#!/bin/bash
#SBATCH --job-name=ctxlora_medium
#SBATCH --partition=a3
#SBATCH --nodes=1
#SBATCH --gpus=4
#SBATCH --output=outputs/%x-%j.out
#SBATCH --error=outputs/%x-%j.out

# module load
# module load cuda/12.1
# module load cudnn/8.9.7
# module load nccl/cuda-12.1/2.18.3
# module load hpcx/2.20

# export OMP_NUM_THREADS=24
# export TRITON_CACHE_DIR=/tmp/.triton/
# . ~/miniconda3/etc/profile.d/conda.sh
# conda activate /home/rujikorn_sakana_ai/.conda/envs/ctx-to-lora
# eval "$@"

uv run accelerate launch --num_processes=4 --gradient_accumulation_steps=32 --gradient_clipping=1.0 \
--gpu_ids all --main_process_port 29560 intx_sft.py configs/fw_qa_pretrain_small.yaml \
--model_name_or_path=google/gemma-2-2b-it --num_train_epochs=2 --per_device_train_batch_size=4 \
--gradient_accumulation_steps=32 --per_device_eval_batch_size=8 --exp_setup=hyper_lora --aggregator_type=perceiver \
--target_modules=down_proj \
--num_self_attends_per_block=8 --num_latent_factor=2 \
--lora_r=8 \
--eval_steps=5000 --save_steps=5000 --learning_rate=4e-5 --lora_dropout=0.0 \
--neftune_noise_alpha=5 --use_light_weight_lora=False \
--add_negative_prompt=False \
--add_repeat_prompt=False \
--use_sequence_packing=True --per_rank_gen=True \
--per_layer_processing=True \
--gen_lora_l1_reg_coef=0.1 \