---
title: DeepSpeedStrategy
description: >-
  Provides capabilities to run training using the DeepSpeed library, with
  training optimizations for large

  billion parameter models. For more information: DeepSpeed.
canonical: >-
  https://lightning.ai/docs/pytorch/stable/optional-api/api/lightning.pytorch.strategies.DeepSpeedStrategy
---
# DeepSpeedStrategy {#deepspeedstrategy}

### `class lightning.pytorch.strategies.DeepSpeedStrategy(accelerator=None, zero_optimization=True, stage=2, remote_device=None, offload_optimizer=False, offload_parameters=False, offload_params_device='cpu', nvme_path='/local_nvme', params_buffer_count=5, params_buffer_size=100000000, max_in_cpu=1000000000, offload_optimizer_device='cpu', optimizer_buffer_count=4, block_size=1048576, queue_depth=8, single_submit=False, overlap_events=True, thread_count=1, pin_memory=False, sub_group_size=1000000000000, contiguous_gradients=True, overlap_comm=True, allgather_partitions=True, reduce_scatter=True, allgather_bucket_size=200000000, reduce_bucket_size=200000000, zero_allow_untested_optimizer=True, logging_batch_size_per_gpu='auto', config=None, logging_level=30, parallel_devices=None, cluster_environment=None, loss_scale=0, initial_scale_power=16, loss_scale_window=1000, hysteresis=2, min_loss_scale=1, partition_activations=False, cpu_checkpointing=False, contiguous_memory_optimization=False, synchronize_checkpoint_boundary=False, load_full_weights=False, precision_plugin=None, process_group_backend=None, timeout=datetime.timedelta(seconds=1800), exclude_frozen_parameters=False)` {#lightning.pytorch.strategies.DeepSpeedStrategy}

Bases: [`DDPStrategy`](/docs/pytorch/2.6.6/optional-api/api/lightning.pytorch.strategies.DDPStrategy.md#lightning.pytorch.strategies.DDPStrategy "lightning.pytorch.strategies.ddp.DDPStrategy")

Provides capabilities to run training using the DeepSpeed library, with training optimizations for large billion parameter models. _For more information:_ [DeepSpeed](/docs/pytorch/2.6.6/reference/advanced/model_parallel/deepspeed.md#deepspeed-advanced).

Warning

This is an [experimental](/docs/pytorch/2.6.6/reference/versioning.md#experimental-api) feature.

Defaults have been set to enable ZeRO-Offload and some have been taken from the link below. These defaults have been set generally, but may require tuning for optimum performance based on your model size. _For more information:_ [https://www.deepspeed.ai/docs/config-json/#zero-optimizations-for-fp16-training](https://www.deepspeed.ai/docs/config-json/#zero-optimizations-for-fp16-training).

Parameters:

*   **zero\_optimization** ([`bool`](https://docs.python.org/3/builtins/functions.html#bool "(in Python v3.15)")) – Enable ZeRO optimization. This is compatible with either precision=”16-mixed” or precision=”bf16-mixed”.
    
*   **stage** ([`int`](https://docs.python.org/3/builtins/functions.html#int "(in Python v3.15)")) – Different stages of the ZeRO Optimizer. 0 is disabled, 1 is optimizer state partitioning, 2 is optimizer+gradient state partitioning, 3 is optimizer+gradient\_parameter partitioning using the infinity engine.
    
*   **remote\_device** ([`Optional`](https://docs.python.org/3/library/typing.html#typing.Optional "(in Python v3.15)")\[[`str`](https://docs.python.org/3/builtins/stdtypes.html#str "(in Python v3.15)")\]) – Device to instantiate the model on initially (`cpu` or `nvme`). Defaults to GPU.
    
*   **offload\_optimizer** ([`bool`](https://docs.python.org/3/builtins/functions.html#bool "(in Python v3.15)")) – Enable offloading optimizer memory and computation to CPU or NVMe based on `offload_optimizer_device`.
    
*   **offload\_parameters** ([`bool`](https://docs.python.org/3/builtins/functions.html#bool "(in Python v3.15)")) – When using ZeRO Stage 3, Enable offloading parameter memory and computation to CPU or NVMe based on `offload_params_device`.
    
*   **offload\_params\_device** ([`str`](https://docs.python.org/3/builtins/stdtypes.html#str "(in Python v3.15)")) – When offloading parameters choose the device to offload to, `cpu` or `nvme`.
    
*   **offload\_optimizer\_device** ([`str`](https://docs.python.org/3/builtins/stdtypes.html#str "(in Python v3.15)")) – When offloading optimizer state choose the device to offload to, `cpu` or `nvme`.
    
*   **params\_buffer\_count** ([`int`](https://docs.python.org/3/builtins/functions.html#int "(in Python v3.15)")) – Number of buffers in buffer pool for parameter offloading when `offload_params_device` is `nvme`.
    
