source /usr/local/Ascend/ascend-toolkit/set_env.sh
export NON_MEGATRON=true
export HCCL_CONNECT_TIMEOUT=1200
export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
export MULTI_STREAM_MEMORY_REUSE=2
export TASK_QUEUE_ENABLE=1
export CPU_AFFINITY_CONF=1
NPUS_PER_NODE=16
MASTER_ADDR=localhost
MASTER_PORT=6000
NNODES=1
NODE_RANK=0
WORLD_SIZE=$(($NPUS_PER_NODE*$NNODES))
DISTRIBUTED_ARGS="
--nproc_per_node $NPUS_PER_NODE \
--nnodes $NNODES \
--node_rank $NODE_RANK \
--master_addr $MASTER_ADDR \
--master_port $MASTER_PORT
"
logfile=$(date +%Y%m%d)_$(date +%H%M%S)
mkdir -p logs
torchrun $DISTRIBUTED_ARGS mindspeed_mm/fsdp/train/trainer.py \
examples/qwen3vl/qwen3vl_30B_config_v1.yaml \
2>&1 | tee logs/train_${logfile}.log