Path: blob/main/recipes_source/distributed_async_checkpoint_recipe.rst
3707 views
Asynchronous Saving with Distributed Checkpoint (DCP)
=====================================================
**Author:** `Lucas Pasqualin <https://github.com/lucasllc>`__, `Iris Zhang <https://github.com/wz337>`__, `Rodrigo Kumpera <https://github.com/kumpera>`__, `Chien-Chin Huang <https://github.com/fegin>`__, `Yunsheng Ni <https://github.com/niyunsheng>`__
Checkpointing is often a bottleneck in the critical path for distributed training workloads, incurring larger and larger costs as both model and world sizes grow.
One excellent strategy for offsetting this cost is to checkpoint in parallel, asynchronously. Below, we expand the save example
from the `Getting Started with Distributed Checkpoint Tutorial <https://github.com/pytorch/tutorials/blob/main/recipes_source/distributed_checkpoint_recipe.rst>`__
to show how this can be integrated quite easily with ``torch.distributed.checkpoint.async_save``.
.. grid:: 2
.. grid-item-card:: :octicon:`mortar-board;1em;` What you will learn
:class-card: card-prerequisites
* How to use DCP to generate checkpoints in parallel
* Effective strategies to optimize performance
.. grid-item-card:: :octicon:`list-unordered;1em;` Prerequisites
:class-card: card-prerequisites
* PyTorch v2.4.0 or later
* `Getting Started with Distributed Checkpoint Tutorial <https://github.com/pytorch/tutorials/blob/main/recipes_source/distributed_checkpoint_recipe.rst>`__
Asynchronous Checkpointing Overview
------------------------------------
Before getting started with Asynchronous Checkpointing, it's important to understand its differences and limitations as compared to synchronous checkpointing.
Specifically:
* Memory requirements - Asynchronous checkpointing works by first copying models into internal CPU-buffers.
This is helpful since it ensures model and optimizer weights are not changing while the model is still checkpointing,
but does raise CPU memory by a factor of ``checkpoint_size_per_rank X number_of_ranks``. Additionally, users should take care to understand
the memory constraints of their systems. Specifically, pinned memory implies the usage of ``page-lock`` memory, which can be scarce as compared to
``pageable`` memory.
* Checkpoint Management - Since checkpointing is asynchronous, it is up to the user to manage concurrently run checkpoints.
In general, users can employ their own management strategies by handling the future object returned form ``async_save``. For most users, we recommend limiting
checkpoints to one asynchronous request at a time, avoiding additional memory pressure per request.
.. code-block:: python
import os
import torch
import torch.distributed as dist
import torch.distributed.checkpoint as dcp
import torch.multiprocessing as mp
import torch.nn as nn
from torch.distributed.fsdp import fully_shard
from torch.distributed.checkpoint.state_dict import get_state_dict, set_state_dict
from torch.distributed.checkpoint.stateful import Stateful
CHECKPOINT_DIR = "checkpoint"
class AppState(Stateful):
"""This is a useful wrapper for checkpointing the Application State. Since this object is compliant
with the Stateful protocol, DCP will automatically call state_dict/load_stat_dict as needed in the
dcp.save/load APIs.
Note: We take advantage of this wrapper to hande calling distributed state dict methods on the model
and optimizer.
