#!/bin/bash

export SAMA_CONFIG=./config/sama_cms_lla13b.yaml
export TOKENIZERS_PARALLELISM=true

# CUDA Include (/cuda.h)
CUDA_INCLUDE_PATH="/home/work/miniconda3/envs/allm/include"

export CPATH=$CPATH:$CUDA_INCLUDE_PATH
export CPLUS_INCLUDE_PATH=$CPLUS_INCLUDE_PATH:$CUDA_INCLUDE_PATH

export WANDB_PROJECT="SAMA_CMS_Llama13B"

date +"%F %T"

# STEP=300
# accelerate launch --dynamo_backend=inductor --dynamo_mode=max-autotune --main_process_port 41353 -m src.cms_main \
#     --config_path $SAMA_CONFIG --trainer_args.learning_rate=1e-3 --trainer_args.output_dir "./Llama13B" \
#     --trainer_args.load_best_model_at_end True --trainer_args.save_strategy '"steps"' \
#     --sama_adapter.col_L 32 --sama_adapter.row_R 32 \
#     --trainer_args.num_train_epochs 2 --trainer_args.report_to wandb \
#     --trainer_args.save_steps $STEP --trainer_args.eval_steps $STEP --trainer_args.logging_steps $STEP \
#     --sama_adapter.num_unique_blocks_L 32 --sama_adapter.num_unique_blocks_R 32 \
#     --sama_adapter.target_modules '["q_proj", "v_proj", "k_proj", "up_proj","down_proj"]' \
#     --data.path ft_training_set/commonsense_147k.json --trainer_args.eval_delay 9000

# # date +"%F %T"

# STEP=100
# accelerate launch --dynamo_backend=inductor --dynamo_mode=max-autotune --main_process_port 41353 -m src.cms_main \
#     --config_path $SAMA_CONFIG --trainer_args.learning_rate=2e-4 --trainer_args.output_dir "./Llama13B" \
#     --trainer_args.load_best_model_at_end True --trainer_args.save_strategy '"steps"' \
#     --sama_adapter.col_L 32 --sama_adapter.row_R 32 \
#     --trainer_args.num_train_epochs 2 --trainer_args.report_to none \
#     --trainer_args.save_steps $STEP --trainer_args.eval_steps $STEP --trainer_args.logging_steps $STEP \
#     --sama_adapter.num_unique_blocks_L 32 --sama_adapter.num_unique_blocks_R 32 \
#     --sama_adapter.target_modules '["q_proj", "v_proj", "k_proj", "up_proj","down_proj"]' \
#     --data.path ft_training_set/commonsense_15k.json --trainer_args.eval_delay 300

# date +"%F %T"

# STEP=100
# accelerate launch --dynamo_backend=inductor --dynamo_mode=max-autotune --main_process_port 41353 -m src.cms_main \
#     --config_path $SAMA_CONFIG --trainer_args.learning_rate=2e-5 --trainer_args.output_dir "./Llama13B" \
#     --trainer_args.load_best_model_at_end True --trainer_args.save_strategy '"steps"' \
#     --sama_adapter.col_L 32 --sama_adapter.row_R 32 \
#     --trainer_args.num_train_epochs 2 --trainer_args.report_to none \
#     --trainer_args.save_steps $STEP --trainer_args.eval_steps $STEP --trainer_args.logging_steps $STEP \
#     --sama_adapter.num_unique_blocks_L 32 --sama_adapter.num_unique_blocks_R 32 \
#     --sama_adapter.target_modules '["q_proj", "v_proj", "k_proj", "up_proj","down_proj"]' \
#     --data.path ft_training_set/commonsense_15k.json --trainer_args.eval_delay 300

