diff --git "a/B9E0T4oBgHgl3EQfyAKb/content/tmp_files/load_file.txt" "b/B9E0T4oBgHgl3EQfyAKb/content/tmp_files/load_file.txt"
new file mode 100644--- /dev/null
+++ "b/B9E0T4oBgHgl3EQfyAKb/content/tmp_files/load_file.txt"
@@ -0,0 +1,1880 @@
+filepath=/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf,len=1879
+page_content='DOES COMPRESSING ACTIVATIONS HELP MODEL PARALLEL TRAINING?' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Song Bian * 1 Dacheng Li * 2 Hongyi Wang 2 Eric P.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Xing 2 3 4 Shivaram Venkataraman 1  ABSTRACT  Large-scale Transformer models are known for their exceptional performance in a range of tasks, but training  them can be difficult due to the requirement for communication-intensive model parallelism.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' One way to improve  training speed is to compress the message size in communication.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Previous approaches have primarily focused on  compressing gradients in a data parallelism setting, but compression in a model-parallel setting is an understudied  area.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' We have discovered that model parallelism has fundamentally different characteristics than data parallelism.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  In this work, we present the first empirical study on the effectiveness of compression methods for model parallelism.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  We implement and evaluate three common classes of compression algorithms - pruning-based, learning-based,  and quantization-based - using a popular Transformer training framework.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' We evaluate these methods across  more than 160 settings and 8 popular datasets, taking into account different hyperparameters, hardware, and both  fine-tuning and pre-training stages.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' We also provide analysis when the model is scaled up.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Finally, we provide  insights for future development of model parallelism compression algorithms.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  1  INTRODUCTION  Transformer models have become the dominant model for  many machine learning tasks (Devlin et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', 2018;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Radford  et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', 2018;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Yang et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', 2019;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Dosovitskiy et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', 2020;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Gong  et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', 2021;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Sharir et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', 2021;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Gong et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', 2021).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' However,  state-of-the-art Transformer models have a large number  of parameters, making it difficult for a single GPU to hold  the entire model.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' As a result, training large Transformer  models often requires partitioning the model parameters  among multiple GPUs, a technique known as model paral-  lelism (Shoeybi et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', 2019;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Rasley et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', 2020).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Model  parallelism strategies often introduce significant commu-  nication overhead, as demonstrated in Figure 1 (Li et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=',  2022).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' For instance, the most commonly used tensor model  parallelism strategy requires two all-reduce operations over  a large tensor in each Transformer encoder block per iter-  ation.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' This can greatly increase the overall computational  cost of training the model (Shoeybi et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', 2019).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  To address the issue of high communication overhead in  model parallelism, one approach is to compress the mes-  sages communicated among GPUs, such as activation val-  ues.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' In the data-parallel setting, several prior works have  explored compressing gradients to reduce the communica-  tion cost of training (Seide et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', 2014;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Bernstein et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=',  2018;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Dettmers, 2015;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Lin et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', 2017;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Wang et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', 2018b;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Equal contribution  1Department of Computer Science, Uni-  versity of Wisconsin-Madison 2Machine Learning Department,  Carnegie Mellon University 3MBZUAI 4Petuum Inc.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='. Correspon-  dence to: Song Bian <songbian@cs.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='wisc.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='edu>.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  (8, 128)  (32, 128)  (32, 512)  Hyper-parameters  0  5  10  15  20  25  30  35  40  Comm.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Overhead  (% of total time)  TP=1, PP=4  TP=2, PP=2  TP=4, PP=1  Figure 1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Communication overhead of model parallelism with dif-  ferent batch sizes and sequence lengths on BERTLarge using Py-  torch 1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='12, NCCL, fp16 and 4 GPUs.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' The x-axis is (batch size,  sequence length)  Vogels et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', 2019).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' However, there has been limited ex-  ploration of compression methods specifically designed for  model parallelism.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Furthermore, it is important to note that  compression in model parallelism is fundamentally different  from compression in data parallelism for two main reasons.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Firstly, as shown in Figure 2, gradients tend to be low-rank,  while activations do not.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Therefore, low-rank gradient com-  pression methods, which have been shown to provide state-  of-the-art end-to-end speedup in communication-efficient  data-parallel training, may not directly apply to model paral-  lelism (Vogels et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', 2019).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Secondly, the performance ben-  efits of gradient compression methods can be significantly  affected by system optimizations in data parallelism (Agar-  wal et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', 2022).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' However, model parallelism has a different  arXiv:2301.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='02654v1  [cs.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='LG]  6 Jan 2023  Does compressing activations help model parallel training?' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  0.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='0  0.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='2  0.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='4  0.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='6  0.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='8  1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='0  Dimension Percentage  0.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='0  0.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='2  0.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='4  0.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='6  0.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='8  1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='0  Sigma Value Percentage  Activation  Gradient  Figure 2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Low-Rank analysis: Curves are drawn by ordering the  singular values of the SVD decomposition.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' The result shows that  the gradient is low-rank but the activation is not.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' The activation is  the output of the 12th transformer layer in BERTLarge model.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  set of system optimization techniques than data parallelism,  so it is unclear how these optimizations would impact the  performance of compression methods in model parallelism.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  In this paper, we present the first systematic study of model  parallelism compression for large Transformer models.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' We  evaluate the impact of different compression methods in  terms of both throughput and accuracy.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' We conduct ex-  periments for both pre-training and fine-tuning tasks.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' (De-  vlin et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', 2018;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Gururangan et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', 2020).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' In particular,  we implement and evaluate popular gradient compression  methods, e.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='g.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Top-K and Random-K as well as a learning-  based compression method, i.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='e.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', auto-encoders (Hinton &  Zemel, 1993), which can not directly be applied to gradi-  ent compression but is compatible with activation compres-  sion.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' To assist researchers and practitioners training new  Transformer-based models (Liu et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', 2019;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Izsak et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=',  2021), we study compression methods using different train-  ing hyper-parameters and hardware setups.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' We also develop  a performance model that can be conveniently used to under-  stand how compression methods would affect throughput  at larger scales.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' In total, we evaluate compression methods  across over 160 different settings with various compression  algorithms, training stages, hyper-parameters, and hardware,  and over 8 datasets (Wang et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', 2018a).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Our findings in-  clude the following takeaways.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Our takeaways.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' 1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Learning-based compression meth-  ods are most suitable for model-parallelism.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' On the fine-  tuning stage(§4.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='2, §4.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='3), only auto-encoders (AEs) can pro-  vide end-to-end speedup (upto 18%) while preserving the  model’s accuracy (within ∼3 GLUE score (Wang et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=',  2018a)).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Top-K, Random-K, and quantization methods  slow down training because their message encoding and de-  coding overhead is larger than the communication time they  reduce.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Top-K and Random-K also hurt model’s accuracy.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  For the pre-training stage (§4.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='4), only AE provides speedup  (upto 16%) while preserving the model’s accuracy (similar  GLUE score).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Top-K marginally improves training time,  but degrades the accuracy.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Quantization slows down the  training time, and degrades the accuracy.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Training hyper-parameters affect the performance  benefits of compression methods.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' None of the compres-  sion methods can improve performance when the batch size  and sequence length are small because the cost of message  encoding and decoding becomes relatively higher (as dis-  cussed in section §4.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='6).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' In practice, we have found that  the batch size and sequence length need to be at least 32  and 512, respectively, for the compression methods to pro-  vide throughput gains.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' The same is true when fine-tuning is  performed on a machine with high-bandwidth NVLink con-  nections between all GPUs (as described in section §4.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='2).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  3.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Early model layers are more sensitive to compression.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Our observations show that compressing the early layers or  too many layers significantly decreases the model’s accuracy  (as discussed in section §4.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='5), which is consistent with the  findings of previous research (Wang et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', 2021).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' In practice,  we have found that compressing the final 12 layers of a 24-  layer Transformer model is an effective approach.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Contributions.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' We make the following contributions:  We conduct the first empirical study on model paral-  lelism compression methods for Transformer models,  considering different compression methods, training  stages, hyper-parameters, and hardware configurations.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  We implement several popular compression algorithms,  including Top-K, Random-K, quantization, and auto-  encoders (AEs), and integrate them into an existing  distributed training system.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  We extensively evaluate these algorithms across over  160 different settings and eight popular datasets.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Based  on our experimental results, we provide several take-  aways for future model parallelism compression stud-  ies.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' We also analyze the speedup when the model size  and cluster size are scaled up.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  2  BACKGROUND AND CHALLENGES  In this section, we first introduce data parallelism and model  parallelism (§2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='1).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Then we introduce the challenges in  model parallelism compression (§2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='2).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='1  Data Parallelism and Model Parallelism  Data Parallelism (DP).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  DP divides the training examples  among multiple workers (Li et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', 2014;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Ho et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', 2013) and  replicates the model at each worker.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' During each iteration,  Does compressing activations help model parallel training?' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  each worker calculates the model gradient based on its as-  signed examples and then synchronizes the gradient with the  other workers (Sergeev & Del Balso, 2018).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' However, DP  requires each worker to compute and synchronize gradients  for the entire model, which can become challenging as the  model size increases.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' One issue is that the large gradients  can create a communication bottleneck, and several previous  studies have proposed gradient compression methods (Seide  et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', 2014;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Bernstein et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', 2018;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Dettmers, 2015;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Lin  et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', 2017;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Wang et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', 2018b) to address this.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Addition-  ally, the worker may not have enough memory to train with  the entire model using even one example, in which case  model parallelism may be necessary.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Model Parallelism (MP).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Model parallelism (MP) di-  vides the model among multiple workers, allowing large  models to be trained by only requiring each worker to main-  tain a portion of the entire model in memory.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' There are two  main paradigms for MP: inter-layer pipeline model paral-  lelism (PP) and intra-layer tensor model parallelism (TP).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' PP  divides the layers among workers, with each worker execut-  ing the forward and backward computations in a pipelined  fashion across different training examples (Narayanan et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=',  2019;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Li et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', 2021).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  For example, a mini-batch of  training examples can be partitioned into smaller micro-  batches (Huang et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', 2019), with the forward computation  of the first micro-batch taking place on one worker while  the forward computation of the second micro-batch hap-  pens on another worker in parallel.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' TP (Lu et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', 2017;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Shazeer et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', 2018;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Kim et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', 2016) divides the tensor  computations among workers.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' In particular, we consider  a specialized strategy developed for Transformer models  that divides the two GEMM layers in the attention module  column-wise and then row-wise, with the same partitioning  applied to the MLP module (Shoeybi et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', 2019).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' However,  TP still involves a communication bottleneck due to the  need for two all-to-all collective operations in each layer,  motivating the use of compression to reduce the communi-  cation overhead of MP (Shoeybi et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', 2019).