"""
def __init__(self, model, optimizer=None):
self.model = model
self.optimizer = optimizer
def state_dict(self):
# this line automatically manages FSDP FQNs, as well as sets the default state dict type to FSDP.SHARDED_STATE_DICT
model_state_dict, optimizer_state_dict = get_state_dict(self.model, self.optimizer)
return {
"model": model_state_dict,
"optim": optimizer_state_dict
}
def load_state_dict(self, state_dict):
# sets our state dicts on the model and optimizer, now that we've loaded
set_state_dict(
self.model,
self.optimizer,
model_state_dict=state_dict["model"],
optim_state_dict=state_dict["optim"]
)
class ToyModel(nn.Module):
def __init__(self):
super(ToyModel, self).__init__()
self.net1 = nn.Linear(16, 16)
self.relu = nn.ReLU()
self.net2 = nn.Linear(16, 8)
def forward(self, x):
return self.net2(self.relu(self.net1(x)))
def setup(rank, world_size):
os.environ["MASTER_ADDR"] = "localhost"
os.environ["MASTER_PORT"] = "12355 "
# initialize the process group
dist.init_process_group("gloo", rank=rank, world_size=world_size)
torch.cuda.set_device(rank)
def cleanup():
dist.destroy_process_group()
def run_fsdp_checkpoint_save_example(rank, world_size):
print(f"Running basic FSDP checkpoint saving example on rank {rank}.")
setup(rank, world_size)
# create a model and move it to GPU with id rank
model = ToyModel().to(rank)
model = fully_shard(model)
loss_fn = nn.MSELoss()
optimizer = torch.optim.Adam(model.parameters(), lr=0.1)
checkpoint_future = None
for step in range(10):
optimizer.zero_grad()
model(torch.rand(8, 16, device="cuda")).sum().backward()
optimizer.step()
# waits for checkpointing to finish if one exists, avoiding queuing more then one checkpoint request at a time
if checkpoint_future is not None:
checkpoint_future.result()
state_dict = { "app": AppState(model, optimizer) }
checkpoint_future = dcp.async_save(state_dict, checkpoint_id=f"{CHECKPOINT_DIR}_step{step}")
cleanup()
if __name__ == "__main__":
world_size = torch.cuda.device_count()
print(f"Running async checkpoint example on {world_size} devices.")
mp.spawn(
run_fsdp_checkpoint_save_example,
args=(world_size,),
nprocs=world_size,
join=True,
)
Even more performance with Pinned Memory
-----------------------------------------
If the above optimization is still not performant enough, you can take advantage of an additional optimization for GPU models which utilizes a pinned memory buffer for checkpoint staging.
Specifically, this optimization attacks the main overhead of asynchronous checkpointing, which is the in-memory copying to checkpointing buffers. By maintaining a pinned memory buffer between
checkpoint requests users can take advantage of direct memory access to speed up this copy.
.. note::
The main drawback of this optimization is the persistence of the buffer in between checkpointing steps. Without
the pinned memory optimization (as demonstrated above), any checkpointing buffers are released as soon as
checkpointing is finished. With the pinned memory implementation, this buffer is maintained between steps,
leading to the same
peak memory pressure being sustained through the application life.
.. code-block:: python
import os
import torch
import torch.distributed as dist
import torch.distributed.checkpoint as dcp
import torch.multiprocessing as mp
import torch.nn as nn
from torch.distributed.fsdp import fully_shard
from torch.distributed.checkpoint.state_dict import get_state_dict, set_state_dict
from torch.distributed.checkpoint.stateful import Stateful
from torch.distributed.checkpoint import FileSystemWriter as StorageWriter
CHECKPOINT_DIR = "checkpoint"
class AppState(Stateful):
"""This is a useful wrapper for checkpointing the Application State. Since this object is compliant
with the Stateful protocol, DCP will automatically call state_dict/load_stat_dict as needed in the
dcp.save/load APIs.
Note: We take advantage of this wrapper to hande calling distributed state dict methods on the model
and optimizer.
