# What changed in accelerate

> 69 changes in accelerate 1 from its official release notes, upgrade guides and advisories: 5 breaking, 1 deprecations, 6 new APIs and options. Each one cited. Sources checked 2026-10-24. HTML: https://unstale.sh/pypi/accelerate/

To see which of these apply to a repository, run `npx @unstale/cli init` in it.

## accelerate 1

### Breaking changes

- Refuses context parallelism for models with sliding-window or chunked attention layers. (1.15.0) — [release notes](https://github.com/huggingface/accelerate/releases/tag/v1.15.0)
- FSDP now raises a clear error when enabled on a mesh without a shard dimension. (1.15.0) — [release notes](https://github.com/huggingface/accelerate/releases/tag/v1.15.0)
- Leaves an already-DTensor-sharded model on its existing device instead of moving it. (1.15.0) — [release notes](https://github.com/huggingface/accelerate/releases/tag/v1.15.0)
- Removes the 8bit force hook for bnb. (1.13.0) — [release notes](https://github.com/huggingface/accelerate/releases/tag/v1.13.0)
- XPU device handling uses spawn instead of fork, removes the IPEX dependency, and improves device-agnostic code. (1.13.0) — [release notes](https://github.com/huggingface/accelerate/releases/tag/v1.13.0)

### Deprecations

- Adds deprecation warnings for MS-AMP. (1.13.0) — [release notes](https://github.com/huggingface/accelerate/releases/tag/v1.13.0)

### New APIs and options

- Adds an FSDP2-only option to move checkpointed layer inputs to pinned CPU memory, reducing activation memory while preserving plain activation-checkpointing gradients. (1.15.0) — [release notes](https://github.com/huggingface/accelerate/releases/tag/v1.15.0)
- Supports dispatching an entire model to disk, including tied weights, for offloading large models. (1.15.0) — [release notes](https://github.com/huggingface/accelerate/releases/tag/v1.15.0)
- Adds a Neuron backend for torch dynamo, enabling --torch-compile with the Transformers Trainer. (1.15.0) — [release notes](https://github.com/huggingface/accelerate/releases/tag/v1.15.0)
- Registers custom trackers by name so they can be selected through Accelerator's log_with argument. (1.15.0) — [release notes](https://github.com/huggingface/accelerate/releases/tag/v1.15.0)
- Offloading support is added for Torchao. (1.14.0) — [release notes](https://github.com/huggingface/accelerate/releases/tag/v1.14.0)
- Adds support for AWS Neuron Trainium and Inferentia devices. (1.13.0) — [release notes](https://github.com/huggingface/accelerate/releases/tag/v1.13.0)

### Performance

- FSDP2 activation checkpointing now wraps the matched transformer layer itself rather than each child, so inter-child activations are not all kept for backward. (1.15.0) — [release notes](https://github.com/huggingface/accelerate/releases/tag/v1.15.0)
- Padded allgather and broadcast reduce recompilation on Neuron devices. (1.14.0) — [release notes](https://github.com/huggingface/accelerate/releases/tag/v1.14.0)
- Accelerate imports faster by deferring heavy dependencies. (1.13.0) — [release notes](https://github.com/huggingface/accelerate/releases/tag/v1.13.0)

### Behavior and default changes

- Accelerate now works end-to-end on AMD ROCm devices. (1.14.0) — [release notes](https://github.com/huggingface/accelerate/releases/tag/v1.14.0)
- Supports dynamic batch sizes when even_batches is enabled. (1.14.0) — [release notes](https://github.com/huggingface/accelerate/releases/tag/v1.14.0)
- FSDP2 now shards embeddings and norms. (1.14.0) — [release notes](https://github.com/huggingface/accelerate/releases/tag/v1.14.0)
- The step parameter defaults to None. (1.14.0) — [release notes](https://github.com/huggingface/accelerate/releases/tag/v1.14.0)
- Batches can contain non-Tensor values when dispatch_batches=True. (1.13.0) — [release notes](https://github.com/huggingface/accelerate/releases/tag/v1.13.0)
- Changes the default value of fsdp_min_num_params to an integer. (1.13.0) — [release notes](https://github.com/huggingface/accelerate/releases/tag/v1.13.0)
- Disables compile hooks lazily. (1.13.0) — [release notes](https://github.com/huggingface/accelerate/releases/tag/v1.13.0)

