2021-04-19 09:19:12 -07:00
ZeRO
####
2021-03-08 12:54:54 -08:00
The Zero Redundancy Optimizer (ZeRO) removes the memory redundancies across
data-parallel processes by partitioning the three model states (optimizer
states, gradients, and parameters) across data-parallel processes instead of
replicating them. By doing this, it boosts memory efficiency compared to
classic data-parallelism while retaining its computational granularity and
communication efficiency.
2021-04-19 09:19:12 -07:00
#. **ZeRO Stage 1** : The optimizer states (e.g., for `Adam optimizer <https://arxiv.org/abs/1412.6980> `_ , 32-bit weights, and the first, and second moment estimates) are partitioned across the processes, so that each process updates only its partition.
#. **ZeRO Stage 2** : The reduced 32-bit gradients for updating the model weights are also partitioned such that each process retains only the gradients corresponding to its portion of the optimizer states.
#. **ZeRO Stage 3** : The 16-bit model parameters are partitioned across the processes. ZeRO-3 will automatically collect and partition them during the forward and backward passes.
In addition, ZeRO-3 includes the *infinity offload engine* to form
ZeRO-Infinity ([paper](https://arxiv.org/abs/2104.07857)), which can offload
all model states to both CPU and NVMe memory for huge memory savings.
For a deep dive of our algorithms, please see our `papers <https://www.deepspeed.ai/#publications> `_ on `ZeRO
<https://arxiv.org/abs/1910.02054> `_, `
<https://arxiv.org/abs/2101.06840> `_,
and `ZeRO-Infinity <https://arxiv.org/abs/2104.07857> `_ .
.. note ::
DeepSpeed first included offloading capabilities with **ZeRO-Offload** , a
system for offloading optimizer and gradient states to CPU memory within
ZeRO-2. **ZeRO-Infinity** is the next generation of offloading
capabilities, accessible to ZeRO-3. ZeRO-Infinity has all of the savings
of ZeRO-Offload, plus is able to offload more the model weights and has
more effective bandwidth utilization and overlapping of computation and
communication.
2021-03-08 12:54:54 -08:00
2021-04-18 23:45:37 -07:00
2021-03-08 12:54:54 -08:00
Getting Started
---------------
If you are new to DeepSpeed, check out our `Getting Started <https://www.deepspeed.ai/getting-started/> `_ page.
2021-04-18 23:45:37 -07:00
Once you are training with DeepSpeed, enabling ZeRO-3 offload is as simple as enabling it
2021-03-08 12:54:54 -08:00
in your DeepSpeed configuration! Below are a few examples of ZeRO-3 configurations. Please see
our `config guide <https://www.deepspeed.ai/docs/config-json/#zero-optimizations-for-fp16-training> `_
for a complete list of options for configuration and performance tuning.
.. note ::
2021-04-19 09:19:12 -07:00
ZeRO-Infinity and ZeRO-Offload work best with our heavily optimized
2021-03-08 12:54:54 -08:00
:class: `deepspeed.ops.adam.DeepSpeedCPUAdam` optimizer. We recommend using
our `optimizer config <https://www.deepspeed.ai/docs/config-json/#optimizer-parameters> `_
to instruct :meth: `deepspeed.initialize` to build the optimizer for you.
2022-11-15 13:27:22 -08:00
ZeRO Configurations
===================
All the settings for DeepSpeed ZeRO are set with the `DeepSpeedZeroConfig`_ .
The dictionary provided under the `` zero_optimization `` entry of the main
DeepSpeed configuration dict will be parsed and validated with this class.
2023-07-11 00:24:01 +08:00
Sub-configurations for parameter offload and optimizer offload settings are
2022-11-15 13:27:22 -08:00
parsed by `DeepSpeedZeroOffloadParamConfig`_ and
`DeepSpeedZeroOffloadOptimizerConfig`_ .
.. _DeepSpeedZeroConfig:
.. autopydantic_model :: deepspeed.runtime.zero.config.DeepSpeedZeroConfig
.. _DeepSpeedZeroOffloadParamConfig:
.. autopydantic_model :: deepspeed.runtime.zero.config.DeepSpeedZeroOffloadParamConfig
.. _DeepSpeedZeroOffloadOptimizerConfig:
.. autopydantic_model :: deepspeed.runtime.zero.config.DeepSpeedZeroOffloadOptimizerConfig
2021-03-08 12:54:54 -08:00
2021-04-19 09:19:12 -07:00
Example ZeRO-3 Configurations
=============================
2021-03-08 12:54:54 -08:00
#. Use ZeRO to partition the optimizer states (stage 1), gradients (stage 2),
and parameters (stage 3).
.. code-block :: python
:emphasize-lines: 3
{
"zero_optimization": {
"stage": 3,
},
"fp16": {
"enabled": true
},
"optimizer": {
"type": "AdamW",
"params": {
"lr": 0.001,
"betas": [
0.8,
0.999
],
"eps": 1e-8,
"weight_decay": 3e-7
}
},
...
}
2021-04-19 09:19:12 -07:00
#. Additionally offload the optimizer states and computations to the CPU with ZeRO-Infinity.