"""
def __init__(self, model, optimizer=None):
self.model = model
self.optimizer = optimizer
def state_dict(self):
# this line automatically manages FSDP FQNs, as well as sets the default state dict type to FSDP.SHARDED_STATE_DICT
model_state_dict, optimizer_state_dict = get_state_dict(self.model, self.optimizer)
return {
"model": model_state_dict,
"optim": optimizer_state_dict
}
def load_state_dict(self, state_dict):
# sets our state dicts on the model and optimizer, now that we've loaded
set_state_dict(
self.model,
self.optimizer,
model_state_dict=state_dict["model"],
optim_state_dict=state_dict["optim"]
)
class ToyModel(nn.Module):
def __init__(self):
super(ToyModel, self).__init__()
self.net1 = nn.Linear(16, 16)
self.relu = nn.ReLU()
self.net2 = nn.Linear(16, 8)
def forward(self, x):
return self.net2(self.relu(self.net1(x)))
def setup(rank, world_size):
os.environ["MASTER_ADDR"] = "localhost"
os.environ["MASTER_PORT"] = "12355 "
# initialize the process group
dist.init_process_group("gloo", rank=rank, world_size=world_size)
torch.cuda.set_device(rank)
def cleanup():
dist.destroy_process_group()
def run_fsdp_checkpoint_save_example(rank, world_size):
print(f"Running basic FSDP checkpoint saving example on rank {rank}.")
setup(rank, world_size)
# create a model and move it to GPU with id rank
model = ToyModel().to(rank)
model = fully_shard(model)
loss_fn = nn.MSELoss()
optimizer = torch.optim.Adam(model.parameters(), lr=0.1)
# The storage writer defines our 'staging' strategy, where staging is considered the process of copying
# checkpoints to in-memory buffers. By setting `cached_state_dict=True`, we enable efficient memory copying
# into a persistent buffer with pinned memory enabled.
# Note: It's important that the writer persists in between checkpointing requests, since it maintains the
# pinned memory buffer.
writer = StorageWriter(cache_staged_state_dict=True, path=CHECKPOINT_DIR)
checkpoint_future = None
for step in range(10):
optimizer.zero_grad()
model(torch.rand(8, 16, device="cuda")).sum().backward()
optimizer.step()
state_dict = { "app": AppState(model, optimizer) }
if checkpoint_future is not None:
# waits for checkpointing to finish, avoiding queuing more then one checkpoint request at a time
checkpoint_future.result()
checkpoint_future = dcp.async_save(state_dict, storage_writer=writer, checkpoint_id=f"{CHECKPOINT_DIR}_step{step}")
cleanup()
if __name__ == "__main__":
world_size = torch.cuda.device_count()
print(f"Running fsdp checkpoint example on {world_size} devices.")
mp.spawn(
run_fsdp_checkpoint_save_example,
args=(world_size,),
nprocs=world_size,
join=True,
)
Fully Asynchronous Staging with DefaultStager
--------------------------------------------
.. versionadded:: 2.9
The ``async_stager`` argument and ``DefaultStager`` class were introduced in PyTorch 2.9.
While ``async_save`` handles the disk write asynchronously, the process of copying data from GPU to CPU (known as "staging") typically happens on the main thread. Even with Pinned Memory, this Device-to-Host (D2H) copy can block the training loop for large models.
To achieve maximum overlap between computation and checkpointing, we can use the ``DefaultStager``. This component offloads the state dictionary creation and the D2H copy to a background thread.
**Timeline Comparison:**
* **Standard async_save:** ``[GPU Compute] -> [CPU Copy (Blocking)] -> [Disk Write (Async)]``
* **With AsyncStager:** ``[GPU Compute] || [CPU Copy (Async)] -> [Disk Write (Async)]``
.. note::
Using ``AsyncStager`` introduces a background thread that consumes CPU resources. Ensure your environment has sufficient CPU cores to handle this without impacting the main training process.
.. code-block:: python
import os
import torch
import torch.distributed as dist
import torch.distributed.checkpoint as dcp
import torch.multiprocessing as mp
import torch.nn as nn
from torch.distributed.fsdp import fully_shard
from torch.distributed.checkpoint.state_dict import get_state_dict, set_state_dict
from torch.distributed.checkpoint.stateful import Stateful
from torch.distributed.checkpoint.staging import DefaultStager
from torch.nn.modules.linear import NonDynamicallyQuantizableLinear
CHECKPOINT_DIR = "checkpoint"
class AppState(Stateful):
"""This is a useful wrapper for checkpointing the Application State. Since this object is compliant
with the Stateful protocol, DCP will automatically call state_dict/load_stat_dict as needed in the
dcp.save/load APIs.