# date +"%F %T"

# STEP=50
# accelerate launch --dynamo_backend=inductor --dynamo_mode=max-autotune --main_process_port 41353 -m src.cms_main \
#     --config_path $SAMA_CONFIG --trainer_args.learning_rate=1e-3 --trainer_args.output_dir "./Llama13B" \
#     --trainer_args.load_best_model_at_end False --trainer_args.save_strategy '"no"' \
#     --sama_adapter.col_L 32 --sama_adapter.row_R 32 \
#     --trainer_args.num_train_epochs 2 --trainer_args.report_to none \
#     --trainer_args.save_steps $STEP --trainer_args.eval_steps $STEP --trainer_args.logging_steps $STEP \
#     --sama_adapter.num_unique_blocks_L 32 --sama_adapter.num_unique_blocks_R 32 \
#     --sama_adapter.target_modules '["q_proj", "v_proj", "k_proj", "up_proj","down_proj"]' \
#     --data.path ft_training_set/commonsense_15k.json --trainer_args.eval_delay 300
# test
# date +"%F %T"
# ====

# STEP=300
# accelerate launch --dynamo_backend=inductor --dynamo_mode=max-autotune --main_process_port 41353 -m src.cms_main \
#     --config_path $SAMA_CONFIG --trainer_args.learning_rate=5e-4 --trainer_args.output_dir "./Llama13B" \
#     --trainer_args.load_best_model_at_end True --trainer_args.save_strategy '"steps"' \
#     --sama_adapter.col_L 16 --sama_adapter.row_R 32 \
#     --trainer_args.num_train_epochs 2 --trainer_args.report_to wandb \
#     --trainer_args.save_steps $STEP --trainer_args.eval_steps $STEP --trainer_args.logging_steps $STEP \
#     --sama_adapter.num_unique_blocks_L 32 --sama_adapter.num_unique_blocks_R 16 \
#     --sama_adapter.target_modules '["q_proj", "v_proj", "k_proj", "up_proj","down_proj"]' \
#     --data.path ft_training_set/commonsense_147k.json --trainer_args.eval_delay 600

# date +"%F %T"

# Llama13B/CMS/t60108d07h46m12,ep=2.0,mlr5.0e-04,b8,nb8,8,cL8,rR8,s1,initdef,dr0.0,size146627,5
# STEP=300
# accelerate launch --dynamo_backend=inductor --dynamo_mode=max-autotune --main_process_port 41353 -m src.cms_main \
#     --config_path $SAMA_CONFIG --trainer_args.learning_rate=5e-4 --trainer_args.output_dir "./Llama13B" \
#     --trainer_args.load_best_model_at_end True --trainer_args.save_strategy '"steps"' \
#     --sama_adapter.col_L 8 --sama_adapter.row_R 8 \
#     --trainer_args.num_train_epochs 2 --trainer_args.report_to wandb \
#     --trainer_args.save_steps $STEP --trainer_args.eval_steps $STEP --trainer_args.logging_steps $STEP \
#     --sama_adapter.num_unique_blocks_L 8 --sama_adapter.num_unique_blocks_R 8 \
#     --sama_adapter.target_modules '["q_proj", "v_proj", "k_proj", "up_proj","down_proj"]' \
#     --data.path ft_training_set/commonsense_147k.json --trainer_args.eval_delay 6000

# date +"%F %T"
# Llama13B/CMS/t60108d17h22m55,ep=2.0,mlr5.0e-04,b8,nb16,8,cL8,rR16,s1,initdef,dr0.0,size146627,5
# STEP=200
# accelerate launch --dynamo_backend=inductor --dynamo_mode=max-autotune --main_process_port 41353 -m src.cms_main \
#     --config_path $SAMA_CONFIG --trainer_args.learning_rate=5e-4 --trainer_args.output_dir "./Llama13B" \
#     --trainer_args.load_best_model_at_end True --trainer_args.save_strategy '"steps"' \
#     --sama_adapter.col_L 8 --sama_adapter.row_R 16 \
#     --trainer_args.num_train_epochs 2 --trainer_args.report_to wandb \
#     --trainer_args.save_steps $STEP --trainer_args.eval_steps $STEP --trainer_args.logging_steps $STEP \
#     --sama_adapter.num_unique_blocks_L 16 --sama_adapter.num_unique_blocks_R 8 \
#     --sama_adapter.target_modules '["q_proj", "v_proj", "k_proj", "up_proj","down_proj"]' \
#     --data.path ft_training_set/commonsense_147k.json --trainer_args.eval_delay 6000

