#!/bin/bash

# GPUS_PER_NODE=8
# NNODES=1
# NODE_RANK=0
# MASTER_ADDR=localhost
# MASTER_PORT=6001
 
MODEL="/data1/speech/anhnmt2/Speech2Speech/half-streaming-speech-nlp/checkpoints/minicpmo_sft_asr"
TOKENIZER_PATH="/data1/speech/anhnmt2/Speech2Speech/half-streaming-speech-nlp/omni_speech/model/minicpmo/MiniCPM-o-2_6"
# or openbmb/MiniCPM-V-2, openbmb/MiniCPM-Llama3-V-2_5, openbmb/MiniCPM-V-2_6
# ATTENTION: specify the path to your training data, which should be a json file consisting of a list of conversations.
# See the section for finetuning in README for more information.
DATA="/data1/speech/anhnmt2/dataset/s2s/minicpmo/asr/train_asr_mixed_500k.jsonl"
EVAL_DATA="/data1/speech/anhnmt2/dataset/s2s/minicpmo/asr/dev_asr_mixed.jsonl"

# if use openbmb/MiniCPM-V-2, please set LLM_TYPE=minicpm, if use openbmb/MiniCPM-Llama3-V-2_5, please set LLM_TYPE="llama3",
# if use openbmb/MiniCPM-o-2_6 or openbmb/MiniCPM-V-2_6, please set LLM_TYPE=qwen
LLM_TYPE="qwen" 
MODEL_MAX_Length=2048 # if conduct multi-images sft, please set MODEL_MAX_Length=4096


# DISTRIBUTED_ARGS="
#     --nproc_per_node $GPUS_PER_NODE \
#     --nnodes $NNODES \
#     --node_rank $NODE_RANK \
#     --master_addr $MASTER_ADDR \
#     --master_port $MASTER_PORT
# "

deepspeed ../omni_speech/train/train_minicpmo.py  \
    --deepspeed zero2.json \
    --model_name_or_path $MODEL \
    --tokenizer_path $TOKENIZER_PATH \
    --llm_type $LLM_TYPE \
    --data_path $DATA \
    --eval_data_path $EVAL_DATA \
    --remove_unused_columns false \
    --label_names "labels" \
    --prediction_loss_only false \
    --bf16 true \
    --do_train \
    --do_eval \
    --tune_speech true \
    --tune_llm false \
    --model_max_length $MODEL_MAX_Length \
    --eval_steps 2000 \
    --output_dir ../checkpoints/minicpmo_sft_asr \
    --num_train_epochs 2 \
    --logging_strategy "steps" \
    --per_device_train_batch_size 1 \
    --per_device_eval_batch_size 1 \
    --gradient_accumulation_steps 4 \
    --evaluation_strategy "steps" \
    --save_strategy "steps" \
    --save_steps 5000 \
    --save_total_limit 1 \
    --learning_rate 1e-5 \
    --max_grad_norm 20. \
    --weight_decay 0. \
    --warmup_ratio 0.03 \
    --lr_scheduler_type "cosine" \
    --logging_steps 1 \
    --tf32 True \
    --gradient_checkpointing true