tensorkelechi/arxiv-cstext · Datasets at Hugging Face

text stringlengths 0 89.3k
42 Techniques for Efficiency and Effectiveness
In this section we discuss some techniques to improve the efficiency
and effectiveness of HiCom The efficiency perspective is important
as LLMs can have multibillion parameters with considerable GPU
memory requirements We thus introduce implementation tech
niques for batch processing of neighborhoods with different sizes
and reducing the memory consumption of unnecessary padding
These techniques are not restricted to the HiCom algorithm and can
be used for general LLMs to optimize GPU utilization on graph data
From the effectiveness perspective we discuss the idea of summary
accumulation which shares a similar idea to skip connection
421 Batch Processing of Neighborhoods with Different Sizes Algo
rithm 1 presents a straightforward version of HiCom for individual
nodes offering a fundamental understanding of the algorithm In
practice batch processing is essential for maximizing the efficiency
of GPU utilization but it also introduces specific challenges Since
LLMs are innately designed for sequential data the challenge in
adapting HiCom with LLMs for batch processing primarily arises
from the variablelength inputs produced by graph data which
complicates the batching of data instances To address these chal
lenges a more sophisticated padding strategy is needed to uniform
both the length of the text sequence on each node and the number
of neighbors of each node for all nodes in the same level of the
hierarchy Algorithm 2 outlines HiCom for processing nodes in
batches We discuss the algorithm in detail in Appendix A1
422 Pretokenization and Rearrange and Trim Another effective
implementation technique employed in the HiCom framework to
reduce the computational overhead associated with LLMs is pre
tokenization which involves preprocessing the text data into to
kens and padding them with dummy tokens to ensure uniformity in
sequence lengths before the training starts Pretokenization saves
the redundant time cost of raw text encoding when a node belongsAlgorithm 1 HiCom for processing individual nodes
1InputGV𝑿 fanouts𝑛1𝑛𝐿𝐶𝑜𝑚𝑝 a target
node𝑣V
2Output Summary vectors 𝑠𝑣of node𝑣
3V𝐵𝐿𝑣 a batch with a single node ie index 𝐵𝐿𝑣
4for𝑙𝐿to1do
5V𝐵𝑙1Sample up to 𝑛𝑙neighbors of 𝑣𝑣𝑉𝐵𝑙
6 𝑿𝐵𝑙1Collect𝑥𝑣𝑣𝑉𝐵𝑙1
7end for
8𝑺𝐵0
9for𝑙1to𝐿do
10 𝑿𝐶
𝐵𝑙 Concat 𝑿𝐵𝑙1
11 𝑺𝐶
𝐵𝑙 Concat 𝑺𝐵𝑙1
12 𝑺𝐵𝑙𝐶𝑜𝑚𝑝𝑺𝐶
𝐵𝑙𝑿𝐶
𝐵𝑙
13end for
14return 𝑺𝐵𝐿 which is𝑠𝑣
Algorithm 2 HiCom for processing nodes in batches
1InputGV𝑿 fanouts𝑛1𝑛𝐿𝐶𝑜𝑚𝑝 a batch of
target nodesV𝐵V indexed by𝐵 The sequence length is 𝑡
and the number of summary tokens is 𝑘
2Output Summary vectors 𝑺𝐵of nodesV𝐵
3V𝐵𝐿V𝐵
4for𝑙𝐿to1do
5V𝐵𝑙1Sample up to 𝑛𝑙neighbors of 𝑣𝑣𝑉𝐵𝑙
6 𝑿𝐵𝑙1Collect𝑥𝑣and pad to length 𝑡𝑣𝑉𝐵𝑙1
7end for
8𝑺𝐵0 A placeholder of all zeros with shape 𝑉𝐵0𝑘
9for𝑙1to𝐿do
10 𝑿𝐶
𝐵𝑙 A placeholder of all zeros with shape 𝑉𝐵𝑙𝑛𝑙𝑡
11 𝑺𝐶
𝐵𝑙 A placeholder of all zeros with shape 𝑉𝐵𝑙𝑛𝑙𝑘
12 𝑿𝐶
𝐵𝑙𝐵𝑙1𝑿𝐵𝑙1
13 𝑺𝐶
𝐵𝑙𝐵𝑙1𝑺𝐵𝑙1
14 𝑿𝐶
𝐵𝑙 Reshape 𝑿𝐶
𝐵𝑙to𝑉𝐵𝑙𝑛𝑙𝑡
15 𝑺𝐶
𝐵𝑙 Reshape 𝑺𝐶
𝐵𝑙to𝑉𝐵𝑙𝑛𝑙𝑘
16 𝑺𝐵𝑙𝐶𝑜𝑚𝑝𝑺𝐶
𝐵𝑙𝑿𝐶
𝐵𝑙
17end for
18return 𝑺𝐵𝐿 which is 𝑺𝐵
to multiple neighborhoods and is thus involved in multiple compu
tations However one inefficiency that arises from pretokenization
is the presence of dummy tokens within the concatenated sequences
for compression and batching further complicates this problem
For instance in the example in Figure 2 when sequences 𝑥4𝑥8
𝑥9 and 𝑥5𝑥6 are batched into a 29tensor only 7 tokens in
each sequence are meaningful with the remainder being dummy
tokens but appear in different places of the tensor
To optimize memory usage and computation time especially
when processing batched sequences a technique to rearrange and
trim the sequences is employed This technique pushes all dummy
tokens to one end of the sequence allowing them to be trimmed off
thereby conserving memory and speeding up computation In theConference acronym XX June 0305 2018 Woodstock NY Zhang et al
same example the 29tensor can be rearranged and trimmed to 27
to save computation This rearrangeandtrim method optimizes
the efficiency of the HiCom framework It is particularly effective
when node text has diverse lengths or when the fanout parameter
is large We show a pseudocode of this function in Appendix A2