*   **params\_buffer\_size** ([`int`](https://docs.python.org/3/builtins/functions.html#int "(in Python v3.15)")) – Size of buffers in buffer pool for parameter offloading when `offload_params_device` is `nvme`.
    
*   **max\_in\_cpu** ([`int`](https://docs.python.org/3/builtins/functions.html#int "(in Python v3.15)")) – Number of parameter elements to maintain in CPU memory when offloading to NVMe is enabled.
    
*   **nvme\_path** ([`str`](https://docs.python.org/3/builtins/stdtypes.html#str "(in Python v3.15)")) – Filesystem path for NVMe device for optimizer/parameter state offloading.
    
*   **optimizer\_buffer\_count** ([`int`](https://docs.python.org/3/builtins/functions.html#int "(in Python v3.15)")) – Number of buffers in buffer pool for optimizer state offloading when `offload_optimizer_device` is set to `nvme`. This should be at least the number of states maintained per parameter by the optimizer. For example, Adam optimizer has 4 states (parameter, gradient, momentum, and variance).
    
*   **block\_size** ([`int`](https://docs.python.org/3/builtins/functions.html#int "(in Python v3.15)")) – When using NVMe Offloading, the I/O block size in bytes.
    
*   **queue\_depth** ([`int`](https://docs.python.org/3/builtins/functions.html#int "(in Python v3.15)")) – When using NVMe Offloading, the I/O queue depth.
    
*   **single\_submit** ([`bool`](https://docs.python.org/3/builtins/functions.html#bool "(in Python v3.15)")) – When using NVMe Offloading, submit requests to storage device as multiple individual requests, as opposed to one block of requests.
    
*   **overlap\_events** ([`bool`](https://docs.python.org/3/builtins/functions.html#bool "(in Python v3.15)")) – When using NVMe Offloading, submit requests to storage device in an overlapped fashion without waiting for completion of earlier requests.
    
*   **thread\_count** ([`int`](https://docs.python.org/3/builtins/functions.html#int "(in Python v3.15)")) – When using NVMe Offloading, Intra-request parallelism for each read/write submitted by a user thread.
    
*   **pin\_memory** ([`bool`](https://docs.python.org/3/builtins/functions.html#bool "(in Python v3.15)")) – When using ZeRO stage 3, pin optimizer state memory on CPU. This could boost throughput at the cost of extra memory overhead.
    
*   **sub\_group\_size** ([`int`](https://docs.python.org/3/builtins/functions.html#int "(in Python v3.15)")) – When using ZeRO stage 3, defines the number of parameters within a sub group to offload at a time. Smaller numbers require more communication, but improve memory efficiency.
    
*   **contiguous\_gradients** ([`bool`](https://docs.python.org/3/builtins/functions.html#bool "(in Python v3.15)")) – Copies gradients to a continuous buffer as they are produced. Avoids memory fragmentation during backwards. Useful when training large models.
    
*   **overlap\_comm** ([`bool`](https://docs.python.org/3/builtins/functions.html#bool "(in Python v3.15)")) – Overlap the reduction (synchronization) of gradients with the backwards computation. This is a speed optimization when training across multiple GPUs/machines.
    
*   **allgather\_partitions** ([`bool`](https://docs.python.org/3/builtins/functions.html#bool "(in Python v3.15)")) – All gather updated parameters at the end of training step, instead of using a series of broadcast collectives.
    
*   **reduce\_scatter** ([`bool`](https://docs.python.org/3/builtins/functions.html#bool "(in Python v3.15)")) – Use reduce/scatter instead of allreduce to average gradients.
    
*   **allgather\_bucket\_size** ([`int`](https://docs.python.org/3/builtins/functions.html#int "(in Python v3.15)")) – Number of elements to allgather at once. Used to limit the memory required for larger model sizes, with a tradeoff with speed.
    
*   **reduce\_bucket\_size** ([`int`](https://docs.python.org/3/builtins/functions.html#int "(in Python v3.15)")) – Number of elements to reduce at once. Used to limit the memory required for larger model sizes, with a tradeoff with speed.
    
*   **zero\_allow\_untested\_optimizer** ([`bool`](https://docs.python.org/3/builtins/functions.html#bool "(in Python v3.15)")) – Allow untested optimizers to be used with ZeRO. Currently only Adam is a DeepSpeed supported optimizer when using ZeRO.
    
*   **logging\_batch\_size\_per\_gpu** (`Union`\[[`str`](https://docs.python.org/3/builtins/stdtypes.html#str "(in Python v3.15)"), [`int`](https://docs.python.org/3/builtins/functions.html#int "(in Python v3.15)")\]) – Config used in DeepSpeed to calculate verbose timing for logging on a per sample per second basis (only displayed if logging=logging.INFO). If set to “auto”, the strategy tries to infer this from the train DataLoader’s BatchSampler, else defaults to 1. To obtain accurate logs when using datasets that do not support batch samplers, set this to the actual per gpu batch size (trainer.batch\_size).
    