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' two all-to-all  collective operations in each layer (Shoeybi et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', 2019).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  This bottleneck motivates our study to use compression for  reducing the communication of model parallelism.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='2  Challenges in Model Parallelism Compression  In data parallelism, synchronizing gradients in large models  is a major bottleneck, and several gradient compression al-  gorithms have been proposed (Seide et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', 2014;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Bernstein  et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', 2018;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Dettmers, 2015;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Lin et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', 2017;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Wang et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=',  2018b) to reduce the communication volume.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' These algo-  rithms often rely on the observation that the gradient matrix  is low-rank.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' In model parallelism, we have observed that  communicating activations becomes the bottleneck.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' How-  ever, we have identified three challenges when adapting  gradient compression algorithms for use in model paral-  lelism.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  First, the low-rank observation for gradient matrices does  not hold for activation matrices, as shown in Figure 2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' The  sigma value percentage for activation matrices increases  nearly linearly with the dimension percentage, indicating  that the activation matrix is not low-rank.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Therefore, ap-  plying gradient compression techniques to activations is  likely to result in a significant loss of accuracy.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Second, the  performance of compression methods is heavily influenced  by system optimizations (Li et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', 2020), and many gradi-  ent compression methods do not lead to speed-ups for data  parallelism (Zhang et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', 2017;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Agarwal et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', 2022) due  to competition for GPU resources between gradient encod-  ing computation and backward computation.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' However, the  impact of these optimizations on compression methods in  model parallelism has not been studied.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Third, model par-  allelism introduces the possibility of using learning-based  compression methods, such as autoencoders (AE) (Hinton &  Zemel, 1993), which have not been examined in the gradient  compression literature because they require gradient com-  putations and raise new considerations.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Given these three  challenges, there is a need for a thorough study of the effects  of different compression methods in model parallelism.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  3  IMPLEMENTATION  In this section, we first introduce the compression algo-  rithms we evaluate in this work (§ 3.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='1).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Then, we discuss  implementation details in Sections 3.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='2 and 3.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='3.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  3.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='1  Compression Algorithms  In this work, we evaluate a range of popular compres-  sion algorithms, including sparsification-based approaches,  learning-based approaches, and quantization-based ap-  proaches (as illustrated in Figure 3).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' We use Top-K and  Random-K as sparsification-based approaches, as they have  been well-studied in gradient compression (Stich et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=',  2018).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' We also implement AEs, which compress messages  using a small neural network (Hinton & Zemel, 1993).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' For  quantization, we use the same scheme as in previous re-  search (Wang et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', 2022), but compare its performance to  other compression algorithms in the context of model paral-  lelism, as the prior work only considered pipeline compres-  sion over slow networks.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Since the activation matrices for  models are not low-rank (as shown in Figure 2), low-rank  based compression algorithms (such as PowerSGD (Vo-  gels et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', 2019)) are not suitable for model parallelism  compression, and we do not evaluate any low-rank based  compression algorithms in this work.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Does compressing activations help model parallel training?' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  3.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='2  Tensor Parallelism Compression  We base our implementation on Megatron-LM (Shoeybi  et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', 2019), a popular Transformer models training system  that supports tensor and pipeline model parallelism.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' To  integrate the compression algorithms into Megatron-LM,  we make the following modifications.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' For AE, we compress  the activation before the all-reduce step and invoke the  all-reduce function as usual.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' The implementation of AE  is shown here: for each layer, we have a learnable matrix  w ∈ Rh×c, and the activation X ∈ Rb×s×h, where b is the  batch size, s is the sequence length, h is the hidden size,  and c < h is the compressed size.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' By using the matrix  w, we output the compressed activation Xw ∈ Rb×s×c.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Then, we use a similar technique(a decoder as opposed to  an encoder) to decompress the compressed activation and  propagate it to the next layer.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' However, since the Top-K,  Random-K, and quantization can output two independent  tensors with different types (e.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='g.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', for Top-K values and  their indices), we cannot use torch.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='distributed.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='all-reduce  to sum the tensors up directly.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  In light of this, we  replace the all-reduce step with the all-gather function:  gather-from-tensor-model-parallel-region,  which  is  implemented  by  Megatron-LM.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  We  use  torch.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='topk function to select the k largest absolute  values of the activation and random.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='sample function to  randomly select k values from the activation.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Finally, our  implementation of quantization is based on code released  by (Wang et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', 2022).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  3.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='3  Pipeline Parallelism Compression  Megatron-LM can only send one tensor to the next pipeline  stage per round, so we modify its communication functions  to allow for the transmission of multiple tensors per round  in order to integrate Top-K, Random-K, and quantization.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Since we compress the activation in the forward step, us-  ing compression also reduces the size of the gradient for  activation and thus the communication cost in the backward  step.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' However, this is not the case when using quantization  to compress the activation for models.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' This is because, as  previously noted (Wang et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', 2022), the Pytorch backward  engine only supports gradients for floating point tensors,  and therefore the size of the gradient is the same as the size  of the decompressed activation.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Our implementation also  allows the integration of error-feedback compression algo-  rithms by retaining the error information from the previous  compression step.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  4  EXPERIMENTS  We next perform experiments using our implementation to  answer the following questions:  What is the impact of activation compression on system  throughput and which compression method achieves  the best throughput?' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  What is the impact on the model’s accuracy?' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  How different network bandwidths affect the best com-  pression method?' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  How do hyper-parameters such as the batch size and  sequence length affect the benefits of compression?' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  We answer these questions in the context of two commonly  used scenarios in language modeling: fine-tuning on the  GLUE benchmark (Wang et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', 2018a), and pre-training  on the Wikipedia (Devlin et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', 2018) dataset and the  BooksCorpus (Zhu et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', 2015) dataset.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  4.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='1  Experimental Setup  In this section, we briefly describe the hardware, model, and  other experiment settings.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  System Configuration.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' To measure the performance of  compression algorithms over different hardware, our ex-  periments are conducted on two different setups.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Our  first setup uses AWS p3.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='8xlarge machines which have 4  Tesla V100 GPUs with all GPUs connected by NVLink.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  AWS p3.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='8xlarge instances have 10 Gbps network band-  width across instances.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Our second setup uses a local ma-  chine which also has 4 Tesla V100 GPUs but does not have  NVLink.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' All the GPUs are connected by a single PCIe  bridge.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' The local server runs Ubuntu 18.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='04 LTS and the  server has 125GB of memory.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Model.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  We use the BERTLARGE model provided by  Megatron-LM (Shoeybi et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', 2019) which has 345M pa-  rameters.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' We configure the model to have 24 layers with  each layer having a hidden size of 1024 and 16 attention  heads.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' We use fp16 training to train the BERTLARGE model.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Experimental Settings.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' For fine-tuning, we follow the  previous work (Devlin et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', 2018;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Liu et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', 2019), and  use micro-batch size 32 and sequence length 512 by de-  fault.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' We use one machine with 4 V100 GPUs and vary  the tensor model-parallel size and the pipeline model-  parallel size across the following three parallelism degrees:  {(1, 4), (2, 2), (4, 1)}, where the first number of the tuple  represents the tensor model-parallel degree and the second  number of the tuple stands for the pipeline model-parallel  degree.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' To investigate the impact of hyper-parameters, we  conduct experiments that vary the batch size from {8, 32},  and sequence length from {128, 512} on fine-tuning.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  For pre-training, we use micro-batch size 128, global batch  size 1024, and sequence length 128.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' To study the impact of  the distributed settings, we use the following three different  parallelism degrees: {(2, 8), (4, 4), (8, 2)}, where the first  Does compressing activations help model parallel training?' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  g  g  C  C  C  C  Machine 1  Machine 2  C  Transformer Layer  Transformer Layer  Activation  Transformer Layer  Machine 1,2  Machine 3,4  DC  DC  DC  DC  Micro-batch  C  DC  Transformer Layer  Transformer Layer  Transformer Layer  Activation  DC  Figure 3.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Illustration of compression on a 6-Layer Transformer model with 4 machines.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Machine 1 and Machine 2 maintain the first three  layers according to the TP strategy (pipeline stage 1).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' g stands for an all-reduce operation in the forward pass.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' A compression method C  is used to reduce the message size for the all-reduce operation to reduce TP communication time.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Correspondingly, a de-compression  method DC is used after the communication.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' For instance, if AE are used, then C is an encoder, and DC is a decoder.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Machine 3 and  Machine 4 are responsible for the last three layers (pipeline stage 2).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' A compression method is used before Machine 1 sends the activation  to Machine 3, and before Machine 2 sends the activation to Machine 4 to reduce PP communication time.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' The goal of this paper is to  study the effect of different pairs of C and DC.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  number of the tuple represents the tensor model-parallel  degree and the second number of the tuple represents the  pipeline model-parallel degree.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  We also evaluate compression algorithms with different  parameters.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' For AE, we use different dimension after com-  pression from {50, 100}.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' For Top-K and Random-K algo-  rithms, we use two comparable settings: (1) Keep the same  compression ratio as AE (i.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='e.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', we compress the activation  around 10 and 20 times.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=') (2) Keep the same communica-  tion cost as AE.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Finally, we also tune the parameters for  quantization and compress the activation to {2, 4, 8} bits.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  By default, we perform experiments on BERTLarge model  with 24 layers and compress the activation for the last 12  layers.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' For instance, when the pipeline model-parallel de-  gree is 2 and tensor model-parallel degree is 2, we compress  the activation between two pipeline stages and the communi-  cation cost over tensor parallelism in the last 12 layers.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' We  also vary the number of layers compressed in Section 4.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='5.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  4.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='2  Throughput Benefits for Fine-Tuning  Takeaway 1 Using non-learning-based compression tech-  niques to compress activations only slightly improves system  throughput (by 1% or less) due to the large overhead of these  methods.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' However, we see end-to-end speedups of up to  Notation  Description  A1  AE with encoder output dimension 50  A2  AE with encoder output dimension 100  T1/R1  Top/Rand-K: same comm.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' cost as A1  T2/R2  Top/Rand-K: same comm.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' cost as A2  T3/R3  Top/Rand-K: same comp.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' ratio as A1  T4/R4  Top/Rand-K: same comp.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' ratio as A2  Q1  Quantization: reduce the precision to 2 bits  Q2  Quantization: reduce the precision to 4 bits  TP  Tensor model-parallelism degree  PP  Pipeline model-parallelism degree  Table 1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Notation Table.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' For ease of notation, we use TP/PP to  denote the degree of tensor/pipeline model parallelism.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' ‘comm’  and ‘comp’ are short for ‘communication’ and ‘compression’.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  17.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='8% when using learning-based compression methods on  a machine without NVLink.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  When running fine-tune experiments on a p3.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='8xlarge in-  stance on Amazon EC2, we cannot improve system through-  put by using non-learning-based compression algorithms  (Table 2).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Comparing Tables 2 and 3, we can see that the net-  Does compressing activations help model parallel training?' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Distributed Setting  w/o  A1  A2  T1  T2  T3  T4  TP=1, PP=4  591.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='96  591.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='36  591.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='47  594.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='81  595.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='53  599.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='65  605.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='05  TP=2, PP=2  440.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='71  437.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='98  444.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='02  465.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='73  473.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='64  493.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='16  528.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='93  TP=4, PP=1  261.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='48  270.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='22  275.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='54  314.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='37  323.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='90  356.