42 Techniques for Efficiency and Effectiveness

In this section we discuss some techniques to improve the efficiency

and effectiveness of HiCom The efficiency perspective is important

as LLMs can have multibillion parameters with considerable GPU

memory requirements We thus introduce implementation tech

niques for batch processing of neighborhoods with different sizes

and reducing the memory consumption of unnecessary padding

These techniques are not restricted to the HiCom algorithm and can

be used for general LLMs to optimize GPU utilization on graph data

From the effectiveness perspective we discuss the idea of summary

accumulation which shares a similar idea to skip connection

421 Batch Processing of Neighborhoods with Different Sizes Algo

rithm 1 presents a straightforward version of HiCom for individual

nodes offering a fundamental understanding of the algorithm In

practice batch processing is essential for maximizing the efficiency

of GPU utilization but it also introduces specific challenges Since

LLMs are innately designed for sequential data the challenge in

adapting HiCom with LLMs for batch processing primarily arises

from the variablelength inputs produced by graph data which

complicates the batching of data instances To address these chal

lenges a more sophisticated padding strategy is needed to uniform

both the length of the text sequence on each node and the number

of neighbors of each node for all nodes in the same level of the

hierarchy Algorithm 2 outlines HiCom for processing nodes in

batches We discuss the algorithm in detail in Appendix A1

422 Pretokenization and Rearrange and Trim Another effective

implementation technique employed in the HiCom framework to

reduce the computational overhead associated with LLMs is pre

tokenization which involves preprocessing the text data into to

kens and padding them with dummy tokens to ensure uniformity in

sequence lengths before the training starts Pretokenization saves

the redundant time cost of raw text encoding when a node belongsAlgorithm 1 HiCom for processing individual nodes

