accelerate breaking changes

1.15.0

33 breaking changes, first published September 9, 2026.

• Removed accelerate.utils.launch.parse_flag_from_env
  accelerate.utils.launch.parse_flag_from_env was removed in accelerate 1.15.0 pypi.org
• Behavior changed
  SHARDED_STATE_DICT checkpoints under FSDP2 now use per-rank torch.save and torch.load. Code that reads or writes these checkpoints directly must account for the per-rank files. github.com
• Behavior changed
  Activation checkpointing under FSDP2 now wraps the matched transformer layer itself rather than each child module. Memory use and recomputation change as a result. github.com
• Behavior changed
  Enabling FSDP on a device mesh without a shard dimension now raises a clear error. Setups that previously ran this way must add a shard dimension to the mesh. github.com
• Behavior changed
  Context parallelism now raises an error for models that have sliding-window or chunked attention layers. Such models must run without context parallelism. github.com
• Behavior changed prepare_model
  prepare_model leaves models already sharded with DTensor where they are instead of moving them to the device. Code that relied on the move must place the model itself. github.com
• Signature changed accelerate.accelerator.convert_model_to_fp8_ao
  Was def convert_model_to_fp8_ao(model: torch.nn.Module, config: Optional[Float8LinearConfig] = None, module_filter_func: Optional[Callable] = filter_first_and_last_linear_layers); now def convert_model_to_fp8_ao(model: torch.nn.Module... pypi.org
• Signature changed accelerate.accelerator.DataLoaderDispatcher.__init__
  Was def __init__(self, dataset, split_batches: bool = False, skip_batches=0, use_stateful_dataloader=False, _drop_last: bool = False, _non_blocking: bool = False, slice_fn=None, torch_device_mesh=None, **kwargs); now def __init__(self... pypi.org
• Signature changed accelerate.accelerator.load_accelerator_state
  Was def load_accelerator_state(input_dir, models, optimizers, schedulers, dataloaders, process_index, scaler=None, map_location=None, load_kwargs=None, **load_model_func_kwargs); now def load_accelerator_state(input_dir: str, models: list... pypi.org
• Signature changed accelerate.accelerator.load_custom_state
  Was def load_custom_state(obj, path, index: int = 0); now def load_custom_state(obj, path: str, index: int = 0) -> None. pypi.org

And 23 more in this version.

1.14.0

7 breaking changes, first published June 11, 2026.

• Default changed tracker.log
  The tracker log method's step parameter defaults to None instead of its previous default. Code that relies on the old default step value should pass step explicitly. github.com
• Signature changed accelerate.accelerator.GeneralTracker.log
  Was def log(self, values: dict, step: Optional[int], **kwargs); now def log(self, values: dict, step: Optional[int] = None, **kwargs). pypi.org
• Signature changed accelerate.Accelerator.init_trackers
  Was def init_trackers(self, project_name: str, config: dict | None = None, init_kwargs: dict | None = {}); now def init_trackers(self, project_name: str, config: dict | None = None, init_kwargs: dict | None = None). pypi.org
• Signature changed accelerate.Accelerator.log
  Was def log(self, values: dict, step: int | None = None, log_kwargs: dict | None = {}); now def log(self, values: dict, step: int | None = None, log_kwargs: dict | None = None). pypi.org
• Signature changed accelerate.tracking.MLflowTracker.log
  Was def log(self, values: dict, step: Optional[int]); now def log(self, values: dict, step: Optional[int] = None, **kwargs). pypi.org
• Signature changed accelerate.tracking.TensorBoardTracker.log_images
  Was def log_images(self, values: dict, step: Optional[int], **kwargs); now def log_images(self, values: dict, step: Optional[int] = None, **kwargs). pypi.org
• Signature changed accelerate.utils.fsdp_utils.fsdp2_prepare_auto_wrap_policy
  Was def fsdp2_prepare_auto_wrap_policy(fsdp2_plugin, model: torch.nn.Module) -> Callable[[torch.nn.Module], bool]; now def fsdp2_prepare_auto_wrap_policy(fsdp2_plugin, model: torch.nn.Module) -> Callable[[torch.nn.Module], bool] | None. pypi.org

1.13.0

36 breaking changes, first published March 4, 2026.

• Removed
  Accelerate no longer depends on Intel Extension for PyTorch (IPEX) for XPU support. Code and environments that relied on IPEX being installed may need to change. github.com
• Removed accelerate.accelerator.AcceleratorState.use_ipex
  AcceleratorState.use_ipex was removed in accelerate 1.13.0 pypi.org
• Removed accelerate.accelerator.DynamoBackend.IPEX
  accelerate.accelerator.DynamoBackend.IPEX was removed in accelerate 1.13.0 pypi.org
• Removed accelerate.accelerator.is_ipex_available
  accelerate.accelerator.is_ipex_available was removed in accelerate 1.13.0 pypi.org
• Removed accelerate.accelerator.MegatronEngine.megatron_generate
  MegatronEngine.megatron_generate was removed in accelerate 1.13.0 pypi.org
• Removed accelerate.commands.config.cluster.ClusterConfig.ipex_config
  ClusterConfig.ipex_config was removed in accelerate 1.13.0 pypi.org
• Removed accelerate.commands.config.config_utils.DynamoBackend.IPEX
  accelerate.commands.config.config_utils.DynamoBackend.IPEX was removed in accelerate 1.13.0 pypi.org
• Removed accelerate.state.DynamoBackend.IPEX
  accelerate.state.DynamoBackend.IPEX was removed in accelerate 1.13.0 pypi.org
• Removed accelerate.state.is_ccl_available
  accelerate.state.is_ccl_available was removed in accelerate 1.13.0 pypi.org
• Removed accelerate.state.is_ipex_available
  accelerate.state.is_ipex_available was removed in accelerate 1.13.0 pypi.org

And 26 more in this version.