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='57  409.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='23  Distributed Setting  w/o  R1  R2  R3  R4  Q1  Q2  TP=1, PP=4  591.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='96  749.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='56  1,008.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='64  1,824.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='36  5,572.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='87  595.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='29  595.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='45  TP=2, PP=2  440.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='71  3,377.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='59  6,616.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='30  17,117.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='01  71,058.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='64  489.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='27  486.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='54  TP=4, PP=1  261.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='48  3,254.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='01  6,561.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='22  16,990.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='88  65,121.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='79  347.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='68  350.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='50  Table 2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' The average iteration time (ms) for fine-tuning with various compression techniques by varying the distributed setting.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' The  results are collected from the AWS p3.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='8xlarge machine with NVLink by using batch size 32, and sequence length 512.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' The best setting is  bolded in the table.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' And the settings which see benefits compared with the baseline, are underlined.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  With NVLink  w/o  A1  A2  TP=1, PP=4  591.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='96  591.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='36  591.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='47  TP=2, PP=2  440.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='71  437.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='98  444.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='02  TP=4, PP=1  261.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='48  270.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='22  275.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='54  Without NVLink  w/o  A1  A2  TP=1, PP=4  633.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='17  620.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='10  620.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='44  TP=2, PP=2  646.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='14  586.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='65  595.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='25  TP=4, PP=1  736.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='01  624.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='62  636.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='15  Table 3.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' The  average  iteration  time  (ms)  for  fine-tuning  with/without NVLink.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' We compare time without compression  and with AE on different distributed settings, with batch size 32,  and sequence length 512.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' The best setting on each machine is  bolded.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' And the settings, under which we can gain benefits com-  pared with the baseline, are underlined.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  work bandwidth across the GPUs can affect the performance  benefits from compression.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' In other words, we can improve  system throughput by at most 17.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='8% when compressing  activation for fine-tuning tasks on a 4-GPU machine without  NVLink.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' That’s because, without NVLink, the communica-  tion time for model parallelism is much longer.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Thus, while  the message encoding and decoding time remain unchanged,  compression methods can provide more throughput benefits  across lower bandwidth links.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Furthermore, from Tables 2 and 3, we observe that AE out-  performs other compression methods.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' In Table 4, we break-  down the time taken by each algorithm and find that Top-K,  Random-K and quantization have large encoding/decoding  overheads and thus cannot provide end-to-end throughput  improvements.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Although AE slightly increases the time  taken by the backward step, the ∼ 2× reduction in commu-  nication time and the limited encoding/decoding overhead  lead to better overall throughput.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  4.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='3  Effect of Compression on Model Accuracy while  Fine-tuning  Takeaway 2 Among all evaluated compression algorithms,  only AE and quantization preserve fine-tuning accuracy.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  From Table 5, we can see that, when using AE and quan-  tization algorithm for compression, the accuracy loss is  within 3% except for CoLA and RTE.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' In Figure 2, we have  shown that the activation for models is not low-rank.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' There-  fore, sparsification-based compression algorithms (Top-  K/Random-K) lose important information and do not pre-  serve model accuracy.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Given that there is significant accu-  racy difference for CoLA and RTE, we study the impact  of varying the number and range of layers compressed for  these two datasets in Section 4.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='5.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  4.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='4  Throughput Benefits for Pre-training  Takeaway 3 Only AE and Top-K algorithms improve  throughput when performing distributed pre-training.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  First, we recap the experimental environment here.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' For pre-  training, we use 4 p3.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='8xlarge instances on Amazon EC2  and each instance has 4 GPUs with NVLink.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' From Table 6,  we can see that using Top-K and AE can speed up pre-  training by 7% and 16% respectively.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Among the three  distributed settings, TP=4, PP=4 is the best setting for  pre-training.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' That is because the communication cost of  tensor parallelism is larger than that of pipeline parallelism  and with TP=4, tensor parallel communication happens over  faster NVLinks.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Takeaway 4 Compressing activation for models can im-  prove throughput for pre-training by 16%.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  From Table 7, we notice that using AE and Top-K can  reduce the waiting time and pipeline communication time  of pre-training.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' This is because the inter-node bandwidth  (10Gbps) is smaller than the intra-node bandwidth (40GB/s  Does compressing activations help model parallel training?' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Compression  Algorithm  Forward  Backward  Optimizer  Waiting &  Pipeline Comm.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Total Time  Tensor Enc.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Tensor Dec.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Tensor  Comm.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  w/o  276.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='34  354.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='16  5.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='80  9.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='83  646.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='14  \\  \\  150.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='72  A1  213.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='83  362.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='61  6.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='16  4.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='06  586.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='65  2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='16  3.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='12  80.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='88  A2  219.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='01  366.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='51  5.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='67  4.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='07  595.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='25  3.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='12  4.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='56  84.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='48  T1  298.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='93  355.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='71  6.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='79  4.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='38  665.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='81  70.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='08  13.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='68  85.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='20  T2  305.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='47  355.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='51  6.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='36  3.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='91  671.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='24  70.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='32  16.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='80  87.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='84  T3  331.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='70  356.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='80  5.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='78  5.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='00  699.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='27  72.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='24  27.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='36  100.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='80  T4  376.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='72  359.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='19  5.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='89  6.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='60  748.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='41  74.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='88  45.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='36  124.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='56  R1  2,408.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='68  357.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='02  6.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='10  7.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='68  2,779.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='49  2,040.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='24  15.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='84  104.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='16  R2  4,696.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='99  356.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='33  6.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='28  6.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='20  5,065.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='80  4,244.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='64  19.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='44  135.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='84  R3  12,603.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='79  362.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='13  6.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='81  25.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='28  12,998.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='01  11,499.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='12  29.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='76  139.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='92  R4  46,968.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='21  365.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='36  7.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='61  22.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='81  47,363.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='98  44,038.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='56  47.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='52  567.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='36  Q1  274.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='03  354.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='56  5.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='88  7.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='98  642.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='46  20.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='64  32.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='16  91.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='68  Q2  282.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='64  354.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='55  5.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='58  7.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='58  650.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='36  19.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='92  30.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='24  104.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='64  Table 4.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' We breakdown the average iteration time (ms) for fine-tuning with various compression techniques when using TP=2 and PP=2,  batch size 32, and sequence length 512.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' The results are collected from the local machine without NVLink.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' The total time (ms) is divided  into following parts: forward step, backward step, optimizer, and waiting & pipeline communication.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' The last three columns further  breakdown the tensor encoder/decoder and communication times which are considered part of the forward step.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Compression  Algorithm  MNLI-(m/mm)  QQP  SST-2  MRPC  CoLA  QNLI  RTE  STS-B  Avg.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  w/o  88.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='07/88.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='70  92.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='02  95.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='07  88.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='46  62.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='22  93.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='39  82.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='67  89.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='16  86.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='64  A1  85.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='42/85.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='43  91.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='07  92.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='09  86.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='14  54.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='18  91.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='31  70.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='04  87.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='61  82.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='59  A2  85.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='53/85.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='65  91.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='24  93.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='23  85.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='86  55.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='93  91.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='01  65.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='34  87.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='76  82.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='40  T1  32.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='05/32.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='18  74.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='31  83.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='60  70.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='78  0.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='00  58.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='37  51.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='99  0.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='00  44.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='81  T2  44.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='12/45.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='67  39.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='68  90.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='83  78.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='09  0.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='00  84.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='42  49.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='82  62.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='70  55.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='04  T3  36.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='12/36.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='08  74.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='75  90.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='25  81.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='51  0.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='00  85.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='41  54.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='15  0.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='00  50.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='92  T4  83.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='85/84.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='41  56.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='39  93.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='69  83.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='65  0.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='00  90.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='54  59.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='21  86.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='02  70.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='86  Q1  87.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='25/87.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='81  91.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='71  93.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='46  87.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='01  55.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='99  61.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='38  67.