Note
torch.distributed.pipelining is currently in alpha state and under
development. API changes may be possible. It was migrated from the PiPPy project.
Pipeline Parallelism is one of the primitive parallelism for deep learning. It allows the execution of a model to be partitioned such that multiple micro-batches can execute different parts of the model code concurrently. Pipeline parallelism can be an effective technique for:
- large-scale training
- bandwidth-limited clusters
- large model inference
The above scenarios share a commonality that the computation per device cannot hide the communication of conventional parallelism, for example, the weight all-gather of FSDP.
While promising for scaling, pipelining is often difficult to implement because it needs to partition the execution of a model in addition to model weights. The partitioning of execution often requires intrusive code changes to your model. Another aspect of complexity comes from scheduling micro-batches in a distributed environment, with data flow dependency considered.
The pipelining package provides a toolkit that does said things
automatically which allows easy implementation of pipeline parallelism
on general models.
It consists of two parts: a splitting frontend and a distributed runtime. The splitting frontend takes your model code as-is, splits it up into "model partitions", and captures the data-flow relationship. The distributed runtime executes the pipeline stages on different devices in parallel, handling things like micro-batch splitting, scheduling, communication, and gradient propagation, etc.
Overall, the pipelining package provides the following features:
- Splitting of model code based on simple specification.
- Rich support for pipeline schedules, including GPipe, 1F1B, Interleaved 1F1B and Looped BFS, and providing the infrastructure for writing customized schedules.
- First-class support for cross-host pipeline parallelism, as this is where PP is typically used (over slower interconnects).
- Composability with other PyTorch parallel techniques such as data parallel (DDP, FSDP) or tensor parallel. The TorchTitan project demonstrates a "3D parallel" application on the Llama model.
Before we can use a PipelineSchedule, we need to create PipelineStage
objects that wrap the part of the model running in that stage. The
PipelineStage is responsible for allocating communication buffers and
creating send/recv ops to communicate with its peers. It manages intermediate
buffers e.g. for the outputs of forward that have not been consumed yet, and it
provides a utility for running the backwards for the stage model.
A PipelineStage needs to know the input and output shapes for the stage
model, so that it can correctly allocate communication buffers. The shapes must
be static, e.g. at runtime the shapes can not change from step to step. A class
PipeliningShapeError will be raised if runtime shapes do not match the
expected shapes. When composing with other paralleisms or applying mixed
precision, these techniques must be taken into account so the PipelineStage
knows the correct shape (and dtype) for the output of the stage module at
runtime.
Users may construct a PipelineStage instance directly, by passing in an
nn.Module representing the portion of the model that should run on the
stage. This may require changes to the original model code. See the example
in :ref:`option_1_manual`.
Alternatively, the splitting frontend can use graph partitioning to split your
model into a series of nn.Module automatically. This technique requires the
model is traceable with torch.Export. Composability of the resulting
nn.Module with other parallelism techniques is experimental, and may require
some workarounds. Usage of this frontend may be more appealing if the user
cannot easily change the model code. See :ref:`option_2_tracer` for more
information.
We can now attach the PipelineStage to a pipeline schedule, and run the
schedule with input data. Here is a GPipe example:
from torch.distributed.pipelining import ScheduleGPipe
# Create a schedule
schedule = ScheduleGPipe(stage, n_microbatches)
# Input data (whole batch)
x = torch.randn(batch_size, in_dim, device=device)
# Run the pipeline with input `x`
# `x` will be divided into microbatches automatically
if rank == 0:
schedule.step(x)
else:
output = schedule.step()Note that the above code needs to be launched for each worker, thus we use a launcher service to launch multiple processes:
torchrun --nproc_per_node=2 example.pyTo directly construct a PipelineStage, the user is responsible for providing
a single nn.Module instance that owns the relevant nn.Parameters and
nn.Buffers, and defines a forward() method that executes the operations
relevant for that stage. For example, a condensed version of the Transformer
class defined in Torchtitan shows a pattern of building an easily partitionable
model.
class Transformer(nn.Module):
def __init__(self, model_args: ModelArgs):
super().__init__()
self.tok_embeddings = nn.Embedding(...)