### Fixes and other changes

- Reports MPS as the accelerator and reads the MPS chip name. (1.15.0) — [release notes](https://github.com/huggingface/accelerate/releases/tag/v1.15.0)
- Handles a missing accelerate executable in the env command. (1.15.0) — [release notes](https://github.com/huggingface/accelerate/releases/tag/v1.15.0)
- Fixes the --cpu option so it sets the KMP variables. (1.15.0) — [release notes](https://github.com/huggingface/accelerate/releases/tag/v1.15.0)
- Reaches through the DeepSpeed optimizer wrapper when calling eval(). (1.15.0) — [release notes](https://github.com/huggingface/accelerate/releases/tag/v1.15.0)
- Fixes compile_regions so it does not run the uncompiled module. (1.15.0) — [release notes](https://github.com/huggingface/accelerate/releases/tag/v1.15.0)
- Fixes conversion so the first and last linear layers are not converted. (1.15.0) — [release notes](https://github.com/huggingface/accelerate/releases/tag/v1.15.0)
- Preserves dataloader iteration state when resuming from a checkpoint. (1.15.0) — [release notes](https://github.com/huggingface/accelerate/releases/tag/v1.15.0)
- Fixes a crash when disk offloading FP8 tensors. (1.15.0) — [release notes](https://github.com/huggingface/accelerate/releases/tag/v1.15.0)
- Fixes the command for timm >= 1.0.29 by adding the hf-hub: prefix. (1.15.0) — [release notes](https://github.com/huggingface/accelerate/releases/tag/v1.15.0)
- Treats MPS out-of-memory errors as OOM when finding an executable batch size. (1.15.0) — [release notes](https://github.com/huggingface/accelerate/releases/tag/v1.15.0)
- Forwards reduce_batch_size_fn through the decorator form of find_executable_batch_size. (1.15.0) — [release notes](https://github.com/huggingface/accelerate/releases/tag/v1.15.0)
- Fixes path resolution for FSDP sharded checkpoints. (1.15.0) — [release notes](https://github.com/huggingface/accelerate/releases/tag/v1.15.0)
- Fixes tied-embedding models on torch >= 2.13 by placing the output embedding in the same fully_shard group. (1.15.0) — [release notes](https://github.com/huggingface/accelerate/releases/tag/v1.15.0)
- Fixes FSDP2/PEFT checkpointing that dropped every rank's adapter shard except rank 0. (1.15.0) — [release notes](https://github.com/huggingface/accelerate/releases/tag/v1.15.0)
- Stops the function from mutating module._non_persistent_buffers_set. (1.15.0) — [release notes](https://github.com/huggingface/accelerate/releases/tag/v1.15.0)
- Gradient clipping now supports DTensor-sharded models when plain tensors and DTensors are mixed, avoiding a foreach-op failure. (1.15.0) — [release notes](https://github.com/huggingface/accelerate/releases/tag/v1.15.0)
- Fixes memory accounting for integrated CUDA devices. (1.15.0) — [release notes](https://github.com/huggingface/accelerate/releases/tag/v1.15.0)
- Restores all RNG backends rather than only one. (1.15.0) — [release notes](https://github.com/huggingface/accelerate/releases/tag/v1.15.0)
- Fixes the setting so MLflow run nesting can be turned off. (1.15.0) — [release notes](https://github.com/huggingface/accelerate/releases/tag/v1.15.0)
- Uses per-rank torch.save/load for FSDP2 sharded state dictionaries to fix checkpoint timeouts on large NPU setups. (1.15.0) — [release notes](https://github.com/huggingface/accelerate/releases/tag/v1.15.0)
- Corrects byte-size handling for FP8 fnuz and e8m0fnu dtypes. (1.14.0) — [release notes](https://github.com/huggingface/accelerate/releases/tag/v1.14.0)
- FSDP2 casts models to a uniform dtype before sharding and excludes non-floating frozen Params4bit parameters to prevent QLoRA crashes. (1.14.0) — [release notes](https://github.com/huggingface/accelerate/releases/tag/v1.14.0)
- Fixes full-state-dict loading, including dtype mismatches and key-based matching, and adds a guard for Params4bit models. (1.14.0) — [release notes](https://github.com/huggingface/accelerate/releases/tag/v1.14.0)
- The FSDP2 auto-wrap policy no longer ignores the _no_split_modules fallback. (1.14.0) — [release notes](https://github.com/huggingface/accelerate/releases/tag/v1.14.0)