# date +"%F %T"
# Llama13B/CMS/t60108d19h26m54,ep=2.0,mlr2.0e-04,b8,nb8,4,cL4,rR8,s1,initdef,dr0.0,size146627,5
# Llama13B/CMS/t60108d23h38m34,ep=2.0,mlr2.0e-04,b8,nb8,4,cL4,rR8,s1,initdef,dr0.0,size146627,5
# STEP=200
# accelerate launch --dynamo_backend=inductor --dynamo_mode=max-autotune --main_process_port 41353 -m src.cms_main \
#     --config_path $SAMA_CONFIG --trainer_args.learning_rate=2e-4 --trainer_args.output_dir "./Llama13B" \
#     --trainer_args.load_best_model_at_end True --trainer_args.save_strategy '"steps"' \
#     --sama_adapter.col_L 4 --sama_adapter.row_R 8 \
#     --trainer_args.num_train_epochs 2 --trainer_args.report_to wandb \
#     --trainer_args.save_steps $STEP --trainer_args.eval_steps $STEP --trainer_args.logging_steps $STEP \
#     --sama_adapter.num_unique_blocks_L 8 --sama_adapter.num_unique_blocks_R 4 \
#     --sama_adapter.target_modules '["q_proj", "v_proj", "k_proj", "up_proj","down_proj"]' \
#     --data.path ft_training_set/commonsense_147k.json --trainer_args.eval_delay 6000

# date +"%F %T"

# Llama13B/CMS/t60109d06h50m53,ep=2.0,mlr2.0e-04,b8,nb4,8,cL8,rR4,s1,initdef,dr0.0,size146627,5
# STEP=200
# accelerate launch --dynamo_backend=inductor --dynamo_mode=max-autotune --main_process_port 41353 -m src.cms_main \
#     --config_path $SAMA_CONFIG --trainer_args.learning_rate=2e-4 --trainer_args.output_dir "./Llama13B" \
#     --trainer_args.load_best_model_at_end True --trainer_args.save_strategy '"steps"' \
#     --sama_adapter.col_L 8 --sama_adapter.row_R 4 \
#     --trainer_args.num_train_epochs 2 --trainer_args.report_to wandb \
#     --trainer_args.save_steps $STEP --trainer_args.eval_steps $STEP --trainer_args.logging_steps $STEP \
#     --sama_adapter.num_unique_blocks_L 4 --sama_adapter.num_unique_blocks_R 8 \
#     --sama_adapter.target_modules '["q_proj", "v_proj", "k_proj", "up_proj","down_proj"]' \
#     --data.path ft_training_set/commonsense_147k.json --trainer_args.eval_delay 6000

# date +"%F %T"

# Llama13B/CMS/t60109d16h12m04,ep=2.0,mlr2.0e-04,b8,nb4,4,cL4,rR4,s1,initdef,dr0.0,size146627,5
# STEP=300
# accelerate launch --dynamo_backend=inductor --dynamo_mode=max-autotune --main_process_port 41353 -m src.cms_main \
#     --config_path $SAMA_CONFIG --trainer_args.learning_rate=2e-4 --trainer_args.output_dir "./Llama13B" \
#     --trainer_args.load_best_model_at_end True --trainer_args.save_strategy '"steps"' \
#     --sama_adapter.col_L 4 --sama_adapter.row_R 4 \
#     --trainer_args.num_train_epochs 2 --trainer_args.report_to wandb \
#     --trainer_args.save_steps $STEP --trainer_args.eval_steps $STEP --trainer_args.logging_steps $STEP \
#     --sama_adapter.num_unique_blocks_L 4 --sama_adapter.num_unique_blocks_R 4 \
#     --sama_adapter.target_modules '["q_proj", "v_proj", "k_proj", "up_proj","down_proj"]' \
#     --data.path ft_training_set/commonsense_147k.json --trainer_args.eval_delay 6000

