scripts/math_gemma9_train.sh

#!/bin/bash

export SAMA_CONFIG=./config/sama_math_gemma9.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_MATH"

date +"%F %T"

# test
STEP=50
# accelerate launch --dynamo_backend=inductor --dynamo_mode=max-autotune --main_process_port 41353 -m src.math_train \
#     --config_path $SAMA_CONFIG --trainer_args.learning_rate=1e-3 --trainer_args.output_dir "./Gemma7B" \
#     --trainer_args.load_best_model_at_end False --trainer_args.save_strategy '"no"' \
#     --sama_adapter.col_L 16 --sama_adapter.row_R 16 \
#     --sama_adapter.num_unique_blocks_L 16 --sama_adapter.num_unique_blocks_R 16 \
#     --trainer_args.num_train_epochs 16 --trainer_args.report_to none --trainer_args.eval_delay 200 \
#     --sama_adapter.target_modules '["q_proj", "v_proj", "k_proj", "o_proj", "gate_proj","up_proj","down_proj"]' \
#     --sama_adapter.scaling 2

# accelerate launch --dynamo_backend=inductor --dynamo_mode=max-autotune --main_process_port 41353 -m src.math_train \
#     --config_path $SAMA_CONFIG --trainer_args.learning_rate=5e-4 --trainer_args.output_dir "./MGemma9B" \
#     --trainer_args.load_best_model_at_end False --trainer_args.save_strategy '"no"' \
#     --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", "o_proj", "gate_proj","up_proj","down_proj"]' \
#     --data.dataset_split train[:20000] --trainer_args.eval_delay 0 \
#     --sama_adapter.scaling 1
# date +"%F %T"

# accelerate launch --dynamo_backend=inductor --dynamo_mode=max-autotune --main_process_port 41353 -m src.math_train \
#     --config_path $SAMA_CONFIG --trainer_args.learning_rate=5e-4 --trainer_args.output_dir "./MGemma9B" \
#     --trainer_args.load_best_model_at_end False --trainer_args.save_strategy '"no"' \
#     --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", "o_proj", "gate_proj","up_proj","down_proj"]' \
#     --data.dataset_split train[:20000] --trainer_args.eval_delay 0 \
#     --sama_adapter.scaling 1

# accelerate launch --dynamo_backend=inductor --dynamo_mode=max-autotune --main_process_port 41353 -m src.math_train \
#     --config_path $SAMA_CONFIG --trainer_args.learning_rate=5e-4 --trainer_args.output_dir "./MGemma9B" \
#     --trainer_args.load_best_model_at_end False --trainer_args.save_strategy '"no"' \
#     --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", "o_proj", "gate_proj","up_proj","down_proj"]' \
#     --data.dataset_split train[:20000] --trainer_args.eval_delay 0 \
#     --sama_adapter.scaling 2
# date +"%F %T"

# accelerate launch --dynamo_backend=inductor --dynamo_mode=max-autotune --main_process_port 41353 -m src.math_train \
#     --config_path $SAMA_CONFIG --trainer_args.learning_rate=5e-4 --trainer_args.output_dir "./MGemma9B" \
#     --trainer_args.load_best_model_at_end False --trainer_args.save_strategy '"no"' \
#     --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 4 --sama_adapter.num_unique_blocks_R 4 \
#     --sama_adapter.target_modules '["q_proj", "v_proj", "k_proj", "o_proj", "gate_proj","up_proj","down_proj"]' \
#     --data.dataset_split train[:20000] --trainer_args.eval_delay 0 \
#     --sama_adapter.scaling 4
# date +"%F %T"

# STEP=500
# accelerate launch --dynamo_backend=inductor --dynamo_mode=max-autotune --main_process_port 41353 -m src.math_train \
#     --config_path $SAMA_CONFIG --trainer_args.learning_rate=5e-4 --trainer_args.output_dir "./MGemma9B" \
#     --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 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", "o_proj", "gate_proj","up_proj","down_proj"]' \
#     --data.dataset_split train[:20000] --trainer_args.eval_delay 0 \
#     --sama_adapter.scaling 2
# date +"%F %T"