*   **config** (`Union`\[[`str`](https://docs.python.org/3/builtins/stdtypes.html#str "(in Python v3.15)"), [`Path`](https://docs.python.org/3/library/pathlib.html#pathlib.Path "(in Python v3.15)"), [`dict`](https://docs.python.org/3/builtins/stdtypes.html#dict "(in Python v3.15)")\[[`str`](https://docs.python.org/3/builtins/stdtypes.html#str "(in Python v3.15)"), [`Any`](https://docs.python.org/3/library/typing.html#typing.Any "(in Python v3.15)")\], [`None`](https://docs.python.org/3/builtins/constants.html#None "(in Python v3.15)")\]) – Pass in a deepspeed formatted config dict, or path to a deepspeed config: [https://www.deepspeed.ai/docs/config-json](https://www.deepspeed.ai/docs/config-json). All defaults will be ignored if a config is passed in.
    
*   **logging\_level** ([`int`](https://docs.python.org/3/builtins/functions.html#int "(in Python v3.15)")) – Set logging level for deepspeed.
    
*   **loss\_scale** ([`float`](https://docs.python.org/3/builtins/functions.html#float "(in Python v3.15)")) – Loss scaling value for FP16 training. 0.0 results in dynamic loss scaling, otherwise static.
    
*   **initial\_scale\_power** ([`int`](https://docs.python.org/3/builtins/functions.html#int "(in Python v3.15)")) – Power of the initial dynamic loss scale value. Loss scale is computed by `2^initial_scale_power`.
    
*   **loss\_scale\_window** ([`int`](https://docs.python.org/3/builtins/functions.html#int "(in Python v3.15)")) – Window in which to raise/lower the dynamic FP16 loss scaling value.
    
*   **hysteresis** ([`int`](https://docs.python.org/3/builtins/functions.html#int "(in Python v3.15)")) – FP16 Delay shift in Dynamic Loss scaling.
    
*   **min\_loss\_scale** ([`int`](https://docs.python.org/3/builtins/functions.html#int "(in Python v3.15)")) – The minimum FP16 dynamic loss scaling value.
    
*   **partition\_activations** ([`bool`](https://docs.python.org/3/builtins/functions.html#bool "(in Python v3.15)")) – Enables partition activation when used with ZeRO stage 3 and model parallelism. Still requires you to wrap your forward functions in deepspeed.checkpointing.checkpoint. See [deepspeed tutorial](https://www.deepspeed.ai/tutorials/megatron/#deepspeed-activation-checkpoints-optional).
    
*   **cpu\_checkpointing** ([`bool`](https://docs.python.org/3/builtins/functions.html#bool "(in Python v3.15)")) – Offloads partitioned activations to CPU if `partition_activations` is enabled.
    
*   **contiguous\_memory\_optimization** ([`bool`](https://docs.python.org/3/builtins/functions.html#bool "(in Python v3.15)")) – Copies partitioned activations so that they are contiguous in memory. Not supported by all models.
    
*   **synchronize\_checkpoint\_boundary** ([`bool`](https://docs.python.org/3/builtins/functions.html#bool "(in Python v3.15)")) – Insert [`torch.cuda.synchronize()`](https://docs.pytorch.org/docs/stable/generated/torch.cuda.synchronize.html#torch.cuda.synchronize "(in PyTorch v2.15)") at each checkpoint boundary.
    
*   **load\_full\_weights** ([`bool`](https://docs.python.org/3/builtins/functions.html#bool "(in Python v3.15)")) – True when loading a single checkpoint file containing the model state dict when using ZeRO Stage 3. This differs from the DeepSpeed checkpoint which contains shards per worker.
    
*   **exclude\_frozen\_parameters** ([`bool`](https://docs.python.org/3/builtins/functions.html#bool "(in Python v3.15)")) – Exclude frozen parameters when saving checkpoints.
    

#### `model_sharded_context()` {#lightning.pytorch.strategies.DeepSpeedStrategy.model_sharded_context}

Provide hook to create modules in a distributed aware context. This is useful for when we’d like to shard the model instantly, which is useful for extremely large models which can save memory and initialization time.

Returns: Model parallel context.

Return type:

[`Generator`](https://docs.python.org/3/library/collections.abc.html#collections.abc.Generator "(in Python v3.15)")\[[`None`](https://docs.python.org/3/builtins/constants.html#None "(in Python v3.15)"), [`None`](https://docs.python.org/3/builtins/constants.html#None "(in Python v3.15)"), [`None`](https://docs.python.org/3/builtins/constants.html#None "(in Python v3.15)")\]

#### `save_checkpoint(checkpoint, filepath, storage_options=None)` {#lightning.pytorch.strategies.DeepSpeedStrategy.save_checkpoint}

Save model/training states as a checkpoint file through state-dump and file-write.

