dataset_code/sft_sftnews/offload/run_wan.sh

export OMNISTORE_LOAD_STRICT_MODE=0
export OMNISTORE_LOGGING_LEVEL=ERROR
#################################################################
## Torch
#################################################################
export TOKENIZERS_PARALLELISM=false
export TORCH_LOGS="+dynamo,recompiles,graph_breaks"
export TORCHDYNAMO_VERBOSE=1
export TORCH_NCCL_ENABLE_MONITORING=1
export PYTORCH_CUDA_ALLOC_CONF="expandable_segments:True,garbage_collection_threshold:0.9"
#################################################################


#################################################################
## NCCL
#################################################################
export NCCL_IB_GID_INDEX=3
export NCCL_IB_HCA=$ARNOLD_RDMA_DEVICE
export NCCL_SOCKET_IFNAME=eth0
export NCCL_SOCKET_TIMEOUT=3600000

export NCCL_DEBUG=WARN  # disable the verbose NCCL logs
export NCCL_P2P_DISABLE=0
export NCCL_IB_DISABLE=0  # was 1
export NCCL_SHM_DISABLE=0  # was 1
export NCCL_P2P_LEVEL=NVL

export NCCL_PXN_DISABLE=0
export NCCL_NET_GDR_LEVEL=2
export NCCL_IB_QPS_PER_CONNECTION=4
export NCCL_IB_TC=160
export NCCL_IB_TIMEOUT=22
#################################################################

#################################################################
## DIST
#################################################################
MASTER_ADDR=$ARNOLD_WORKER_0_HOST
ports=(`echo $METIS_WORKER_0_PORT | tr ',' ' '`)
MASTER_PORT=${ports[0]}
NNODES=$ARNOLD_WORKER_NUM
NODE_RANK=$ARNOLD_ID
GPUS_PER_NODE=$ARNOLD_WORKER_GPU
# GPUS_PER_NODE=1
# NNODES=1
# NODE_RANK=0
WORLD_SIZE=$(($GPUS_PER_NODE*$NNODES))

DISTRIBUTED_ARGS="--nproc_per_node $GPUS_PER_NODE --nnodes $NNODES --node_rank $NODE_RANK --master_addr $MASTER_ADDR --master_port $MASTER_PORT"
if [ ! -z $RDZV_BACKEND ]; then
    DISTRIBUTED_ARGS="${DISTRIBUTED_ARGS} --rdzv_endpoint $MASTER_ADDR:$MASTER_PORT --rdzv_id 9863 --rdzv_backend c10d"
    export NCCL_SHM_DISABLE=1
fi

echo -e "\033[31mDISTRIBUTED_ARGS: ${DISTRIBUTED_ARGS}\033[0m"

#################################################################

#TODO: prefetching
for i in {0..1}
do
    torchrun $DISTRIBUTED_ARGS offoload_features_wan.py \
        --batch_size 1 \
        --config_path part0.yaml || true
done

# torchrun $DISTRIBUTED_ARGS offoload_features.py \
#     --batch_size 1 \
#     --config_path part1.yaml

# torchrun $DISTRIBUTED_ARGS offoload_features.py \
#     --batch_size 1 \
#     --config_path part2.yaml

# torchrun $DISTRIBUTED_ARGS offoload_features.py \
#     --batch_size 1 \
#     --config_path part3.yaml

# torchrun $DISTRIBUTED_ARGS offoload_features.py \
#     --batch_size 1 \
#     --config_path part4.yaml

# torchrun $DISTRIBUTED_ARGS offoload_features.py \
#     --batch_size 1 \
#     --config_path part5.yaml

# torchrun $DISTRIBUTED_ARGS offoload_features.py \
#     --batch_size 1 \
#     --config_path part6.yaml

# torchrun $DISTRIBUTED_ARGS offoload_features.py \
#     --batch_size 1 \
#     --config_path part7.yaml