Installation Guide

Pre-requirement: Python3

Step 1. git clone https://github.com/baidu-research/DeepFlow.git

Step 2. cd DeepFlow

Step 3. Setup the environment:

* python3 -m venv [/path/to/new/virtual/environment]
* source [/path/to/new/virtual/environment]/bin/activate
* pip install --upgrade pip
* pip install click ruamel.yaml numpy
* mkdir -p [/path/to/output/directory] 

Step 4. Test if the installation has been successful:

* python perf.py --exp_config configs/[config.yaml] --exp_dir output
* check the output result: vim [/path/to/output/directory]/summary.txt

Execution Modes

DeepFlow can be used in 5 different modes:

(1) Peformance Prediction Mode (GEMM)

• When to use: Use for distributed GEMM prediction
• How: python perf.py --exp_config configs/[config.yaml] --exp_dir output --debug [False|True] --gemm True --t [RC|CR] --kp1 [kp1 dim.] --kp2 [kp2 dim.] --m [input dim.] --n [output dim.] --k [inner dim.]

(2) Performance Prediction Mode (End-2-End Application)

• Specify the application parameters in configs/[config.yaml]
• python perf.py --exp_config configs/[config.yaml] --exp_dir [/path/to/output/directory]
• python perf.py --exp_config configs/[config.yaml] --exp_dir [/path/to/output/directory] --batch_size [batch] --seq_len [seq_len] --hidden_dim [lstm_dim] --vocab_size [vocab_size] --dp [data parallel] --kp1 [kernel parallel dim1.] --kp2 [kernel parallel dim2.] --t [RC|CR] --args_input [True|False]

(3) Performance Prediction Mode (using main.py standalone argument; this is somewhat equivalent of option 2, for running on slurm)

• python main.py stand_alone --exp_dir [/path/to/output/result/directory] --exp_config configs/[config.yaml]

(4) Architecture search for a fixed parallelism strategy

• python GD_search.py --exp_config configs/[config.yaml] --exp_dir [/path/to/output/directory] --debug False --index [index] --batch_size [batch] --hidden_dim [lstm_dim] --data_scale [dataset_scaling_factor] --dp [data parallel dim.] --lp [layer parallel dim.] --kp_type [0|1] --kp1 [kp1 dim.] --kp2 [kp2 dim.] --inter_derate [derate_factor_for_inter_package_bandwidth] --intra_derate [derate_factor_for_intra_package_bandwidth] --kp1_inter [False|True] --kp2_inter [False|True] --dp_inter [False|True] --lp_inter [False|True] --wafer_dim [package dim.]
• Example: python GD_search.py --exp_config configs/exp_config.yaml --exp_dir output --debug False --index 40 --batch_size 256 --hidden_dim 19968 --data_scale 1 --dp 64 --lp 1 --kp_type 1 --kp1 1 --kp2 1 --inter_derate 0 --intra_derate 2 --kp1_inter False --kp2_inter False --dp_inter False --lp_inter False --wafer_dim 8

(5) Architecture Search mode for all types of parallelism strategies

• python main.py arch_search --exp_dir [/path/to/output/directory] --exp_config configs/[config.yaml]

Tips

• Use --no_launch True to see the command that would be used to launch the application w/o running
• Check config directory for different architecture templates and technology node configurations
• Use --debug True to activate debugging mode