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=32  # Kimi-K2.5模型需要32机A3运行，可参考readme介绍配置多机训练
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
"
torchrun $DISTRIBUTED_ARGS mindspeed_mm/fsdp/train/trainer.py \
    examples/kimik2_5/kimik2_5_config.yaml