# Using a ModuleDict lets us delete layers without affecting names,
# ensuring checkpoints will correctly save and load.
self.layers = torch.nn.ModuleDict()
for layer_id in range(model_args.n_layers):
self.layers[str(layer_id)] = TransformerBlock(...)
self.output = nn.Linear(...)
def forward(self, tokens: torch.Tensor):
# Handling layers being 'None' at runtime enables easy pipeline splitting
h = self.tok_embeddings(tokens) if self.tok_embeddings else tokens
for layer in self.layers.values():
h = layer(h, self.freqs_cis)
h = self.norm(h) if self.norm else h
output = self.output(h).float() if self.output else h
return outputA model defined in this manner can be easily configured per stage by first initializing the whole model (using meta-device to avoid OOM errors), deleting undesired layers for that stage, and then creating a PipelineStage that wraps the model. For example:
with torch.device("meta"):
assert num_stages == 2, "This is a simple 2-stage example"
# we construct the entire model, then delete the parts we do not need for this stage
# in practice, this can be done using a helper function that automatically divides up layers across stages.
model = Transformer()
if stage_index == 0:
# prepare the first stage model
del model.layers["1"]
model.norm = None
model.output = None
elif stage_index == 1:
# prepare the second stage model
model.tok_embeddings = None
del model.layers["0"]
from torch.distributed.pipelining import PipelineStage
stage = PipelineStage(
model,
stage_index,
num_stages,
device,
input_args=example_input_microbatch,
)The PipelineStage requires an example argument input_args representing
the runtime input to the stage, which would be one microbatch worth of input
data. This argument is passed through the forward method of the stage module to
determine the input and output shapes required for communication.
When composing with other Data or Model parallelism techniques, output_args
may also be required, if the output shape/dtype of the model chunk will be
affected.
If you have a full model and do not want to spend time on modifying it into a
sequence of "model partitions", the pipeline API is here to help.
Here is a brief example:
class Model(torch.nn.Module):
def __init__(self) -> None:
super().__init__()
self.emb = torch.nn.Embedding(10, 3)
self.layers = torch.nn.ModuleList(
Layer() for _ in range(2)
)
self.lm = LMHead()
def forward(self, x: torch.Tensor) -> torch.Tensor:
x = self.emb(x)
for layer in self.layers:
x = layer(x)
x = self.lm(x)
return xIf we print the model, we can see multiple hierarchies, which makes it hard to split by hand:
Model(
(emb): Embedding(10, 3)
(layers): ModuleList(
(0-1): 2 x Layer(
(lin): Linear(in_features=3, out_features=3, bias=True)
)
)
(lm): LMHead(
(proj): Linear(in_features=3, out_features=3, bias=True)
)
)
Let us see how the pipeline API works:
from torch.distributed.pipelining import pipeline, SplitPoint
# An example micro-batch input
x = torch.LongTensor([1, 2, 4, 5])
pipe = pipeline(
module=mod,
mb_args=(x,),
split_spec={
"layers.1": SplitPoint.BEGINNING,
}
)The pipeline API splits your model given a split_spec, where
SplitPoint.BEGINNING stands for adding a split point
before execution of certain submodule in the forward function, and
similarly, SplitPoint.END for split point after such.
If we print(pipe), we can see:
GraphModule(
(submod_0): GraphModule(
(emb): InterpreterModule()
(layers): Module(
(0): InterpreterModule(
(lin): InterpreterModule()
)
)
)
(submod_1): GraphModule(
(layers): Module(
(1): InterpreterModule(
(lin): InterpreterModule()
)
)
(lm): InterpreterModule(
(proj): InterpreterModule()
)
)
)
def forward(self, x):
submod_0 = self.submod_0(x); x = None
submod_1 = self.submod_1(submod_0); submod_0 = None
return (submod_1,)
The "model partitions" are represented by submodules (submod_0,
submod_1), each of which is reconstructed with original model operations, weights
and hierarchies. In addition, a "root-level" forward function is
reconstructed to capture the data flow between those partitions. Such data flow
will be replayed by the pipeline runtime later, in a distributed fashion.
The Pipe object provides a method for retrieving the "model partitions":
stage_mod : nn.Module = pipe.get_stage_module(stage_idx)The returned stage_mod is a nn.Module, with which you can create an
optimizer, save or load checkpoints, or apply other parallelisms.
