import AutoModelForSequenceClassification, AutoTokenizer, get_linear_schedule_with_warmup, set_seed
+
 from accelerate import Accelerator, DistributedType
 from accelerate.utils import find_executable_batch_size
-from datasets import load_dataset
-from transformers import AutoModelForSequenceClassification, AutoTokenizer, get_linear_schedule_with_warmup, set_seed
 
 
 ########################################################################

examples/by_feature/checkpointing.py

@@ -15,15 +15,15 @@
 import argparse
 import os
 
+import evaluate
 import torch
+from datasets import load_dataset
 from torch.optim import AdamW
 from torch.utils.data import DataLoader
-
-import evaluate
-from accelerate import Accelerator, DistributedType
-from datasets import load_dataset
 from transformers import AutoModelForSequenceClassification, AutoTokenizer, get_linear_schedule_with_warmup, set_seed
 
+from accelerate import Accelerator, DistributedType
+
 
 ########################################################################
 # This is a fully working simple example to use Accelerate,

examples/by_feature/cross_validation.py

@@ -15,20 +15,20 @@
 import argparse
 from typing import List
 
+import evaluate
 import numpy as np
 import torch
-from torch.optim import AdamW
-from torch.utils.data import DataLoader
-
-import evaluate
-from accelerate import Accelerator, DistributedType
 from datasets import DatasetDict, load_dataset
 
 # New Code #
 # We'll be using StratifiedKFold for this example
 from sklearn.model_selection import StratifiedKFold
+from torch.optim import AdamW
+from torch.utils.data import DataLoader
 from transformers import AutoModelForSequenceClassification, AutoTokenizer, get_linear_schedule_with_warmup, set_seed
 
+from accelerate import Accelerator, DistributedType
+
 
 ########################################################################
 # This is a fully working simple example to use Accelerate,

examples/by_feature/deepspeed_with_config_support.py

@@ -31,16 +31,12 @@ import random
 from itertools import chain
 from pathlib import Path
 
-import torch
-from torch.utils.data import DataLoader
-
 import datasets
+import torch
 import transformers
-from accelerate import Accelerator, DistributedType
-from accelerate.logging import get_logger
-from accelerate.utils import DummyOptim, DummyScheduler, set_seed
 from datasets import load_dataset
 from huggingface_hub import Repository
+from torch.utils.data import DataLoader
 from tqdm.auto import tqdm
 from transformers import (
     CONFIG_MAPPING,
@@ -55,6 +51,10 @@ from transformers import (
 from transformers.utils import get_full_repo_name
 from transformers.utils.versions import require_version
 
+from accelerate import Accelerator, DistributedType
+from accelerate.logging import get_logger
+from accelerate.utils import DummyOptim, DummyScheduler, set_seed
+
 
 logger = get_logger(__name__)
 

examples/by_feature/fsdp_with_peak_mem_tracking.py

@@ -16,14 +16,14 @@ import argparse
 import gc
 import os
 
-import torch
-from torch.utils.data import DataLoader
-
 import evaluate
-from accelerate import Accelerator, DistributedType
+import torch
 from datasets import load_dataset
+from torch.utils.data import DataLoader
 from transformers import AutoModelForSequenceClassification, AutoTokenizer, get_linear_schedule_with_warmup, set_seed
 
+from accelerate import Accelerator, DistributedType
+
 
 ########################################################################
 # This is a fully working simple example to use Accelerate

examples/by_feature/gradient_accumulation.py

@@ -15,15 +15,15 @@
 import argparse
 import os
 
+import evaluate
 import torch
+from datasets import load_dataset
 from torch.optim import AdamW
 from torch.utils.data import DataLoader
-
-import evaluate
-from accelerate import Accelerator, DistributedType
-from datasets import load_dataset
 from transformers import AutoModelForSequenceClassification, AutoTokenizer, get_linear_schedule_with_warmup, set_seed
 
+from accelerate import Accelerator, DistributedType
+
 
 ########################################################################
 # This is a fully working simple example to use Accelerate

examples/by_feature/megatron_lm_gpt_pretraining.py

@@ -31,16 +31,12 @@ import random
 from itertools import chain
 from pathlib import Path
 
-import torch
-from torch.utils.data import DataLoader
-
 import datasets
+import torch
 import transformers
-from accelerate import Accelerator, DistributedType
-from accelerate.logging import get_logger
-from accelerate.utils import MegatronLMDummyScheduler, set_seed
 from datasets import load_dataset
 from huggingface_hub import Repository
+from torch.utils.data import DataLoader
 from tqdm.auto import tqdm
 from transformers import (
     CONFIG_MAPPING,
@@ -55,6 +51,10 @@ from transformers import (
 from transformers.utils import check_min_version, get_full_repo_name, send_example_telemetry
 from transformers.utils.versions import require_version
 
