News!!!
[12/03/2023] Manuscript Accelerating Heterogeneous Tensor Parallelism via Flexible Workload Control is submitted to ICDE 2024 for under review!!!
[10/21/2023] Some bugs about priority pruning is solved.
[09/16/2023] SEMI-migration is available on top of resizing and migration.
[07/29/2023] Data migration is online.
[06/29/2013] Priority pruning is online.
[06/15/2013] ViT with 1.2B and 2.7B parameters is successfully run.
[05/24/2023] ZERO-resizing/pruning technique is implemented and its effectiveness is validated.
[03/24/2023] Colossal-AI is succefully deployed and tested on 8 V100 GPUs.
[03/12/2023] This project starts.

FlexTP

FlexTP consists of a series of optimization techniques tailored for efficient Tensor Parallelism in heterogeneous environments. It is particularly suitable for training foundation models with billions of parameters.

1. Introduction

Recently, transformer-based models are becoming deeper and larger. For better scalability, an underlying training solution is to split billions of parameters into many tasks for parallel computation, i.e., tensor parallelism. The computing cluster, however, usually consists of heterogeneous devices, and is shared by multi-tenants. The heterogeneous compute power and resource contention lead to the heavy straggling problem. Existing solutions for it are all tailored for traditional data parallelism. That cannot work under tensor parallelism due to the correctness and convergence constraint. FlexTP is the first work that attempts to flexibly control workloads for balanced tensor parallelism.

The FlexTP project started at Ocean University of China in March 2023. The backbone members are Xu Zhang (master student) and Zhigang Wang (associate professor). We implement all related techniques on the newly released platform Colossal-AI, which has integrated many existing optimizations for training foundation models. Features of Flex-TP are listed as below.

• ZERO-resizing: We temporarily resize matrices involved in core tensor computations to dynamically balance workloads, which employs data imputation and priority selection policies to guarantee consistency in computations and mitigate the loss of accuracy.
• Lightweight migration: We re-assign matrix data to cope with heavily heterogeneous environments without accuracy loss, which reduces the migration runtime latency via existing tree-based broadcasting/reducing efforts and our novel reducing merging optimization.
• SEMI-migration: We build this hybrid solution on top of the two approaches, which can smartly run the reasonable one in different scenarios based on the cost-benefit analysis.

2. Quick Start

FlexTP is developed on top of Colossal-AI. Before running it, some softwares must be installed, which is beyond the scope of this document.

2.1 Requirements

• Colossal-AI 0.2.7
• PyTorch >= 1.13
• Python 3.7 or higher version

2.2 Testing FlexTP

We should first prepare the input training samples.

Second, we can submit a training job for two foundation models using the following commands. (The config file is stored in ./examples/hybrid_parallel)

• For ViT-1B(1.2B):
  colossalai run --nproc_per_node 8 timer_trainer_with_cifar10.py --config tpconfig.py
  About arguments:
  [1] GPUs num [2] trainer file [3] config file

• For ViT-3B(2.7B):
  colossalai run --nproc_per_node 8 timer_trainer_with_cifar10.py --config finalconfig.py
  About arguments:
  [1] GPUs num [2] trainer file [3] config file

3. Contact

If you encounter any problem with FlexTP, please feel free to contact [email protected] and [email protected].