# WANDB_MODE=online CUDA_VISIBLE_DEVICES=6  python src/lgmodeling/train_model.py \
#     --seed 0 --tgt_len 512 --mem_len 512 --eval_tgt_len 128 --position-embeddings "learnable" \
#     --num-shared-experts 1 --num-routed-experts 0  --topk 0 --rotary-dim 64 \
#     --n_layer 12 --n_embd 512 --n_head 8 --n_inner 2048 --attention-bias \
#     --learning-rate 0.00025 --batch-size 24 --max_step 60000 --warmup_step 0 --dataset enwik8 \
#     --wandb-project LMC-Attention --wandb-group "GPT2-Enwik8-FFN" --wandb-entity "vinh-bui0512-hcmut"\
#     --model-save-dir /mnt/data/vinhbk/weights/enwik8 --data-path /mnt/data/vinhbk/datasets/enwik8
# WANDB_MODE=online CUDA_VISIBLE_DEVICES=6  python src/lgmodeling/train_model.py \
#     --seed 0 --tgt_len 512 --mem_len 512 --eval_tgt_len 128 --position-embeddings "rope" \
#     --num-shared-experts 1 --num-routed-experts 0  --topk 0 --rotary-dim 64 \
#     --n_layer 12 --n_embd 512 --n_head 8 --n_inner 2048 --attention-bias \
#     --learning-rate 0.00025 --batch-size 24 --max_step 60000 --warmup_step 0 --dataset enwik8 \
#     --wandb-project LMC-Attention --wandb-group "GPT2-Enwik8-FFN" --wandb-entity "vinh-bui0512-hcmut"\
#     --model-save-dir /mnt/data/vinhbk/weights/enwik8 --data-path /mnt/data/vinhbk/datasets/enwik8
# WANDB_MODE=online CUDA_VISIBLE_DEVICES=6  python src/lgmodeling/train_model.py \
#     --seed 0 --tgt_len 512 --mem_len 512 --eval_tgt_len 128 --position-embeddings "sinusoidal" \
#     --num-shared-experts 1 --num-routed-experts 0  --topk 0 --rotary-dim 64 \
#     --n_layer 12 --n_embd 512 --n_head 8 --n_inner 2048 --attention-bias \
#     --learning-rate 0.00025 --batch-size 24 --max_step 60000 --warmup_step 0 --dataset enwik8 \
#     --wandb-project LMC-Attention --wandb-group "GPT2-Enwik8-FFN" --wandb-entity "vinh-bui0512-hcmut"\
#     --model-save-dir /mnt/data/vinhbk/weights/enwik8 --data-path /mnt/data/vinhbk/datasets/enwik8


WANDB_MODE=online CUDA_VISIBLE_DEVICES=6  python src/lgmodeling/train_model.py \
    --seed 0 --tgt_len 512 --mem_len 512 --eval_tgt_len 128 --position-embeddings "learnable" \
    --num-shared-experts 0 --num-routed-experts 4  --topk 4 --rotary-dim 64 \
    --n_layer 12 --n_embd 512 --n_head 8 --n_inner 2048 --attention-bias \
    --learning-rate 0.00025 --batch-size 24 --max_step 60000 --warmup_step 0 --dataset enwik8 \
    --wandb-project LMC-Attention --wandb-group "GPT2-Enwik8-MoE" --wandb-entity "vinh-bui0512-hcmut"\
    --model-save-dir /mnt/data/vinhbk/weights/enwik8 --data-path /mnt/data/vinhbk/datasets/enwik8
WANDB_MODE=online CUDA_VISIBLE_DEVICES=6  python src/lgmodeling/train_model.py \
    --seed 0 --tgt_len 512 --mem_len 512 --eval_tgt_len 128 --position-embeddings "rope" \
    --num-shared-experts 0 --num-routed-experts 4  --topk 4 --rotary-dim 64 \
    --n_layer 12 --n_embd 512 --n_head 8 --n_inner 2048 --attention-bias \