+from accelerate import Accelerator, DistributedType
+from accelerate.logging import get_logger
+from accelerate.utils import MegatronLMDummyScheduler, set_seed
+
 
 # Will error if the minimal version of Transformers is not installed. Remove at your own risks.
 check_min_version("4.23.0.dev0")

examples/by_feature/memory.py

@@ -14,16 +14,16 @@
 import argparse
 import os
 
-import torch
-from torch.optim import AdamW
-from torch.utils.data import DataLoader
-
 # New Code #
 import evaluate
+import torch
+from datasets import load_dataset
+from torch.optim import AdamW
+from torch.utils.data import DataLoader
+from transformers import AutoModelForSequenceClassification, AutoTokenizer, get_linear_schedule_with_warmup, set_seed
+
 from accelerate import Accelerator, DistributedType
 from accelerate.utils import find_executable_batch_size
-from datasets import load_dataset
-from transformers import AutoModelForSequenceClassification, AutoTokenizer, get_linear_schedule_with_warmup, set_seed
 
 
 ########################################################################

examples/by_feature/multi_process_metrics.py

@@ -15,15 +15,15 @@
 import argparse
 import os
 
+import evaluate
 import torch
+from datasets import load_dataset
 from torch.optim import AdamW
 from torch.utils.data import DataLoader
-
-import evaluate
-from accelerate import Accelerator, DistributedType
-from datasets import load_dataset
 from transformers import AutoModelForSequenceClassification, AutoTokenizer, get_linear_schedule_with_warmup, set_seed
 
+from accelerate import Accelerator, DistributedType
+
 
 ########################################################################
 # This is a fully working simple example to use Accelerate,

examples/by_feature/tracking.py

@@ -15,15 +15,15 @@
 import argparse
 import os
 
+import evaluate
 import torch
+from datasets import load_dataset
 from torch.optim import AdamW
 from torch.utils.data import DataLoader
-
-import evaluate
-from accelerate import Accelerator, DistributedType
-from datasets import load_dataset
 from transformers import AutoModelForSequenceClassification, AutoTokenizer, get_linear_schedule_with_warmup, set_seed
 
+from accelerate import Accelerator, DistributedType
+
 
 ########################################################################
 # This is a fully working simple example to use Accelerate,

examples/complete_cv_example.py

@@ -17,15 +17,15 @@ import os
 import re
 
 import numpy as np
+import PIL
 import torch
+from timm import create_model
 from torch.optim.lr_scheduler import OneCycleLR
 from torch.utils.data import DataLoader, Dataset
-
-import PIL
-from accelerate import Accelerator
-from timm import create_model
 from torchvision.transforms import Compose, RandomResizedCrop, Resize, ToTensor
 
+from accelerate import Accelerator
+
 
 ########################################################################
 # This is a fully working simple example to use Accelerate

examples/complete_nlp_example.py

@@ -15,15 +15,15 @@
 import argparse
 import os
 
+import evaluate
 import torch
+from datasets import load_dataset
 from torch.optim import AdamW
 from torch.utils.data import DataLoader
-
-import evaluate
-from accelerate import Accelerator, DistributedType
-from datasets import load_dataset
 from transformers import AutoModelForSequenceClassification, AutoTokenizer, get_linear_schedule_with_warmup, set_seed
 
+from accelerate import Accelerator, DistributedType
+
 
 ########################################################################
 # This is a fully working simple example to use Accelerate

examples/cv_example.py

@@ -17,15 +17,15 @@ import os
 import re
 
 import numpy as np
+import PIL
 import torch
+from timm import create_model
 from torch.optim.lr_scheduler import OneCycleLR
 from torch.utils.data import DataLoader, Dataset
-
-import PIL
-from accelerate import Accelerator
-from timm import create_model
 from torchvision.transforms import Compose, RandomResizedCrop, Resize, ToTensor
 
+from accelerate import Accelerator
+
 
 ########################################################################
 # This is a fully working simple example to use Accelerate

examples/nlp_example.py

@@ -14,15 +14,15 @@
 # limitations under the License.
 import argparse
 
+import evaluate
 import torch
+from datasets import load_dataset
 from torch.optim import AdamW
 from torch.utils.data import DataLoader
-
-import evaluate
-from accelerate import Accelerator, DistributedType
-from datasets import load_dataset
 from transformers import AutoModelForSequenceClassification, AutoTokenizer, get_linear_schedule_with_warmup, set_seed
 