2021-03-08 12:54:54 -08:00
.. code-block :: python
{
" zero_optimization " : {
" stage " : 3 ,
2021-04-18 23:45:37 -07:00
" offload_optimizer " : {
" device " : " cpu "
}
2021-03-08 12:54:54 -08:00
} ,
. . .
}
#. Save even more memory by offloading parameters to the CPU memory.
.. code-block :: python
{
" zero_optimization " : {
" stage " : 3 ,
2021-04-18 23:45:37 -07:00
" offload_optimizer " : {
" device " : " cpu "
}
" offload_param " : {
" device " : " cpu "
}
} ,
. . .
}
2021-04-19 09:19:12 -07:00
#. Save even MORE memory by offloading to NVMe (if available on your system):
2021-04-18 23:45:37 -07:00
.. code-block :: python
{
" zero_optimization " : {
" stage " : 3 ,
" offload_optimizer " : {
" device " : " nvme " ,
" nvme_path " : " /nvme_data "
}
" offload_param " : {
" device " : " nvme " ,
" nvme_path " : " /nvme_data "
}
2021-03-08 12:54:54 -08:00
} ,
. . .
}
2023-04-25 17:23:00 -07:00
MiCS Configurations
===================
All MiCS configurations are set with `DeepSpeedZeroConfig`
stage 3 optimization is enabled. For now, there are two configuration fields of
MiCS `mics_shard_size` `mics_hierarchical_params_gather` `mics_shard_size`
controls how many devices are used for partitioning the model states.
`mics_hierarchical_params_gather`
hierarchical way to gather parameters in the forward computation.
`mics_hierarchical_params_gather`
across multiple nodes and the cross-node bandwidth is slow. By default this is
turned off.
Example MiCS Configurations
===========================
#. Use MiCS to partition the model states (including optimizer states,
gradients, and parameters). The following config example partitions the model
states to eight devices, and assumes the eight devices are located within a
single node (`mics_hierarchical_params_gather` `False`
.. code-block :: python
:emphasize-lines: 3
2021-03-08 12:54:54 -08:00
2023-04-25 17:23:00 -07:00
{
"zero_optimization": {
"stage": 3,
"mics_shard_size": 8,
"mics_hierarchical_params_gather": False,
},
...
}
2021-03-08 12:54:54 -08:00
Assumptions
===========
DeepSpeed automatically coordinates the collection (*i.e.,* all-gather),
partitioning (*i.e.,* scatter), and offloading of parameters at the
granularity of (sub)module `` forward() `` methods. The backward pass is
handled similarly. This strategy has two underlying assumptions:
#. The forward and backward passes of submodules must individually fit in device memory.
2021-04-19 09:19:12 -07:00
If this not the case, :class: `deepspeed.zero.TiledLinear` implements
**memory-centric tiling** and works with ZeRO-3 to break linear layers
into a sequence of smaller submodules that can fit in memory.
2021-03-08 12:54:54 -08:00
#. A module's parameters are only accessed within its own `` __init__ `` and `` forward() `` methods.
Otherwise, DeepSpeed must be instructed to collect and re-partition the parameter.
See :ref: `external-parameters` for manually coordinating parameters.
Constructing Massive Models
---------------------------
ZeRO-3 enables massive models whose parameters exceed the size of individual
nodes in a system. For the typical case of training without model parallelism,
you can simply allocate your model in our context:
.. code-block :: python
with deepspeed . zero . Init ( ) :
model = MyLargeModel ( )
2021-04-19 09:19:12 -07:00
.. autoclass :: deepspeed.zero.Init
2021-03-08 12:54:54 -08:00
:members:
.. _external-parameters:
Manual Parameter Coordination
-----------------------------
Most models require no modification to be trained with ZeRO-3. However, in
some cases one may need to access model weights outside of the training loop,
or to share weights across submodules during training. DeepSpeed has
several mechanisms to coordinate partitioned weights for ZeRO-3.
Gathering Parameters
====================
DeepSpeed provides mechanisms for collecting (or *gathering* ) a partitioned parameter.
Some models partitioned with :class: `deepspeed.zero.Init` may need to access
a module’ `` forward() ``
2021-03-11 10:36:34 -08:00
method. We refer to these weights as **external parameters** , since these
parameters are accessed outside of the module that created them. To do so, use
2021-03-08 12:54:54 -08:00
:class: `deepspeed.zero.GatheredParameters` or :meth: `deepspeed.zero.register_external_parameter` .
.. autoclass :: deepspeed.zero.GatheredParameters
:members:
Registering External Parameters
===============================
2021-04-19 09:19:12 -07:00
ZeRO-3 will automatically collect and partition the model parameters as they
are needed during the forward and backward passes. However, in some cases a
parameter may be used outside of its module's forward pass. We call these
*external* parameters. ZeRO-3 can coordinate these parameters if they are
registered either automatically or manually.
2021-03-08 12:54:54 -08:00
2021-04-19 09:19:12 -07:00
.. note ::
DeepSpeed version `` 0.3.15 `` includes automatic external parameter
discovery and registration to support the most common cases. Parameters
can still be manually registered if they cannot be automatically
detected.