#
# STEP=500
# accelerate launch --dynamo_backend=inductor --dynamo_mode=max-autotune --main_process_port 41353 -m src.math_train \
#     --config_path $SAMA_CONFIG --trainer_args.learning_rate=5e-4 --trainer_args.output_dir "./MGemma9B" \
#     --trainer_args.load_best_model_at_end False --trainer_args.save_strategy '"no"' \
#     --sama_adapter.col_L 2 --sama_adapter.row_R 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 2 --sama_adapter.num_unique_blocks_R 2 \
#     --sama_adapter.target_modules '["q_proj", "v_proj", "k_proj", "o_proj", "gate_proj","up_proj","down_proj"]' \
#     --data.dataset_split train[:20000] --trainer_args.eval_delay 0 \
#     --sama_adapter.scaling 1
# date +"%F %T"

# STEP=500
# accelerate launch --dynamo_backend=inductor --dynamo_mode=max-autotune --main_process_port 41353 -m src.math_train \
#     --config_path $SAMA_CONFIG --trainer_args.learning_rate=5e-4 --trainer_args.output_dir "./MGemma9B" \
#     --trainer_args.load_best_model_at_end False --trainer_args.save_strategy '"no"' \
#     --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 2 --sama_adapter.num_unique_blocks_R 2 \
#     --sama_adapter.target_modules '["q_proj", "v_proj", "k_proj", "o_proj", "gate_proj","up_proj","down_proj"]' \
#     --data.dataset_split train[:20000] --trainer_args.eval_delay 0 \
#     --sama_adapter.scaling 2
# date +"%F %T"
# 
# wandb sync wandb/latest-run
# date +"%F %T"

# STEP=500
# accelerate launch --dynamo_backend=inductor --dynamo_mode=max-autotune --main_process_port 41353 -m src.math_train \
#     --config_path $SAMA_CONFIG --trainer_args.learning_rate=5e-4 --trainer_args.output_dir "./MGemma9B" \
#     --trainer_args.load_best_model_at_end False --trainer_args.save_strategy '"no"' \
#     --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", "o_proj", "gate_proj","up_proj","down_proj"]' \
#     --data.dataset_split train[:20000] --trainer_args.eval_delay 0 \
#     --sama_adapter.scaling 1
# date +"%F %T"

# STEP=500
# accelerate launch --dynamo_backend=inductor --dynamo_mode=max-autotune --main_process_port 41353 -m src.math_train \
#     --config_path $SAMA_CONFIG --trainer_args.learning_rate=5e-4 --trainer_args.output_dir "./MGemma9B" \
#     --trainer_args.load_best_model_at_end False --trainer_args.save_strategy '"no"' \
#     --sama_adapter.col_L 2 --sama_adapter.row_R 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 2 --sama_adapter.num_unique_blocks_R 2 \
#     --sama_adapter.target_modules '["q_proj", "v_proj", "k_proj", "o_proj", "gate_proj","up_proj","down_proj"]' \
#     --data.dataset_split train[:20000] --trainer_args.eval_delay 0 \
#     --sama_adapter.scaling 1
# date +"%F %T"


# STEP=500
# accelerate launch --dynamo_backend=inductor --dynamo_mode=max-autotune --main_process_port 41353 -m src.math_train \
#     --config_path $SAMA_CONFIG --trainer_args.learning_rate=5e-4 --trainer_args.output_dir "./MGemma9B" \
#     --trainer_args.load_best_model_at_end False --trainer_args.save_strategy '"no"' \
#     --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", "o_proj", "gate_proj","up_proj","down_proj"]' \
#     --data.dataset_split train[:20000] --trainer_args.eval_delay 0 \
#     --sama_adapter.scaling 1

# STEP=500
# accelerate launch --dynamo_backend=inductor --dynamo_mode=max-autotune --main_process_port 41353 -m src.math_train \
#     --config_path $SAMA_CONFIG --trainer_args.learning_rate=5e-4 --trainer_args.output_dir "./MGemma9B" \
#     --trainer_args.load_best_model_at_end False --trainer_args.save_strategy '"no"' \
#     --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 2 --sama_adapter.num_unique_blocks_R 2 \
#     --sama_adapter.target_modules '["q_proj", "v_proj", "k_proj", "o_proj", "gate_proj","up_proj","down_proj"]' \
#     --data.dataset_split train[:20000] --trainer_args.eval_delay 0 \
#     --sama_adapter.scaling 2
# date +"%F %T"