+from accelerate import Accelerator, DistributedType
+
 
 ########################################################################
 # This is a fully working simple example to use Accelerate

pyproject.toml

@@ -1,3 +1,17 @@
 [tool.black]
 line-length = 119
-target-version = ['py36']
+target-version = ['py37']
+
+[tool.ruff]
+# Never enforce `E501` (line length violations).
+ignore = ["E501", "E741", "W605"]
+select = ["E", "F", "I", "W"]
+line-length = 119
+
+# Ignore import violations in all `__init__.py` files.
+[tool.ruff.per-file-ignores]
+"__init__.py" = ["E402", "F401", "F403", "F811"]
+
+[tool.ruff.isort]
+lines-after-imports = 2
+known-first-party = ["accelerate"]

setup.cfg

@@ -4,11 +4,6 @@ ensure_newline_before_comments = True
 force_grid_wrap = 0
 include_trailing_comma = True
 known_first_party = accelerate
-known_third_party =
-    numpy
-    torch
-    torch_xla
-
 line_length = 119
 lines_after_imports = 2
 multi_line_output = 3

setup.py

@@ -16,7 +16,7 @@ from setuptools import setup
 from setuptools import find_packages
 
 extras = {}
-extras["quality"] = ["black ~= 22.0", "isort >= 5.5.4", "flake8 >= 3.8.3", "hf-doc-builder >= 0.3.0"]
+extras["quality"] = ["black ~= 23.1", "ruff >= 0.0.241", "hf-doc-builder >= 0.3.0"]
 extras["docs"] = []
 extras["test_prod"] = ["pytest", "pytest-xdist", "pytest-subtests", "parameterized"]
 extras["test_dev"] = ["datasets", "evaluate", "transformers", "scipy", "scikit-learn", "deepspeed<0.7.0", "tqdm"]

src/accelerate/__init__.py

@@ -1,7 +1,3 @@
-# flake8: noqa
-# There's no way to ignore "F401 '...' imported but unused" warnings in this
-# module, but to preserve other warnings. So, don't check this module at all.
-
 __version__ = "0.17.0.dev0"
 
 from .accelerator import Accelerator

src/accelerate/accelerator.py

@@ -1017,7 +1017,6 @@ class Accelerator:
         return model
 
     def _prepare_deepspeed(self, *args):
-
         deepspeed_plugin = self.state.deepspeed_plugin
 
         if deepspeed_plugin.deepspeed_config["train_micro_batch_size_per_gpu"] == "auto":
@@ -1469,7 +1468,7 @@ class Accelerator:
         >>> accelerator = Accelerator(gradient_accumulation_steps=2)
         >>> dataloader, model, optimizer, scheduler = accelerator.prepare(dataloader, model, optimizer, scheduler)
 
-        >>> for (input, target) in dataloader:
+        >>> for input, target in dataloader:
         ...     optimizer.zero_grad()
         ...     output = model(input)
         ...     loss = loss_func(output, target)
@@ -1504,7 +1503,7 @@ class Accelerator:
         >>> accelerator = Accelerator(gradient_accumulation_steps=2)
         >>> dataloader, model, optimizer, scheduler = accelerator.prepare(dataloader, model, optimizer, scheduler)
 
-        >>> for (input, target) in dataloader:
+        >>> for input, target in dataloader:
         ...     optimizer.zero_grad()
         ...     output = model(input)
         ...     loss = loss_func(output, target)
@@ -1594,7 +1593,7 @@ class Accelerator:
                 else:
                     # Not at the end of the dataloader, no need to adjust the tensors
                     return tensor
-            except:
+            except Exception:
                 # Dataset had no length or raised an error
                 return tensor
         return tensor
@@ -2349,7 +2348,7 @@ class Accelerator:
         >>> accelerator = Accelerator()
         >>> dataloader, model, optimizer, scheduler = accelerator.prepare(dataloader, model, optimizer, scheduler)
 
-        >>> for (input, target) in accelerator.skip_first_batches(dataloader, num_batches=2):
+        >>> for input, target in accelerator.skip_first_batches(dataloader, num_batches=2):
         ...     optimizer.zero_grad()
         ...     output = model(input)
         ...     loss = loss_func(output, target)

src/accelerate/checkpointing.py

@@ -169,7 +169,7 @@ def load_accelerator_state(
         if is_tpu_available():
             xm.set_rng_state(states["xm_seed"])
         logger.info("All random states loaded successfully")
-    except:
+    except Exception:
         logger.info("Could not load random states")
 