1.12.0

1 breaking change, first published November 21, 2025.

• Signature changed accelerate.ParallelismConfig.__init__
  Was def __init__(self, dp_replicate_size: Optional[int] = None, dp_shard_size: Optional[int] = None, tp_size: Optional[int] = None, cp_size: Optional[int] = None, tp_handler: Union[None, TorchTensorParallelConfig] = None, cp_handler... pypi.org

1.11.0

87 breaking changes, first published October 20, 2025.

• Removed
  Python 3.9 reached end of life and is no longer supported; users on Python 3.9 must upgrade to Python 3.10 or later. github.com
• Removed accelerate.test_utils.RegressionModel4XPU
  accelerate.test_utils.RegressionModel4XPU was removed in accelerate 1.11.0 pypi.org
• Removed accelerate.test_utils.scripts.external_deps.test_ds_multiple_model.MAX_GPU_BATCH_SIZE
  accelerate.test_utils.scripts.external_deps.test_ds_multiple_model.MAX_GPU_BATCH_SIZE was removed in accelerate 1.11.0 pypi.org
• Removed accelerate.test_utils.scripts.test_script.is_xpu_available
  accelerate.test_utils.scripts.test_script.is_xpu_available was removed in accelerate 1.11.0 pypi.org
• Removed accelerate.utils.modeling.FindTiedParametersResult
  The previously deprecated FindTiedParametersResult was removed; code that imports or uses it will fail. pypi.org
• Renamed accelerate.test_utils.scripts.test_ops.test_gather_non_contigous
  Use accelerate.test_utils.scripts.test_ops.test_gather_non_contiguous instead. pypi.org
• Signature changed accelerate.accelerator.AcceleratorState.__init__
  Was def __init__(self, mixed_precision: str = None, cpu: bool = False, dynamo_plugin=None, deepspeed_plugin=None, fsdp_plugin=None, torch_tp_plugin=None, megatron_lm_plugin=None, parallelism_config=None, _from_accelerator: bool = False... pypi.org
• Signature changed accelerate.accelerator.AcceleratorState.select_deepspeed_plugin
  Was def select_deepspeed_plugin(self, name: str = None); now def select_deepspeed_plugin(self, name: str | None = None). pypi.org
• Signature changed accelerate.accelerator.filter_trackers
  Was def filter_trackers(log_with: list[Union[str, LoggerType, GeneralTracker]], logging_dir: Union[str, os.PathLike] = None); now def filter_trackers(log_with: list[Union[str, LoggerType, GeneralTracker]], logging_dir: Optional[Union[str... pypi.org
• Signature changed accelerate.accelerator.FP8RecipeKwargs.__init__
  Was def __init__(self, opt_level: OptLevel = None, use_autocast_during_eval: bool = None, margin: int = None, interval: int = None, fp8_format: FP8Format = None, amax_history_len: int = None, amax_compute_algo: AmaxComputeAlgorithm = None... pypi.org

And 77 more in this version.

1.10.0

39 breaking changes, first published August 7, 2025.

• Removed accelerate.accelerator.BETA_TP_AVAILABLE_PYTORCH_VERSION
  accelerate.accelerator.BETA_TP_AVAILABLE_PYTORCH_VERSION was removed in accelerate 1.10.0 pypi.org
• Removed accelerate.accelerator.BETA_TP_AVAILABLE_TRANSFORMERS_VERSION
  accelerate.accelerator.BETA_TP_AVAILABLE_TRANSFORMERS_VERSION was removed in accelerate 1.10.0 pypi.org
• Removed accelerate.accelerator.DistributedType.TP
  accelerate.accelerator.DistributedType.TP was removed in accelerate 1.10.0 pypi.org
• Removed accelerate.accelerator.TorchTensorParallelPlugin.__post_init__
  TorchTensorParallelPlugin.__post_init__ was removed in accelerate 1.10.0 pypi.org
• Removed accelerate.commands.config.cluster.DistributedType.TP
  accelerate.commands.config.cluster.DistributedType.TP was removed in accelerate 1.10.0 pypi.org
• Removed accelerate.commands.config.config_args.DistributedType.TP
  accelerate.commands.config.config_args.DistributedType.TP was removed in accelerate 1.10.0 pypi.org
• Removed accelerate.commands.config.config_utils.DistributedType.TP
  accelerate.commands.config.config_utils.DistributedType.TP was removed in accelerate 1.10.0 pypi.org
• Removed accelerate.commands.launch.DistributedType.TP
  accelerate.commands.launch.DistributedType.TP was removed in accelerate 1.10.0 pypi.org
• Removed accelerate.data_loader.DistributedType.TP
  accelerate.data_loader.DistributedType.TP was removed in accelerate 1.10.0 pypi.org
• Removed accelerate.DistributedType.TP
  accelerate.DistributedType.TP was removed in accelerate 1.10.0 pypi.org

And 29 more in this version.

Check your code

These are the changes. To see which of them reach your code, scan a public repository, or have your coding agent call check_code with your manifest or lockfile over MCP. The API reads every change in every version, not only the newest. On Team and up, a fix pull request opens for the changes that reach a connected repository. Nothing merges or deploys.