1InputGV𝑿 fanouts𝑛1𝑛𝐿𝐶𝑜𝑚𝑝 a target

node𝑣V

2Output Summary vectors 𝑠𝑣of node𝑣

3V𝐵𝐿𝑣 a batch with a single node ie index 𝐵𝐿𝑣

4for𝑙𝐿to1do

5V𝐵𝑙1Sample up to 𝑛𝑙neighbors of 𝑣𝑣𝑉𝐵𝑙

6 𝑿𝐵𝑙1Collect𝑥𝑣𝑣𝑉𝐵𝑙1

7end for

8𝑺𝐵0

9for𝑙1to𝐿do

10 𝑿𝐶

𝐵𝑙 Concat 𝑿𝐵𝑙1

11 𝑺𝐶

𝐵𝑙 Concat 𝑺𝐵𝑙1

12 𝑺𝐵𝑙𝐶𝑜𝑚𝑝𝑺𝐶

𝐵𝑙𝑿𝐶

𝐵𝑙

13end for

14return 𝑺𝐵𝐿 which is𝑠𝑣

Algorithm 2 HiCom for processing nodes in batches

1InputGV𝑿 fanouts𝑛1𝑛𝐿𝐶𝑜𝑚𝑝 a batch of

target nodesV𝐵V indexed by𝐵 The sequence length is 𝑡

and the number of summary tokens is 𝑘

2Output Summary vectors 𝑺𝐵of nodesV𝐵

3V𝐵𝐿V𝐵

4for𝑙𝐿to1do

5V𝐵𝑙1Sample up to 𝑛𝑙neighbors of 𝑣𝑣𝑉𝐵𝑙

6 𝑿𝐵𝑙1Collect𝑥𝑣and pad to length 𝑡𝑣𝑉𝐵𝑙1

7end for

8𝑺𝐵0 A placeholder of all zeros with shape 𝑉𝐵0𝑘

9for𝑙1to𝐿do

10 𝑿𝐶

𝐵𝑙 A placeholder of all zeros with shape 𝑉𝐵𝑙𝑛𝑙𝑡

11 𝑺𝐶

𝐵𝑙 A placeholder of all zeros with shape 𝑉𝐵𝑙𝑛𝑙𝑘

12 𝑿𝐶

𝐵𝑙𝐵𝑙1𝑿𝐵𝑙1

13 𝑺𝐶

𝐵𝑙𝐵𝑙1𝑺𝐵𝑙1

14 𝑿𝐶

𝐵𝑙 Reshape 𝑿𝐶

𝐵𝑙to𝑉𝐵𝑙𝑛𝑙𝑡

15 𝑺𝐶

𝐵𝑙 Reshape 𝑺𝐶

𝐵𝑙to𝑉𝐵𝑙𝑛𝑙𝑘

16 𝑺𝐵𝑙𝐶𝑜𝑚𝑝𝑺𝐶

𝐵𝑙𝑿𝐶

𝐵𝑙

17end for

18return 𝑺𝐵𝐿 which is 𝑺𝐵

to multiple neighborhoods and is thus involved in multiple compu

tations However one inefficiency that arises from pretokenization

is the presence of dummy tokens within the concatenated sequences

for compression and batching further complicates this problem

For instance in the example in Figure 2 when sequences 𝑥4𝑥8

𝑥9 and 𝑥5𝑥6 are batched into a 29tensor only 7 tokens in

each sequence are meaningful with the remainder being dummy

tokens but appear in different places of the tensor

To optimize memory usage and computation time especially

when processing batched sequences a technique to rearrange and

trim the sequences is employed This technique pushes all dummy

tokens to one end of the sequence allowing them to be trimmed off

thereby conserving memory and speeding up computation In theConference acronym XX June 0305 2018 Woodstock NY Zhang et al

same example the 29tensor can be rearranged and trimmed to 27

to save computation This rearrangeandtrim method optimizes

the efficiency of the HiCom framework It is particularly effective

when node text has diverse lengths or when the fanout parameter

is large We show a pseudocode of this function in Appendix A2