 

src/accelerate/commands/config/config_utils.py

@@ -48,7 +48,7 @@ def _ask_field(input_text, convert_value=None, default=None, error_message=None)
             if default is not None and len(result) == 0:
                 return default
             return convert_value(result) if convert_value is not None else result
-        except:
+        except Exception:
             if error_message is not None:
                 print(error_message)
 

src/accelerate/commands/launch.py

@@ -25,9 +25,9 @@ from ast import literal_eval
 from pathlib import Path
 from typing import Dict, List
 
+import psutil
 import torch
 
-import psutil
 from accelerate.commands.config import default_config_file, load_config_from_file
 from accelerate.commands.config.config_args import SageMakerConfig
 from accelerate.commands.config.config_utils import DYNAMO_BACKENDS
@@ -644,7 +644,7 @@ def multi_gpu_launcher(args):
     with patch_environment(**current_env):
         try:
             distrib_run.run(args)
-        except:
+        except Exception:
             if is_rich_available() and debug:
                 console = get_console()
                 console.print("\n[bold red]Using --debug, `torch.distributed` Stack Trace:[/bold red]")
@@ -770,7 +770,7 @@ def deepspeed_launcher(args):
         with patch_environment(**current_env):
             try:
                 distrib_run.run(args)
-            except:
+            except Exception:
                 if is_rich_available() and debug:
                     console = get_console()
                     console.print("\n[bold red]Using --debug, `torch.distributed` Stack Trace:[/bold red]")

src/accelerate/commands/menu/__init__.py

@@ -1,5 +1 @@
-# flake8: noqa
-# There's no way to ignore "F401 '...' imported but unused" warnings in this
-# module, but to preserve other warnings. So, don't check this module at all
-
 from .selection_menu import BulletMenu

src/accelerate/commands/tpu.py

@@ -18,9 +18,10 @@ import argparse
 import os
 import subprocess
 
-from accelerate.commands.config.config_args import default_config_file, load_config_from_file
 from packaging.version import Version, parse
 
+from accelerate.commands.config.config_args import default_config_file, load_config_from_file
+
 
 _description = "Run commands across TPU VMs for initial setup before running `accelerate launch`."
 

src/accelerate/memory_utils.py

@@ -12,11 +12,6 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-# flake8: noqa
-# There's no way to ignore "F401 '...' imported but unused" warnings in this
-# module, but to preserve other warnings. So, don't check this module at all
-
-
 import warnings
 
 
@@ -25,5 +20,3 @@ warnings.warn(
     "`from accelerate import find_executable_batch_size` to avoid this warning.",
     FutureWarning,
 )
-
-from .utils.memory import find_executable_batch_size

src/accelerate/test_utils/__init__.py

@@ -1,7 +1,3 @@
-# flake8: noqa
-# There's no way to ignore "F401 '...' imported but unused" warnings in this
-# module, but to preserve other warnings. So, don't check this module at all.
-
 from .testing import (
     are_the_same_tensors,
     execute_subprocess_async,
@@ -19,4 +15,4 @@ from .testing import (
 from .training import RegressionDataset, RegressionModel
 
 
-from .scripts import test_script, test_sync  # isort:skip
+from .scripts import test_script, test_sync  # isort: skip

src/accelerate/test_utils/scripts/external_deps/test_checkpointing.py

@@ -16,15 +16,15 @@ import argparse
 import json
 import os
 
+import evaluate
 import torch
+from datasets import load_dataset
 from torch.optim import AdamW
 from torch.utils.data import DataLoader
+from transformers import AutoModelForSequenceClassification, AutoTokenizer, get_linear_schedule_with_warmup, set_seed
 
-import evaluate
 from accelerate import Accelerator, DistributedType
 from accelerate.utils.deepspeed import DummyOptim, DummyScheduler
-from datasets import load_dataset
-from transformers import AutoModelForSequenceClassification, AutoTokenizer, get_linear_schedule_with_warmup, set_seed
 