    --learning-rate 0.00025 --batch-size 24 --max_step 60000 --warmup_step 0 --dataset enwik8 \
    --wandb-project LMC-Attention --wandb-group "GPT2-Enwik8-MoE" --wandb-entity "vinh-bui0512-hcmut"\
    --model-save-dir /mnt/data/vinhbk/weights/enwik8 --data-path /mnt/data/vinhbk/datasets/enwik8
WANDB_MODE=online CUDA_VISIBLE_DEVICES=6  python src/lgmodeling/train_model.py \
    --seed 0 --tgt_len 512 --mem_len 512 --eval_tgt_len 128 --position-embeddings "sinusoidal" \
    --num-shared-experts 0 --num-routed-experts 4  --topk 4 --rotary-dim 64 \
    --n_layer 12 --n_embd 512 --n_head 8 --n_inner 2048 --attention-bias \
    --learning-rate 0.00025 --batch-size 24 --max_step 60000 --warmup_step 0 --dataset enwik8 \
    --wandb-project LMC-Attention --wandb-group "GPT2-Enwik8-MoE" --wandb-entity "vinh-bui0512-hcmut"\
    --model-save-dir /mnt/data/vinhbk/weights/enwik8 --data-path /mnt/data/vinhbk/datasets/enwik8

WANDB_MODE=online CUDA_VISIBLE_DEVICES=6  python src/lgmodeling/train_model.py \
    --seed 0 --tgt_len 512 --mem_len 512 --eval_tgt_len 128 --position-embeddings "learnable" \
    --num-shared-experts 0 --num-routed-experts 4  --topk 2 --rotary-dim 64 \
    --n_layer 12 --n_embd 512 --n_head 8 --n_inner 2048 --attention-bias \
    --learning-rate 0.00025 --batch-size 24 --max_step 60000 --warmup_step 0 --dataset enwik8 \
    --wandb-project LMC-Attention --wandb-group "GPT2-Enwik8-SMoE" --wandb-entity "vinh-bui0512-hcmut"\
    --model-save-dir /mnt/data/vinhbk/weights/enwik8 --data-path /mnt/data/vinhbk/datasets/enwik8
WANDB_MODE=online CUDA_VISIBLE_DEVICES=6  python src/lgmodeling/train_model.py \
    --seed 0 --tgt_len 512 --mem_len 512 --eval_tgt_len 128 --position-embeddings "rope" \
    --num-shared-experts 0 --num-routed-experts 4  --topk 2 --rotary-dim 64 \
    --n_layer 12 --n_embd 512 --n_head 8 --n_inner 2048 --attention-bias \
    --learning-rate 0.00025 --batch-size 24 --max_step 60000 --warmup_step 0 --dataset enwik8 \
    --wandb-project LMC-Attention --wandb-group "GPT2-Enwik8-SMoE" --wandb-entity "vinh-bui0512-hcmut"\
    --model-save-dir /mnt/data/vinhbk/weights/enwik8 --data-path /mnt/data/vinhbk/datasets/enwik8
WANDB_MODE=online CUDA_VISIBLE_DEVICES=6  python src/lgmodeling/train_model.py \
    --seed 0 --tgt_len 512 --mem_len 512 --eval_tgt_len 128 --position-embeddings "sinusoidal" \
    --num-shared-experts 0 --num-routed-experts 4  --topk 2 --rotary-dim 64 \
    --n_layer 12 --n_embd 512 --n_head 8 --n_inner 2048 --attention-bias \
    --learning-rate 0.00025 --batch-size 24 --max_step 60000 --warmup_step 0 --dataset enwik8 \
    --wandb-project LMC-Attention --wandb-group "GPT2-Enwik8-SMoE" --wandb-entity "vinh-bui0512-hcmut"\
    --model-save-dir /mnt/data/vinhbk/weights/enwik8 --data-path /mnt/data/vinhbk/datasets/enwik8