# date +"%F %T"

# Llama13B/CMS/t60109d22h46m19,ep=2.0,mlr2.0e-04,b8,nb4,4,cL8,rR8,s1,initdef,dr0.0,size146627,5
# STEP=300
# accelerate launch --dynamo_backend=inductor --dynamo_mode=max-autotune --main_process_port 41353 -m src.cms_main \
#     --config_path $SAMA_CONFIG --trainer_args.learning_rate=2e-4 --trainer_args.output_dir "./Llama13B" \
#     --trainer_args.load_best_model_at_end True --trainer_args.save_strategy '"steps"' \
#     --sama_adapter.col_L 8 --sama_adapter.row_R 8 \
#     --trainer_args.num_train_epochs 2 --trainer_args.report_to wandb \
#     --trainer_args.save_steps $STEP --trainer_args.eval_steps $STEP --trainer_args.logging_steps $STEP \
#     --sama_adapter.num_unique_blocks_L 4 --sama_adapter.num_unique_blocks_R 4 \
#     --sama_adapter.target_modules '["q_proj", "v_proj", "k_proj", "up_proj","down_proj"]' \
#     --data.path ft_training_set/commonsense_147k.json --trainer_args.eval_delay 6000

# date +"%F %T"

# Llama13B/CMS/t60110d05h52m12,ep=2.0,mlr2.0e-04,b8,nb16,16,cL16,rR16,s1,initdef,dr0.0,size146627,5
# STEP=300
# accelerate launch --dynamo_backend=inductor --dynamo_mode=max-autotune --main_process_port 41353 -m src.cms_main \
#     --config_path $SAMA_CONFIG --trainer_args.learning_rate=2e-4 --trainer_args.output_dir "./Llama13B" \
#     --trainer_args.load_best_model_at_end True --trainer_args.save_strategy '"steps"' \
#     --sama_adapter.col_L 16 --sama_adapter.row_R 16 \
#     --trainer_args.num_train_epochs 2 --trainer_args.report_to wandb \
#     --trainer_args.save_steps $STEP --trainer_args.eval_steps $STEP --trainer_args.logging_steps $STEP \
#     --sama_adapter.num_unique_blocks_L 16 --sama_adapter.num_unique_blocks_R 16 \
#     --sama_adapter.target_modules '["q_proj", "v_proj", "k_proj", "up_proj","down_proj"]' \
#     --data.path ft_training_set/commonsense_147k.json --trainer_args.eval_delay 6000

# date +"%F %T"

# Llama13B/CMS/t60110d14h39m36,ep=2.0,mlr5.0e-04,b8,nb16,16,cL16,rR16,s1,initdef,dr0.0,size146627,5
# STEP=300
# accelerate launch --dynamo_backend=inductor --dynamo_mode=max-autotune --main_process_port 41353 -m src.cms_main \
#     --config_path $SAMA_CONFIG --trainer_args.learning_rate=5e-4 --trainer_args.output_dir "./Llama13B" \
#     --trainer_args.load_best_model_at_end True --trainer_args.save_strategy '"steps"' \
#     --sama_adapter.col_L 16 --sama_adapter.row_R 16 \
#     --trainer_args.num_train_epochs 2 --trainer_args.report_to wandb \
#     --trainer_args.save_steps $STEP --trainer_args.eval_steps $STEP --trainer_args.logging_steps $STEP \
#     --sama_adapter.num_unique_blocks_L 16 --sama_adapter.num_unique_blocks_R 16 \
#     --sama_adapter.target_modules '["q_proj", "v_proj", "k_proj", "up_proj","down_proj"]' \
#     --data.path ft_training_set/commonsense_147k.json --trainer_args.eval_delay 6000
# date +"%F %T"

### QV only
# Llama13B/CMS/t60111d02h06m52,ep=2.0,mlr5.0e-04,b8,nb4,4,cL4,rR4,s1.0,initdef,dr0.0,size146627,2
# STEP=300
# accelerate launch --dynamo_backend=inductor --dynamo_mode=max-autotune --main_process_port 41353 -m src.cms_main \
#     --config_path $SAMA_CONFIG --trainer_args.learning_rate=5e-4 --trainer_args.output_dir "./Llama13B" \
#     --trainer_args.load_best_model_at_end True --trainer_args.save_strategy '"steps"' \
#     --sama_adapter.col_L 4 --sama_adapter.row_R 4 \
#     --trainer_args.num_train_epochs 2 --trainer_args.report_to wandb \
#     --trainer_args.save_steps $STEP --trainer_args.eval_steps $STEP --trainer_args.logging_steps $STEP \
#     --sama_adapter.num_unique_blocks_L 4 --sama_adapter.num_unique_blocks_R 4 \
#     --sama_adapter.target_modules '["q_proj", "v_proj"]' \
#     --data.path ft_training_set/commonsense_147k.json --trainer_args.eval_delay 6000
# date +"%F %T"

# STEP=50
# accelerate launch --dynamo_backend=inductor --dynamo_mode=max-autotune --main_process_port 41353 -m src.cms_main \
#     --config_path $SAMA_CONFIG --trainer_args.learning_rate=5e-4 --trainer_args.output_dir "./Llama13B" \
#     --trainer_args.load_best_model_at_end True --trainer_args.save_strategy '"steps"' \
#     --sama_adapter.col_L 8 --sama_adapter.row_R 8 \
#     --trainer_args.num_train_epochs 2 --trainer_args.report_to wandb \
#     --trainer_args.save_steps $STEP --trainer_args.eval_steps $STEP --trainer_args.logging_steps $STEP \
#     --sama_adapter.num_unique_blocks_L 4 --sama_adapter.num_unique_blocks_R 4 \
#     --sama_adapter.target_modules '["q_proj", "v_proj"]' \
#     --data.path ft_training_set/commonsense_147k.json --trainer_args.eval_delay 6000 --sama_adapter.scaling 2
# date +"%F %T"

# bash scripts/drop_train.sh
# bash scripts/cms_13bl_merge_eval.sh
# bash scripts/cms_l3_train.sh


# lora
# STEP=300
# accelerate launch --dynamo_backend=no --dynamo_mode=max-autotune --main_process_port 41353 -m src.cms_main \
#     --config_path $SAMA_CONFIG --trainer_args.learning_rate=3e-4 --trainer_args.output_dir "./Llama13B" \
#     --trainer_args.load_best_model_at_end True --trainer_args.save_strategy '"steps"' \
#     --sama_adapter.col_L 32 --sama_adapter.row_R 32 \
#     --trainer_args.num_train_epochs 2 \
#     --trainer_args.save_steps $STEP --trainer_args.eval_steps $STEP --trainer_args.logging_steps $STEP \
#     --sama_adapter.num_unique_blocks_L 32 --sama_adapter.num_unique_blocks_R 32 \
#     --sama_adapter.target_modules '["q_proj", "v_proj"]' \
#     --data.path ft_training_set/commonsense_147k.json --trainer_args.report_to wandb --run_text LORA4
# date +"%F %T"

# STEP=300
# accelerate launch --dynamo_backend=no --dynamo_mode=max-autotune --main_process_port 41353 -m src.cms_main \
#     --config_path $SAMA_CONFIG --trainer_args.learning_rate=1e-4 --trainer_args.output_dir "./Llama13B" \
#     --trainer_args.load_best_model_at_end True --trainer_args.save_strategy '"steps"' \
#     --sama_adapter.col_L 32 --sama_adapter.row_R 32 \
#     --trainer_args.num_train_epochs 2 \
#     --trainer_args.save_steps $STEP --trainer_args.eval_steps $STEP --trainer_args.logging_steps $STEP \
#     --sama_adapter.num_unique_blocks_L 32 --sama_adapter.num_unique_blocks_R 32 \
#     --sama_adapter.target_modules '["q_proj", "v_proj"]' \
#     --data.path ft_training_set/commonsense_147k.json --trainer_args.report_to wandb --run_text LORA4
# date +"%F %T"


STEP=300
accelerate launch --dynamo_backend=inductor --dynamo_mode=max-autotune --main_process_port 41353 -m src.cms_main \
    --config_path $SAMA_CONFIG --trainer_args.learning_rate=5e-4 --trainer_args.output_dir "./Llama13B" \
    --trainer_args.load_best_model_at_end True --trainer_args.save_strategy '"steps"' \
    --sama_adapter.col_L 4 --sama_adapter.row_R 4 \
    --trainer_args.num_train_epochs 2 --trainer_args.report_to wandb \
    --trainer_args.save_steps $STEP --trainer_args.eval_steps $STEP --trainer_args.logging_steps $STEP \
    --sama_adapter.num_unique_blocks_L 4 --sama_adapter.num_unique_blocks_R 4 \
    --sama_adapter.target_modules '["q_proj", "v_proj", "k_proj", "up_proj","down_proj"]' \
    --data.path ft_training_set/commonsense_147k.json --trainer_args.eval_delay 00

#######
# STEP=300
# accelerate launch --dynamo_backend=no --dynamo_mode=max-autotune --main_process_port 41353 -m src.cms_main \
#     --config_path $SAMA_CONFIG --trainer_args.learning_rate=3e-4 --trainer_args.output_dir "./Llama13B" \
#     --trainer_args.load_best_model_at_end True --trainer_args.save_strategy '"steps"' \
#     --sama_adapter.col_L 32 --sama_adapter.row_R 32 \
#     --trainer_args.num_train_epochs 2 \
#     --trainer_args.save_steps $STEP --trainer_args.eval_steps $STEP --trainer_args.logging_steps $STEP \
#     --sama_adapter.num_unique_blocks_L 32 --sama_adapter.num_unique_blocks_R 32 \
#     --sama_adapter.target_modules '["q_proj", "v_proj"]' \
#     --data.path ft_training_set/commonsense_147k.json --trainer_args.report_to wandb --run_text LORA32
# date +"%F %T"

# STEP=300
# accelerate launch --dynamo_backend=inductor --dynamo_mode=max-autotune --main_process_port 41353 -m src.cms_main \
#     --config_path $SAMA_CONFIG --trainer_args.learning_rate=7e-4 --trainer_args.output_dir "./Llama13B" \
#     --trainer_args.load_best_model_at_end True --trainer_args.save_strategy '"steps"' \
#     --sama_adapter.col_L 32 --sama_adapter.row_R 32 \
#     --trainer_args.num_train_epochs 2 \
#     --trainer_args.save_steps $STEP --trainer_args.eval_steps $STEP --trainer_args.logging_steps $STEP \
#     --sama_adapter.num_unique_blocks_L 32 --sama_adapter.num_unique_blocks_R 32 \
#     --sama_adapter.target_modules '["q_proj", "v_proj"]' \
#     --data.path ft_training_set/commonsense_147k.json --trainer_args.report_to wandb --run_text lora32
# date +"%F %T"

# STEP=300
# accelerate launch --dynamo_backend=inductor --dynamo_mode=max-autotune --main_process_port 41353 -m src.cms_main \
#     --config_path $SAMA_CONFIG --trainer_args.learning_rate=5e-4 --trainer_args.output_dir "./Llama13B" \
#     --trainer_args.load_best_model_at_end True --trainer_args.save_strategy '"steps"' \
#     --sama_adapter.col_L 32 --sama_adapter.row_R 16 \
#     --trainer_args.num_train_epochs 2 --trainer_args.report_to wandb \
#     --trainer_args.save_steps $STEP --trainer_args.eval_steps $STEP --trainer_args.logging_steps $STEP \
#     --sama_adapter.num_unique_blocks_L 16 --sama_adapter.num_unique_blocks_R 32 \
#     --sama_adapter.target_modules '["q_proj", "v_proj", "k_proj", "up_proj","down_proj"]' \
#     --data.path ft_training_set/commonsense_147k.json --trainer_args.eval_delay 6000

# date +"%F %T"