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='51  88.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='02  80.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='02  Q2  87.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='85/88.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='47  91.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='93  93.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='23  87.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='42  57.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='67  93.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='01  78.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='34  87.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='43  85.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='04  Table 5.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Fine-tuning results over GLUE dataset under the setting that the tensor model-parallel size is 2 and pipeline model-parallel size is  2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' F1 scores are reported for QQP and MRPC, Matthews correlation coefficients are reported for CoLA, and Spearman correlations are  reported for STS-B, and accuracy scores are reported for the other tasks.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  with NVLink), so compression is effective at reducing the  communication time between two pipeline stages.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' From  Table 9, we can observe that, by using A2 to compress  the activation over the last 12 layers, we can reduce the  communication cost between two pipeline stages effectively.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Takeaway 5 Among all evaluated methods, AE is the  best strategy to compress activation over pre-training.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' It  achieves higher pre-training throughput and preserves the  model’s accuracy.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  From Table 8, compared with the baseline (without com-  pression), we can observe that using AE is able to keep  the accuracy when compared to the uncompressed model.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  In addition, we observe that we can use the AE at the pre-  training phase and remove it during the fine-tuning phase.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  In other words, we only need to load the parameter of the  BERTLarge model to do fine-tuning, and the parameters of  the AE can be ignored.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Furthermore, Table 8 shows that pre-  trained models suffer significant accuracy loss when using  Top-K for compression.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Finally, quantization can preserve  the model’s accuracy, but we cannot achieve end-to-end  speedup by using quantization as strategy to compress ac-  Does compressing activations help model parallel training?' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Distributed Setting  w/o  A1  A2  T1  T2  T3  T4  TP=2, PP=8  1,625.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='16  1,550.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='18  1,579.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='70  1,508.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='34  1,503.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='54  1,593.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='37  1,682.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='87  TP=4, PP=4  1,422.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='40  1,242.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='97  1,223.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='20  1,360.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='37  1,352.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='61  1,410.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='47  1,721.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='87  TP=8, PP=2  15,642.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='30  14,577.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='29  14,073.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='45  14,308.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='12  14,543.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='81  18,919.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='92  27,152.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='07  Distributed Setting  w/o  R1  R2  R3  R4  Q1  Q2  TP=2, PP=8  1,625.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='16  10,308.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='03  20,814.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='20  55,925.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='28  >100,000  1,759.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='27  1,752.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='24  TP=4, PP=4  1,422.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='40  15,433.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='12  31,565.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='19  87,421.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='46  >100,000  2,435.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='03  2,594.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='94  TP=8, PP=2  15,642.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='30  32,522.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='47  61,049.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='87  >100,000  >100,000  16,414.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='57  16,517.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='44  Table 6.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' The average iteration time (ms) for pre-training with various compression techniques by varying the distributed setting.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' The  results are collected from 4 AWS p3.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='8xlarge machines with NVLink by using micro-batch size 128, global batch size 1024, and sequence  length 128.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' The best setting is bolded in the table.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' And the settings, under which we can gain benefits compared with the baseline, are  underlined.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Compression  Algorithm  Forward  Backward  Optimizer  Waiting &  Pipeline Comm.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Total Time  Tensor Enc.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Tensor Dec.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Tensor  Comm.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  w/o  467.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='73  419.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='26  7.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='42  527.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='99  1,422.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='40  \\  \\  91.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='08  A1  546.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='95  455.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='26  7.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='29  233.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='47  1,242.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='97  8.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='64  16.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='20  32.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='76  A2  459.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='26  467.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='51  9.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='64  286.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='78  1,223.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='20  12.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='96  20.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='52  43.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='56  T1  712.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='22  423.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='91  7.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='21  217.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='03  1,360.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='37  73.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='44  140.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='4  80.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='28  T2  671.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='19  424.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='27  7.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='35  249.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='80  1,352.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='61  81.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='00  170.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='64  81.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='36  T3  813.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='03  433.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='42  7.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='35  156.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='67  1,410.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='47  108.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='00  268.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='92  115.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='92  T4  1,068.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='38  444.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='26  6.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='75  202.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='48  1,721.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='87  153.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='36  427.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='68  151.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='56  R1  14,199.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='56  421.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='40  4.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='23  807.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='93  15,433.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='12  13,185.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='72  181.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='44  193.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='68  R2  29,344.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='85  427.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='18  3.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='91  1,789.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='25  31,565.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='19  27,975.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='24  181.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='44  187.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='20  R3  78,906.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='91  444.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='88  6.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='08  3,707.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='37  83,065.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='23  73,847.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='16  279.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='72  649.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='44  Q1  803.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='63  417.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='33  8.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='61  1,205.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='46  2,435.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='03  90.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='72  304.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='56  193.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='68  Q2  805.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='33  417.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='74  7.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='55  1,364.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='32  2,594.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='94  85.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='32  271.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='08  111.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='60  Table 7.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' We breakdown the average iteration time (ms) for pre-training with various compression techniques when using tensor model-  parallel size 4, pipeline model-parallel size 4, micro batch size 128, global batch size 1024, and sequence length 128.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' The results are  collected from 4 AWS p3.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='8xlarge machines with NVLink.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Compression  Algorithm  MNLI-(m/mm)  QQP  SST-2  MRPC  CoLA  QNLI  RTE  STS-B  Avg.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  w/o  84.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='87/84.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='79  91.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='25  92.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='43  86.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='84  56.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='36  92.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='26  70.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='40  86.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='83  82.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='89  A2  83.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='77/84.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='32  91.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='14  91.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='63  86.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='55  58.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='61  91.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='96  71.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='48  87.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='16  82.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='96  T2  61.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='06/60.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='93  80.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='74  80.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='16  63.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='83  10.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='01  59.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='55  47.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='29  0.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='37  51.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='55  Q2  84.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='47/85.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='32  91.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='36  93.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='23  85.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='10  58.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='84  91.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='69  71.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='84  86.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='39  83.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='14  Table 8.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Fine-tuning results over GLUE dataset by using the checkpoint obtained by pre-training.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' F1 scores are reported for QQP and  MRPC, Matthews correlation coefficient is reported for CoLA, and Spearman correlations are reported for STS-B, and accuracy scores  are reported for the other tasks.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  tivation over pre-training.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' In conclusion, it is not a good  choice to compress the activation by using quantization or  Top-K.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  4.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='5  Varying Compression Layers and Location  Takeaway 6 When the number of compressed layers in-  creases, the model accuracy decreases.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  From Figure 4(a), we can observe that the accuracy for RTE  Does compressing activations help model parallel training?' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Pipeline Stages  Comm.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' (w/o)  Comm.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' (A2)  0 ↔ 1  77.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='82  76.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='13  1 ↔ 2  88.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='69  13.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='19  2 ↔ 3  97.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='67  14.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='09  Table 9.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' The average communication time (ms) per iteration be-  tween two pipeline stages.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' The first column indicates the pipeline  stage.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' And the second column shows the communication time  per iteration without compression.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Moreover, the third column  presents the communication time with A2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' We only compress  the activation in the last 12 layers and thus the time for the first  pipeline stage is unchanged.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  w/o  6  8  10  12  14  16  18  Number of Layers Compressed  0  20  40  60  80  100  Metrics (%)  CoLA  RTE  (a) Vary Number of Layers Compressed  1-12  4-15  7-18  10-21  13-24  w/o  Compression Location  0  20  40  60  80  100  Metrics (%)  CoLA  RTE  (b) Vary Compression Location  Figure 4.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Fine-tuning results over CoLA and RTE datasets by vary-  ing the compression location and number of layers compressed.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  The above figure shows that model performance vs the number  of layers compressed.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' The below figure shows that model per-  formance versus the compression location.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' We use tensor model-  parallel degree 2, pipeline model-parallel degree 2, batch size 32,  and sequence length 512.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  and the matthews correlation coefficient for CoLA decreases  as we increase the number of layers compressed.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' This is  because as we increase number of layers compressed, we  lose more information in the activations leading to a loss in  accuracy.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' From Figure 4(a), we observe that compressing  activations of the last 8 layers is the best strategy to keep  the accuracy loss within 3% for both datasets.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Takeaway 7 Compressing the activation for the initial lay-  ers harms the accuracy of the model.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  We keep the number of layers compressed constant and  vary the location where we apply compression (Figure 4(b)).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  The results indicate that compressing activations of the first  few layers of the model significantly harms the model’s  accuracy.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' This is because compressing activations generates  error and the error in the early layers can be accumulated  and propagated to later layers.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  4.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='6  Impact of Model Hyper-parameters  Takeaway 8 Using a smaller batch size or sequence length  for fine-tuning negates the throughput benefits from com-  pression because of the smaller communication cost.