src/scripts/3b_run_grpo_vllm_self_eval_thenNoDesEval.sh

#!/bin/bash

./move.sh

cd src/r1-v

export DEBUG_MODE="true"
export LOG_PATH="./vllm_run.txt"

## Take self eval then do no des eval
QWEN_PATH='/apdcephfs_sh2/share_300000800/user/zongxia/Video-R1/src/r1-v/log/3B-Video-GRPO-SelfEval-Train/pool_numerical_chunk_01/checkpoint-25'
DATA_FILE='video_pool_multiple_choice_chunk_01'

# QWEN_PATH='/apdcephfs_sh2/share_300000800/user/zongxia/Video-R1/src/r1-v/log/3B-Video-GRPO-SelfEval-Train/pool_numerical_chunk_01/checkpoint-25'
# # DATA_FILE="pool_numerical_chunk_01"
# DATA_FILE='pool_multiple_choice_chunk_02'

# QWEN_PATH='/apdcephfs_sh2/share_300000800/user/zongxia/Video-R1/src/r1-v/log/3B-Video-GRPO-AnswerBERT/video_pool_multiple_choice_chunk_01/checkpoint-46'
# DATA_FILE='pool_numerical_chunk_02'

QWEN_PATH='/apdcephfs_sh2/share_300000800/user/zongxia/Video-R1/src/r1-v/log/3B-Video-GRPO-AnswerBERT/video_pool_multiple_choice_chunk_01/checkpoint-46'
DATA_FILE="pool_multiple_choice_chunk_02"


HF_DATASET="./Video-R1-data/${DATA_FILE}.json"
OUTPUT_DIR="./log/3B-Video-GRPO-selfEval-ThenNoDesEval/${DATA_FILE}"
if [ ! -d "$OUTPUT_DIR" ]; then
 mkdir -p "$OUTPUT_DIR"
fi
RUN_NAME="3B-Video-GRPO-selfEval-ThenNoDesEval"
DS_CONFIG="local_scripts/zero3.json"  

# Set temporal to choose between T-GRPO and GRPO, and len_control to enable or disable the length control reward.
# NOTE: you are expected to use X + 1 cards for X training proc and 1 vLLM proc 
# e.g., the visible devices should be 0,1,2,3,4 for 5 cards, and  --nproc_per_node="4"
export WANDB_API_KEY="5e11bfa8cf4062940486d279ecd9e70617d4ac7a"
CUDA_VISIBLE_DEVICES="0,1,2,3,4,5,6,7" torchrun \
    --nproc_per_node="7" \
    --nnodes="1" \
    --node_rank="0" \
    --master_addr="127.0.0.1" \
    --master_port="12345" \
    src/open_r1/grpo-cot-noDesEval.py \
    --use_vllm true \
    --output_dir ${OUTPUT_DIR} \
    --model_name_or_path ${QWEN_PATH} \
    --dataset_name ${HF_DATASET} \
    --max_prompt_length 16384 \
    --max_completion_length 1600 \
    --per_device_train_batch_size 1 \
    --gradient_accumulation_steps 32 \
    --learning_rate 1e-6 \
    --lr_scheduler_type "cosine" \
    --weight_decay 0.01 \
    --logging_steps 1 \
    --bf16 true \
    --gradient_checkpointing true \
    --attn_implementation flash_attention_2 \
    --min_pixels 3136 \
    --max_pixels 501760 \
    --num_train_epochs 1 \
    --run_name ${RUN_NAME} \
    --save_steps 14 \
    --save_only_model false \
    --temporal true \
    --len_control false \
    --report_to wandb \
    --beta 0.04 \
    --max_grad_norm 5 \
    --temperature 1.0 \
    --num_generations 8 \
    --vllm_device "cuda:7" \
    --vllm_gpu_memory_utilization 0.7 \
    --deepspeed ${DS_CONFIG} \
    2>&1 | tee "${OUTPUT_DIR}/training_log.txt"


python /cq_1/share_1603164/user/zongxia/workspace/gpu_burn.py