run_performance_4rank.sh
ATB_PATH=/usr/local/Ascend/atb-models source /usr/local/Ascend/ascend-toolkit/set_env.sh source /usr/local/Ascend/nnal/atb/set_env.sh # source ${ATB_PATH}/ascend-transformer-boost/output/atb/set_env.sh source /usr/local/Ascend/atb-models/set_env.sh # source ${ATB_PATH}/MindIE-LLM/examples/atb_models/output/atb_models/set_env.sh source /usr/local/Ascend/mindie/set_env.sh export ASCEND_RT_VISIBLE_DEVICES=0,1,2,3 export ATB_LLM_BENCHMARK_ENABLE=1 export ATB_LLM_ENABLE_AUTO_TRANSPOSE=0 export HCCL_CONNECT_TIMEOUT=7200 export HCCL_EXEC_TIMEOUT=1600 export ATB_LLM_HCCL_ENABLE=1 export HCCL_OP_EXPANSION_MODE="AIV" # export ATB_PROFILING_ENABLE=1 # export PROFILING_LEVEL=Level2 # export HCCL_BUFFSIZE=512 # export MASTER_IP=90.90.122.1 # export ALL_IN_OUT_SETS=[[16384,1024],[512,2048],[1024,128],[128,1024],[512,512]] ALL_IN_OUT_SETS=[[8192,1024]] #ALL_IN_OUT_SETS=[[8192,128]] export BS_GROUP=8 export P_MAX_BS=1 #!/bin/bash numbers=${ALL_IN_OUT_SETS//[\[\]]/} IFS=',' read -r encode decode <<< "$numbers" # clear formal log mkdir -p /root/atb/log.bak cp -rf /root/atb/log/* /root/atb/log.bak/* rm -r /root/atb/log/* mkdir -p /root/mindie/log.bak cp -rf /root/mindie/log/* /root/mindie/log.bak/* rm -r /root/mindie/log/* LOG_LEVEL=ERROR export ASDOPS_LOG_LEVEL=${LOG_LEVEL} export ASDOPS_LOG_TO_FILE=1 export ATB_LOG_TO_FILE=1 export MINDIE_LOG_TO_FILE=1 export MINDIE_LOG_TO_STDOUT=1 # export ATB_LOG_TO_STDOUT=1 export ATB_LOG_LEVEL=${LOG_LEVEL} export MINDIE_LOG_LEVEL=INFO export SPDLOG_LEVEL=${LOG_LEVEL} # export ASCEND_GLOBAL_LOG_LEVEL=0 # export HCCL_BUFFSIZE=1024 # export ATB_LLM_HCCL_ENABLE=1 export ATB_WORKSPACE_MEM_ALLOC_ALG_TYPE=3 export ATB_WORKSPACE_MEM_ALLOC_GLOBAL=1 export ASDOPS_LOG_LEVEL=ERROR export ASDOPS_LOG_TO_STDOUT=1 cd ${ATB_PATH}/tests/modeltest bash run.sh pa_fp16 performance $ALL_IN_OUT_SETS $BS_GROUP $P_MAX_BS llama /xx/xx/models/DeepSeek-R1-Distill-Llama-70B-W8A8SC-full 4 [1,4,-1,-1,-1,-1]