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  We vary the batch size from {8, 32} and sequence length  from {128, 512}, and report the results in Table 11-14.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' We  provide more detailed experimental results in Appendix A.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  We notice that when the communication cost over model par-  allelism is small, the overhead of the compression methods  can become the bottleneck.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Therefore, we cannot improve  system throughput when using compression algorithms with  batch size 8 and sequence length 128.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  4.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='7  Performance Analysis  In this section, we develop an analytical cost model to an-  swer the question:  What will happen if we scale up the model size and the  cluster size?' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Given that prior works (Li et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', 2022) have analyzed the  complexity of various model parallelism strategies, we only  consider a fixed strategy of using tensor model parallelism  here.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Concretely, we use tensor model parallelism in the  same node, and pipeline model parallelism across the node,  a suggested strategy according to (Narayanan et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', 2021).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  In particular, we build the performance analysis for real-  world settings similar to (Narayanan et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', 2019) in two  steps.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' First, we develop our own model on a single-node  scenario, and we scale up the model size on a single node.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Second, we increase the cluster size and, according to the  model-parallelism strategy we choose, assign additional  GPUs to pipeline parallelism, and use off-the-shelf pipeline  parallelism cost models to predict the performance (Li et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=',  2022;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Zheng et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', 2022).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Denote the vocabulary size as V , hidden size as h, sequence  length as s, and batch size as B.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' From (Narayanan et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=',  2021), we know that the number of floating points opera-  tions (FLOPs) and all-reduce message size in a Transformer  layer is 96Bsh2 + 16Bs2h, and Bsh respectively.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  If we do not use compression methods, the total time of a  Transformer layer can be modeled as a sum of the all-reduce  communication step and the computation time step.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' We note  Does compressing activations help model parallel training?' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='2500  ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='5000  ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='7500  ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='10000 12500  ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='Hidden size  ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='0  ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='50  ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='100  ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='150  ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='200  ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='Run-time (ms)  ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='bs16(pred)  ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='bs32(pred)  ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='bs64(pred)  ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='bs128(pred)  ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='bs16  ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='bs32  ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='bs64  ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='bs128  ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='(a) Tcomp  ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='2500  ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='5000  ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='7500  ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='10000  ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='12500  ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='Hidden size  ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='0  ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='20  ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='40  ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='60  ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='Run-time (ms)  ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='bs16(pred)  ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='bs32(pred)  ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='bs64(pred)  ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='bs128(pred)  ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='bs16  ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='bs32  ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='bs64  ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='bs128  ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='(b) Tcomm  ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='2500  ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='5000  ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='7500  ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='10000  ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='12500  ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='Hidden size  ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='0  ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='1  ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='2  ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='3  ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='4  ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='5  ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='Run-time (ms)  ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='bs16(pred)  ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='bs32(pred)  ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='bs64(pred)  ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='bs128(pred)  ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='bs16  ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='bs32  ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='bs64  ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='bs128  ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='(c) Toverhead  ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='2500  ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='5000  ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='7500  ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='10000 12500  ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='Hidden Size  ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='1  ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='2  ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='3  ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='4  ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='5  ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='Speedup  ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='bs16(pred)  ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='bs32(pred)  ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='bs64(pred)  ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='bs128(pred)  ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='bs16  ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='bs32  ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='bs64  ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='bs128  ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='(d) Speedup  ' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='Figure 5.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' We vary the batch size and the hidden size to show that our prediction model is accurate compared with the real experimental  results.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' The model we use here has only one transformer layer and the tensor model-parallel size is 4.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' In specific, Figure (a) shows the  real and predicted computation time with the increase of the hidden size.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Figure (b) presents the real and predicted communication time  between tensor parallelism by varying the hidden size.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' As for the Figure (c), it presents the computation time of AE with the increase of  hidden size.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' In the end, Figure (d) show the total speedup when we use AE to compress activations over tensor parallelism.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  that these two steps can hardly overlap because , the reason  behind it is that the all-reduce communication depends on  the previous computational results:  T = Tcomp(96Bsh2 + 16Bs2h) + Tcomm(Bsh)  (1)  Modeling Tcomp  We model Tcomp as a linear function of  FLOPs with the coefficient α that corresponds to the peak  performance of the GPU.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' In particular, we estimate α using  ground truth wall clock time of the largest hidden size we  can fit, where the GPU is more likely to be of the peak  utilization (Williams et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', 2009).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' During experiments, we  found that fitting α using time of smaller hidden sizes can  result in a 30x higher prediction time when we scale up the  hidden size because of low GPU utilization.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Our prediction  versus the ground truth time is plotted in Figure 5(a).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Modeling Tcomm  we model Tcomm as a piece-wise func-  tion of the message size (Agarwal et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', 2022).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Formally,  Tcomm(Bsh) =  �  c  if Bsh < d  βBsh  if Bsh ≥ d  If the message size is smaller than a threshold d, then  Tcomm(Bsh) is a constant c because the worker needs to  launch one communication round (Li et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', 2020).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Other-  wise, the number of communication rounds is proportional  to the message size.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' The fitting result is in Figure 5(b).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Using AE as the compression method and a fixed encoder  dimension e (we set e to 100 in this section), the total time  of a single Transformer layer is:  TAE = Tcomp(96Bsh2 + 16Bs2h) + Tcomm(Bse)  + Toverhead  Compared with the setting without compression, the compu-  tation time remains unchanged.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' In addition, Tcomm(Bse)  is roughly equal to c because Bse is usually smaller than  the threshold d.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' In our experiments, the threshold d =  16 × 128 × 100 = 409600 and c ≈ 0.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='2.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Modeling Toverhead  In AE, Toverhead is the encoder and  decoder computation time.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' It is a batched matrix multiplica-  tion with input dimension B × s × h and h × e.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Assuming  e is kept constant, it can be modeled as Toverhead = γBsh.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  The fitting result is shown in 5(c).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Since each Transformer layer has identical configurations in  popular Transformer models (Devlin et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', 2018;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Radford  et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', 2018), the overall speedup ratio is the same as we vary  the number of layers.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Thus, we can estimate the speedup of  different hidden sizes of any number of Transformer layers  using  T  TAE .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' We provide the fitting result in Figure 5(d).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Understanding the trend  We consider the asymptotic  behavior of large hidden size h:  T  TAE  ≈  α(96Bsh2 + 16Bs2h) + βBsh  α(96Bsh2 + 16Bs2h) + γBsh + c  (2)  Thus, we can see that as hidden layer size increases, the  benefits from compression diminish.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Scaling up the cluster size  Next we analyze the speedup  when scaling up the cluster size by combining the pipeline  parallelism cost model developed in (Li et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', 2022;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Zheng  et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', 2022).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Formally, the running time is modeled as a  sum of per-micro-batch pipeline communication time, per-  micro-batch of non-straggler pipeline execution time, and  the per-mini-batch straggler pipeline execution time.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' To use  the cost model, we denote the micro-batch size as m, the  number of nodes n, the number of layers L, the pipeline  communication time p or pAE.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  We use the default pipeline layer assignment strategy  in (Shoeybi et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', 2019), which balances the number of  transformer layers.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Thus, every stage takes the same time in  our scenario: L  nT or L  nTAE.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' We use the pipeline communi-  cation model in (Jia et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', 2019;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Li et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', 2022), p = Bsh  w ,  pAE = Bse  w , where w is the bandwidth.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Thus the overall  speedup can be written as:  ( m−1  n  + 1) × LT + (n − 1) × Bsh  w  ( m−1  n  + 1) × LTAE + (n − 1) × Bse  w  (3)  Does compressing activations help model parallel training?' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  From the Table 10, we see that we can maintain a ∼1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='5x  speedup as we scale the hidden size to 25600.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' This shows  that if we increase the number of nodes when we increase in  hidden size, AE compression retains its benefits.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' However,  it is possible to avoid the diminishing speedup by properly  scaling up the number of nodes n, where the speedup will  asymptotically converge to h  e .' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  In summary, compression in model parallelism has dimin-  ishing returns if we only scale up the model on a fixed  cluster.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' To gain benefits from compression methods, one  needs to also properly manage other parameters in the  cost model, e.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='g.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' also scaling up the number of nodes and  use the pipeline parallelism.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  5  RELATED WORK  In this section, we first introduce work related to the de-  velopment of large Transformer models.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Then, we discuss  strategies to train these models at scale.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' In the end, we  discuss prior work that accelerates distributed ML models  training by using compression techniques.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Transformer Models.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Transformer models were first intro-  duced by Vaswani et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' (2017) in the machine translation  context.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' It has been shown to be effective in various other  language understanding tasks such as text generation, text  classification and question answering (Devlin et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', 2018;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Radford et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', 2018;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Wang et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', 2018a;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Rajpurkar et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=',  2016).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Recent research has also successfully applied Trans-  former models to images (Dosovitskiy et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', 2020;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Touvron  et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', 2021), audio (Gong et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', 2021) and beyond (Sharir  et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', 2021).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' An N-layers transformer model is composed  of three major components: (1) An embedding layer that  maps an input token to a hidden state, (2) A stack of N  transformer layers, and (3) a prediction layer that maps the  hidden state proceeded by transformer layers to the task  output.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' A transformer layer is composed of an attention  module (Bahdanau et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', 2014) and several matrix multipli-  cations.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Several optimizations have been proposed to speed  up Transformer model training such as carefully managing  the I/O (Dao et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', 2022) and reducing the complexity of the  attention module (Wang et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', 2020).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' In this work, we speed  up the Transformer model training in the distributed setting,  where we reduce the communication between workers.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Training Large Transformer models.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Several parallelism  strategies have been proposed to train Transformer mod-  els.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Megatron (Shoeybi et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', 2019) proposes tensor model  parallelism, which parallelizes the computation in attention  layers and in the following matrix multiplications.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Deep-  Speed (Rasley et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', 2020) uses a specialized form of  pipeline parallelism (Huang et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', 2019;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Narayanan et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=',  2019) that treats a transformer layer as the smallest unit  in pipeline stages.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' It further combines the tensor model  parallelism developed in Megatron and data parallelism to  train Transformer models at the scale of trillion parame-  ters.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' (Li et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', 2022) considers a more sophisticated model  parallelism strategy space for Transformer models and uses  a cost model to automatically search for the optimal one.