add benchmark for sft, dpo, simpo, orpo. Add benchmarking result. Support lora with gradient checkpoint

applications/ColossalChat/benchmarks/benchmark_sft.sh

@@ -0,0 +1,43 @@
+set_n_least_used_CUDA_VISIBLE_DEVICES() {
+    local n=${1:-"9999"}
+    echo "GPU Memory Usage:"
+    local FIRST_N_GPU_IDS=$(nvidia-smi --query-gpu=memory.used --format=csv |
+        tail -n +2 |
+        nl -v 0 |
+        tee /dev/tty |
+        sort -g -k 2 |
+        awk '{print $1}' |
+        head -n $n)
+    export CUDA_VISIBLE_DEVICES=$(echo $FIRST_N_GPU_IDS | sed 's/ /,/g')
+    echo "Now CUDA_VISIBLE_DEVICES is set to:"
+    echo "CUDA_VISIBLE_DEVICES=$CUDA_VISIBLE_DEVICES"
+}
+
+set_n_least_used_CUDA_VISIBLE_DEVICES 4
+# export CUDA_VISIBLE_DEVICES=3,4
+PROJECT_NAME="sft"
+PARENT_CONFIG_FILE="./benchmark_config" # Path to a folder to save training config logs
+PRETRAINED_MODEL_PATH="/root/commonData/Llama-2-7b-hf" # huggingface or local model path
+PRETRAINED_TOKENIZER_PATH="/root/commonData/Llama-2-7b-hf" # huggingface or local tokenizer path
+
+TIMESTAMP=$(date +%Y-%m-%d-%H-%M-%S)
+FULL_PROJECT_NAME="${PROJECT_NAME}-${TIMESTAMP}"
+CONFIG_FILE="${PARENT_CONFIG_FILE}-${FULL_PROJECT_NAME}.json"
+
+echo $(which colossalai)
+echo $(which python)
+# the real batch size for gradient descent is number_of_node_in_hostfile * nproc_per_node * train_batch_size
+colossalai run --nproc_per_node 4 --master_port 31312 benchmark_sft.py \
+    --pretrain $PRETRAINED_MODEL_PATH \
+    --tokenizer_dir $PRETRAINED_TOKENIZER_PATH \
+    --config_file $CONFIG_FILE \
+    --plugin zero2 \
+    --batch_size 8 \
+    --max_epochs 1 \
+    --accumulation_steps 1 \
+    --lr 5e-5 \
+    --lora_rank 32 \
+    --max_len 2048 \
+    --dataset_size 640 \
+    --grad_checkpoint \
+    --use_flash_attn