 
 MAX_GPU_BATCH_SIZE = 16

src/accelerate/test_utils/scripts/external_deps/test_metrics.py

@@ -15,17 +15,17 @@
 import math
 from copy import deepcopy
 
-import torch
-from torch.utils.data import DataLoader
-
 import datasets
 import evaluate
+import torch
 import transformers
+from datasets import load_dataset
+from torch.utils.data import DataLoader
+from transformers import AutoModelForSequenceClassification, AutoTokenizer
+
 from accelerate import Accelerator
 from accelerate.test_utils import RegressionDataset, RegressionModel
 from accelerate.utils import is_tpu_available, set_seed
-from datasets import load_dataset
-from transformers import AutoModelForSequenceClassification, AutoTokenizer
 
 
 def get_basic_setup(accelerator, num_samples=82, batch_size=16):
@@ -84,7 +84,7 @@ def generate_predictions(model, dataloader, accelerator):
             logit, target = accelerator.gather_for_metrics((logit, target))
             logits_and_targets.append((logit, target))
     logits, targs = [], []
-    for (logit, targ) in logits_and_targets:
+    for logit, targ in logits_and_targets:
         logits.append(logit)
         targs.append(targ)
     logits, targs = torch.cat(logits), torch.cat(targs)

src/accelerate/test_utils/scripts/external_deps/test_peak_memory_usage.py

@@ -18,13 +18,13 @@ import json
 import os
 
 import torch
+from datasets import load_dataset
 from torch.optim import AdamW
 from torch.utils.data import DataLoader
+from transformers import AutoModelForSequenceClassification, AutoTokenizer, get_linear_schedule_with_warmup, set_seed
 
 from accelerate import Accelerator, DistributedType
 from accelerate.utils.deepspeed import DummyOptim, DummyScheduler
-from datasets import load_dataset
-from transformers import AutoModelForSequenceClassification, AutoTokenizer, get_linear_schedule_with_warmup, set_seed
 
 
 MAX_GPU_BATCH_SIZE = 16

src/accelerate/test_utils/scripts/external_deps/test_performance.py

@@ -16,15 +16,15 @@ import argparse
 import json
 import os
 
+import evaluate
 import torch
+from datasets import load_dataset
 from torch.optim import AdamW
 from torch.utils.data import DataLoader
+from transformers import AutoModelForSequenceClassification, AutoTokenizer, get_linear_schedule_with_warmup, set_seed
 
-import evaluate
 from accelerate import Accelerator, DistributedType
 from accelerate.utils.deepspeed import DummyOptim, DummyScheduler
-from datasets import load_dataset
-from transformers import AutoModelForSequenceClassification, AutoTokenizer, get_linear_schedule_with_warmup, set_seed
 
 
 MAX_GPU_BATCH_SIZE = 16

src/accelerate/test_utils/scripts/test_distributed_data_loop.py

@@ -77,7 +77,6 @@ def verify_dataloader_batch_sizes(
 
 
 def test_default_ensures_even_batch_sizes():
-
     accelerator = create_accelerator()
 
     # without padding, we would expect a different number of batches
@@ -144,7 +143,6 @@ def test_can_join_uneven_inputs():
 
 
 def test_join_raises_warning_for_non_ddp_distributed(accelerator):
-
     with warnings.catch_warnings(record=True) as w:
         with accelerator.join_uneven_inputs([Mock()]):
             pass

src/accelerate/test_utils/testing.py

@@ -338,7 +338,6 @@ async def _stream_subprocess(cmd, env=None, stdin=None, timeout=None, quiet=Fals
 
 
 def execute_subprocess_async(cmd, env=None, stdin=None, timeout=180, quiet=False, echo=True) -> _RunOutput:
-
     loop = asyncio.get_event_loop()
     result = loop.run_until_complete(
         _stream_subprocess(cmd, env=env, stdin=stdin, timeout=timeout, quiet=quiet, echo=echo)

src/accelerate/utils/__init__.py

@@ -1,7 +1,3 @@
-# flake8: noqa
-# There's no way to ignore "F401 '...' imported but unused" warnings in this
-# module, but to preserve other warnings. So, don't check this module at all
-
 from .constants import MODEL_NAME, OPTIMIZER_NAME, RNG_STATE_NAME, SCALER_NAME, SCHEDULER_NAME, TORCH_LAUNCH_PARAMS
 from .dataclasses import (
     ComputeEnvironment,

src/accelerate/utils/deepspeed.py

@@ -40,7 +40,6 @@ class HfDeepSpeedConfig:
     """
 
     def __init__(self, config_file_or_dict):
-
         if isinstance(config_file_or_dict, dict):
             # Don't modify user's data should they want to reuse it (e.g. in tests), because once we
             # modified it, it will not be accepted here again, since `auto` values would have been overridden