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Our work is orthogonal to the direction of developing new  parallel training strategies.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' In this work, we study how to  compress communication on existing parallel strategies.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Distributed training with Compression.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Distributed ML  model training requires frequent and heavy synchronization  between workers.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Several directions have been proposed  to reduce the communication bottleneck by compressing  the message size.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' One direction is developed on the data  parallelism setting, where workers communicate model gra-  dients (Wang et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', 2021;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Agarwal et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', 2022) during  backward propagation.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Common techniques to reduce the  gradient communication include low-rank updates (Wang  et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', 2018b), sparsification (Lin et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', 2017), and quanti-  zation (Seide et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', 2014;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Bernstein et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', 2018;' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Dettmers,  2015).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' A more recent direction find that the activation pro-  duced during the forward propagation in neural networks is  large, and thus compressing them is beneficial (Wang et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=',  2022).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' In particular, they use quantization to compress the  activation volume between pipeline parallelism workers.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  However, they focus on the geo-distributed setting where  the network bandwidth is very low.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' In this paper, we study  the effect of a rich set of popular compression techniques  on tensor and pipeline parallelism, and in a typical cloud  computing setting.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  6  CONCLUSION  In this work, we studied the impact of compressing acti-  vations for models trained using model parallelism.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' We  implemented and integrated several popular compression  algorithms into an existing distributed training framework  (Megatron-LM) and evaluated their performance in terms  of throughput and accuracy under various settings.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Our re-  sults show that learning-based compression algorithms are  the most effective approach for compressing activations in  model parallelism.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' We also developed a performance model  to analyze the speedup when scaling up the model.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Our ex-  periments provide valuable insights for the development of  improved activation compression algorithms in the future.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Acknowledgments  Shivaram Venkataraman is supported by the Office of the  Vice Chancellor for Research and Graduate Education at  UW-Madison with funding from the Wisconsin Alumni  Research Foundation.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Eric Xing is supported by NSF  IIS1563887, NSF CCF1629559, NSF IIS1617583, NGA  HM04762010002, NSF IIS1955532, NSF CNS2008248,  NSF IIS2123952, and NSF BCS2040381.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Does compressing activations help model parallel training?' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  hidden size  number of layers  number of nodes  batch size  speedup  6144  40  1  1024  1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='91×  8192  48  2  1536  1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='75×  10240  60  4  1792  1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='63×  12288  80  8  2304  1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='55×  16384  96  16  2176  1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='46×  20480  105  35  2528  1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='46×  25600  128  64  3072  1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='47×  Table 10.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Weak-scaling speedup for the Transformer models.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' The number of tensor model parallelism is 4, and the micro-batch size is 16.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  As for the change of the hidden size, the number of layers, and the batch size, we follow the setting of Table 1 in (Narayanan et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', 2021).' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  REFERENCES  Agarwal, S.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Wang, H.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Venkataraman, S.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', and Papailiopou-  los, D.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' On the utility of gradient compression in dis-  tributed training systems.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Proceedings of Machine Learn-  ing and Systems, 4:652–672, 2022.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Bahdanau, D.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Cho, K.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', and Bengio, Y.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Neural machine  translation by jointly learning to align and translate.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' arXiv  preprint arXiv:1409.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='0473, 2014.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Bernstein, J.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Wang, Y.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='-X.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Azizzadenesheli, K.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', and Anand-  kumar, A.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' signsgd: Compressed optimisation for non-  convex problems.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' In International Conference on Ma-  chine Learning, pp.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' 560–569.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' PMLR, 2018.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Dao, T.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Fu, D.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Y.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Ermon, S.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Rudra, A.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', and R´e, C.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Flashat-  tention: Fast and memory-efficient exact attention with  io-awareness.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' arXiv preprint arXiv:2205.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='14135, 2022.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Dettmers, T.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' 8-bit approximations for parallelism in deep  learning.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' arXiv preprint arXiv:1511.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='04561, 2015.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Devlin, J.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Chang, M.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='-W.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Lee, K.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', and Toutanova, K.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Bert:  Pre-training of deep bidirectional transformers for lan-  guage understanding.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' arXiv preprint arXiv:1810.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='04805,  2018.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Dosovitskiy, A.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Beyer, L.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Kolesnikov, A.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Weissenborn,  D.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Zhai, X.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Unterthiner, T.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Dehghani, M.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Minderer, M.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=',  Heigold, G.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Gelly, S.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' An image is worth 16x16  words: Transformers for image recognition at scale.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' arXiv  preprint arXiv:2010.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='11929, 2020.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Gong, Y.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Chung, Y.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='-A.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', and Glass, J.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Ast: Audio spec-  trogram transformer.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' arXiv preprint arXiv:2104.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='01778,  2021.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Gururangan, S.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Marasovi´c, A.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Swayamdipta, S.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Lo, K.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=',  Beltagy, I.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Downey, D.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', and Smith, N.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' A.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Don’t stop  pretraining: adapt language models to domains and tasks.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  arXiv preprint arXiv:2004.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='10964, 2020.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Hinton, G.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' E.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' and Zemel, R.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Autoencoders, minimum de-  scription length and helmholtz free energy.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Advances in  neural information processing systems, 6, 1993.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Ho, Q.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Cipar, J.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Cui, H.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Lee, S.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Kim, J.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' K.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Gibbons, P.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' B.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=',  Gibson, G.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' A.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Ganger, G.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', and Xing, E.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' P.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' More effective  distributed ml via a stale synchronous parallel parameter  server.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' In Advances in neural information processing  systems, pp.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' 1223–1231, 2013.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Huang, Y.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Cheng, Y.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Bapna, A.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Firat, O.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Chen, D.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Chen,  M.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Lee, H.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Ngiam, J.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Le, Q.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' V.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Wu, Y.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Gpipe:  Efficient training of giant neural networks using pipeline  parallelism.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Advances in neural information processing  systems, 32, 2019.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Izsak, P.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Berchansky, M.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', and Levy, O.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  How to  train bert with an academic budget.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  arXiv preprint  arXiv:2104.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='07705, 2021.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Jia, Z.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Zaharia, M.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', and Aiken, A.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Beyond data and model  parallelism for deep neural networks.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Proceedings of  Machine Learning and Systems, 1:1–13, 2019.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Kim, J.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' K.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Ho, Q.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Lee, S.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Zheng, X.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Dai, W.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Gibson,  G.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' A.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', and Xing, E.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' P.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Strads: A distributed framework for  scheduled model parallel machine learning.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' In Proceed-  ings of the Eleventh European Conference on Computer  Systems, pp.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' 1–16, 2016.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Li, D.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Wang, H.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Xing, E.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', and Zhang, H.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Amp: Automati-  cally finding model parallel strategies with heterogeneity  awareness.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' arXiv preprint arXiv:2210.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='07297, 2022.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Li, M.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Andersen, D.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' G.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Park, J.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' W.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Smola, A.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' J.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Ahmed,  A.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Josifovski, V.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Long, J.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Shekita, E.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' J.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', and Su, B.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='-Y.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Scaling distributed machine learning with the parameter  server.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  In 11th {USENIX} Symposium on Operating  Systems Design and Implementation ({OSDI} 14), pp.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  583–598, 2014.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Does compressing activations help model parallel training?' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Li, S.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Zhao, Y.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Varma, R.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Salpekar, O.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Noordhuis, P.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=',  Li, T.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Paszke, A.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Smith, J.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Vaughan, B.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Damania, P.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=',  et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Pytorch distributed: Experiences on accelerating  data parallel training.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' arXiv preprint arXiv:2006.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='15704,  2020.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Li, Z.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Zhuang, S.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Guo, S.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Zhuo, D.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Zhang, H.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Song, D.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=',  and Stoica, I.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Terapipe: Token-level pipeline parallelism  for training large-scale language models.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  In Interna-  tional Conference on Machine Learning, pp.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' 6543–6552.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  PMLR, 2021.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Lin, Y.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Han, S.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Mao, H.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Wang, Y.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', and Dally, W.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' J.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Deep gradient compression: Reducing the communica-  tion bandwidth for distributed training.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' arXiv preprint  arXiv:1712.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='01887, 2017.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Liu, Y.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Ott, M.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Goyal, N.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Du, J.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Joshi, M.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Chen, D.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=',  Levy, O.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Lewis, M.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Zettlemoyer, L.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', and Stoyanov, V.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Roberta: A robustly optimized bert pretraining approach.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  arXiv preprint arXiv:1907.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='11692, 2019.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Lu, W.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Yan, G.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Li, J.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Gong, S.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Han, Y.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', and Li, X.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Flexflow: A flexible dataflow accelerator architecture  for convolutional neural networks.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' In 2017 IEEE In-  ternational Symposium on High Performance Computer  Architecture (HPCA), pp.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' 553–564.