2D Parallelism combines Tensor Parallelism (TP) and Fully Sharded Data Parallelism (FSDP) to leverage the memory efficiency of FSDP and the computational scalability of TP.
This hybrid approach balances the trade-offs of each method, optimizing memory usage and minimizing communication overhead, enabling the training of extremely large models on large GPU clusters.
The :doc:`Tensor Parallelism documentation <tp>` and a general understanding of `FSDP <https://pytorch.org/tutorials/intermediate/FSDP_tutorial.html>`_ are a prerequisite for this tutorial.
..note:: This is an experimental feature.
----
*********************
Enable 2D parallelism
*********************
We will start off with the same feed forward example model as in the :doc:`Tensor Parallelism tutorial <tp>`.
..code-block:: python
import torch.nn as nn
import torch.nn.functional as F
class FeedForward(nn.Module):
def __init__(self, dim, hidden_dim):
super().__init__()
self.w1 = nn.Linear(dim, hidden_dim, bias=False)
self.w2 = nn.Linear(hidden_dim, dim, bias=False)
self.w3 = nn.Linear(dim, hidden_dim, bias=False)
def forward(self, x):
return self.w2(F.silu(self.w1(x)) * self.w3(x))
Next, we define a function that applies the desired parallelism to our model.
The function must take as first argument the model and as second argument the a :class:`~torch.distributed.device_mesh.DeviceMesh`.
More on how the device mesh works later.
..code-block:: python
from torch.distributed.tensor.parallel import ColwiseParallel, RowwiseParallel
from torch.distributed.tensor.parallel import parallelize_module
from torch.distributed._composable.fsdp.fully_shard import fully_shard
def parallelize_feedforward(model, device_mesh):
# Lightning will set up a device mesh for you
# Here, it is 2-dimensional
tp_mesh = device_mesh["tensor_parallel"]
dp_mesh = device_mesh["data_parallel"]
if tp_mesh.size() > 1:
# Use PyTorch's distributed tensor APIs to parallelize the model
plan = {
"w1": ColwiseParallel(),
"w2": RowwiseParallel(),
"w3": ColwiseParallel(),
}
parallelize_module(model, tp_mesh, plan)
if dp_mesh.size() > 1:
# Use PyTorch's FSDP2 APIs to parallelize the model
fully_shard(model.w1, mesh=dp_mesh)
fully_shard(model.w2, mesh=dp_mesh)
fully_shard(model.w3, mesh=dp_mesh)
fully_shard(model, mesh=dp_mesh)
return model
By writing the parallelization code in a separate function rather than hardcoding it into the model, we keep the original source code clean and maintainable.
In addition to the tensor-parallel code from the :doc:`Tensor Parallelism tutorial <tp>`, this function also shards the model's parameters using FSDP along the data-parallel dimension.
Finally, pass the parallelization function to the :class:`~lightning.fabric.strategies.model_parallel.ModelParallelStrategy` and configure the data-parallel and tensor-parallel sizes:
..code-block:: python
import lightning as L
from lightning.fabric.strategies import ModelParallelStrategy
strategy = ModelParallelStrategy(
parallelize_fn=parallelize_feedforward,
# Define the size of the 2D parallelism
# Set these to "auto" (default) to apply TP intra-node and FSDP inter-node
In this example with 4 GPUs, Fabric will create a device mesh that groups GPU 0-1 and GPU 2-3 (2 groups because ``data_parallel_size=2``, and 2 GPUs per group because ``tensor_parallel_size=2``).
Later on when ``fabric.setup(model)`` is called, each layer wrapped with FSDP (``fully_shard``) will be split into two shards, one for the GPU 0-1 group, and one for the GPU 2-3 group.
Finally, the tensor parallelism will apply to each group, splitting the sharded tensor across the GPUs within each group.
..collapse:: Full training example (requires at least 4 GPUs).
..code-block:: python
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.distributed.tensor.parallel import ColwiseParallel, RowwiseParallel
from torch.distributed.tensor.parallel import parallelize_module
from torch.distributed._composable.fsdp.fully_shard import fully_shard
import lightning as L
from lightning.pytorch.demos.boring_classes import RandomDataset
from lightning.fabric.strategies import ModelParallelStrategy
class FeedForward(nn.Module):
def __init__(self, dim, hidden_dim):
super().__init__()
self.w1 = nn.Linear(dim, hidden_dim, bias=False)
self.w2 = nn.Linear(hidden_dim, dim, bias=False)
self.w3 = nn.Linear(dim, hidden_dim, bias=False)
def forward(self, x):
return self.w2(F.silu(self.w1(x)) * self.w3(x))
def parallelize_feedforward(model, device_mesh):
# Lightning will set up a device mesh for you
# Here, it is 2-dimensional
tp_mesh = device_mesh["tensor_parallel"]
dp_mesh = device_mesh["data_parallel"]
if tp_mesh.size() > 1:
# Use PyTorch's distributed tensor APIs to parallelize the model
plan = {
"w1": ColwiseParallel(),
"w2": RowwiseParallel(),
"w3": ColwiseParallel(),
}
parallelize_module(model, tp_mesh, plan)
if dp_mesh.size() > 1:
# Use PyTorch's FSDP2 APIs to parallelize the model
Beyond this toy example, we recommend you study our `LLM 2D Parallel Example (Llama 3) <https://github.com/Lightning-AI/pytorch-lightning/tree/master/examples/fabric/tensor_parallel>`_.
----
*******************
Effective use cases
*******************
In the toy example above, the parallelization is configured to work within a single machine across multiple GPUs.
However, in practice the main use case for 2D parallelism is in multi-node training, where one can effectively combine both methods to maximize throughput and model scale.
Since tensor-parallelism requires blocking collective calls, fast GPU data transfers are essential to keep throughput high and therefore TP is typically applied across GPUs within a machine.
On the other hand, FSDP by design has the advantage that it can overlap GPU transfers with the computation (it can prefetch layers).
Hence, combining FSDP for inter-node parallelism and TP for intra-node parallelism is generally a good strategy to minimize both the latency and network bandwidth usage, making it possible to scale to much larger models than is possible with FSDP alone.
..code-block:: python
from lightning.fabric.strategies import ModelParallelStrategy
strategy = ModelParallelStrategy(
# Default is "auto"
# Applies TP intra-node and DP inter-node
data_parallel_size="auto",
tensor_parallel_size="auto",
)
----
***************************
Data-loading considerations
***************************
In a tensor-parallelized model, it is important that the model receives an identical input on each GPU that participates in the same tensor-parallel group.
However, across the data-parallel dimension, the inputs should be different.
In other words, if TP is applied within a node, and FSDP across nodes, each node must receive a different batch, but every GPU within the node gets the same batch of data.
If you use a PyTorch data loader and set it up using :meth:`~lightning.fabric.fabric.Fabric.setup_dataloaders`, Fabric will automatically handle this for you by configuring the distributed sampler.
However, when you shuffle data in your dataset or data loader, or when applying randomized transformations/augmentations in your data, you must still ensure that the seed is set appropriately.
..code-block:: python
import lightning as L
fabric = L.Fabric(...)
# Define dataset/dataloader
# If there is randomness/augmentation in the dataset, fix the seed