Parameters:

*   **checkpoint** ([`dict`](https://docs.python.org/3/builtins/stdtypes.html#dict "(in Python v3.15)")) – The checkpoint state dictionary
    
*   **filepath** (`Union`\[[`str`](https://docs.python.org/3/builtins/stdtypes.html#str "(in Python v3.15)"), [`Path`](https://docs.python.org/3/library/pathlib.html#pathlib.Path "(in Python v3.15)")\]) – write-target file’s path
    
*   **storage\_options** ([`Optional`](https://docs.python.org/3/library/typing.html#typing.Optional "(in Python v3.15)")\[[`Any`](https://docs.python.org/3/library/typing.html#typing.Any "(in Python v3.15)")\]) – not used for `DeepSpeedStrategy` as `CheckpointIO` is not used
    

Raises:

[**TypeError**](https://docs.python.org/3/builtins/exceptions.html#TypeError "(in Python v3.15)") – If `storage_options` arg is passed in

Return type:

[`None`](https://docs.python.org/3/builtins/constants.html#None "(in Python v3.15)")

#### `setup(trainer)` {#lightning.pytorch.strategies.DeepSpeedStrategy.setup}

Sets up the accelerator, plugins and initializes the optimizers (if needed).

Parameters:

**trainer** ([`Trainer`](/docs/pytorch/2.6.6/optional-api/api/lightning.pytorch.trainer.trainer.Trainer.md#lightning.pytorch.trainer.trainer.Trainer "lightning.pytorch.trainer.trainer.Trainer")) – the trainer instance

Return type:

[`None`](https://docs.python.org/3/builtins/constants.html#None "(in Python v3.15)")

#### `setup_environment()` {#lightning.pytorch.strategies.DeepSpeedStrategy.setup_environment}

Setup any processes or distributed connections.

This is called before the LightningModule/DataModule setup hook which allows the user to access the accelerator environment before setup is complete.

Return type:

[`None`](https://docs.python.org/3/builtins/constants.html#None "(in Python v3.15)")

#### `setup_optimizers(trainer)` {#lightning.pytorch.strategies.DeepSpeedStrategy.setup_optimizers}

Creates optimizers and schedulers.

Parameters:

**trainer** ([`Trainer`](/docs/pytorch/2.6.6/optional-api/api/lightning.pytorch.trainer.trainer.Trainer.md#lightning.pytorch.trainer.trainer.Trainer "lightning.pytorch.trainer.trainer.Trainer")) – the Trainer, these optimizers should be connected to

Return type:

[`None`](https://docs.python.org/3/builtins/constants.html#None "(in Python v3.15)")

#### `tensor_init_context(empty_init=None)` {#lightning.pytorch.strategies.DeepSpeedStrategy.tensor_init_context}

Controls how tensors get created (device, dtype).

Parameters:

**empty\_init** ([`Optional`](https://docs.python.org/3/library/typing.html#typing.Optional "(in Python v3.15)")\[[`bool`](https://docs.python.org/3/builtins/functions.html#bool "(in Python v3.15)")\]) – Whether to initialize the model with empty weights (uninitialized memory). If `None`, the strategy will decide. Some strategies may not support all options.

Return type:

[`Generator`](https://docs.python.org/3/library/collections.abc.html#collections.abc.Generator "(in Python v3.15)")\[[`None`](https://docs.python.org/3/builtins/constants.html#None "(in Python v3.15)"), [`None`](https://docs.python.org/3/builtins/constants.html#None "(in Python v3.15)"), [`None`](https://docs.python.org/3/builtins/constants.html#None "(in Python v3.15)")\]

#### `property handles_gradient_accumulation: bool` {#lightning.pytorch.strategies.DeepSpeedStrategy.handles_gradient_accumulation}

Whether the strategy handles gradient accumulation internally.

#### `property lightning_restore_optimizer: bool` {#lightning.pytorch.strategies.DeepSpeedStrategy.lightning_restore_optimizer}

Override to disable Lightning restoring optimizers/schedulers.

This is useful for strategies which manage restoring optimizers/schedulers.

#### `property restore_checkpoint_after_setup: bool` {#lightning.pytorch.strategies.DeepSpeedStrategy.restore_checkpoint_after_setup}

Override to delay restoring from checkpoint till after the setup phase has completed. This is useful when the strategy requires all the setup hooks to run before loading checkpoint.

Returns:

If `True`, restore checkpoint after strategy setup.

---

**Navigation**

- Parent: [Api](/docs/pytorch/2.6.6/optional-api/api.md)