reproduction_script.sh

#!/bin/bash

set -e

# Install the requirements
pip install -e .
pip install -r requirements.txt
pip install flash-attn --no-build-isolation

accelerate config

# Install the flash_gemv package
cd flash_gemv
pip install -e .
cd ..

# Get the root folder for faster_transformer
export PYTHONPATH=${PYTHONPATH}:$PWD
project_path=$PWD
ckpt_path=$1
result_path=$2
export PROJECT_PATH=$project_path
export CATS_CKPTPATH="$project_path/$ckpt_path"
export CATS_RESPATH="$project_path/$result_path"

# 1. Collect the statistics before General finetuning & plot
bash scripts/plot_mlp_histogram.sh $ckpt_path $result_path

# 2. Run general finetuning
bash scripts/general_finetuning_llama_cats.sh $ckpt_path $result_path
bash scripts/general_finetuning_llama_relufication.sh $ckpt_path $result_path

bash scripts/general_finetuning_mistral_cats.sh $ckpt_path $result_path
bash scripts/general_finetuning_mistral_relufication.sh $ckpt_path $result_path

python experiments/plot_zero_shot.py # Plot figure 1

# 3. Plot figure 5: activation sparsity (figure 5) after general finetuning
python experiments/plot_act_sparsity.py

# 4. Experiments for Table 1. Results are saved in "output/"
bash scripts/zero_shot_evaluation_without_general_finetuning.sh $ckpt_path $result_path

# 5. Experiments for Table 2 and 3
bash scripts/instruction_tuning.sh $ckpt_path $result_path

# Benchmark MLP Block
cd flash_gemv/bench/]
bash final_profile_llama7B.sh
bash final_profile_mistral7B.sh
# Plot fig 3 and fig 6
python plot_latency_sub.py

# Benchmark Generation (figure 4)
cd $project_path
bash scripts/bench_generation_llama7B.sh $project_path $ckpt_path $result_path
bash scripts/bench_generation_mistral7B.sh $project_path $ckpt_path $result_path
python experiments/plot_throughput.py