- Guards against None kwargs in tracker initialization and logging. (1.14.0) — [release notes](https://github.com/huggingface/accelerate/releases/tag/v1.14.0)
- Fixes hook detachment statistics restoration for int8 offload. (1.14.0) — [release notes](https://github.com/huggingface/accelerate/releases/tag/v1.14.0)
- Fixes the sharding condition when the number of shards equals the number of processes. (1.14.0) — [release notes](https://github.com/huggingface/accelerate/releases/tag/v1.14.0)
- Fixes keep_in_fp32_modules handling for tied weights. (1.14.0) — [release notes](https://github.com/huggingface/accelerate/releases/tag/v1.14.0)
- Prevents models from being wrapped twice. (1.14.0) — [release notes](https://github.com/huggingface/accelerate/releases/tag/v1.14.0)
- Avoids implicit padding when apply_padding is false and num_samples is smaller than num_processes. (1.14.0) — [release notes](https://github.com/huggingface/accelerate/releases/tag/v1.14.0)
- Drops REMOVED and NOT_YET_IMPLEMENTED FSDP1 keys instead of leaking them during conversion. (1.14.0) — [release notes](https://github.com/huggingface/accelerate/releases/tag/v1.14.0)
- Passes the step parameter through instead of ignoring it. (1.14.0) — [release notes](https://github.com/huggingface/accelerate/releases/tag/v1.14.0)
- Fixes async compatibility across Python versions. (1.13.0) — [release notes](https://github.com/huggingface/accelerate/releases/tag/v1.13.0)
- Fixes the CPU offload check and a bug when cpu_ram_efficient_loading and cpu_offload are both enabled. (1.13.0) — [release notes](https://github.com/huggingface/accelerate/releases/tag/v1.13.0)
- Improves DeepSpeed sequence-parallel integration, including evaluation during training and process-group handling; errors when context and sequence parallelism are both enabled. (1.13.0) — [release notes](https://github.com/huggingface/accelerate/releases/tag/v1.13.0)
- Fixes a KeyError when extracting a model with partial torch.compile. (1.13.0) — [release notes](https://github.com/huggingface/accelerate/releases/tag/v1.13.0)
- Fixes torchao default configuration with padding and FSDP2 all-gather support, and execution with Transformer Engine. (1.13.0) — [release notes](https://github.com/huggingface/accelerate/releases/tag/v1.13.0)
- Fixes parameter upcasting, tied embedding errors, DCP optimizer-state loading, bfloat16 optimizer-step crashes, and ignored_params crashes with torch versions below 2.7.0. (1.13.0) — [release notes](https://github.com/huggingface/accelerate/releases/tag/v1.13.0)
- Fixes setting and loading the HPU RNG state. (1.13.0) — [release notes](https://github.com/huggingface/accelerate/releases/tag/v1.13.0)
- Fixes logging logic when in_order is set to True. (1.13.0) — [release notes](https://github.com/huggingface/accelerate/releases/tag/v1.13.0)
- Fixes KMP_AFFINITY being incorrectly set for non-CPU training. (1.13.0) — [release notes](https://github.com/huggingface/accelerate/releases/tag/v1.13.0)
- Fixes a mutable default in Megatron initialization and an IndexError on an empty ModuleList. (1.13.0) — [release notes](https://github.com/huggingface/accelerate/releases/tag/v1.13.0)
- Avoids using nvidia-smi on CPU-only Colab instances. (1.13.0) — [release notes](https://github.com/huggingface/accelerate/releases/tag/v1.13.0)
- Fixes an error when parallelism_config is None. (1.13.0) — [release notes](https://github.com/huggingface/accelerate/releases/tag/v1.13.0)
- Fixes device-index comparison for hf_device_map. (1.13.0) — [release notes](https://github.com/huggingface/accelerate/releases/tag/v1.13.0)
- Fixes a KeyError with num_workers greater than zero and fixes stateful dataloader behavior with DDP. (1.13.0) — [release notes](https://github.com/huggingface/accelerate/releases/tag/v1.13.0)
- Avoids duplicate W&B initialization in offline mode. (1.13.0) — [release notes](https://github.com/huggingface/accelerate/releases/tag/v1.13.0)