2021-03-08 12:54:54 -08:00
2021-04-19 09:19:12 -07:00
DeepSpeed can automatically detect the following external parameter scenarios:
2021-03-08 12:54:54 -08:00
2021-04-18 23:45:37 -07:00
2021-04-19 09:19:12 -07:00
#. Parameter access: consider the following pattern common in language models such as GPT:
2021-04-18 23:45:37 -07:00
2021-04-19 09:19:12 -07:00
The tensor `` embeddings.weight `` is used in both `` embeddings.forward() `` and
`` compute_logits() `` . We call `` embeddings.weight `` an *external* parameter
because it is used in the training loop outside of its owning module's
forward pass.
2021-04-18 23:45:37 -07:00
2021-04-19 09:19:12 -07:00
.. code-block :: python
class LanguageModel ( torch . nn . Module ) :
. . .
def forward ( self , inputs ) :
embeds = self . embeddings ( inputs )
. . .
logits = compute_logits ( output , self . embeddings . weight )
. . .
#. Returning a parameter:
`` CustomLinear `` returns both an output and its own `` bias `` parameter. DeepSpeed
will detect the external `` bias `` parameter and register it with submodules that
use `` CustomLinear `` .
.. code-block :: python
class CustomLinear ( torch . nn . Linear ) :
def forward ( self , * input ) :
output = super ( ) . forward ( * input )
return output , self . bias
2021-04-18 23:45:37 -07:00
2021-03-08 12:54:54 -08:00
.. autofunction :: deepspeed.zero.register_external_parameter
.. autofunction :: deepspeed.zero.unregister_external_parameter
2021-04-18 23:45:37 -07:00
Memory-Centric Tiling
---------------------
2021-04-19 09:19:12 -07:00
To reduce the working memory requirements of DL training for large models,
ZeRO-Infinity includes technique called *memory-centric tiling* that exploits
the data fetch and release pattern of ZeRO-3 to reduce the working memory
requirements by breaking down a large operator into smaller tiles that can be
executed sequentially. When combined with ZeRO-3, the parameter and gradients
of each tile can be fetched and released one at a time, reducing the working
memory proportional to the number of tiles. Therefore, ZeRO-Infinity can
support operators of arbitrary sizes, without refactoring for model
parallelism to fit them in limited GPU memory.
2021-04-18 23:45:37 -07:00
.. autoclass :: deepspeed.zero.TiledLinear
:members:
2023-02-27 23:40:49 -05:00
Debugging
---------
Debugging ZeRO training is complicated by the partitioning of parameters, gradients, and optimizer states. None of these 3 groups of tensors (model states) can be normally accessed because of that. To overcome that DeepSpeed provides the following routines for accessing individual model states in their unpartitioned form.
Important: Please note that these utilities must be called by all processes participating in the training, even if you decide to do something with the result only in the main process. If all processes don't participate these utilities will hang waiting for all processes to send their contribution.
Additionally, you must be aware that these routines return correct data only in specific phases of the training. So for examples the gradients are valid after `` backward `` and before `` step `` . The optimizer states are updated after `` step `` . Same goes for fp32 master weights.
.. autofunction :: deepspeed.utils.safe_get_full_fp32_param
.. autofunction :: deepspeed.utils.safe_get_full_grad
.. autofunction :: deepspeed.utils.safe_get_full_optimizer_state
These routines can be used in a training loop as shown in the following snippet.
.. code-block :: python
backward ( loss )
[ . . . ]
from deepspeed . utils import safe_get_full_fp32_param , safe_get_full_grad , safe_get_full_optimizer_state
for n , lp in model . named_parameters ( ) :
# 1. gradient lookup
# For zero1 and zero2, gradient lookup must be called after `backward` and before `step`
# For zero3, gradient lookup must be called after `backward`
hp_grad = safe_get_full_grad ( lp )
# 2. fp32 and optim states can probably be called anywhere in the training loop, but will be updated after `step`
hp = safe_get_full_fp32_param ( lp )
exp_avg = safe_get_full_optimizer_state ( lp , " exp_avg " )
exp_avg_sq = safe_get_full_optimizer_state ( lp , " exp_avg_sq " )
[ . . . ]
optimizer . step ( )
2023-03-23 20:15:34 -04:00
GPU Memory Management
---------------------
By default at the end of training with ZeRO stage 3 some parameters could remain unpartitioned and use up some gpu memory.
This is done on purpose as an optimization should you resume training again. If you'd like to clear out the cached
parameters that use up gpu memory, you can call `` empty_partition_cache `` method of a DeepSpeed engine.
.. autofunction :: deepspeed.DeepSpeedEngine.empty_partition_cache
The following code snippet illustrates this functionality.
.. code-block :: python
with zero . Init ( ) :
model = MyLargeModel ( )
ds_engine , _ , _ , _ = deepspeed . initialize ( model , . . . )
for batch in . . . :
loss = ds_engine ( batch )
ds_engine . backward ( batch )
ds_engine . step ( )
# Free GPU memory consumed by model parameters
ds_engine . empty_partition_cache ( )