Pipe also allows you to create a distributed stage runtime on a device given
a ProcessGroup:
stage = pipe.build_stage(stage_idx, device, group)Alternatively, if you would like to build the stage runtime later after some
modification to the stage_mod, you can use a functional version of the
build_stage API. For example:
from torch.distributed.pipelining import build_stage
from torch.nn.parallel import DistributedDataParallel
dp_mod = DistributedDataParallel(stage_mod)
info = pipe.info()
stage = build_stage(dp_mod, stage_idx, info, device, group)Note
The pipeline frontend uses a tracer (torch.export) to capture your
model into a single graph. If your model is not full-graph'able, you can use
our manual frontend below.
In the PiPPy repo where this package was original created, we kept examples based on unmodified Hugging Face models. See the examples/huggingface directory.
Examples include:
First, the pipeline API turns our model into a directed acyclic graph (DAG)
by tracing the model. It traces the model using torch.export -- a PyTorch 2
full-graph capturing tool.
Then, it groups together the operations and parameters needed by a stage
into a reconstructed submodule: submod_0, submod_1, ...
Different from conventional submodule access methods like Module.children(),
the pipeline API does not only cut the module structure of your model, but
also the forward function of your model.
This is necessary because model structure like Module.children() merely
captures information during Module.__init__(), and does not capture any
information about Module.forward(). Said differently, Module.children()
lacks information about the following aspects key to pipelininig:
- Execution order of child modules in
forward - Activation flows between child modules
- Whether there are any functional operators between child modules (for example,
reluoraddoperations will not be captured byModule.children()).
The pipeline API, on the contrary, makes sure that the forward behavior
is truly preserved. It also captures the activation flow between the partitions,
helping the distributed runtime to make correct send/receive calls without human
intervention.
Another flexibility of the pipeline API is that split points can be at
arbitrary levels within your model hierarchy. In the split partitions, the original model
hierarchy related to that partition will be reconstructed at no cost to you.
At a result, fully-qualified names (FQNs) pointing to a submodule or parameter
would be still valid, and services that relies on FQNs (such as FSDP, TP or
checkpointing) can still run with your partitioned modules with almost zero code
change.
You can implement your own pipeline schedule by extending one of the following two class:
PipelineScheduleSinglePipelineScheduleMulti
PipelineScheduleSingle is for schedules that assigns only one stage per rank.
PipelineScheduleMulti is for schedules that assigns multiple stages per rank.
For example, ScheduleGPipe and Schedule1F1B are subclasses of PipelineScheduleSingle.
Whereas, ScheduleInterleaved1F1B, ScheduleLoopedBFS, and ScheduleInterleavedZeroBubble
are subclasses of PipelineScheduleMulti.
You can turn on additional logging using the TORCH_LOGS environment variable from [torch._logging](https://pytorch.org/docs/main/logging.html#module-torch._logging):
- TORCH_LOGS=+pp will display logging.DEBUG messages and all levels above it.
- TORCH_LOGS=pp will display logging.INFO messages and above.
- TORCH_LOGS=-pp will display logging.WARNING messages and above.
.. automodule:: torch.distributed.pipelining
The following set of APIs transform your model into a pipeline representation.
.. currentmodule:: torch.distributed.pipelining
.. autoclass:: SplitPoint
.. autofunction:: pipeline
.. autoclass:: Pipe
.. autofunction:: pipe_split
.. automodule:: torch.distributed.pipelining.microbatch
.. currentmodule:: torch.distributed.pipelining.microbatch
.. autoclass:: TensorChunkSpec
.. autofunction:: split_args_kwargs_into_chunks
.. autofunction:: merge_chunks
.. automodule:: torch.distributed.pipelining.stage
.. currentmodule:: torch.distributed.pipelining.stage
.. autoclass:: PipelineStage
.. autofunction:: build_stage
.. automodule:: torch.distributed.pipelining.schedules
.. currentmodule:: torch.distributed.pipelining.schedules
.. autoclass:: ScheduleGPipe
.. autoclass:: Schedule1F1B
.. autoclass:: ScheduleInterleaved1F1B
.. autoclass:: ScheduleLoopedBFS
.. autoclass:: ScheduleInterleavedZeroBubble
.. autoclass:: PipelineScheduleSingle :members:
.. autoclass:: PipelineScheduleMulti :members: