scripts/cms_13b2_train.sh

#!/bin/bash

# export SAMA_CONFIG=./config/sama_cms_lla.yaml
export SAMA_CONFIG=./config/sama_cms_2lla13b.yaml
# export SAMA_CONFIG=./config/sama_cms_fb.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_13B2"

date +"%F %T"
# QUANTA
# STEP=300
# accelerate launch --dynamo_backend=no --dynamo_mode=default --main_process_port 41353 -m src.cms_main \
#     --config_path $SAMA_CONFIG --trainer_args.learning_rate=8e-4 --trainer_args.output_dir "./Llama2_13B" \
#     --trainer_args.load_best_model_at_end True --trainer_args.save_strategy '"steps"' \
#     --sama_adapter.col_L 4 --sama_adapter.row_R 4 --sama_adapter.scaling 1.0 \
#     --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 --seed 43
# wandb sync wandb/latest-run
# date +"%F %T"

# STEP=300
# accelerate launch --dynamo_backend=no --dynamo_mode=default --main_process_port 41353 -m src.cms_main \
#     --config_path $SAMA_CONFIG --trainer_args.learning_rate=5e-4 --trainer_args.output_dir "./Llama2_13B" \
#     --trainer_args.load_best_model_at_end True --trainer_args.save_strategy '"steps"' \
#     --sama_adapter.col_L 4 --sama_adapter.row_R 4 --sama_adapter.scaling 1 \
#     --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 --seed 43
# wandb sync wandb/latest-run
# date +"%F %T"

# STEP=300
# accelerate launch --dynamo_backend=no --dynamo_mode=default --main_process_port 41353 -m src.cms_main \
#     --config_path $SAMA_CONFIG --trainer_args.learning_rate=8e-4 --trainer_args.output_dir "./Llama2_13B" \
#     --trainer_args.load_best_model_at_end True --trainer_args.save_strategy '"steps"' \
#     --sama_adapter.col_L 8 --sama_adapter.row_R 8 --sama_adapter.scaling 2 \
#     --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 --seed 43
# wandb sync wandb/latest-run
# date +"%F %T"

# STEP=300
# accelerate launch --dynamo_backend=no --dynamo_mode=default --main_process_port 41353 -m src.cms_main \
#     --config_path $SAMA_CONFIG --trainer_args.learning_rate=8e-4 --trainer_args.output_dir "./Llama2_13B" \
#     --trainer_args.load_best_model_at_end True --trainer_args.save_strategy '"steps"' \
#     --sama_adapter.col_L 16 --sama_adapter.row_R 16 --sama_adapter.scaling 1 \
#     --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"]' \
#     --data.path ft_training_set/commonsense_147k.json --trainer_args.eval_delay 6000 --seed 43
# wandb sync wandb/latest-run
# date +"%F %T"

# run_text lora
# STEP=300
# accelerate launch --dynamo_backend=no --dynamo_mode=default --main_process_port 41353 -m src.cms_main \
#     --config_path $SAMA_CONFIG --trainer_args.learning_rate=8e-4 --trainer_args.output_dir "./Llama2_13B" \
#     --trainer_args.load_best_model_at_end True --trainer_args.save_strategy '"steps"' \
#     --sama_adapter.col_L 8 --sama_adapter.row_R 8 --sama_adapter.scaling 2 \
#     --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 0 --seed 43 --run_text lora32
# date +"%F %T"

# wandb sync wandb/latest-run



#### SAMA 5 modules
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=8e-4 --trainer_args.output_dir "./Llama2_13B" \
    --trainer_args.load_best_model_at_end True --trainer_args.save_strategy '"steps"' \
    --sama_adapter.col_L 4 --sama_adapter.row_R 4 --sama_adapter.scaling 1 \
    --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 0 --seed 42 --run_text s42
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=8e-4 --trainer_args.output_dir "./Llama2_13B" \
#     --trainer_args.load_best_model_at_end True --trainer_args.save_strategy '"steps"' \
#     --sama_adapter.col_L 16 --sama_adapter.row_R 16 --sama_adapter.scaling 1 \
#     --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 0 --seed 42 --run_text s42
date +"%F %T"

# 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=6e-4 --trainer_args.output_dir "./Llama2_13B" \
#     --trainer_args.load_best_model_at_end True --trainer_args.save_strategy '"steps"' \
#     --sama_adapter.col_L 16 --sama_adapter.row_R 16 --sama_adapter.scaling 1 \
#     --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 0 --seed 42 --run_text s42
# date +"%F %T"

# 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=9e-4 --trainer_args.output_dir "./Llama2_13B" \
#     --trainer_args.load_best_model_at_end True --trainer_args.save_strategy '"steps"' \
#     --sama_adapter.col_L 16 --sama_adapter.row_R 16 --sama_adapter.scaling 1 \
#     --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 0 --seed 42 --run_text s42
# date +"%F %T"

# 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=4e-4 --trainer_args.output_dir "./Llama2_13B" \
#     --trainer_args.load_best_model_at_end True --trainer_args.save_strategy '"steps"' \
#     --sama_adapter.col_L 16 --sama_adapter.row_R 16 --sama_adapter.scaling 1 \
#     --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 0 --seed 42 --run_text s42
# date +"%F %T"