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' IEEE, 2017.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Narayanan, D.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Harlap, A.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Phanishayee, A.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Seshadri, V.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=',  Devanur, N.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' R.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Ganger, G.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' R.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Gibbons, P.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' B.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', and Za-  haria, M.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Pipedream: generalized pipeline parallelism for  dnn training.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' In Proceedings of the 27th ACM Symposium  on Operating Systems Principles, pp.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' 1–15, 2019.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Narayanan, D.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Shoeybi, M.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Casper, J.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', LeGresley, P.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Pat-  wary, M.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Korthikanti, V.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Vainbrand, D.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Kashinkunti, P.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=',  Bernauer, J.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Catanzaro, B.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Efficient large-scale  language model training on gpu clusters using megatron-  lm.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' In Proceedings of the International Conference for  High Performance Computing, Networking, Storage and  Analysis, pp.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' 1–15, 2021.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Radford, A.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Narasimhan, K.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Salimans, T.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Sutskever, I.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=',  et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Improving language understanding by generative  pre-training.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' 2018.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Rajpurkar, P.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Zhang, J.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Lopyrev, K.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', and Liang, P.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Squad:  100,000+ questions for machine comprehension of text.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  arXiv preprint arXiv:1606.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='05250, 2016.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Rasley, J.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Rajbhandari, S.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Ruwase, O.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', and He, Y.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Deep-  speed: System optimizations enable training deep learn-  ing models with over 100 billion parameters.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' In Proceed-  ings of the 26th ACM SIGKDD International Conference  on Knowledge Discovery & Data Mining, pp.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' 3505–3506,  2020.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Seide, F.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Fu, H.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Droppo, J.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Li, G.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', and Yu, D.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' 1-bit stochas-  tic gradient descent and its application to data-parallel  distributed training of speech dnns.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' In Fifteenth annual  conference of the international speech communication  association.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Citeseer, 2014.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Sergeev, A.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' and Del Balso, M.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Horovod: fast and easy  distributed deep learning in tensorflow.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' arXiv preprint  arXiv:1802.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='05799, 2018.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Sharir, G.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Noy, A.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', and Zelnik-Manor, L.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' An image is worth  16x16 words, what is a video worth?' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  arXiv preprint  arXiv:2103.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='13915, 2021.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Shazeer, N.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Cheng, Y.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Parmar, N.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Tran, D.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Vaswani, A.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=',  Koanantakool, P.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Hawkins, P.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Lee, H.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Hong, M.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Young,  C.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Mesh-tensorflow: Deep learning for super-  computers.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Advances in neural information processing  systems, 31, 2018.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Shoeybi, M.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Patwary, M.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Puri, R.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', LeGresley, P.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Casper,  J.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', and Catanzaro, B.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Megatron-lm: Training multi-  billion parameter language models using model paral-  lelism.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' arXiv preprint arXiv:1909.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='08053, 2019.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Stich, S.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' U.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Cordonnier, J.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='-B.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', and Jaggi, M.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Sparsified sgd  with memory.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Advances in Neural Information Process-  ing Systems, 31, 2018.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Touvron, H.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Cord, M.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Douze, M.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Massa, F.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Sablayrolles,  A.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', and J´egou, H.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Training data-efficient image transform-  ers & distillation through attention.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' In International Con-  ference on Machine Learning, pp.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' 10347–10357.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' PMLR,  2021.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Vaswani, A.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Shazeer, N.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Parmar, N.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Uszkoreit, J.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Jones,  L.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Gomez, A.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' N.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Kaiser, Ł.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', and Polosukhin, I.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' At-  tention is all you need.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Advances in neural information  processing systems, 30, 2017.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Vogels, T.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Karimireddy, S.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' P.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', and Jaggi, M.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Powersgd:  Practical low-rank gradient compression for distributed  optimization.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Advances in Neural Information Processing  Systems, 32, 2019.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Wang, A.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Singh, A.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Michael, J.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Hill, F.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Levy, O.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', and  Bowman, S.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' R.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Glue: A multi-task benchmark and anal-  ysis platform for natural language understanding.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' arXiv  preprint arXiv:1804.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='07461, 2018a.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Wang, H.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Sievert, S.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Liu, S.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Charles, Z.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Papailiopoulos,  D.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', and Wright, S.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Atomo: Communication-efficient  learning via atomic sparsification.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Advances in Neural  Information Processing Systems, 31, 2018b.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Wang, H.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Agarwal, S.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', and Papailiopoulos, D.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Pufferfish:  communication-efficient models at no extra cost.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Pro-  ceedings of Machine Learning and Systems, 3:365–386,  2021.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Does compressing activations help model parallel training?' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Wang, J.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Yuan, B.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Rimanic, L.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', He, Y.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Dao, T.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Chen,  B.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Re, C.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', and Zhang, C.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Fine-tuning language models  over slow networks using activation compression with  guarantees.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' arXiv preprint arXiv:2206.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='01299, 2022.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Wang, S.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Li, B.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Z.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Khabsa, M.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Fang, H.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', and Ma, H.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Linformer: Self-attention with linear complexity.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' arXiv  preprint arXiv:2006.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='04768, 2020.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Williams, S.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Waterman, A.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', and Patterson, D.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Roofline:  an insightful visual performance model for multicore  architectures.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Communications of the ACM, 52(4):65–76,  2009.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Yang, Z.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Dai, Z.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Yang, Y.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Carbonell, J.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Salakhutdinov,  R.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' R.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', and Le, Q.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' V.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Xlnet: Generalized autoregressive  pretraining for language understanding.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Advances in  neural information processing systems, 32, 2019.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Zhang, H.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Zheng, Z.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Xu, S.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Dai, W.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Ho, Q.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Liang, X.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Hu,  Z.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Wei, J.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Xie, P.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', and Xing, E.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' P.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Poseidon: An efficient  communication architecture for distributed deep learning  on GPU clusters.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' In 2017 USENIX Annual Technical  Conference (USENIX ATC 17), pp.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' 181–193, 2017.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Zheng, L.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Li, Z.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Zhang, H.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Zhuang, Y.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Chen, Z.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Huang,  Y.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Wang, Y.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Xu, Y.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Zhuo, D.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Gonzalez, J.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' E.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', et al.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Alpa:  Automating inter-and intra-operator parallelism for dis-  tributed deep learning.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' arXiv preprint arXiv:2201.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='12023,  2022.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Zhu, Y.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Kiros, R.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Zemel, R.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Salakhutdinov, R.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Urta-  sun, R.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', Torralba, A.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=', and Fidler, S.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Aligning books and  movies: Towards story-like visual explanations by watch-  ing movies and reading books.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' In Proceedings of the  IEEE international conference on computer vision, pp.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  19–27, 2015.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Does compressing activations help model parallel training?' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  A  MORE EXPERIMENTAL RESULTS  We provide more experimental results in this section.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Distributed Setting  w/o  A1  A2  T1  T2  T3  T4  TP=1, PP=4  151.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='82  154.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='62  155.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='03  155.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='78  155.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='12  156.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='84  158.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='58  TP=2, PP=2  145.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='58  157.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='49  163.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='63  175.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='67  177.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='39  186.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='71  178.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='91  TP=4, PP=1  136.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='66  155.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='43  145.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='97  170.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='04  176.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='88  186.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='06  190.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='01  Distributed Setting  R1  R2  R3  R4  Q1  Q2  Q3  TP=1, PP=4  206.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='89  273.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='49  449.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='70  1,292.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='15  154.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='30  153.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='65  152.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='33  TP=2, PP=2  844.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='66  1,589.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='66  3,915.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='32  15,732.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='57  178.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='09  175.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='23  172.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='93  TP=4, PP=1  820.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='37  1,588.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='59  3,915.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='52  15,469.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='87  188.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='10  168.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='90  167.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='90  Table 11.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' The total time (ms) for fine-tuning with various compression techniques by varying the distributed setting.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' The results are  collected from the AWS p3.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='8xlarge machine with NVLink by using batch size 32, and sequence length 128.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Distributed Setting  w/o  A1  A2  T1  T2  T3  T4  TP=1, PP=4  106.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='04  113.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='67  106.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='35  109.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='58  109.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='10  109.