Skip to main content

3D Tensor Parallelism

Author: Zhengda Bian, Yongbin Li


Example Code

Related Paper


The 3D tensor parallelism is an approach to parallelize the computation of neural models, hoping to obtain the optimal communication cost.

Let's still take a linear layer Y=XAY = XA as an example. Given P=q×q×qP=q \times q \times q processors (necessary condition), e.g. q=2q=2, we split the input XX and weight AA into

[X000X001X010X011X100X101X110X111] and [A000A001A010A011A100A101A110A111] respectively,\left[\begin{matrix} X_{000} & X_{001} \\ X_{010} & X_{011} \\ X_{100} & X_{101} \\ X_{110} & X_{111} \end{matrix} \right] \text{~and~} \left[\begin{matrix} A_{000} & A_{001} & A_{010} & A_{011} \\ A_{100} & A_{101} & A_{110} & A_{111} \end{matrix} \right] \text{~respectively,}

where each XijlX_{ijl} and AljiA_{lji} are stored at processor (i,j,l)(i,j,l), as shown in the figure below.

Then we all-gather XijlX_{ijl} across (i,0...q,l)(i, 0...q,l), as well as AljiA_{lji} across (0...q,j,l)(0...q, j, l). So, we have XilX_{il} and AljA_{lj} on each processor (i,j,l)(i,j,l) to get XilAljX_{il}A_{lj}. Finally, we reduce-scatter the results across (i,j,0...q)(i, j, 0...q) to get YijlY_{ijl}, which forms

Y=[Y000Y001Y010Y011Y100Y101Y110Y111].Y= \left[\begin{matrix} Y_{000} & Y_{001} \\ Y_{010} & Y_{011} \\ Y_{100} & Y_{101} \\ Y_{110} & Y_{111} \end{matrix} \right].

We also need to note that in the backward pass, we need to all-gather the gradient Yijl˙\dot{Y_{ijl}}, and then reduce-scatter the gradient Xil˙=Yij˙AljT\dot{X_{il}}=\dot{Y_{ij}}A_{lj}^T and Alj˙=XilTYij˙\dot{A_{lj}}=X_{il}^T\dot{Y_{ij}}.


Given P=q×q×qP=q \times q \times q processors, we present the theoretical computation and memory cost, as well as the communication cost based on the ring algorithm in both the forward and backward pass of 3D tensor parallelism.

ComputationMemory (parameters)Memory (activations)Communication (bandwidth)Communication (latency)


To enable 3D tensor parallelism for our model, e.g. on 8 GPUs, we need to configure the parallism setting as below.

CONFIG = dict(parallel=dict(
tensor=dict(size=8, mode='3d'),

Then Colossal-AI will automatically apply 3D parallelism to all the layers from colossalai.nn.

Let's define a model that consists of a two-layer multi-layer perceptron (MLP) as below.

import colossalai
import colossalai.nn as col_nn
import torch
from colossalai.utils import print_rank_0

class MLP(torch.nn.Module):
def __init__(self, dim: int = 256):
intermediate_dim = dim * 4
self.dense_1 = col_nn.Linear(dim, intermediate_dim)
print_rank_0(f'Weight of the first linear layer: {self.dense_1.weight.shape}')
self.activation = torch.nn.GELU()
self.dense_2 = col_nn.Linear(intermediate_dim, dim)
print_rank_0(f'Weight of the second linear layer: {self.dense_2.weight.shape}')
self.dropout = col_nn.Dropout(0.1)

def forward(self, x):
x = self.dense_1(x)
print_rank_0(f'Output of the first linear layer: {x.shape}')
x = self.activation(x)
x = self.dense_2(x)
print_rank_0(f'Output of the second linear layer: {x.shape}')
x = self.dropout(x)
return x

Launch Colossal-AI on 8 GPUs and build the model

parser = colossalai.get_default_parser()

m = MLP()

We will see the shapes of partitioned parameters(e.g. weights) in the MLP model.

Weight of the first linear layer: torch.Size([128, 256])
Weight of the second linear layer: torch.Size([512, 64])

The complete weight of the first linear layer is supposed to have the shape [256, 1024]. After the partitioning of 3D parallelism, it becomes [128, 256] on each GPU. Similarly, the second layer partitions the weight [1024, 256] into [512, 64].

We can run the model with some random inputs.

from colossalai.context import ParallelMode
from colossalai.core import global_context as gpc
from colossalai.utils import get_current_device

x = torch.randn((16, 256), device=get_current_device())
# partition input
torch.distributed.broadcast(x, src=0)
x = torch.chunk(x, 2, dim=0)[gpc.get_local_rank(ParallelMode.PARALLEL_3D_WEIGHT)]
x = torch.chunk(x, 2, dim=0)[gpc.get_local_rank(ParallelMode.PARALLEL_3D_INPUT)]
x = torch.chunk(x, 2, dim=-1)[gpc.get_local_rank(ParallelMode.PARALLEL_3D_OUTPUT)]
print_rank_0(f'Input: {x.shape}')

x = m(x)

Then we can see the shapes of activation results.

Input: torch.Size([4, 128])
Output of the first linear layer: torch.Size([4, 512])
Output of the second linear layer: torch.Size([4, 128])

The activation tensors in 3D parallelism are all split by q2q^2 in the row and qq in the column. E.g. the output of the first linear layer has the shape [4, 512]), while the second layer has the output of [4, 128]. Note, although the results of 3D parallelism have the same shape as that of 2.5D parallelism for weights here, the content of each partition is different.