# STEP=500
# accelerate launch --dynamo_backend=inductor --dynamo_mode=max-autotune --main_process_port 41353 -m src.math_train \
#     --config_path $SAMA_CONFIG --trainer_args.learning_rate=5e-4 --trainer_args.output_dir "./MGemma9B" \
#     --trainer_args.load_best_model_at_end False --trainer_args.save_strategy '"no"' \
#     --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 2 --sama_adapter.num_unique_blocks_R 2 \
#     --sama_adapter.target_modules '["q_proj", "v_proj", "k_proj", "o_proj", "gate_proj","up_proj","down_proj"]' \
#     --data.dataset_split train[:20000] --trainer_args.eval_delay 0 \
#     --sama_adapter.scaling 1.4142
# date +"%F %T"

# STEP=500
# accelerate launch --dynamo_backend=inductor --dynamo_mode=max-autotune --main_process_port 41353 -m src.math_train \
#     --config_path $SAMA_CONFIG --trainer_args.learning_rate=5e-4 --trainer_args.output_dir "./MGemma9B" \
#     --trainer_args.load_best_model_at_end False --trainer_args.save_strategy '"no"' \
#     --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 2 --sama_adapter.num_unique_blocks_R 2 \
#     --sama_adapter.target_modules '["q_proj", "v_proj", "k_proj", "o_proj", "gate_proj","up_proj","down_proj"]' \
#     --data.dataset_split train[:20000] --trainer_args.eval_delay 0 \
#     --sama_adapter.scaling 2.8284
# date +"%F %T"

# STEP=500
# accelerate launch --dynamo_backend=inductor --dynamo_mode=max-autotune --main_process_port 41353 -m src.math_train \
#     --config_path $SAMA_CONFIG --trainer_args.learning_rate=5e-4 --trainer_args.output_dir "./MGemma9B" \
#     --trainer_args.load_best_model_at_end False --trainer_args.save_strategy '"no"' \
#     --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 2 --sama_adapter.num_unique_blocks_R 2 \
#     --sama_adapter.target_modules '["q_proj", "v_proj", "k_proj", "o_proj", "gate_proj","up_proj","down_proj"]' \
#     --data.dataset_split train[:20000] --trainer_args.eval_delay 0 \
#     --sama_adapter.scaling 4
# date +"%F %T"

STEP=500
accelerate launch --dynamo_backend=inductor --dynamo_mode=max-autotune --main_process_port 41353 -m src.math_train \
    --config_path $SAMA_CONFIG --trainer_args.learning_rate=5e-4 --trainer_args.output_dir "./MGemma9B" \
    --trainer_args.load_best_model_at_end False --trainer_args.save_strategy '"no"' \
    --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 2 --sama_adapter.num_unique_blocks_R 2 \
    --sama_adapter.target_modules '["q_proj", "v_proj", "k_proj", "o_proj", "gate_proj","up_proj","down_proj"]' \
    --data.dataset_split train[:20000] --trainer_args.eval_delay 0 \
    --sama_adapter.scaling 1
date +"%F %T"

STEP=500
accelerate launch --dynamo_backend=inductor --dynamo_mode=max-autotune --main_process_port 41353 -m src.math_train \
    --config_path $SAMA_CONFIG --trainer_args.learning_rate=5e-4 --trainer_args.output_dir "./MGemma9B" \
    --trainer_args.load_best_model_at_end False --trainer_args.save_strategy '"no"' \
    --sama_adapter.col_L 2 --sama_adapter.row_R 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 2 --sama_adapter.num_unique_blocks_R 2 \
    --sama_adapter.target_modules '["q_proj", "v_proj", "k_proj", "o_proj", "gate_proj","up_proj","down_proj"]' \
    --data.dataset_split train[:20000] --trainer_args.eval_delay 0 \
    --sama_adapter.scaling 0.7071
date +"%F %T"

# bash scripts/math_mistral7_train.sh