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='18  110.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='57  TP=2, PP=2  121.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='26  142.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='41  140.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='05  152.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='91  154.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='60  162.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='00  157.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='12  TP=4, PP=1  122.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='22  142.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='33  139.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='47  171.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='24  165.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='77  172.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='69  170.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='61  Distributed Setting  R1  R2  R3  R4  Q1  Q2  Q3  TP=1, PP=4  124.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='39  137.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='51  187.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='59  333.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='61  108.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='18  109.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='56  109.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='49  TP=2, PP=2  314.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='51  507.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='00  998.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='51  3,197.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='42  163.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='18  155.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='48  150.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='31  TP=4, PP=1  329.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='33  513.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='89  1,007.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='65  3,406.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='20  171.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='06  163.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='96  152.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='82  Table 12.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' The total time (ms) for fine-tuning with various compression techniques by varying the distributed setting.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' The results are  collected from the AWS p3.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='8xlarge machine with NVLink by using batch size 8, and sequence length 128.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Distributed Setting  w/o  A1  A2  T1  T2  T3  T4  TP=1, PP=4  154.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='82  152.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='50  153.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='47  155.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='56  156.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='01  156.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='81  158.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='37  TP=2, PP=2  184.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='48  175.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='29  180.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='35  206.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='56  204.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='48  207.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='66  214.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='30  TP=4, PP=1  212.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='76  201.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='39  200.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='31  234.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='16  240.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='42  242.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='62  261.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='39  Distributed Setting  R1  R2  R3  R4  Q1  Q2  Q3  TP=1, PP=4  185.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='83  231.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='78  368.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='95  963.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='62  155.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='33  154.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='85  154.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='82  TP=2, PP=2  684.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='28  1,228.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='36  2,900.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='86  10,499.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='14  188.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='82  189.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='14  194.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='25  TP=4, PP=1  722.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='87  1,275.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='57  2,973.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='04  10,891.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='70  225.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='42  230.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='69  242.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='42  Table 13.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' The total time (ms) for fine-tuning with various compression techniques by varying the distributed setting.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' The results are  collected from the local machine without NVLink by using batch size 32, and sequence length 128.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Distributed Setting  w/o  A1  A2  T1  T2  T3  T4  TP=1, PP=4  73.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='19  72.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='94  72.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='58  75.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='98  74.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='15  73.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='62  74.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='86  TP=2, PP=2  100.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='86  107.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='73  100.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='54  113.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='59  117.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='36  114.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='86  112.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='11  TP=4, PP=1  100.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='73  107.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='90  115.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='18  129.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='31  124.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='94  136.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='18  133.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='91  Distributed Setting  R1  R2  R3  R4  Q1  Q2  Q3  TP=1, PP=4  82.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='45  94.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='84  123.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='78  239.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='81  73.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='33  74.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='41  71.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='80  TP=2, PP=2  235.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='02  366.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='59  769.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='47  2,183.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='39  111.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='61  106.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='75  101.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='25  TP=4, PP=1  238.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='28  368.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='45  733.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='03  2,509.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='73  120.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='14  114.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='73  118.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='98  Table 14.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' The total time (ms) for fine-tuning with various compression techniques by varying the distributed setting.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' The results are  collected from the local machine without NVLink by using batch size 8, and sequence length 128.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Does compressing activations help model parallel training?' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Compression  Algorithm  MNLI-(m/mm)  QQP  SST-2  MRPC  CoLA  QNLI  RTE  STS-B  w/o  87.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='87/88.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='02  91.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='96  95.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='18  87.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='71  59.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='40  92.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='99  76.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='90  88.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='43  A1  85.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='30/85.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='33  91.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='28  92.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='32  84.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='58  55.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='18  90.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='87  59.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='93  87.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='92  A2  85.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='25/85.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='19  91.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='41  93.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='23  86.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='72  57.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='02  90.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='92  64.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='26  87.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='74  T1  34.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='38/34.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='01  72.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='29  49.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='54  70.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='38  36.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='64  59.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='89  53.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='43  70.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='81  T2  40.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='10/38.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='97  58.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='91  79.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='24  66.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='49  0.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='00  80.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='40  45.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='49  11.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='32  T3  68.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='76/69.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='23  64.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='58  91.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='40  80.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='93  0.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='00  67.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='34  66.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='43  69.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='24  T4  84.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='24/85.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='23  89.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='17  92.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='09  81.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='68  51.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='54  91.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='71  63.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='54  84.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='80  Q1  86.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='85/87.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='58  91.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='50  93.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='58  86.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='96  59.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='20  92.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='24  59.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='57  86.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='89  Q2  87.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='46/88.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='02  91.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='82  94.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='95  87.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='48  57.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='02  93.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='36  68.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='95  87.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='84  Table 15.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Fintune results over GLUE dataset under the setting using tensor parallelism size 2, pipeline parallelism size 2, batch size  32, and sequence length 128.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' F1 scores are reported for QQP and MRPC, Matthews correlation coefficient is reported for CoLA, and  Spearman correlations are reported for STS-B, and accuracy scores are reported for the other tasks.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='  Compression  Algorithm  MNLI-(m/mm)  QQP  SST-2  MRPC  CoLA  QNLI  RTE  STS-B  w/o  86.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='23/86.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='07  91.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='22  91.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='74  88.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='17  59.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='02  92.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='09  78.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='70  88.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='40  A1  82.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='49/82.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='41  89.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='93  91.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='85  82.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='43  43.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='56  89.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='84  47.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='29  87.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='03  A2  82.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='18/82.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='23  90.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='45  90.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='52  83.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='54  0.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='00  89.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='02  62.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='82  87.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='66  T1  36.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='69/38.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='13  66.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='85  55.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='32  68.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='93  0.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='00  59.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='13  52.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='71  1.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='97  T2  43.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='92/43.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='66  73.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='63  51.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='26  62.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='26  0.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='00  60.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='13  49.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='82  0.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='00  T3  49.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='07/47.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='96  72.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='02  83.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='57  69.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='33  12.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='04  83.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='60  55.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='60  84.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='96  T4  83.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='99/84.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='37  35.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='78  68.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='30  83.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='54  47.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='33  60.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='52  64.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='62  86.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='72  Q1  84.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='91/85.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='18  90.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='54  92.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='43  85.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='91  53.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='25  60.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='68  57.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='04  87.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='91  Q2  85.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='66/86.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='09  90.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='99  91.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='74  86.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='84  53.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='92  91.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='31  75.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='81  88.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content='19  Table 16.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' Fintune results over GLUE dataset under the setting using tensor parallelism size 2, pipeline parallelism size 2, batch size 8, and  sequence length 128.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}
+page_content=' F1 scores are reported for QQP and MRPC, Matthews correlation coefficient is reported for CoLA, and Spearman  correlations are reported for STS-B, and accuracy scores are reported for the other tasks.' metadata={'source': '/home/zjlab/wf/langchain-ChatGLM/knowledge_base/B9E0T4oBgHgl3EQfyAKb/content/2301.02654v1.pdf'}