Note: We take advantage of this wrapper to hande calling distributed state dict methods on the model
and optimizer.
"""
def __init__(self, model, optimizer=None):
self.model = model
self.optimizer = optimizer
def state_dict(self):
# this line automatically manages FSDP FQNs, as well as sets the default state dict type to FSDP.SHARDED_STATE_DICT
model_state_dict, optimizer_state_dict = get_state_dict(self.model, self.optimizer)
return {
"model": model_state_dict,
"optim": optimizer_state_dict
}
def load_state_dict(self, state_dict):
# sets our state dicts on the model and optimizer, now that we've loaded
set_state_dict(
self.model,
self.optimizer,
model_state_dict=state_dict["model"],
optim_state_dict=state_dict["optim"]
)
class ToyModel(nn.Module):
def __init__(self):
super(ToyModel, self).__init__()
self.net1 = nn.Linear(16, 16)
self.relu = nn.ReLU()
self.net2 = nn.Linear(16, 8)
def forward(self, x):
return self.net2(self.relu(self.net1(x)))
def setup(rank, world_size):
os.environ["MASTER_ADDR"] = "localhost"
os.environ["MASTER_PORT"] = "12355 "
# initialize the process group
dist.init_process_group("gloo", rank=rank, world_size=world_size)
torch.cuda.set_device(rank)
def cleanup():
dist.destroy_process_group()
def run_fsdp_checkpoint_save_example(rank, world_size):
print(f"Running basic FSDP checkpoint saving example on rank {rank}.")
setup(rank, world_size)
# create a model and move it to GPU with id rank
model = ToyModel().to(rank)
model = fully_shard(model)
loss_fn = nn.MSELoss()
optimizer = torch.optim.Adam(model.parameters(), lr=0.1)
checkpoint_future = None
for step in range(10):
print(f"Step {step} starting...")
optimizer.zero_grad()
model(torch.rand(8, 16, device="cuda")).sum().backward()
# Critical: We must ensure the previous checkpoint's D2H copy (staging)
# is complete before the optimizer modifies the model parameters.
# Placing this await AFTER the backward pass allows us to overlap
# the D2H copy with the current step's Forward and Backward computation.
if checkpoint_future is not None:
checkpoint_future.staging_completion.result()
optimizer.step()
# waits for checkpointing to finish if one exists, avoiding queuing more then one checkpoint request at a time
if checkpoint_future is not None:
checkpoint_future.upload_completion.result()
state_dict = { "app": AppState(model, optimizer) }
# Pass the DefaultStager to enable fully asynchronous staging.
# This offloads the state_dict creation and GPU-to-CPU copy to a background thread.
# The return object (AsyncSaveResponse) exposes distinct futures for staging and upload.
checkpoint_future = dcp.async_save(
state_dict,
checkpoint_id=f"{CHECKPOINT_DIR}_step{step}",
async_stager=DefaultStager(),
)
# Ensure the last checkpoint completes
if checkpoint_future:
checkpoint_future.upload_completion.result()
cleanup()
if __name__ == "__main__":
world_size = torch.cuda.device_count()
print(f"Running async checkpoint example on {world_size} devices.")
mp.spawn(
run_fsdp_checkpoint_save_example,
args=(world_size,),
nprocs=world_size,
join=True,
)
Conclusion
----------
In conclusion, we have learned how to use DCP's :func:`async_save` API to generate checkpoints off the critical training path. We've also learned about the
additional memory and concurrency overhead introduced by using this API, as well as additional optimizations which utilize pinned memory to speed things up
even further.
- `Saving and loading models tutorial <https://pytorch.org/tutorials/beginner/saving_loading_models.html>`__
- `Getting started with FullyShardedDataParallel tutorial <https://pytorch.org/tutorials/intermediate/FSDP_tutorial.html>`__