src/accelerate/utils/imports.py

@@ -20,7 +20,6 @@ from distutils.util import strtobool
 from functools import lru_cache
 
 import torch
-
 from packaging.version import parse
 
 from .environment import parse_flag_from_env

src/accelerate/utils/megatron_lm.py

@@ -58,9 +58,8 @@ if is_megatron_lm_available():
         set_jit_fusion_options,
         write_args_to_tensorboard,
     )
-    from megatron.model import BertModel
+    from megatron.model import BertModel, Float16Module, GPTModel, ModelType, T5Model
     from megatron.model import DistributedDataParallel as LocalDDP
-    from megatron.model import Float16Module, GPTModel, ModelType, T5Model
     from megatron.model.classification import Classification
     from megatron.optimizer import get_megatron_optimizer
     from megatron.schedules import get_forward_backward_func
@@ -101,7 +100,6 @@ def model_provider_func(pre_process=True, post_process=True, add_encoder=True, a
                 post_process=post_process,
             )
         else:
-
             model = Classification(
                 num_classes=args.num_labels, num_tokentypes=2, pre_process=pre_process, post_process=post_process
             )
@@ -270,7 +268,6 @@ class MegatronLMDummyDataLoader:
 
         # Data loader only on rank 0 of each model parallel group.
         if mpu.get_tensor_model_parallel_rank() == 0:
-
             # Number of train/valid/test samples.
             if args.train_samples:
                 train_samples = args.train_samples

tests/deepspeed/test_deepspeed.py

@@ -22,7 +22,12 @@ from copy import deepcopy
 from pathlib import Path
 
 import torch
+from parameterized import parameterized
 from torch.utils.data import DataLoader
+from transformers import AutoModel, AutoModelForCausalLM, get_scheduler
+from transformers.testing_utils import mockenv_context
+from transformers.trainer_utils import set_seed
+from transformers.utils import is_torch_bf16_available
 
 import accelerate
 from accelerate.accelerator import Accelerator
@@ -47,11 +52,6 @@ from accelerate.utils.deepspeed import (
     DummyScheduler,
 )
 from accelerate.utils.other import patch_environment
-from parameterized import parameterized
-from transformers import AutoModel, AutoModelForCausalLM, get_scheduler
-from transformers.testing_utils import mockenv_context
-from transformers.trainer_utils import set_seed
-from transformers.utils import is_torch_bf16_available
 
 
 set_seed(42)
@@ -133,7 +133,6 @@ class DeepSpeedConfigIntegration(AccelerateTestCase):
 
     @parameterized.expand(stages, name_func=parameterized_custom_name_func)
     def test_deepspeed_plugin(self, stage):
-
         # Test zero3_init_flag will be set to False when ZeRO stage != 3
         deepspeed_plugin = DeepSpeedPlugin(
             gradient_accumulation_steps=1,

tests/fsdp/test_fsdp.py

@@ -17,6 +17,9 @@ import inspect
 import os
 
 import torch
+from transformers import AutoModel
+from transformers.testing_utils import mockenv_context
+from transformers.trainer_utils import set_seed
 
 import accelerate
 from accelerate.accelerator import Accelerator
@@ -38,9 +41,6 @@ from accelerate.utils.constants import (
 )
 from accelerate.utils.dataclasses import FullyShardedDataParallelPlugin
 from accelerate.utils.other import patch_environment
-from transformers import AutoModel
-from transformers.testing_utils import mockenv_context
-from transformers.trainer_utils import set_seed
 
 
 set_seed(42)

tests/test_big_modeling.py

@@ -18,6 +18,7 @@ from tempfile import TemporaryDirectory
 
 import torch
 import torch.nn as nn
+from transformers import AutoModelForCausalLM, AutoTokenizer
 
 from accelerate.big_modeling import (
     cpu_offload,
@@ -31,7 +32,6 @@ from accelerate.big_modeling import (
 from accelerate.hooks import remove_hook_from_submodules
 from accelerate.test_utils import require_cuda, require_mps, require_multi_gpu, require_torch_min_version, slow
 from accelerate.utils import offload_state_dict
-from transformers import AutoModelForCausalLM, AutoTokenizer
 
 
 class ModelForTest(nn.Module):

tests/test_sagemaker.py

@@ -2,6 +2,7 @@ import unittest
 from dataclasses import dataclass
 
 import pytest
+
 from accelerate.commands.config.config_args import SageMakerConfig
 from accelerate.commands.launch import _convert_nargs_to_dict
 from accelerate.utils import ComputeEnvironment